xref: /openbmc/linux/kernel/sched/fair.c (revision 4da722ca)
1 /*
2  * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
3  *
4  *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
5  *
6  *  Interactivity improvements by Mike Galbraith
7  *  (C) 2007 Mike Galbraith <efault@gmx.de>
8  *
9  *  Various enhancements by Dmitry Adamushko.
10  *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
11  *
12  *  Group scheduling enhancements by Srivatsa Vaddagiri
13  *  Copyright IBM Corporation, 2007
14  *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
15  *
16  *  Scaled math optimizations by Thomas Gleixner
17  *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
18  *
19  *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
21  */
22 
23 #include <linux/sched/mm.h>
24 #include <linux/sched/topology.h>
25 
26 #include <linux/latencytop.h>
27 #include <linux/cpumask.h>
28 #include <linux/cpuidle.h>
29 #include <linux/slab.h>
30 #include <linux/profile.h>
31 #include <linux/interrupt.h>
32 #include <linux/mempolicy.h>
33 #include <linux/migrate.h>
34 #include <linux/task_work.h>
35 
36 #include <trace/events/sched.h>
37 
38 #include "sched.h"
39 
40 /*
41  * Targeted preemption latency for CPU-bound tasks:
42  *
43  * NOTE: this latency value is not the same as the concept of
44  * 'timeslice length' - timeslices in CFS are of variable length
45  * and have no persistent notion like in traditional, time-slice
46  * based scheduling concepts.
47  *
48  * (to see the precise effective timeslice length of your workload,
49  *  run vmstat and monitor the context-switches (cs) field)
50  *
51  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
52  */
53 unsigned int sysctl_sched_latency			= 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency		= 6000000ULL;
55 
56 /*
57  * The initial- and re-scaling of tunables is configurable
58  *
59  * Options are:
60  *
61  *   SCHED_TUNABLESCALING_NONE - unscaled, always *1
62  *   SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
63  *   SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
64  *
65  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
66  */
67 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
68 
69 /*
70  * Minimal preemption granularity for CPU-bound tasks:
71  *
72  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
73  */
74 unsigned int sysctl_sched_min_granularity		= 750000ULL;
75 unsigned int normalized_sysctl_sched_min_granularity	= 750000ULL;
76 
77 /*
78  * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
79  */
80 static unsigned int sched_nr_latency = 8;
81 
82 /*
83  * After fork, child runs first. If set to 0 (default) then
84  * parent will (try to) run first.
85  */
86 unsigned int sysctl_sched_child_runs_first __read_mostly;
87 
88 /*
89  * SCHED_OTHER wake-up granularity.
90  *
91  * This option delays the preemption effects of decoupled workloads
92  * and reduces their over-scheduling. Synchronous workloads will still
93  * have immediate wakeup/sleep latencies.
94  *
95  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
96  */
97 unsigned int sysctl_sched_wakeup_granularity		= 1000000UL;
98 unsigned int normalized_sysctl_sched_wakeup_granularity	= 1000000UL;
99 
100 const_debug unsigned int sysctl_sched_migration_cost	= 500000UL;
101 
102 #ifdef CONFIG_SMP
103 /*
104  * For asym packing, by default the lower numbered cpu has higher priority.
105  */
106 int __weak arch_asym_cpu_priority(int cpu)
107 {
108 	return -cpu;
109 }
110 #endif
111 
112 #ifdef CONFIG_CFS_BANDWIDTH
113 /*
114  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
115  * each time a cfs_rq requests quota.
116  *
117  * Note: in the case that the slice exceeds the runtime remaining (either due
118  * to consumption or the quota being specified to be smaller than the slice)
119  * we will always only issue the remaining available time.
120  *
121  * (default: 5 msec, units: microseconds)
122  */
123 unsigned int sysctl_sched_cfs_bandwidth_slice		= 5000UL;
124 #endif
125 
126 /*
127  * The margin used when comparing utilization with CPU capacity:
128  * util * margin < capacity * 1024
129  *
130  * (default: ~20%)
131  */
132 unsigned int capacity_margin				= 1280;
133 
134 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
135 {
136 	lw->weight += inc;
137 	lw->inv_weight = 0;
138 }
139 
140 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
141 {
142 	lw->weight -= dec;
143 	lw->inv_weight = 0;
144 }
145 
146 static inline void update_load_set(struct load_weight *lw, unsigned long w)
147 {
148 	lw->weight = w;
149 	lw->inv_weight = 0;
150 }
151 
152 /*
153  * Increase the granularity value when there are more CPUs,
154  * because with more CPUs the 'effective latency' as visible
155  * to users decreases. But the relationship is not linear,
156  * so pick a second-best guess by going with the log2 of the
157  * number of CPUs.
158  *
159  * This idea comes from the SD scheduler of Con Kolivas:
160  */
161 static unsigned int get_update_sysctl_factor(void)
162 {
163 	unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
164 	unsigned int factor;
165 
166 	switch (sysctl_sched_tunable_scaling) {
167 	case SCHED_TUNABLESCALING_NONE:
168 		factor = 1;
169 		break;
170 	case SCHED_TUNABLESCALING_LINEAR:
171 		factor = cpus;
172 		break;
173 	case SCHED_TUNABLESCALING_LOG:
174 	default:
175 		factor = 1 + ilog2(cpus);
176 		break;
177 	}
178 
179 	return factor;
180 }
181 
182 static void update_sysctl(void)
183 {
184 	unsigned int factor = get_update_sysctl_factor();
185 
186 #define SET_SYSCTL(name) \
187 	(sysctl_##name = (factor) * normalized_sysctl_##name)
188 	SET_SYSCTL(sched_min_granularity);
189 	SET_SYSCTL(sched_latency);
190 	SET_SYSCTL(sched_wakeup_granularity);
191 #undef SET_SYSCTL
192 }
193 
194 void sched_init_granularity(void)
195 {
196 	update_sysctl();
197 }
198 
199 #define WMULT_CONST	(~0U)
200 #define WMULT_SHIFT	32
201 
202 static void __update_inv_weight(struct load_weight *lw)
203 {
204 	unsigned long w;
205 
206 	if (likely(lw->inv_weight))
207 		return;
208 
209 	w = scale_load_down(lw->weight);
210 
211 	if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
212 		lw->inv_weight = 1;
213 	else if (unlikely(!w))
214 		lw->inv_weight = WMULT_CONST;
215 	else
216 		lw->inv_weight = WMULT_CONST / w;
217 }
218 
219 /*
220  * delta_exec * weight / lw.weight
221  *   OR
222  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
223  *
224  * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
225  * we're guaranteed shift stays positive because inv_weight is guaranteed to
226  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
227  *
228  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
229  * weight/lw.weight <= 1, and therefore our shift will also be positive.
230  */
231 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
232 {
233 	u64 fact = scale_load_down(weight);
234 	int shift = WMULT_SHIFT;
235 
236 	__update_inv_weight(lw);
237 
238 	if (unlikely(fact >> 32)) {
239 		while (fact >> 32) {
240 			fact >>= 1;
241 			shift--;
242 		}
243 	}
244 
245 	/* hint to use a 32x32->64 mul */
246 	fact = (u64)(u32)fact * lw->inv_weight;
247 
248 	while (fact >> 32) {
249 		fact >>= 1;
250 		shift--;
251 	}
252 
253 	return mul_u64_u32_shr(delta_exec, fact, shift);
254 }
255 
256 
257 const struct sched_class fair_sched_class;
258 
259 /**************************************************************
260  * CFS operations on generic schedulable entities:
261  */
262 
263 #ifdef CONFIG_FAIR_GROUP_SCHED
264 
265 /* cpu runqueue to which this cfs_rq is attached */
266 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
267 {
268 	return cfs_rq->rq;
269 }
270 
271 /* An entity is a task if it doesn't "own" a runqueue */
272 #define entity_is_task(se)	(!se->my_q)
273 
274 static inline struct task_struct *task_of(struct sched_entity *se)
275 {
276 	SCHED_WARN_ON(!entity_is_task(se));
277 	return container_of(se, struct task_struct, se);
278 }
279 
280 /* Walk up scheduling entities hierarchy */
281 #define for_each_sched_entity(se) \
282 		for (; se; se = se->parent)
283 
284 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
285 {
286 	return p->se.cfs_rq;
287 }
288 
289 /* runqueue on which this entity is (to be) queued */
290 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
291 {
292 	return se->cfs_rq;
293 }
294 
295 /* runqueue "owned" by this group */
296 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
297 {
298 	return grp->my_q;
299 }
300 
301 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
302 {
303 	if (!cfs_rq->on_list) {
304 		struct rq *rq = rq_of(cfs_rq);
305 		int cpu = cpu_of(rq);
306 		/*
307 		 * Ensure we either appear before our parent (if already
308 		 * enqueued) or force our parent to appear after us when it is
309 		 * enqueued. The fact that we always enqueue bottom-up
310 		 * reduces this to two cases and a special case for the root
311 		 * cfs_rq. Furthermore, it also means that we will always reset
312 		 * tmp_alone_branch either when the branch is connected
313 		 * to a tree or when we reach the beg of the tree
314 		 */
315 		if (cfs_rq->tg->parent &&
316 		    cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
317 			/*
318 			 * If parent is already on the list, we add the child
319 			 * just before. Thanks to circular linked property of
320 			 * the list, this means to put the child at the tail
321 			 * of the list that starts by parent.
322 			 */
323 			list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
324 				&(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
325 			/*
326 			 * The branch is now connected to its tree so we can
327 			 * reset tmp_alone_branch to the beginning of the
328 			 * list.
329 			 */
330 			rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
331 		} else if (!cfs_rq->tg->parent) {
332 			/*
333 			 * cfs rq without parent should be put
334 			 * at the tail of the list.
335 			 */
336 			list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
337 				&rq->leaf_cfs_rq_list);
338 			/*
339 			 * We have reach the beg of a tree so we can reset
340 			 * tmp_alone_branch to the beginning of the list.
341 			 */
342 			rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
343 		} else {
344 			/*
345 			 * The parent has not already been added so we want to
346 			 * make sure that it will be put after us.
347 			 * tmp_alone_branch points to the beg of the branch
348 			 * where we will add parent.
349 			 */
350 			list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
351 				rq->tmp_alone_branch);
352 			/*
353 			 * update tmp_alone_branch to points to the new beg
354 			 * of the branch
355 			 */
356 			rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
357 		}
358 
359 		cfs_rq->on_list = 1;
360 	}
361 }
362 
363 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
364 {
365 	if (cfs_rq->on_list) {
366 		list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
367 		cfs_rq->on_list = 0;
368 	}
369 }
370 
371 /* Iterate thr' all leaf cfs_rq's on a runqueue */
372 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)			\
373 	list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list,	\
374 				 leaf_cfs_rq_list)
375 
376 /* Do the two (enqueued) entities belong to the same group ? */
377 static inline struct cfs_rq *
378 is_same_group(struct sched_entity *se, struct sched_entity *pse)
379 {
380 	if (se->cfs_rq == pse->cfs_rq)
381 		return se->cfs_rq;
382 
383 	return NULL;
384 }
385 
386 static inline struct sched_entity *parent_entity(struct sched_entity *se)
387 {
388 	return se->parent;
389 }
390 
391 static void
392 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
393 {
394 	int se_depth, pse_depth;
395 
396 	/*
397 	 * preemption test can be made between sibling entities who are in the
398 	 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
399 	 * both tasks until we find their ancestors who are siblings of common
400 	 * parent.
401 	 */
402 
403 	/* First walk up until both entities are at same depth */
404 	se_depth = (*se)->depth;
405 	pse_depth = (*pse)->depth;
406 
407 	while (se_depth > pse_depth) {
408 		se_depth--;
409 		*se = parent_entity(*se);
410 	}
411 
412 	while (pse_depth > se_depth) {
413 		pse_depth--;
414 		*pse = parent_entity(*pse);
415 	}
416 
417 	while (!is_same_group(*se, *pse)) {
418 		*se = parent_entity(*se);
419 		*pse = parent_entity(*pse);
420 	}
421 }
422 
423 #else	/* !CONFIG_FAIR_GROUP_SCHED */
424 
425 static inline struct task_struct *task_of(struct sched_entity *se)
426 {
427 	return container_of(se, struct task_struct, se);
428 }
429 
430 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
431 {
432 	return container_of(cfs_rq, struct rq, cfs);
433 }
434 
435 #define entity_is_task(se)	1
436 
437 #define for_each_sched_entity(se) \
438 		for (; se; se = NULL)
439 
440 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
441 {
442 	return &task_rq(p)->cfs;
443 }
444 
445 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
446 {
447 	struct task_struct *p = task_of(se);
448 	struct rq *rq = task_rq(p);
449 
450 	return &rq->cfs;
451 }
452 
453 /* runqueue "owned" by this group */
454 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
455 {
456 	return NULL;
457 }
458 
459 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
460 {
461 }
462 
463 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
464 {
465 }
466 
467 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)	\
468 		for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
469 
470 static inline struct sched_entity *parent_entity(struct sched_entity *se)
471 {
472 	return NULL;
473 }
474 
475 static inline void
476 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
477 {
478 }
479 
480 #endif	/* CONFIG_FAIR_GROUP_SCHED */
481 
482 static __always_inline
483 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
484 
485 /**************************************************************
486  * Scheduling class tree data structure manipulation methods:
487  */
488 
489 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
490 {
491 	s64 delta = (s64)(vruntime - max_vruntime);
492 	if (delta > 0)
493 		max_vruntime = vruntime;
494 
495 	return max_vruntime;
496 }
497 
498 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
499 {
500 	s64 delta = (s64)(vruntime - min_vruntime);
501 	if (delta < 0)
502 		min_vruntime = vruntime;
503 
504 	return min_vruntime;
505 }
506 
507 static inline int entity_before(struct sched_entity *a,
508 				struct sched_entity *b)
509 {
510 	return (s64)(a->vruntime - b->vruntime) < 0;
511 }
512 
513 static void update_min_vruntime(struct cfs_rq *cfs_rq)
514 {
515 	struct sched_entity *curr = cfs_rq->curr;
516 
517 	u64 vruntime = cfs_rq->min_vruntime;
518 
519 	if (curr) {
520 		if (curr->on_rq)
521 			vruntime = curr->vruntime;
522 		else
523 			curr = NULL;
524 	}
525 
526 	if (cfs_rq->rb_leftmost) {
527 		struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
528 						   struct sched_entity,
529 						   run_node);
530 
531 		if (!curr)
532 			vruntime = se->vruntime;
533 		else
534 			vruntime = min_vruntime(vruntime, se->vruntime);
535 	}
536 
537 	/* ensure we never gain time by being placed backwards. */
538 	cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
539 #ifndef CONFIG_64BIT
540 	smp_wmb();
541 	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
542 #endif
543 }
544 
545 /*
546  * Enqueue an entity into the rb-tree:
547  */
548 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
549 {
550 	struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
551 	struct rb_node *parent = NULL;
552 	struct sched_entity *entry;
553 	int leftmost = 1;
554 
555 	/*
556 	 * Find the right place in the rbtree:
557 	 */
558 	while (*link) {
559 		parent = *link;
560 		entry = rb_entry(parent, struct sched_entity, run_node);
561 		/*
562 		 * We dont care about collisions. Nodes with
563 		 * the same key stay together.
564 		 */
565 		if (entity_before(se, entry)) {
566 			link = &parent->rb_left;
567 		} else {
568 			link = &parent->rb_right;
569 			leftmost = 0;
570 		}
571 	}
572 
573 	/*
574 	 * Maintain a cache of leftmost tree entries (it is frequently
575 	 * used):
576 	 */
577 	if (leftmost)
578 		cfs_rq->rb_leftmost = &se->run_node;
579 
580 	rb_link_node(&se->run_node, parent, link);
581 	rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
582 }
583 
584 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
585 {
586 	if (cfs_rq->rb_leftmost == &se->run_node) {
587 		struct rb_node *next_node;
588 
589 		next_node = rb_next(&se->run_node);
590 		cfs_rq->rb_leftmost = next_node;
591 	}
592 
593 	rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
594 }
595 
596 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
597 {
598 	struct rb_node *left = cfs_rq->rb_leftmost;
599 
600 	if (!left)
601 		return NULL;
602 
603 	return rb_entry(left, struct sched_entity, run_node);
604 }
605 
606 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
607 {
608 	struct rb_node *next = rb_next(&se->run_node);
609 
610 	if (!next)
611 		return NULL;
612 
613 	return rb_entry(next, struct sched_entity, run_node);
614 }
615 
616 #ifdef CONFIG_SCHED_DEBUG
617 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
618 {
619 	struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
620 
621 	if (!last)
622 		return NULL;
623 
624 	return rb_entry(last, struct sched_entity, run_node);
625 }
626 
627 /**************************************************************
628  * Scheduling class statistics methods:
629  */
630 
631 int sched_proc_update_handler(struct ctl_table *table, int write,
632 		void __user *buffer, size_t *lenp,
633 		loff_t *ppos)
634 {
635 	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
636 	unsigned int factor = get_update_sysctl_factor();
637 
638 	if (ret || !write)
639 		return ret;
640 
641 	sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
642 					sysctl_sched_min_granularity);
643 
644 #define WRT_SYSCTL(name) \
645 	(normalized_sysctl_##name = sysctl_##name / (factor))
646 	WRT_SYSCTL(sched_min_granularity);
647 	WRT_SYSCTL(sched_latency);
648 	WRT_SYSCTL(sched_wakeup_granularity);
649 #undef WRT_SYSCTL
650 
651 	return 0;
652 }
653 #endif
654 
655 /*
656  * delta /= w
657  */
658 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
659 {
660 	if (unlikely(se->load.weight != NICE_0_LOAD))
661 		delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
662 
663 	return delta;
664 }
665 
666 /*
667  * The idea is to set a period in which each task runs once.
668  *
669  * When there are too many tasks (sched_nr_latency) we have to stretch
670  * this period because otherwise the slices get too small.
671  *
672  * p = (nr <= nl) ? l : l*nr/nl
673  */
674 static u64 __sched_period(unsigned long nr_running)
675 {
676 	if (unlikely(nr_running > sched_nr_latency))
677 		return nr_running * sysctl_sched_min_granularity;
678 	else
679 		return sysctl_sched_latency;
680 }
681 
682 /*
683  * We calculate the wall-time slice from the period by taking a part
684  * proportional to the weight.
685  *
686  * s = p*P[w/rw]
687  */
688 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
689 {
690 	u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
691 
692 	for_each_sched_entity(se) {
693 		struct load_weight *load;
694 		struct load_weight lw;
695 
696 		cfs_rq = cfs_rq_of(se);
697 		load = &cfs_rq->load;
698 
699 		if (unlikely(!se->on_rq)) {
700 			lw = cfs_rq->load;
701 
702 			update_load_add(&lw, se->load.weight);
703 			load = &lw;
704 		}
705 		slice = __calc_delta(slice, se->load.weight, load);
706 	}
707 	return slice;
708 }
709 
710 /*
711  * We calculate the vruntime slice of a to-be-inserted task.
712  *
713  * vs = s/w
714  */
715 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
716 {
717 	return calc_delta_fair(sched_slice(cfs_rq, se), se);
718 }
719 
720 #ifdef CONFIG_SMP
721 
722 #include "sched-pelt.h"
723 
724 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
725 static unsigned long task_h_load(struct task_struct *p);
726 
727 /* Give new sched_entity start runnable values to heavy its load in infant time */
728 void init_entity_runnable_average(struct sched_entity *se)
729 {
730 	struct sched_avg *sa = &se->avg;
731 
732 	sa->last_update_time = 0;
733 	/*
734 	 * sched_avg's period_contrib should be strictly less then 1024, so
735 	 * we give it 1023 to make sure it is almost a period (1024us), and
736 	 * will definitely be update (after enqueue).
737 	 */
738 	sa->period_contrib = 1023;
739 	/*
740 	 * Tasks are intialized with full load to be seen as heavy tasks until
741 	 * they get a chance to stabilize to their real load level.
742 	 * Group entities are intialized with zero load to reflect the fact that
743 	 * nothing has been attached to the task group yet.
744 	 */
745 	if (entity_is_task(se))
746 		sa->load_avg = scale_load_down(se->load.weight);
747 	sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
748 	/*
749 	 * At this point, util_avg won't be used in select_task_rq_fair anyway
750 	 */
751 	sa->util_avg = 0;
752 	sa->util_sum = 0;
753 	/* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
754 }
755 
756 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
757 static void attach_entity_cfs_rq(struct sched_entity *se);
758 
759 /*
760  * With new tasks being created, their initial util_avgs are extrapolated
761  * based on the cfs_rq's current util_avg:
762  *
763  *   util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
764  *
765  * However, in many cases, the above util_avg does not give a desired
766  * value. Moreover, the sum of the util_avgs may be divergent, such
767  * as when the series is a harmonic series.
768  *
769  * To solve this problem, we also cap the util_avg of successive tasks to
770  * only 1/2 of the left utilization budget:
771  *
772  *   util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
773  *
774  * where n denotes the nth task.
775  *
776  * For example, a simplest series from the beginning would be like:
777  *
778  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
779  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
780  *
781  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
782  * if util_avg > util_avg_cap.
783  */
784 void post_init_entity_util_avg(struct sched_entity *se)
785 {
786 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
787 	struct sched_avg *sa = &se->avg;
788 	long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
789 
790 	if (cap > 0) {
791 		if (cfs_rq->avg.util_avg != 0) {
792 			sa->util_avg  = cfs_rq->avg.util_avg * se->load.weight;
793 			sa->util_avg /= (cfs_rq->avg.load_avg + 1);
794 
795 			if (sa->util_avg > cap)
796 				sa->util_avg = cap;
797 		} else {
798 			sa->util_avg = cap;
799 		}
800 		sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
801 	}
802 
803 	if (entity_is_task(se)) {
804 		struct task_struct *p = task_of(se);
805 		if (p->sched_class != &fair_sched_class) {
806 			/*
807 			 * For !fair tasks do:
808 			 *
809 			update_cfs_rq_load_avg(now, cfs_rq, false);
810 			attach_entity_load_avg(cfs_rq, se);
811 			switched_from_fair(rq, p);
812 			 *
813 			 * such that the next switched_to_fair() has the
814 			 * expected state.
815 			 */
816 			se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
817 			return;
818 		}
819 	}
820 
821 	attach_entity_cfs_rq(se);
822 }
823 
824 #else /* !CONFIG_SMP */
825 void init_entity_runnable_average(struct sched_entity *se)
826 {
827 }
828 void post_init_entity_util_avg(struct sched_entity *se)
829 {
830 }
831 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
832 {
833 }
834 #endif /* CONFIG_SMP */
835 
836 /*
837  * Update the current task's runtime statistics.
838  */
839 static void update_curr(struct cfs_rq *cfs_rq)
840 {
841 	struct sched_entity *curr = cfs_rq->curr;
842 	u64 now = rq_clock_task(rq_of(cfs_rq));
843 	u64 delta_exec;
844 
845 	if (unlikely(!curr))
846 		return;
847 
848 	delta_exec = now - curr->exec_start;
849 	if (unlikely((s64)delta_exec <= 0))
850 		return;
851 
852 	curr->exec_start = now;
853 
854 	schedstat_set(curr->statistics.exec_max,
855 		      max(delta_exec, curr->statistics.exec_max));
856 
857 	curr->sum_exec_runtime += delta_exec;
858 	schedstat_add(cfs_rq->exec_clock, delta_exec);
859 
860 	curr->vruntime += calc_delta_fair(delta_exec, curr);
861 	update_min_vruntime(cfs_rq);
862 
863 	if (entity_is_task(curr)) {
864 		struct task_struct *curtask = task_of(curr);
865 
866 		trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
867 		cpuacct_charge(curtask, delta_exec);
868 		account_group_exec_runtime(curtask, delta_exec);
869 	}
870 
871 	account_cfs_rq_runtime(cfs_rq, delta_exec);
872 }
873 
874 static void update_curr_fair(struct rq *rq)
875 {
876 	update_curr(cfs_rq_of(&rq->curr->se));
877 }
878 
879 static inline void
880 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
881 {
882 	u64 wait_start, prev_wait_start;
883 
884 	if (!schedstat_enabled())
885 		return;
886 
887 	wait_start = rq_clock(rq_of(cfs_rq));
888 	prev_wait_start = schedstat_val(se->statistics.wait_start);
889 
890 	if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
891 	    likely(wait_start > prev_wait_start))
892 		wait_start -= prev_wait_start;
893 
894 	schedstat_set(se->statistics.wait_start, wait_start);
895 }
896 
897 static inline void
898 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
899 {
900 	struct task_struct *p;
901 	u64 delta;
902 
903 	if (!schedstat_enabled())
904 		return;
905 
906 	delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
907 
908 	if (entity_is_task(se)) {
909 		p = task_of(se);
910 		if (task_on_rq_migrating(p)) {
911 			/*
912 			 * Preserve migrating task's wait time so wait_start
913 			 * time stamp can be adjusted to accumulate wait time
914 			 * prior to migration.
915 			 */
916 			schedstat_set(se->statistics.wait_start, delta);
917 			return;
918 		}
919 		trace_sched_stat_wait(p, delta);
920 	}
921 
922 	schedstat_set(se->statistics.wait_max,
923 		      max(schedstat_val(se->statistics.wait_max), delta));
924 	schedstat_inc(se->statistics.wait_count);
925 	schedstat_add(se->statistics.wait_sum, delta);
926 	schedstat_set(se->statistics.wait_start, 0);
927 }
928 
929 static inline void
930 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
931 {
932 	struct task_struct *tsk = NULL;
933 	u64 sleep_start, block_start;
934 
935 	if (!schedstat_enabled())
936 		return;
937 
938 	sleep_start = schedstat_val(se->statistics.sleep_start);
939 	block_start = schedstat_val(se->statistics.block_start);
940 
941 	if (entity_is_task(se))
942 		tsk = task_of(se);
943 
944 	if (sleep_start) {
945 		u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
946 
947 		if ((s64)delta < 0)
948 			delta = 0;
949 
950 		if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
951 			schedstat_set(se->statistics.sleep_max, delta);
952 
953 		schedstat_set(se->statistics.sleep_start, 0);
954 		schedstat_add(se->statistics.sum_sleep_runtime, delta);
955 
956 		if (tsk) {
957 			account_scheduler_latency(tsk, delta >> 10, 1);
958 			trace_sched_stat_sleep(tsk, delta);
959 		}
960 	}
961 	if (block_start) {
962 		u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
963 
964 		if ((s64)delta < 0)
965 			delta = 0;
966 
967 		if (unlikely(delta > schedstat_val(se->statistics.block_max)))
968 			schedstat_set(se->statistics.block_max, delta);
969 
970 		schedstat_set(se->statistics.block_start, 0);
971 		schedstat_add(se->statistics.sum_sleep_runtime, delta);
972 
973 		if (tsk) {
974 			if (tsk->in_iowait) {
975 				schedstat_add(se->statistics.iowait_sum, delta);
976 				schedstat_inc(se->statistics.iowait_count);
977 				trace_sched_stat_iowait(tsk, delta);
978 			}
979 
980 			trace_sched_stat_blocked(tsk, delta);
981 
982 			/*
983 			 * Blocking time is in units of nanosecs, so shift by
984 			 * 20 to get a milliseconds-range estimation of the
985 			 * amount of time that the task spent sleeping:
986 			 */
987 			if (unlikely(prof_on == SLEEP_PROFILING)) {
988 				profile_hits(SLEEP_PROFILING,
989 						(void *)get_wchan(tsk),
990 						delta >> 20);
991 			}
992 			account_scheduler_latency(tsk, delta >> 10, 0);
993 		}
994 	}
995 }
996 
997 /*
998  * Task is being enqueued - update stats:
999  */
1000 static inline void
1001 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1002 {
1003 	if (!schedstat_enabled())
1004 		return;
1005 
1006 	/*
1007 	 * Are we enqueueing a waiting task? (for current tasks
1008 	 * a dequeue/enqueue event is a NOP)
1009 	 */
1010 	if (se != cfs_rq->curr)
1011 		update_stats_wait_start(cfs_rq, se);
1012 
1013 	if (flags & ENQUEUE_WAKEUP)
1014 		update_stats_enqueue_sleeper(cfs_rq, se);
1015 }
1016 
1017 static inline void
1018 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1019 {
1020 
1021 	if (!schedstat_enabled())
1022 		return;
1023 
1024 	/*
1025 	 * Mark the end of the wait period if dequeueing a
1026 	 * waiting task:
1027 	 */
1028 	if (se != cfs_rq->curr)
1029 		update_stats_wait_end(cfs_rq, se);
1030 
1031 	if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1032 		struct task_struct *tsk = task_of(se);
1033 
1034 		if (tsk->state & TASK_INTERRUPTIBLE)
1035 			schedstat_set(se->statistics.sleep_start,
1036 				      rq_clock(rq_of(cfs_rq)));
1037 		if (tsk->state & TASK_UNINTERRUPTIBLE)
1038 			schedstat_set(se->statistics.block_start,
1039 				      rq_clock(rq_of(cfs_rq)));
1040 	}
1041 }
1042 
1043 /*
1044  * We are picking a new current task - update its stats:
1045  */
1046 static inline void
1047 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1048 {
1049 	/*
1050 	 * We are starting a new run period:
1051 	 */
1052 	se->exec_start = rq_clock_task(rq_of(cfs_rq));
1053 }
1054 
1055 /**************************************************
1056  * Scheduling class queueing methods:
1057  */
1058 
1059 #ifdef CONFIG_NUMA_BALANCING
1060 /*
1061  * Approximate time to scan a full NUMA task in ms. The task scan period is
1062  * calculated based on the tasks virtual memory size and
1063  * numa_balancing_scan_size.
1064  */
1065 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1066 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1067 
1068 /* Portion of address space to scan in MB */
1069 unsigned int sysctl_numa_balancing_scan_size = 256;
1070 
1071 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1072 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1073 
1074 static unsigned int task_nr_scan_windows(struct task_struct *p)
1075 {
1076 	unsigned long rss = 0;
1077 	unsigned long nr_scan_pages;
1078 
1079 	/*
1080 	 * Calculations based on RSS as non-present and empty pages are skipped
1081 	 * by the PTE scanner and NUMA hinting faults should be trapped based
1082 	 * on resident pages
1083 	 */
1084 	nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1085 	rss = get_mm_rss(p->mm);
1086 	if (!rss)
1087 		rss = nr_scan_pages;
1088 
1089 	rss = round_up(rss, nr_scan_pages);
1090 	return rss / nr_scan_pages;
1091 }
1092 
1093 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1094 #define MAX_SCAN_WINDOW 2560
1095 
1096 static unsigned int task_scan_min(struct task_struct *p)
1097 {
1098 	unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1099 	unsigned int scan, floor;
1100 	unsigned int windows = 1;
1101 
1102 	if (scan_size < MAX_SCAN_WINDOW)
1103 		windows = MAX_SCAN_WINDOW / scan_size;
1104 	floor = 1000 / windows;
1105 
1106 	scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1107 	return max_t(unsigned int, floor, scan);
1108 }
1109 
1110 static unsigned int task_scan_max(struct task_struct *p)
1111 {
1112 	unsigned int smin = task_scan_min(p);
1113 	unsigned int smax;
1114 
1115 	/* Watch for min being lower than max due to floor calculations */
1116 	smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1117 	return max(smin, smax);
1118 }
1119 
1120 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1121 {
1122 	rq->nr_numa_running += (p->numa_preferred_nid != -1);
1123 	rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1124 }
1125 
1126 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1127 {
1128 	rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1129 	rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1130 }
1131 
1132 struct numa_group {
1133 	atomic_t refcount;
1134 
1135 	spinlock_t lock; /* nr_tasks, tasks */
1136 	int nr_tasks;
1137 	pid_t gid;
1138 	int active_nodes;
1139 
1140 	struct rcu_head rcu;
1141 	unsigned long total_faults;
1142 	unsigned long max_faults_cpu;
1143 	/*
1144 	 * Faults_cpu is used to decide whether memory should move
1145 	 * towards the CPU. As a consequence, these stats are weighted
1146 	 * more by CPU use than by memory faults.
1147 	 */
1148 	unsigned long *faults_cpu;
1149 	unsigned long faults[0];
1150 };
1151 
1152 /* Shared or private faults. */
1153 #define NR_NUMA_HINT_FAULT_TYPES 2
1154 
1155 /* Memory and CPU locality */
1156 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1157 
1158 /* Averaged statistics, and temporary buffers. */
1159 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1160 
1161 pid_t task_numa_group_id(struct task_struct *p)
1162 {
1163 	return p->numa_group ? p->numa_group->gid : 0;
1164 }
1165 
1166 /*
1167  * The averaged statistics, shared & private, memory & cpu,
1168  * occupy the first half of the array. The second half of the
1169  * array is for current counters, which are averaged into the
1170  * first set by task_numa_placement.
1171  */
1172 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1173 {
1174 	return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1175 }
1176 
1177 static inline unsigned long task_faults(struct task_struct *p, int nid)
1178 {
1179 	if (!p->numa_faults)
1180 		return 0;
1181 
1182 	return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1183 		p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1184 }
1185 
1186 static inline unsigned long group_faults(struct task_struct *p, int nid)
1187 {
1188 	if (!p->numa_group)
1189 		return 0;
1190 
1191 	return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1192 		p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1193 }
1194 
1195 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1196 {
1197 	return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1198 		group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1199 }
1200 
1201 /*
1202  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1203  * considered part of a numa group's pseudo-interleaving set. Migrations
1204  * between these nodes are slowed down, to allow things to settle down.
1205  */
1206 #define ACTIVE_NODE_FRACTION 3
1207 
1208 static bool numa_is_active_node(int nid, struct numa_group *ng)
1209 {
1210 	return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1211 }
1212 
1213 /* Handle placement on systems where not all nodes are directly connected. */
1214 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1215 					int maxdist, bool task)
1216 {
1217 	unsigned long score = 0;
1218 	int node;
1219 
1220 	/*
1221 	 * All nodes are directly connected, and the same distance
1222 	 * from each other. No need for fancy placement algorithms.
1223 	 */
1224 	if (sched_numa_topology_type == NUMA_DIRECT)
1225 		return 0;
1226 
1227 	/*
1228 	 * This code is called for each node, introducing N^2 complexity,
1229 	 * which should be ok given the number of nodes rarely exceeds 8.
1230 	 */
1231 	for_each_online_node(node) {
1232 		unsigned long faults;
1233 		int dist = node_distance(nid, node);
1234 
1235 		/*
1236 		 * The furthest away nodes in the system are not interesting
1237 		 * for placement; nid was already counted.
1238 		 */
1239 		if (dist == sched_max_numa_distance || node == nid)
1240 			continue;
1241 
1242 		/*
1243 		 * On systems with a backplane NUMA topology, compare groups
1244 		 * of nodes, and move tasks towards the group with the most
1245 		 * memory accesses. When comparing two nodes at distance
1246 		 * "hoplimit", only nodes closer by than "hoplimit" are part
1247 		 * of each group. Skip other nodes.
1248 		 */
1249 		if (sched_numa_topology_type == NUMA_BACKPLANE &&
1250 					dist > maxdist)
1251 			continue;
1252 
1253 		/* Add up the faults from nearby nodes. */
1254 		if (task)
1255 			faults = task_faults(p, node);
1256 		else
1257 			faults = group_faults(p, node);
1258 
1259 		/*
1260 		 * On systems with a glueless mesh NUMA topology, there are
1261 		 * no fixed "groups of nodes". Instead, nodes that are not
1262 		 * directly connected bounce traffic through intermediate
1263 		 * nodes; a numa_group can occupy any set of nodes.
1264 		 * The further away a node is, the less the faults count.
1265 		 * This seems to result in good task placement.
1266 		 */
1267 		if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1268 			faults *= (sched_max_numa_distance - dist);
1269 			faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1270 		}
1271 
1272 		score += faults;
1273 	}
1274 
1275 	return score;
1276 }
1277 
1278 /*
1279  * These return the fraction of accesses done by a particular task, or
1280  * task group, on a particular numa node.  The group weight is given a
1281  * larger multiplier, in order to group tasks together that are almost
1282  * evenly spread out between numa nodes.
1283  */
1284 static inline unsigned long task_weight(struct task_struct *p, int nid,
1285 					int dist)
1286 {
1287 	unsigned long faults, total_faults;
1288 
1289 	if (!p->numa_faults)
1290 		return 0;
1291 
1292 	total_faults = p->total_numa_faults;
1293 
1294 	if (!total_faults)
1295 		return 0;
1296 
1297 	faults = task_faults(p, nid);
1298 	faults += score_nearby_nodes(p, nid, dist, true);
1299 
1300 	return 1000 * faults / total_faults;
1301 }
1302 
1303 static inline unsigned long group_weight(struct task_struct *p, int nid,
1304 					 int dist)
1305 {
1306 	unsigned long faults, total_faults;
1307 
1308 	if (!p->numa_group)
1309 		return 0;
1310 
1311 	total_faults = p->numa_group->total_faults;
1312 
1313 	if (!total_faults)
1314 		return 0;
1315 
1316 	faults = group_faults(p, nid);
1317 	faults += score_nearby_nodes(p, nid, dist, false);
1318 
1319 	return 1000 * faults / total_faults;
1320 }
1321 
1322 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1323 				int src_nid, int dst_cpu)
1324 {
1325 	struct numa_group *ng = p->numa_group;
1326 	int dst_nid = cpu_to_node(dst_cpu);
1327 	int last_cpupid, this_cpupid;
1328 
1329 	this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1330 
1331 	/*
1332 	 * Multi-stage node selection is used in conjunction with a periodic
1333 	 * migration fault to build a temporal task<->page relation. By using
1334 	 * a two-stage filter we remove short/unlikely relations.
1335 	 *
1336 	 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1337 	 * a task's usage of a particular page (n_p) per total usage of this
1338 	 * page (n_t) (in a given time-span) to a probability.
1339 	 *
1340 	 * Our periodic faults will sample this probability and getting the
1341 	 * same result twice in a row, given these samples are fully
1342 	 * independent, is then given by P(n)^2, provided our sample period
1343 	 * is sufficiently short compared to the usage pattern.
1344 	 *
1345 	 * This quadric squishes small probabilities, making it less likely we
1346 	 * act on an unlikely task<->page relation.
1347 	 */
1348 	last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1349 	if (!cpupid_pid_unset(last_cpupid) &&
1350 				cpupid_to_nid(last_cpupid) != dst_nid)
1351 		return false;
1352 
1353 	/* Always allow migrate on private faults */
1354 	if (cpupid_match_pid(p, last_cpupid))
1355 		return true;
1356 
1357 	/* A shared fault, but p->numa_group has not been set up yet. */
1358 	if (!ng)
1359 		return true;
1360 
1361 	/*
1362 	 * Destination node is much more heavily used than the source
1363 	 * node? Allow migration.
1364 	 */
1365 	if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1366 					ACTIVE_NODE_FRACTION)
1367 		return true;
1368 
1369 	/*
1370 	 * Distribute memory according to CPU & memory use on each node,
1371 	 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1372 	 *
1373 	 * faults_cpu(dst)   3   faults_cpu(src)
1374 	 * --------------- * - > ---------------
1375 	 * faults_mem(dst)   4   faults_mem(src)
1376 	 */
1377 	return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1378 	       group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1379 }
1380 
1381 static unsigned long weighted_cpuload(const int cpu);
1382 static unsigned long source_load(int cpu, int type);
1383 static unsigned long target_load(int cpu, int type);
1384 static unsigned long capacity_of(int cpu);
1385 
1386 /* Cached statistics for all CPUs within a node */
1387 struct numa_stats {
1388 	unsigned long nr_running;
1389 	unsigned long load;
1390 
1391 	/* Total compute capacity of CPUs on a node */
1392 	unsigned long compute_capacity;
1393 
1394 	/* Approximate capacity in terms of runnable tasks on a node */
1395 	unsigned long task_capacity;
1396 	int has_free_capacity;
1397 };
1398 
1399 /*
1400  * XXX borrowed from update_sg_lb_stats
1401  */
1402 static void update_numa_stats(struct numa_stats *ns, int nid)
1403 {
1404 	int smt, cpu, cpus = 0;
1405 	unsigned long capacity;
1406 
1407 	memset(ns, 0, sizeof(*ns));
1408 	for_each_cpu(cpu, cpumask_of_node(nid)) {
1409 		struct rq *rq = cpu_rq(cpu);
1410 
1411 		ns->nr_running += rq->nr_running;
1412 		ns->load += weighted_cpuload(cpu);
1413 		ns->compute_capacity += capacity_of(cpu);
1414 
1415 		cpus++;
1416 	}
1417 
1418 	/*
1419 	 * If we raced with hotplug and there are no CPUs left in our mask
1420 	 * the @ns structure is NULL'ed and task_numa_compare() will
1421 	 * not find this node attractive.
1422 	 *
1423 	 * We'll either bail at !has_free_capacity, or we'll detect a huge
1424 	 * imbalance and bail there.
1425 	 */
1426 	if (!cpus)
1427 		return;
1428 
1429 	/* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1430 	smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1431 	capacity = cpus / smt; /* cores */
1432 
1433 	ns->task_capacity = min_t(unsigned, capacity,
1434 		DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1435 	ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1436 }
1437 
1438 struct task_numa_env {
1439 	struct task_struct *p;
1440 
1441 	int src_cpu, src_nid;
1442 	int dst_cpu, dst_nid;
1443 
1444 	struct numa_stats src_stats, dst_stats;
1445 
1446 	int imbalance_pct;
1447 	int dist;
1448 
1449 	struct task_struct *best_task;
1450 	long best_imp;
1451 	int best_cpu;
1452 };
1453 
1454 static void task_numa_assign(struct task_numa_env *env,
1455 			     struct task_struct *p, long imp)
1456 {
1457 	if (env->best_task)
1458 		put_task_struct(env->best_task);
1459 	if (p)
1460 		get_task_struct(p);
1461 
1462 	env->best_task = p;
1463 	env->best_imp = imp;
1464 	env->best_cpu = env->dst_cpu;
1465 }
1466 
1467 static bool load_too_imbalanced(long src_load, long dst_load,
1468 				struct task_numa_env *env)
1469 {
1470 	long imb, old_imb;
1471 	long orig_src_load, orig_dst_load;
1472 	long src_capacity, dst_capacity;
1473 
1474 	/*
1475 	 * The load is corrected for the CPU capacity available on each node.
1476 	 *
1477 	 * src_load        dst_load
1478 	 * ------------ vs ---------
1479 	 * src_capacity    dst_capacity
1480 	 */
1481 	src_capacity = env->src_stats.compute_capacity;
1482 	dst_capacity = env->dst_stats.compute_capacity;
1483 
1484 	/* We care about the slope of the imbalance, not the direction. */
1485 	if (dst_load < src_load)
1486 		swap(dst_load, src_load);
1487 
1488 	/* Is the difference below the threshold? */
1489 	imb = dst_load * src_capacity * 100 -
1490 	      src_load * dst_capacity * env->imbalance_pct;
1491 	if (imb <= 0)
1492 		return false;
1493 
1494 	/*
1495 	 * The imbalance is above the allowed threshold.
1496 	 * Compare it with the old imbalance.
1497 	 */
1498 	orig_src_load = env->src_stats.load;
1499 	orig_dst_load = env->dst_stats.load;
1500 
1501 	if (orig_dst_load < orig_src_load)
1502 		swap(orig_dst_load, orig_src_load);
1503 
1504 	old_imb = orig_dst_load * src_capacity * 100 -
1505 		  orig_src_load * dst_capacity * env->imbalance_pct;
1506 
1507 	/* Would this change make things worse? */
1508 	return (imb > old_imb);
1509 }
1510 
1511 /*
1512  * This checks if the overall compute and NUMA accesses of the system would
1513  * be improved if the source tasks was migrated to the target dst_cpu taking
1514  * into account that it might be best if task running on the dst_cpu should
1515  * be exchanged with the source task
1516  */
1517 static void task_numa_compare(struct task_numa_env *env,
1518 			      long taskimp, long groupimp)
1519 {
1520 	struct rq *src_rq = cpu_rq(env->src_cpu);
1521 	struct rq *dst_rq = cpu_rq(env->dst_cpu);
1522 	struct task_struct *cur;
1523 	long src_load, dst_load;
1524 	long load;
1525 	long imp = env->p->numa_group ? groupimp : taskimp;
1526 	long moveimp = imp;
1527 	int dist = env->dist;
1528 
1529 	rcu_read_lock();
1530 	cur = task_rcu_dereference(&dst_rq->curr);
1531 	if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1532 		cur = NULL;
1533 
1534 	/*
1535 	 * Because we have preemption enabled we can get migrated around and
1536 	 * end try selecting ourselves (current == env->p) as a swap candidate.
1537 	 */
1538 	if (cur == env->p)
1539 		goto unlock;
1540 
1541 	/*
1542 	 * "imp" is the fault differential for the source task between the
1543 	 * source and destination node. Calculate the total differential for
1544 	 * the source task and potential destination task. The more negative
1545 	 * the value is, the more rmeote accesses that would be expected to
1546 	 * be incurred if the tasks were swapped.
1547 	 */
1548 	if (cur) {
1549 		/* Skip this swap candidate if cannot move to the source cpu */
1550 		if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1551 			goto unlock;
1552 
1553 		/*
1554 		 * If dst and source tasks are in the same NUMA group, or not
1555 		 * in any group then look only at task weights.
1556 		 */
1557 		if (cur->numa_group == env->p->numa_group) {
1558 			imp = taskimp + task_weight(cur, env->src_nid, dist) -
1559 			      task_weight(cur, env->dst_nid, dist);
1560 			/*
1561 			 * Add some hysteresis to prevent swapping the
1562 			 * tasks within a group over tiny differences.
1563 			 */
1564 			if (cur->numa_group)
1565 				imp -= imp/16;
1566 		} else {
1567 			/*
1568 			 * Compare the group weights. If a task is all by
1569 			 * itself (not part of a group), use the task weight
1570 			 * instead.
1571 			 */
1572 			if (cur->numa_group)
1573 				imp += group_weight(cur, env->src_nid, dist) -
1574 				       group_weight(cur, env->dst_nid, dist);
1575 			else
1576 				imp += task_weight(cur, env->src_nid, dist) -
1577 				       task_weight(cur, env->dst_nid, dist);
1578 		}
1579 	}
1580 
1581 	if (imp <= env->best_imp && moveimp <= env->best_imp)
1582 		goto unlock;
1583 
1584 	if (!cur) {
1585 		/* Is there capacity at our destination? */
1586 		if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1587 		    !env->dst_stats.has_free_capacity)
1588 			goto unlock;
1589 
1590 		goto balance;
1591 	}
1592 
1593 	/* Balance doesn't matter much if we're running a task per cpu */
1594 	if (imp > env->best_imp && src_rq->nr_running == 1 &&
1595 			dst_rq->nr_running == 1)
1596 		goto assign;
1597 
1598 	/*
1599 	 * In the overloaded case, try and keep the load balanced.
1600 	 */
1601 balance:
1602 	load = task_h_load(env->p);
1603 	dst_load = env->dst_stats.load + load;
1604 	src_load = env->src_stats.load - load;
1605 
1606 	if (moveimp > imp && moveimp > env->best_imp) {
1607 		/*
1608 		 * If the improvement from just moving env->p direction is
1609 		 * better than swapping tasks around, check if a move is
1610 		 * possible. Store a slightly smaller score than moveimp,
1611 		 * so an actually idle CPU will win.
1612 		 */
1613 		if (!load_too_imbalanced(src_load, dst_load, env)) {
1614 			imp = moveimp - 1;
1615 			cur = NULL;
1616 			goto assign;
1617 		}
1618 	}
1619 
1620 	if (imp <= env->best_imp)
1621 		goto unlock;
1622 
1623 	if (cur) {
1624 		load = task_h_load(cur);
1625 		dst_load -= load;
1626 		src_load += load;
1627 	}
1628 
1629 	if (load_too_imbalanced(src_load, dst_load, env))
1630 		goto unlock;
1631 
1632 	/*
1633 	 * One idle CPU per node is evaluated for a task numa move.
1634 	 * Call select_idle_sibling to maybe find a better one.
1635 	 */
1636 	if (!cur) {
1637 		/*
1638 		 * select_idle_siblings() uses an per-cpu cpumask that
1639 		 * can be used from IRQ context.
1640 		 */
1641 		local_irq_disable();
1642 		env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1643 						   env->dst_cpu);
1644 		local_irq_enable();
1645 	}
1646 
1647 assign:
1648 	task_numa_assign(env, cur, imp);
1649 unlock:
1650 	rcu_read_unlock();
1651 }
1652 
1653 static void task_numa_find_cpu(struct task_numa_env *env,
1654 				long taskimp, long groupimp)
1655 {
1656 	int cpu;
1657 
1658 	for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1659 		/* Skip this CPU if the source task cannot migrate */
1660 		if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1661 			continue;
1662 
1663 		env->dst_cpu = cpu;
1664 		task_numa_compare(env, taskimp, groupimp);
1665 	}
1666 }
1667 
1668 /* Only move tasks to a NUMA node less busy than the current node. */
1669 static bool numa_has_capacity(struct task_numa_env *env)
1670 {
1671 	struct numa_stats *src = &env->src_stats;
1672 	struct numa_stats *dst = &env->dst_stats;
1673 
1674 	if (src->has_free_capacity && !dst->has_free_capacity)
1675 		return false;
1676 
1677 	/*
1678 	 * Only consider a task move if the source has a higher load
1679 	 * than the destination, corrected for CPU capacity on each node.
1680 	 *
1681 	 *      src->load                dst->load
1682 	 * --------------------- vs ---------------------
1683 	 * src->compute_capacity    dst->compute_capacity
1684 	 */
1685 	if (src->load * dst->compute_capacity * env->imbalance_pct >
1686 
1687 	    dst->load * src->compute_capacity * 100)
1688 		return true;
1689 
1690 	return false;
1691 }
1692 
1693 static int task_numa_migrate(struct task_struct *p)
1694 {
1695 	struct task_numa_env env = {
1696 		.p = p,
1697 
1698 		.src_cpu = task_cpu(p),
1699 		.src_nid = task_node(p),
1700 
1701 		.imbalance_pct = 112,
1702 
1703 		.best_task = NULL,
1704 		.best_imp = 0,
1705 		.best_cpu = -1,
1706 	};
1707 	struct sched_domain *sd;
1708 	unsigned long taskweight, groupweight;
1709 	int nid, ret, dist;
1710 	long taskimp, groupimp;
1711 
1712 	/*
1713 	 * Pick the lowest SD_NUMA domain, as that would have the smallest
1714 	 * imbalance and would be the first to start moving tasks about.
1715 	 *
1716 	 * And we want to avoid any moving of tasks about, as that would create
1717 	 * random movement of tasks -- counter the numa conditions we're trying
1718 	 * to satisfy here.
1719 	 */
1720 	rcu_read_lock();
1721 	sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1722 	if (sd)
1723 		env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1724 	rcu_read_unlock();
1725 
1726 	/*
1727 	 * Cpusets can break the scheduler domain tree into smaller
1728 	 * balance domains, some of which do not cross NUMA boundaries.
1729 	 * Tasks that are "trapped" in such domains cannot be migrated
1730 	 * elsewhere, so there is no point in (re)trying.
1731 	 */
1732 	if (unlikely(!sd)) {
1733 		p->numa_preferred_nid = task_node(p);
1734 		return -EINVAL;
1735 	}
1736 
1737 	env.dst_nid = p->numa_preferred_nid;
1738 	dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1739 	taskweight = task_weight(p, env.src_nid, dist);
1740 	groupweight = group_weight(p, env.src_nid, dist);
1741 	update_numa_stats(&env.src_stats, env.src_nid);
1742 	taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1743 	groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1744 	update_numa_stats(&env.dst_stats, env.dst_nid);
1745 
1746 	/* Try to find a spot on the preferred nid. */
1747 	if (numa_has_capacity(&env))
1748 		task_numa_find_cpu(&env, taskimp, groupimp);
1749 
1750 	/*
1751 	 * Look at other nodes in these cases:
1752 	 * - there is no space available on the preferred_nid
1753 	 * - the task is part of a numa_group that is interleaved across
1754 	 *   multiple NUMA nodes; in order to better consolidate the group,
1755 	 *   we need to check other locations.
1756 	 */
1757 	if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1758 		for_each_online_node(nid) {
1759 			if (nid == env.src_nid || nid == p->numa_preferred_nid)
1760 				continue;
1761 
1762 			dist = node_distance(env.src_nid, env.dst_nid);
1763 			if (sched_numa_topology_type == NUMA_BACKPLANE &&
1764 						dist != env.dist) {
1765 				taskweight = task_weight(p, env.src_nid, dist);
1766 				groupweight = group_weight(p, env.src_nid, dist);
1767 			}
1768 
1769 			/* Only consider nodes where both task and groups benefit */
1770 			taskimp = task_weight(p, nid, dist) - taskweight;
1771 			groupimp = group_weight(p, nid, dist) - groupweight;
1772 			if (taskimp < 0 && groupimp < 0)
1773 				continue;
1774 
1775 			env.dist = dist;
1776 			env.dst_nid = nid;
1777 			update_numa_stats(&env.dst_stats, env.dst_nid);
1778 			if (numa_has_capacity(&env))
1779 				task_numa_find_cpu(&env, taskimp, groupimp);
1780 		}
1781 	}
1782 
1783 	/*
1784 	 * If the task is part of a workload that spans multiple NUMA nodes,
1785 	 * and is migrating into one of the workload's active nodes, remember
1786 	 * this node as the task's preferred numa node, so the workload can
1787 	 * settle down.
1788 	 * A task that migrated to a second choice node will be better off
1789 	 * trying for a better one later. Do not set the preferred node here.
1790 	 */
1791 	if (p->numa_group) {
1792 		struct numa_group *ng = p->numa_group;
1793 
1794 		if (env.best_cpu == -1)
1795 			nid = env.src_nid;
1796 		else
1797 			nid = env.dst_nid;
1798 
1799 		if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1800 			sched_setnuma(p, env.dst_nid);
1801 	}
1802 
1803 	/* No better CPU than the current one was found. */
1804 	if (env.best_cpu == -1)
1805 		return -EAGAIN;
1806 
1807 	/*
1808 	 * Reset the scan period if the task is being rescheduled on an
1809 	 * alternative node to recheck if the tasks is now properly placed.
1810 	 */
1811 	p->numa_scan_period = task_scan_min(p);
1812 
1813 	if (env.best_task == NULL) {
1814 		ret = migrate_task_to(p, env.best_cpu);
1815 		if (ret != 0)
1816 			trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1817 		return ret;
1818 	}
1819 
1820 	ret = migrate_swap(p, env.best_task);
1821 	if (ret != 0)
1822 		trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1823 	put_task_struct(env.best_task);
1824 	return ret;
1825 }
1826 
1827 /* Attempt to migrate a task to a CPU on the preferred node. */
1828 static void numa_migrate_preferred(struct task_struct *p)
1829 {
1830 	unsigned long interval = HZ;
1831 
1832 	/* This task has no NUMA fault statistics yet */
1833 	if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1834 		return;
1835 
1836 	/* Periodically retry migrating the task to the preferred node */
1837 	interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1838 	p->numa_migrate_retry = jiffies + interval;
1839 
1840 	/* Success if task is already running on preferred CPU */
1841 	if (task_node(p) == p->numa_preferred_nid)
1842 		return;
1843 
1844 	/* Otherwise, try migrate to a CPU on the preferred node */
1845 	task_numa_migrate(p);
1846 }
1847 
1848 /*
1849  * Find out how many nodes on the workload is actively running on. Do this by
1850  * tracking the nodes from which NUMA hinting faults are triggered. This can
1851  * be different from the set of nodes where the workload's memory is currently
1852  * located.
1853  */
1854 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1855 {
1856 	unsigned long faults, max_faults = 0;
1857 	int nid, active_nodes = 0;
1858 
1859 	for_each_online_node(nid) {
1860 		faults = group_faults_cpu(numa_group, nid);
1861 		if (faults > max_faults)
1862 			max_faults = faults;
1863 	}
1864 
1865 	for_each_online_node(nid) {
1866 		faults = group_faults_cpu(numa_group, nid);
1867 		if (faults * ACTIVE_NODE_FRACTION > max_faults)
1868 			active_nodes++;
1869 	}
1870 
1871 	numa_group->max_faults_cpu = max_faults;
1872 	numa_group->active_nodes = active_nodes;
1873 }
1874 
1875 /*
1876  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1877  * increments. The more local the fault statistics are, the higher the scan
1878  * period will be for the next scan window. If local/(local+remote) ratio is
1879  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1880  * the scan period will decrease. Aim for 70% local accesses.
1881  */
1882 #define NUMA_PERIOD_SLOTS 10
1883 #define NUMA_PERIOD_THRESHOLD 7
1884 
1885 /*
1886  * Increase the scan period (slow down scanning) if the majority of
1887  * our memory is already on our local node, or if the majority of
1888  * the page accesses are shared with other processes.
1889  * Otherwise, decrease the scan period.
1890  */
1891 static void update_task_scan_period(struct task_struct *p,
1892 			unsigned long shared, unsigned long private)
1893 {
1894 	unsigned int period_slot;
1895 	int ratio;
1896 	int diff;
1897 
1898 	unsigned long remote = p->numa_faults_locality[0];
1899 	unsigned long local = p->numa_faults_locality[1];
1900 
1901 	/*
1902 	 * If there were no record hinting faults then either the task is
1903 	 * completely idle or all activity is areas that are not of interest
1904 	 * to automatic numa balancing. Related to that, if there were failed
1905 	 * migration then it implies we are migrating too quickly or the local
1906 	 * node is overloaded. In either case, scan slower
1907 	 */
1908 	if (local + shared == 0 || p->numa_faults_locality[2]) {
1909 		p->numa_scan_period = min(p->numa_scan_period_max,
1910 			p->numa_scan_period << 1);
1911 
1912 		p->mm->numa_next_scan = jiffies +
1913 			msecs_to_jiffies(p->numa_scan_period);
1914 
1915 		return;
1916 	}
1917 
1918 	/*
1919 	 * Prepare to scale scan period relative to the current period.
1920 	 *	 == NUMA_PERIOD_THRESHOLD scan period stays the same
1921 	 *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1922 	 *	 >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1923 	 */
1924 	period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1925 	ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1926 	if (ratio >= NUMA_PERIOD_THRESHOLD) {
1927 		int slot = ratio - NUMA_PERIOD_THRESHOLD;
1928 		if (!slot)
1929 			slot = 1;
1930 		diff = slot * period_slot;
1931 	} else {
1932 		diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1933 
1934 		/*
1935 		 * Scale scan rate increases based on sharing. There is an
1936 		 * inverse relationship between the degree of sharing and
1937 		 * the adjustment made to the scanning period. Broadly
1938 		 * speaking the intent is that there is little point
1939 		 * scanning faster if shared accesses dominate as it may
1940 		 * simply bounce migrations uselessly
1941 		 */
1942 		ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1943 		diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1944 	}
1945 
1946 	p->numa_scan_period = clamp(p->numa_scan_period + diff,
1947 			task_scan_min(p), task_scan_max(p));
1948 	memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1949 }
1950 
1951 /*
1952  * Get the fraction of time the task has been running since the last
1953  * NUMA placement cycle. The scheduler keeps similar statistics, but
1954  * decays those on a 32ms period, which is orders of magnitude off
1955  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1956  * stats only if the task is so new there are no NUMA statistics yet.
1957  */
1958 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1959 {
1960 	u64 runtime, delta, now;
1961 	/* Use the start of this time slice to avoid calculations. */
1962 	now = p->se.exec_start;
1963 	runtime = p->se.sum_exec_runtime;
1964 
1965 	if (p->last_task_numa_placement) {
1966 		delta = runtime - p->last_sum_exec_runtime;
1967 		*period = now - p->last_task_numa_placement;
1968 	} else {
1969 		delta = p->se.avg.load_sum / p->se.load.weight;
1970 		*period = LOAD_AVG_MAX;
1971 	}
1972 
1973 	p->last_sum_exec_runtime = runtime;
1974 	p->last_task_numa_placement = now;
1975 
1976 	return delta;
1977 }
1978 
1979 /*
1980  * Determine the preferred nid for a task in a numa_group. This needs to
1981  * be done in a way that produces consistent results with group_weight,
1982  * otherwise workloads might not converge.
1983  */
1984 static int preferred_group_nid(struct task_struct *p, int nid)
1985 {
1986 	nodemask_t nodes;
1987 	int dist;
1988 
1989 	/* Direct connections between all NUMA nodes. */
1990 	if (sched_numa_topology_type == NUMA_DIRECT)
1991 		return nid;
1992 
1993 	/*
1994 	 * On a system with glueless mesh NUMA topology, group_weight
1995 	 * scores nodes according to the number of NUMA hinting faults on
1996 	 * both the node itself, and on nearby nodes.
1997 	 */
1998 	if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1999 		unsigned long score, max_score = 0;
2000 		int node, max_node = nid;
2001 
2002 		dist = sched_max_numa_distance;
2003 
2004 		for_each_online_node(node) {
2005 			score = group_weight(p, node, dist);
2006 			if (score > max_score) {
2007 				max_score = score;
2008 				max_node = node;
2009 			}
2010 		}
2011 		return max_node;
2012 	}
2013 
2014 	/*
2015 	 * Finding the preferred nid in a system with NUMA backplane
2016 	 * interconnect topology is more involved. The goal is to locate
2017 	 * tasks from numa_groups near each other in the system, and
2018 	 * untangle workloads from different sides of the system. This requires
2019 	 * searching down the hierarchy of node groups, recursively searching
2020 	 * inside the highest scoring group of nodes. The nodemask tricks
2021 	 * keep the complexity of the search down.
2022 	 */
2023 	nodes = node_online_map;
2024 	for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2025 		unsigned long max_faults = 0;
2026 		nodemask_t max_group = NODE_MASK_NONE;
2027 		int a, b;
2028 
2029 		/* Are there nodes at this distance from each other? */
2030 		if (!find_numa_distance(dist))
2031 			continue;
2032 
2033 		for_each_node_mask(a, nodes) {
2034 			unsigned long faults = 0;
2035 			nodemask_t this_group;
2036 			nodes_clear(this_group);
2037 
2038 			/* Sum group's NUMA faults; includes a==b case. */
2039 			for_each_node_mask(b, nodes) {
2040 				if (node_distance(a, b) < dist) {
2041 					faults += group_faults(p, b);
2042 					node_set(b, this_group);
2043 					node_clear(b, nodes);
2044 				}
2045 			}
2046 
2047 			/* Remember the top group. */
2048 			if (faults > max_faults) {
2049 				max_faults = faults;
2050 				max_group = this_group;
2051 				/*
2052 				 * subtle: at the smallest distance there is
2053 				 * just one node left in each "group", the
2054 				 * winner is the preferred nid.
2055 				 */
2056 				nid = a;
2057 			}
2058 		}
2059 		/* Next round, evaluate the nodes within max_group. */
2060 		if (!max_faults)
2061 			break;
2062 		nodes = max_group;
2063 	}
2064 	return nid;
2065 }
2066 
2067 static void task_numa_placement(struct task_struct *p)
2068 {
2069 	int seq, nid, max_nid = -1, max_group_nid = -1;
2070 	unsigned long max_faults = 0, max_group_faults = 0;
2071 	unsigned long fault_types[2] = { 0, 0 };
2072 	unsigned long total_faults;
2073 	u64 runtime, period;
2074 	spinlock_t *group_lock = NULL;
2075 
2076 	/*
2077 	 * The p->mm->numa_scan_seq field gets updated without
2078 	 * exclusive access. Use READ_ONCE() here to ensure
2079 	 * that the field is read in a single access:
2080 	 */
2081 	seq = READ_ONCE(p->mm->numa_scan_seq);
2082 	if (p->numa_scan_seq == seq)
2083 		return;
2084 	p->numa_scan_seq = seq;
2085 	p->numa_scan_period_max = task_scan_max(p);
2086 
2087 	total_faults = p->numa_faults_locality[0] +
2088 		       p->numa_faults_locality[1];
2089 	runtime = numa_get_avg_runtime(p, &period);
2090 
2091 	/* If the task is part of a group prevent parallel updates to group stats */
2092 	if (p->numa_group) {
2093 		group_lock = &p->numa_group->lock;
2094 		spin_lock_irq(group_lock);
2095 	}
2096 
2097 	/* Find the node with the highest number of faults */
2098 	for_each_online_node(nid) {
2099 		/* Keep track of the offsets in numa_faults array */
2100 		int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2101 		unsigned long faults = 0, group_faults = 0;
2102 		int priv;
2103 
2104 		for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2105 			long diff, f_diff, f_weight;
2106 
2107 			mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2108 			membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2109 			cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2110 			cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2111 
2112 			/* Decay existing window, copy faults since last scan */
2113 			diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2114 			fault_types[priv] += p->numa_faults[membuf_idx];
2115 			p->numa_faults[membuf_idx] = 0;
2116 
2117 			/*
2118 			 * Normalize the faults_from, so all tasks in a group
2119 			 * count according to CPU use, instead of by the raw
2120 			 * number of faults. Tasks with little runtime have
2121 			 * little over-all impact on throughput, and thus their
2122 			 * faults are less important.
2123 			 */
2124 			f_weight = div64_u64(runtime << 16, period + 1);
2125 			f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2126 				   (total_faults + 1);
2127 			f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2128 			p->numa_faults[cpubuf_idx] = 0;
2129 
2130 			p->numa_faults[mem_idx] += diff;
2131 			p->numa_faults[cpu_idx] += f_diff;
2132 			faults += p->numa_faults[mem_idx];
2133 			p->total_numa_faults += diff;
2134 			if (p->numa_group) {
2135 				/*
2136 				 * safe because we can only change our own group
2137 				 *
2138 				 * mem_idx represents the offset for a given
2139 				 * nid and priv in a specific region because it
2140 				 * is at the beginning of the numa_faults array.
2141 				 */
2142 				p->numa_group->faults[mem_idx] += diff;
2143 				p->numa_group->faults_cpu[mem_idx] += f_diff;
2144 				p->numa_group->total_faults += diff;
2145 				group_faults += p->numa_group->faults[mem_idx];
2146 			}
2147 		}
2148 
2149 		if (faults > max_faults) {
2150 			max_faults = faults;
2151 			max_nid = nid;
2152 		}
2153 
2154 		if (group_faults > max_group_faults) {
2155 			max_group_faults = group_faults;
2156 			max_group_nid = nid;
2157 		}
2158 	}
2159 
2160 	update_task_scan_period(p, fault_types[0], fault_types[1]);
2161 
2162 	if (p->numa_group) {
2163 		numa_group_count_active_nodes(p->numa_group);
2164 		spin_unlock_irq(group_lock);
2165 		max_nid = preferred_group_nid(p, max_group_nid);
2166 	}
2167 
2168 	if (max_faults) {
2169 		/* Set the new preferred node */
2170 		if (max_nid != p->numa_preferred_nid)
2171 			sched_setnuma(p, max_nid);
2172 
2173 		if (task_node(p) != p->numa_preferred_nid)
2174 			numa_migrate_preferred(p);
2175 	}
2176 }
2177 
2178 static inline int get_numa_group(struct numa_group *grp)
2179 {
2180 	return atomic_inc_not_zero(&grp->refcount);
2181 }
2182 
2183 static inline void put_numa_group(struct numa_group *grp)
2184 {
2185 	if (atomic_dec_and_test(&grp->refcount))
2186 		kfree_rcu(grp, rcu);
2187 }
2188 
2189 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2190 			int *priv)
2191 {
2192 	struct numa_group *grp, *my_grp;
2193 	struct task_struct *tsk;
2194 	bool join = false;
2195 	int cpu = cpupid_to_cpu(cpupid);
2196 	int i;
2197 
2198 	if (unlikely(!p->numa_group)) {
2199 		unsigned int size = sizeof(struct numa_group) +
2200 				    4*nr_node_ids*sizeof(unsigned long);
2201 
2202 		grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2203 		if (!grp)
2204 			return;
2205 
2206 		atomic_set(&grp->refcount, 1);
2207 		grp->active_nodes = 1;
2208 		grp->max_faults_cpu = 0;
2209 		spin_lock_init(&grp->lock);
2210 		grp->gid = p->pid;
2211 		/* Second half of the array tracks nids where faults happen */
2212 		grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2213 						nr_node_ids;
2214 
2215 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2216 			grp->faults[i] = p->numa_faults[i];
2217 
2218 		grp->total_faults = p->total_numa_faults;
2219 
2220 		grp->nr_tasks++;
2221 		rcu_assign_pointer(p->numa_group, grp);
2222 	}
2223 
2224 	rcu_read_lock();
2225 	tsk = READ_ONCE(cpu_rq(cpu)->curr);
2226 
2227 	if (!cpupid_match_pid(tsk, cpupid))
2228 		goto no_join;
2229 
2230 	grp = rcu_dereference(tsk->numa_group);
2231 	if (!grp)
2232 		goto no_join;
2233 
2234 	my_grp = p->numa_group;
2235 	if (grp == my_grp)
2236 		goto no_join;
2237 
2238 	/*
2239 	 * Only join the other group if its bigger; if we're the bigger group,
2240 	 * the other task will join us.
2241 	 */
2242 	if (my_grp->nr_tasks > grp->nr_tasks)
2243 		goto no_join;
2244 
2245 	/*
2246 	 * Tie-break on the grp address.
2247 	 */
2248 	if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2249 		goto no_join;
2250 
2251 	/* Always join threads in the same process. */
2252 	if (tsk->mm == current->mm)
2253 		join = true;
2254 
2255 	/* Simple filter to avoid false positives due to PID collisions */
2256 	if (flags & TNF_SHARED)
2257 		join = true;
2258 
2259 	/* Update priv based on whether false sharing was detected */
2260 	*priv = !join;
2261 
2262 	if (join && !get_numa_group(grp))
2263 		goto no_join;
2264 
2265 	rcu_read_unlock();
2266 
2267 	if (!join)
2268 		return;
2269 
2270 	BUG_ON(irqs_disabled());
2271 	double_lock_irq(&my_grp->lock, &grp->lock);
2272 
2273 	for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2274 		my_grp->faults[i] -= p->numa_faults[i];
2275 		grp->faults[i] += p->numa_faults[i];
2276 	}
2277 	my_grp->total_faults -= p->total_numa_faults;
2278 	grp->total_faults += p->total_numa_faults;
2279 
2280 	my_grp->nr_tasks--;
2281 	grp->nr_tasks++;
2282 
2283 	spin_unlock(&my_grp->lock);
2284 	spin_unlock_irq(&grp->lock);
2285 
2286 	rcu_assign_pointer(p->numa_group, grp);
2287 
2288 	put_numa_group(my_grp);
2289 	return;
2290 
2291 no_join:
2292 	rcu_read_unlock();
2293 	return;
2294 }
2295 
2296 void task_numa_free(struct task_struct *p)
2297 {
2298 	struct numa_group *grp = p->numa_group;
2299 	void *numa_faults = p->numa_faults;
2300 	unsigned long flags;
2301 	int i;
2302 
2303 	if (grp) {
2304 		spin_lock_irqsave(&grp->lock, flags);
2305 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2306 			grp->faults[i] -= p->numa_faults[i];
2307 		grp->total_faults -= p->total_numa_faults;
2308 
2309 		grp->nr_tasks--;
2310 		spin_unlock_irqrestore(&grp->lock, flags);
2311 		RCU_INIT_POINTER(p->numa_group, NULL);
2312 		put_numa_group(grp);
2313 	}
2314 
2315 	p->numa_faults = NULL;
2316 	kfree(numa_faults);
2317 }
2318 
2319 /*
2320  * Got a PROT_NONE fault for a page on @node.
2321  */
2322 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2323 {
2324 	struct task_struct *p = current;
2325 	bool migrated = flags & TNF_MIGRATED;
2326 	int cpu_node = task_node(current);
2327 	int local = !!(flags & TNF_FAULT_LOCAL);
2328 	struct numa_group *ng;
2329 	int priv;
2330 
2331 	if (!static_branch_likely(&sched_numa_balancing))
2332 		return;
2333 
2334 	/* for example, ksmd faulting in a user's mm */
2335 	if (!p->mm)
2336 		return;
2337 
2338 	/* Allocate buffer to track faults on a per-node basis */
2339 	if (unlikely(!p->numa_faults)) {
2340 		int size = sizeof(*p->numa_faults) *
2341 			   NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2342 
2343 		p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2344 		if (!p->numa_faults)
2345 			return;
2346 
2347 		p->total_numa_faults = 0;
2348 		memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2349 	}
2350 
2351 	/*
2352 	 * First accesses are treated as private, otherwise consider accesses
2353 	 * to be private if the accessing pid has not changed
2354 	 */
2355 	if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2356 		priv = 1;
2357 	} else {
2358 		priv = cpupid_match_pid(p, last_cpupid);
2359 		if (!priv && !(flags & TNF_NO_GROUP))
2360 			task_numa_group(p, last_cpupid, flags, &priv);
2361 	}
2362 
2363 	/*
2364 	 * If a workload spans multiple NUMA nodes, a shared fault that
2365 	 * occurs wholly within the set of nodes that the workload is
2366 	 * actively using should be counted as local. This allows the
2367 	 * scan rate to slow down when a workload has settled down.
2368 	 */
2369 	ng = p->numa_group;
2370 	if (!priv && !local && ng && ng->active_nodes > 1 &&
2371 				numa_is_active_node(cpu_node, ng) &&
2372 				numa_is_active_node(mem_node, ng))
2373 		local = 1;
2374 
2375 	task_numa_placement(p);
2376 
2377 	/*
2378 	 * Retry task to preferred node migration periodically, in case it
2379 	 * case it previously failed, or the scheduler moved us.
2380 	 */
2381 	if (time_after(jiffies, p->numa_migrate_retry))
2382 		numa_migrate_preferred(p);
2383 
2384 	if (migrated)
2385 		p->numa_pages_migrated += pages;
2386 	if (flags & TNF_MIGRATE_FAIL)
2387 		p->numa_faults_locality[2] += pages;
2388 
2389 	p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2390 	p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2391 	p->numa_faults_locality[local] += pages;
2392 }
2393 
2394 static void reset_ptenuma_scan(struct task_struct *p)
2395 {
2396 	/*
2397 	 * We only did a read acquisition of the mmap sem, so
2398 	 * p->mm->numa_scan_seq is written to without exclusive access
2399 	 * and the update is not guaranteed to be atomic. That's not
2400 	 * much of an issue though, since this is just used for
2401 	 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2402 	 * expensive, to avoid any form of compiler optimizations:
2403 	 */
2404 	WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2405 	p->mm->numa_scan_offset = 0;
2406 }
2407 
2408 /*
2409  * The expensive part of numa migration is done from task_work context.
2410  * Triggered from task_tick_numa().
2411  */
2412 void task_numa_work(struct callback_head *work)
2413 {
2414 	unsigned long migrate, next_scan, now = jiffies;
2415 	struct task_struct *p = current;
2416 	struct mm_struct *mm = p->mm;
2417 	u64 runtime = p->se.sum_exec_runtime;
2418 	struct vm_area_struct *vma;
2419 	unsigned long start, end;
2420 	unsigned long nr_pte_updates = 0;
2421 	long pages, virtpages;
2422 
2423 	SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2424 
2425 	work->next = work; /* protect against double add */
2426 	/*
2427 	 * Who cares about NUMA placement when they're dying.
2428 	 *
2429 	 * NOTE: make sure not to dereference p->mm before this check,
2430 	 * exit_task_work() happens _after_ exit_mm() so we could be called
2431 	 * without p->mm even though we still had it when we enqueued this
2432 	 * work.
2433 	 */
2434 	if (p->flags & PF_EXITING)
2435 		return;
2436 
2437 	if (!mm->numa_next_scan) {
2438 		mm->numa_next_scan = now +
2439 			msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2440 	}
2441 
2442 	/*
2443 	 * Enforce maximal scan/migration frequency..
2444 	 */
2445 	migrate = mm->numa_next_scan;
2446 	if (time_before(now, migrate))
2447 		return;
2448 
2449 	if (p->numa_scan_period == 0) {
2450 		p->numa_scan_period_max = task_scan_max(p);
2451 		p->numa_scan_period = task_scan_min(p);
2452 	}
2453 
2454 	next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2455 	if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2456 		return;
2457 
2458 	/*
2459 	 * Delay this task enough that another task of this mm will likely win
2460 	 * the next time around.
2461 	 */
2462 	p->node_stamp += 2 * TICK_NSEC;
2463 
2464 	start = mm->numa_scan_offset;
2465 	pages = sysctl_numa_balancing_scan_size;
2466 	pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2467 	virtpages = pages * 8;	   /* Scan up to this much virtual space */
2468 	if (!pages)
2469 		return;
2470 
2471 
2472 	if (!down_read_trylock(&mm->mmap_sem))
2473 		return;
2474 	vma = find_vma(mm, start);
2475 	if (!vma) {
2476 		reset_ptenuma_scan(p);
2477 		start = 0;
2478 		vma = mm->mmap;
2479 	}
2480 	for (; vma; vma = vma->vm_next) {
2481 		if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2482 			is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2483 			continue;
2484 		}
2485 
2486 		/*
2487 		 * Shared library pages mapped by multiple processes are not
2488 		 * migrated as it is expected they are cache replicated. Avoid
2489 		 * hinting faults in read-only file-backed mappings or the vdso
2490 		 * as migrating the pages will be of marginal benefit.
2491 		 */
2492 		if (!vma->vm_mm ||
2493 		    (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2494 			continue;
2495 
2496 		/*
2497 		 * Skip inaccessible VMAs to avoid any confusion between
2498 		 * PROT_NONE and NUMA hinting ptes
2499 		 */
2500 		if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2501 			continue;
2502 
2503 		do {
2504 			start = max(start, vma->vm_start);
2505 			end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2506 			end = min(end, vma->vm_end);
2507 			nr_pte_updates = change_prot_numa(vma, start, end);
2508 
2509 			/*
2510 			 * Try to scan sysctl_numa_balancing_size worth of
2511 			 * hpages that have at least one present PTE that
2512 			 * is not already pte-numa. If the VMA contains
2513 			 * areas that are unused or already full of prot_numa
2514 			 * PTEs, scan up to virtpages, to skip through those
2515 			 * areas faster.
2516 			 */
2517 			if (nr_pte_updates)
2518 				pages -= (end - start) >> PAGE_SHIFT;
2519 			virtpages -= (end - start) >> PAGE_SHIFT;
2520 
2521 			start = end;
2522 			if (pages <= 0 || virtpages <= 0)
2523 				goto out;
2524 
2525 			cond_resched();
2526 		} while (end != vma->vm_end);
2527 	}
2528 
2529 out:
2530 	/*
2531 	 * It is possible to reach the end of the VMA list but the last few
2532 	 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2533 	 * would find the !migratable VMA on the next scan but not reset the
2534 	 * scanner to the start so check it now.
2535 	 */
2536 	if (vma)
2537 		mm->numa_scan_offset = start;
2538 	else
2539 		reset_ptenuma_scan(p);
2540 	up_read(&mm->mmap_sem);
2541 
2542 	/*
2543 	 * Make sure tasks use at least 32x as much time to run other code
2544 	 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2545 	 * Usually update_task_scan_period slows down scanning enough; on an
2546 	 * overloaded system we need to limit overhead on a per task basis.
2547 	 */
2548 	if (unlikely(p->se.sum_exec_runtime != runtime)) {
2549 		u64 diff = p->se.sum_exec_runtime - runtime;
2550 		p->node_stamp += 32 * diff;
2551 	}
2552 }
2553 
2554 /*
2555  * Drive the periodic memory faults..
2556  */
2557 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2558 {
2559 	struct callback_head *work = &curr->numa_work;
2560 	u64 period, now;
2561 
2562 	/*
2563 	 * We don't care about NUMA placement if we don't have memory.
2564 	 */
2565 	if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2566 		return;
2567 
2568 	/*
2569 	 * Using runtime rather than walltime has the dual advantage that
2570 	 * we (mostly) drive the selection from busy threads and that the
2571 	 * task needs to have done some actual work before we bother with
2572 	 * NUMA placement.
2573 	 */
2574 	now = curr->se.sum_exec_runtime;
2575 	period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2576 
2577 	if (now > curr->node_stamp + period) {
2578 		if (!curr->node_stamp)
2579 			curr->numa_scan_period = task_scan_min(curr);
2580 		curr->node_stamp += period;
2581 
2582 		if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2583 			init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2584 			task_work_add(curr, work, true);
2585 		}
2586 	}
2587 }
2588 
2589 /*
2590  * Can a task be moved from prev_cpu to this_cpu without causing a load
2591  * imbalance that would trigger the load balancer?
2592  */
2593 static inline bool numa_wake_affine(struct sched_domain *sd,
2594 				    struct task_struct *p, int this_cpu,
2595 				    int prev_cpu, int sync)
2596 {
2597 	struct numa_stats prev_load, this_load;
2598 	s64 this_eff_load, prev_eff_load;
2599 
2600 	update_numa_stats(&prev_load, cpu_to_node(prev_cpu));
2601 	update_numa_stats(&this_load, cpu_to_node(this_cpu));
2602 
2603 	/*
2604 	 * If sync wakeup then subtract the (maximum possible)
2605 	 * effect of the currently running task from the load
2606 	 * of the current CPU:
2607 	 */
2608 	if (sync) {
2609 		unsigned long current_load = task_h_load(current);
2610 
2611 		if (this_load.load > current_load)
2612 			this_load.load -= current_load;
2613 		else
2614 			this_load.load = 0;
2615 	}
2616 
2617 	/*
2618 	 * In low-load situations, where this_cpu's node is idle due to the
2619 	 * sync cause above having dropped this_load.load to 0, move the task.
2620 	 * Moving to an idle socket will not create a bad imbalance.
2621 	 *
2622 	 * Otherwise check if the nodes are near enough in load to allow this
2623 	 * task to be woken on this_cpu's node.
2624 	 */
2625 	if (this_load.load > 0) {
2626 		unsigned long task_load = task_h_load(p);
2627 
2628 		this_eff_load = 100;
2629 		this_eff_load *= prev_load.compute_capacity;
2630 
2631 		prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2632 		prev_eff_load *= this_load.compute_capacity;
2633 
2634 		this_eff_load *= this_load.load + task_load;
2635 		prev_eff_load *= prev_load.load - task_load;
2636 
2637 		return this_eff_load <= prev_eff_load;
2638 	}
2639 
2640 	return true;
2641 }
2642 #else
2643 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2644 {
2645 }
2646 
2647 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2648 {
2649 }
2650 
2651 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2652 {
2653 }
2654 
2655 #ifdef CONFIG_SMP
2656 static inline bool numa_wake_affine(struct sched_domain *sd,
2657 				    struct task_struct *p, int this_cpu,
2658 				    int prev_cpu, int sync)
2659 {
2660 	return true;
2661 }
2662 #endif /* !SMP */
2663 #endif /* CONFIG_NUMA_BALANCING */
2664 
2665 static void
2666 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2667 {
2668 	update_load_add(&cfs_rq->load, se->load.weight);
2669 	if (!parent_entity(se))
2670 		update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2671 #ifdef CONFIG_SMP
2672 	if (entity_is_task(se)) {
2673 		struct rq *rq = rq_of(cfs_rq);
2674 
2675 		account_numa_enqueue(rq, task_of(se));
2676 		list_add(&se->group_node, &rq->cfs_tasks);
2677 	}
2678 #endif
2679 	cfs_rq->nr_running++;
2680 }
2681 
2682 static void
2683 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2684 {
2685 	update_load_sub(&cfs_rq->load, se->load.weight);
2686 	if (!parent_entity(se))
2687 		update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2688 #ifdef CONFIG_SMP
2689 	if (entity_is_task(se)) {
2690 		account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2691 		list_del_init(&se->group_node);
2692 	}
2693 #endif
2694 	cfs_rq->nr_running--;
2695 }
2696 
2697 #ifdef CONFIG_FAIR_GROUP_SCHED
2698 # ifdef CONFIG_SMP
2699 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2700 {
2701 	long tg_weight, load, shares;
2702 
2703 	/*
2704 	 * This really should be: cfs_rq->avg.load_avg, but instead we use
2705 	 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2706 	 * the shares for small weight interactive tasks.
2707 	 */
2708 	load = scale_load_down(cfs_rq->load.weight);
2709 
2710 	tg_weight = atomic_long_read(&tg->load_avg);
2711 
2712 	/* Ensure tg_weight >= load */
2713 	tg_weight -= cfs_rq->tg_load_avg_contrib;
2714 	tg_weight += load;
2715 
2716 	shares = (tg->shares * load);
2717 	if (tg_weight)
2718 		shares /= tg_weight;
2719 
2720 	/*
2721 	 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2722 	 * of a group with small tg->shares value. It is a floor value which is
2723 	 * assigned as a minimum load.weight to the sched_entity representing
2724 	 * the group on a CPU.
2725 	 *
2726 	 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2727 	 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2728 	 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2729 	 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2730 	 * instead of 0.
2731 	 */
2732 	if (shares < MIN_SHARES)
2733 		shares = MIN_SHARES;
2734 	if (shares > tg->shares)
2735 		shares = tg->shares;
2736 
2737 	return shares;
2738 }
2739 # else /* CONFIG_SMP */
2740 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2741 {
2742 	return tg->shares;
2743 }
2744 # endif /* CONFIG_SMP */
2745 
2746 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2747 			    unsigned long weight)
2748 {
2749 	if (se->on_rq) {
2750 		/* commit outstanding execution time */
2751 		if (cfs_rq->curr == se)
2752 			update_curr(cfs_rq);
2753 		account_entity_dequeue(cfs_rq, se);
2754 	}
2755 
2756 	update_load_set(&se->load, weight);
2757 
2758 	if (se->on_rq)
2759 		account_entity_enqueue(cfs_rq, se);
2760 }
2761 
2762 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2763 
2764 static void update_cfs_shares(struct sched_entity *se)
2765 {
2766 	struct cfs_rq *cfs_rq = group_cfs_rq(se);
2767 	struct task_group *tg;
2768 	long shares;
2769 
2770 	if (!cfs_rq)
2771 		return;
2772 
2773 	if (throttled_hierarchy(cfs_rq))
2774 		return;
2775 
2776 	tg = cfs_rq->tg;
2777 
2778 #ifndef CONFIG_SMP
2779 	if (likely(se->load.weight == tg->shares))
2780 		return;
2781 #endif
2782 	shares = calc_cfs_shares(cfs_rq, tg);
2783 
2784 	reweight_entity(cfs_rq_of(se), se, shares);
2785 }
2786 
2787 #else /* CONFIG_FAIR_GROUP_SCHED */
2788 static inline void update_cfs_shares(struct sched_entity *se)
2789 {
2790 }
2791 #endif /* CONFIG_FAIR_GROUP_SCHED */
2792 
2793 #ifdef CONFIG_SMP
2794 /*
2795  * Approximate:
2796  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2797  */
2798 static u64 decay_load(u64 val, u64 n)
2799 {
2800 	unsigned int local_n;
2801 
2802 	if (unlikely(n > LOAD_AVG_PERIOD * 63))
2803 		return 0;
2804 
2805 	/* after bounds checking we can collapse to 32-bit */
2806 	local_n = n;
2807 
2808 	/*
2809 	 * As y^PERIOD = 1/2, we can combine
2810 	 *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2811 	 * With a look-up table which covers y^n (n<PERIOD)
2812 	 *
2813 	 * To achieve constant time decay_load.
2814 	 */
2815 	if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2816 		val >>= local_n / LOAD_AVG_PERIOD;
2817 		local_n %= LOAD_AVG_PERIOD;
2818 	}
2819 
2820 	val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2821 	return val;
2822 }
2823 
2824 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
2825 {
2826 	u32 c1, c2, c3 = d3; /* y^0 == 1 */
2827 
2828 	/*
2829 	 * c1 = d1 y^p
2830 	 */
2831 	c1 = decay_load((u64)d1, periods);
2832 
2833 	/*
2834 	 *            p-1
2835 	 * c2 = 1024 \Sum y^n
2836 	 *            n=1
2837 	 *
2838 	 *              inf        inf
2839 	 *    = 1024 ( \Sum y^n - \Sum y^n - y^0 )
2840 	 *              n=0        n=p
2841 	 */
2842 	c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
2843 
2844 	return c1 + c2 + c3;
2845 }
2846 
2847 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2848 
2849 /*
2850  * Accumulate the three separate parts of the sum; d1 the remainder
2851  * of the last (incomplete) period, d2 the span of full periods and d3
2852  * the remainder of the (incomplete) current period.
2853  *
2854  *           d1          d2           d3
2855  *           ^           ^            ^
2856  *           |           |            |
2857  *         |<->|<----------------->|<--->|
2858  * ... |---x---|------| ... |------|-----x (now)
2859  *
2860  *                           p-1
2861  * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2862  *                           n=1
2863  *
2864  *    = u y^p +					(Step 1)
2865  *
2866  *                     p-1
2867  *      d1 y^p + 1024 \Sum y^n + d3 y^0		(Step 2)
2868  *                     n=1
2869  */
2870 static __always_inline u32
2871 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
2872 	       unsigned long weight, int running, struct cfs_rq *cfs_rq)
2873 {
2874 	unsigned long scale_freq, scale_cpu;
2875 	u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
2876 	u64 periods;
2877 
2878 	scale_freq = arch_scale_freq_capacity(NULL, cpu);
2879 	scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2880 
2881 	delta += sa->period_contrib;
2882 	periods = delta / 1024; /* A period is 1024us (~1ms) */
2883 
2884 	/*
2885 	 * Step 1: decay old *_sum if we crossed period boundaries.
2886 	 */
2887 	if (periods) {
2888 		sa->load_sum = decay_load(sa->load_sum, periods);
2889 		if (cfs_rq) {
2890 			cfs_rq->runnable_load_sum =
2891 				decay_load(cfs_rq->runnable_load_sum, periods);
2892 		}
2893 		sa->util_sum = decay_load((u64)(sa->util_sum), periods);
2894 
2895 		/*
2896 		 * Step 2
2897 		 */
2898 		delta %= 1024;
2899 		contrib = __accumulate_pelt_segments(periods,
2900 				1024 - sa->period_contrib, delta);
2901 	}
2902 	sa->period_contrib = delta;
2903 
2904 	contrib = cap_scale(contrib, scale_freq);
2905 	if (weight) {
2906 		sa->load_sum += weight * contrib;
2907 		if (cfs_rq)
2908 			cfs_rq->runnable_load_sum += weight * contrib;
2909 	}
2910 	if (running)
2911 		sa->util_sum += contrib * scale_cpu;
2912 
2913 	return periods;
2914 }
2915 
2916 /*
2917  * We can represent the historical contribution to runnable average as the
2918  * coefficients of a geometric series.  To do this we sub-divide our runnable
2919  * history into segments of approximately 1ms (1024us); label the segment that
2920  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2921  *
2922  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2923  *      p0            p1           p2
2924  *     (now)       (~1ms ago)  (~2ms ago)
2925  *
2926  * Let u_i denote the fraction of p_i that the entity was runnable.
2927  *
2928  * We then designate the fractions u_i as our co-efficients, yielding the
2929  * following representation of historical load:
2930  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2931  *
2932  * We choose y based on the with of a reasonably scheduling period, fixing:
2933  *   y^32 = 0.5
2934  *
2935  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2936  * approximately half as much as the contribution to load within the last ms
2937  * (u_0).
2938  *
2939  * When a period "rolls over" and we have new u_0`, multiplying the previous
2940  * sum again by y is sufficient to update:
2941  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2942  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2943  */
2944 static __always_inline int
2945 ___update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2946 		  unsigned long weight, int running, struct cfs_rq *cfs_rq)
2947 {
2948 	u64 delta;
2949 
2950 	delta = now - sa->last_update_time;
2951 	/*
2952 	 * This should only happen when time goes backwards, which it
2953 	 * unfortunately does during sched clock init when we swap over to TSC.
2954 	 */
2955 	if ((s64)delta < 0) {
2956 		sa->last_update_time = now;
2957 		return 0;
2958 	}
2959 
2960 	/*
2961 	 * Use 1024ns as the unit of measurement since it's a reasonable
2962 	 * approximation of 1us and fast to compute.
2963 	 */
2964 	delta >>= 10;
2965 	if (!delta)
2966 		return 0;
2967 
2968 	sa->last_update_time += delta << 10;
2969 
2970 	/*
2971 	 * Now we know we crossed measurement unit boundaries. The *_avg
2972 	 * accrues by two steps:
2973 	 *
2974 	 * Step 1: accumulate *_sum since last_update_time. If we haven't
2975 	 * crossed period boundaries, finish.
2976 	 */
2977 	if (!accumulate_sum(delta, cpu, sa, weight, running, cfs_rq))
2978 		return 0;
2979 
2980 	/*
2981 	 * Step 2: update *_avg.
2982 	 */
2983 	sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
2984 	if (cfs_rq) {
2985 		cfs_rq->runnable_load_avg =
2986 			div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
2987 	}
2988 	sa->util_avg = sa->util_sum / (LOAD_AVG_MAX - 1024 + sa->period_contrib);
2989 
2990 	return 1;
2991 }
2992 
2993 static int
2994 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
2995 {
2996 	return ___update_load_avg(now, cpu, &se->avg, 0, 0, NULL);
2997 }
2998 
2999 static int
3000 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3001 {
3002 	return ___update_load_avg(now, cpu, &se->avg,
3003 				  se->on_rq * scale_load_down(se->load.weight),
3004 				  cfs_rq->curr == se, NULL);
3005 }
3006 
3007 static int
3008 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3009 {
3010 	return ___update_load_avg(now, cpu, &cfs_rq->avg,
3011 			scale_load_down(cfs_rq->load.weight),
3012 			cfs_rq->curr != NULL, cfs_rq);
3013 }
3014 
3015 /*
3016  * Signed add and clamp on underflow.
3017  *
3018  * Explicitly do a load-store to ensure the intermediate value never hits
3019  * memory. This allows lockless observations without ever seeing the negative
3020  * values.
3021  */
3022 #define add_positive(_ptr, _val) do {                           \
3023 	typeof(_ptr) ptr = (_ptr);                              \
3024 	typeof(_val) val = (_val);                              \
3025 	typeof(*ptr) res, var = READ_ONCE(*ptr);                \
3026 								\
3027 	res = var + val;                                        \
3028 								\
3029 	if (val < 0 && res > var)                               \
3030 		res = 0;                                        \
3031 								\
3032 	WRITE_ONCE(*ptr, res);                                  \
3033 } while (0)
3034 
3035 #ifdef CONFIG_FAIR_GROUP_SCHED
3036 /**
3037  * update_tg_load_avg - update the tg's load avg
3038  * @cfs_rq: the cfs_rq whose avg changed
3039  * @force: update regardless of how small the difference
3040  *
3041  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3042  * However, because tg->load_avg is a global value there are performance
3043  * considerations.
3044  *
3045  * In order to avoid having to look at the other cfs_rq's, we use a
3046  * differential update where we store the last value we propagated. This in
3047  * turn allows skipping updates if the differential is 'small'.
3048  *
3049  * Updating tg's load_avg is necessary before update_cfs_share().
3050  */
3051 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3052 {
3053 	long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3054 
3055 	/*
3056 	 * No need to update load_avg for root_task_group as it is not used.
3057 	 */
3058 	if (cfs_rq->tg == &root_task_group)
3059 		return;
3060 
3061 	if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3062 		atomic_long_add(delta, &cfs_rq->tg->load_avg);
3063 		cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3064 	}
3065 }
3066 
3067 /*
3068  * Called within set_task_rq() right before setting a task's cpu. The
3069  * caller only guarantees p->pi_lock is held; no other assumptions,
3070  * including the state of rq->lock, should be made.
3071  */
3072 void set_task_rq_fair(struct sched_entity *se,
3073 		      struct cfs_rq *prev, struct cfs_rq *next)
3074 {
3075 	u64 p_last_update_time;
3076 	u64 n_last_update_time;
3077 
3078 	if (!sched_feat(ATTACH_AGE_LOAD))
3079 		return;
3080 
3081 	/*
3082 	 * We are supposed to update the task to "current" time, then its up to
3083 	 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3084 	 * getting what current time is, so simply throw away the out-of-date
3085 	 * time. This will result in the wakee task is less decayed, but giving
3086 	 * the wakee more load sounds not bad.
3087 	 */
3088 	if (!(se->avg.last_update_time && prev))
3089 		return;
3090 
3091 #ifndef CONFIG_64BIT
3092 	{
3093 		u64 p_last_update_time_copy;
3094 		u64 n_last_update_time_copy;
3095 
3096 		do {
3097 			p_last_update_time_copy = prev->load_last_update_time_copy;
3098 			n_last_update_time_copy = next->load_last_update_time_copy;
3099 
3100 			smp_rmb();
3101 
3102 			p_last_update_time = prev->avg.last_update_time;
3103 			n_last_update_time = next->avg.last_update_time;
3104 
3105 		} while (p_last_update_time != p_last_update_time_copy ||
3106 			 n_last_update_time != n_last_update_time_copy);
3107 	}
3108 #else
3109 	p_last_update_time = prev->avg.last_update_time;
3110 	n_last_update_time = next->avg.last_update_time;
3111 #endif
3112 	__update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3113 	se->avg.last_update_time = n_last_update_time;
3114 }
3115 
3116 /* Take into account change of utilization of a child task group */
3117 static inline void
3118 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
3119 {
3120 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3121 	long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3122 
3123 	/* Nothing to update */
3124 	if (!delta)
3125 		return;
3126 
3127 	/* Set new sched_entity's utilization */
3128 	se->avg.util_avg = gcfs_rq->avg.util_avg;
3129 	se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3130 
3131 	/* Update parent cfs_rq utilization */
3132 	add_positive(&cfs_rq->avg.util_avg, delta);
3133 	cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3134 }
3135 
3136 /* Take into account change of load of a child task group */
3137 static inline void
3138 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
3139 {
3140 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3141 	long delta, load = gcfs_rq->avg.load_avg;
3142 
3143 	/*
3144 	 * If the load of group cfs_rq is null, the load of the
3145 	 * sched_entity will also be null so we can skip the formula
3146 	 */
3147 	if (load) {
3148 		long tg_load;
3149 
3150 		/* Get tg's load and ensure tg_load > 0 */
3151 		tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
3152 
3153 		/* Ensure tg_load >= load and updated with current load*/
3154 		tg_load -= gcfs_rq->tg_load_avg_contrib;
3155 		tg_load += load;
3156 
3157 		/*
3158 		 * We need to compute a correction term in the case that the
3159 		 * task group is consuming more CPU than a task of equal
3160 		 * weight. A task with a weight equals to tg->shares will have
3161 		 * a load less or equal to scale_load_down(tg->shares).
3162 		 * Similarly, the sched_entities that represent the task group
3163 		 * at parent level, can't have a load higher than
3164 		 * scale_load_down(tg->shares). And the Sum of sched_entities'
3165 		 * load must be <= scale_load_down(tg->shares).
3166 		 */
3167 		if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3168 			/* scale gcfs_rq's load into tg's shares*/
3169 			load *= scale_load_down(gcfs_rq->tg->shares);
3170 			load /= tg_load;
3171 		}
3172 	}
3173 
3174 	delta = load - se->avg.load_avg;
3175 
3176 	/* Nothing to update */
3177 	if (!delta)
3178 		return;
3179 
3180 	/* Set new sched_entity's load */
3181 	se->avg.load_avg = load;
3182 	se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3183 
3184 	/* Update parent cfs_rq load */
3185 	add_positive(&cfs_rq->avg.load_avg, delta);
3186 	cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3187 
3188 	/*
3189 	 * If the sched_entity is already enqueued, we also have to update the
3190 	 * runnable load avg.
3191 	 */
3192 	if (se->on_rq) {
3193 		/* Update parent cfs_rq runnable_load_avg */
3194 		add_positive(&cfs_rq->runnable_load_avg, delta);
3195 		cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3196 	}
3197 }
3198 
3199 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3200 {
3201 	cfs_rq->propagate_avg = 1;
3202 }
3203 
3204 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3205 {
3206 	struct cfs_rq *cfs_rq = group_cfs_rq(se);
3207 
3208 	if (!cfs_rq->propagate_avg)
3209 		return 0;
3210 
3211 	cfs_rq->propagate_avg = 0;
3212 	return 1;
3213 }
3214 
3215 /* Update task and its cfs_rq load average */
3216 static inline int propagate_entity_load_avg(struct sched_entity *se)
3217 {
3218 	struct cfs_rq *cfs_rq;
3219 
3220 	if (entity_is_task(se))
3221 		return 0;
3222 
3223 	if (!test_and_clear_tg_cfs_propagate(se))
3224 		return 0;
3225 
3226 	cfs_rq = cfs_rq_of(se);
3227 
3228 	set_tg_cfs_propagate(cfs_rq);
3229 
3230 	update_tg_cfs_util(cfs_rq, se);
3231 	update_tg_cfs_load(cfs_rq, se);
3232 
3233 	return 1;
3234 }
3235 
3236 /*
3237  * Check if we need to update the load and the utilization of a blocked
3238  * group_entity:
3239  */
3240 static inline bool skip_blocked_update(struct sched_entity *se)
3241 {
3242 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3243 
3244 	/*
3245 	 * If sched_entity still have not zero load or utilization, we have to
3246 	 * decay it:
3247 	 */
3248 	if (se->avg.load_avg || se->avg.util_avg)
3249 		return false;
3250 
3251 	/*
3252 	 * If there is a pending propagation, we have to update the load and
3253 	 * the utilization of the sched_entity:
3254 	 */
3255 	if (gcfs_rq->propagate_avg)
3256 		return false;
3257 
3258 	/*
3259 	 * Otherwise, the load and the utilization of the sched_entity is
3260 	 * already zero and there is no pending propagation, so it will be a
3261 	 * waste of time to try to decay it:
3262 	 */
3263 	return true;
3264 }
3265 
3266 #else /* CONFIG_FAIR_GROUP_SCHED */
3267 
3268 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3269 
3270 static inline int propagate_entity_load_avg(struct sched_entity *se)
3271 {
3272 	return 0;
3273 }
3274 
3275 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3276 
3277 #endif /* CONFIG_FAIR_GROUP_SCHED */
3278 
3279 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3280 {
3281 	if (&this_rq()->cfs == cfs_rq) {
3282 		/*
3283 		 * There are a few boundary cases this might miss but it should
3284 		 * get called often enough that that should (hopefully) not be
3285 		 * a real problem -- added to that it only calls on the local
3286 		 * CPU, so if we enqueue remotely we'll miss an update, but
3287 		 * the next tick/schedule should update.
3288 		 *
3289 		 * It will not get called when we go idle, because the idle
3290 		 * thread is a different class (!fair), nor will the utilization
3291 		 * number include things like RT tasks.
3292 		 *
3293 		 * As is, the util number is not freq-invariant (we'd have to
3294 		 * implement arch_scale_freq_capacity() for that).
3295 		 *
3296 		 * See cpu_util().
3297 		 */
3298 		cpufreq_update_util(rq_of(cfs_rq), 0);
3299 	}
3300 }
3301 
3302 /*
3303  * Unsigned subtract and clamp on underflow.
3304  *
3305  * Explicitly do a load-store to ensure the intermediate value never hits
3306  * memory. This allows lockless observations without ever seeing the negative
3307  * values.
3308  */
3309 #define sub_positive(_ptr, _val) do {				\
3310 	typeof(_ptr) ptr = (_ptr);				\
3311 	typeof(*ptr) val = (_val);				\
3312 	typeof(*ptr) res, var = READ_ONCE(*ptr);		\
3313 	res = var - val;					\
3314 	if (res > var)						\
3315 		res = 0;					\
3316 	WRITE_ONCE(*ptr, res);					\
3317 } while (0)
3318 
3319 /**
3320  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3321  * @now: current time, as per cfs_rq_clock_task()
3322  * @cfs_rq: cfs_rq to update
3323  * @update_freq: should we call cfs_rq_util_change() or will the call do so
3324  *
3325  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3326  * avg. The immediate corollary is that all (fair) tasks must be attached, see
3327  * post_init_entity_util_avg().
3328  *
3329  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3330  *
3331  * Returns true if the load decayed or we removed load.
3332  *
3333  * Since both these conditions indicate a changed cfs_rq->avg.load we should
3334  * call update_tg_load_avg() when this function returns true.
3335  */
3336 static inline int
3337 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3338 {
3339 	struct sched_avg *sa = &cfs_rq->avg;
3340 	int decayed, removed_load = 0, removed_util = 0;
3341 
3342 	if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3343 		s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3344 		sub_positive(&sa->load_avg, r);
3345 		sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3346 		removed_load = 1;
3347 		set_tg_cfs_propagate(cfs_rq);
3348 	}
3349 
3350 	if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3351 		long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3352 		sub_positive(&sa->util_avg, r);
3353 		sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3354 		removed_util = 1;
3355 		set_tg_cfs_propagate(cfs_rq);
3356 	}
3357 
3358 	decayed = __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3359 
3360 #ifndef CONFIG_64BIT
3361 	smp_wmb();
3362 	cfs_rq->load_last_update_time_copy = sa->last_update_time;
3363 #endif
3364 
3365 	if (update_freq && (decayed || removed_util))
3366 		cfs_rq_util_change(cfs_rq);
3367 
3368 	return decayed || removed_load;
3369 }
3370 
3371 /*
3372  * Optional action to be done while updating the load average
3373  */
3374 #define UPDATE_TG	0x1
3375 #define SKIP_AGE_LOAD	0x2
3376 
3377 /* Update task and its cfs_rq load average */
3378 static inline void update_load_avg(struct sched_entity *se, int flags)
3379 {
3380 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
3381 	u64 now = cfs_rq_clock_task(cfs_rq);
3382 	struct rq *rq = rq_of(cfs_rq);
3383 	int cpu = cpu_of(rq);
3384 	int decayed;
3385 
3386 	/*
3387 	 * Track task load average for carrying it to new CPU after migrated, and
3388 	 * track group sched_entity load average for task_h_load calc in migration
3389 	 */
3390 	if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3391 		__update_load_avg_se(now, cpu, cfs_rq, se);
3392 
3393 	decayed  = update_cfs_rq_load_avg(now, cfs_rq, true);
3394 	decayed |= propagate_entity_load_avg(se);
3395 
3396 	if (decayed && (flags & UPDATE_TG))
3397 		update_tg_load_avg(cfs_rq, 0);
3398 }
3399 
3400 /**
3401  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3402  * @cfs_rq: cfs_rq to attach to
3403  * @se: sched_entity to attach
3404  *
3405  * Must call update_cfs_rq_load_avg() before this, since we rely on
3406  * cfs_rq->avg.last_update_time being current.
3407  */
3408 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3409 {
3410 	se->avg.last_update_time = cfs_rq->avg.last_update_time;
3411 	cfs_rq->avg.load_avg += se->avg.load_avg;
3412 	cfs_rq->avg.load_sum += se->avg.load_sum;
3413 	cfs_rq->avg.util_avg += se->avg.util_avg;
3414 	cfs_rq->avg.util_sum += se->avg.util_sum;
3415 	set_tg_cfs_propagate(cfs_rq);
3416 
3417 	cfs_rq_util_change(cfs_rq);
3418 }
3419 
3420 /**
3421  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3422  * @cfs_rq: cfs_rq to detach from
3423  * @se: sched_entity to detach
3424  *
3425  * Must call update_cfs_rq_load_avg() before this, since we rely on
3426  * cfs_rq->avg.last_update_time being current.
3427  */
3428 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3429 {
3430 
3431 	sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3432 	sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3433 	sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3434 	sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3435 	set_tg_cfs_propagate(cfs_rq);
3436 
3437 	cfs_rq_util_change(cfs_rq);
3438 }
3439 
3440 /* Add the load generated by se into cfs_rq's load average */
3441 static inline void
3442 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3443 {
3444 	struct sched_avg *sa = &se->avg;
3445 
3446 	cfs_rq->runnable_load_avg += sa->load_avg;
3447 	cfs_rq->runnable_load_sum += sa->load_sum;
3448 
3449 	if (!sa->last_update_time) {
3450 		attach_entity_load_avg(cfs_rq, se);
3451 		update_tg_load_avg(cfs_rq, 0);
3452 	}
3453 }
3454 
3455 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3456 static inline void
3457 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3458 {
3459 	cfs_rq->runnable_load_avg =
3460 		max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3461 	cfs_rq->runnable_load_sum =
3462 		max_t(s64,  cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3463 }
3464 
3465 #ifndef CONFIG_64BIT
3466 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3467 {
3468 	u64 last_update_time_copy;
3469 	u64 last_update_time;
3470 
3471 	do {
3472 		last_update_time_copy = cfs_rq->load_last_update_time_copy;
3473 		smp_rmb();
3474 		last_update_time = cfs_rq->avg.last_update_time;
3475 	} while (last_update_time != last_update_time_copy);
3476 
3477 	return last_update_time;
3478 }
3479 #else
3480 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3481 {
3482 	return cfs_rq->avg.last_update_time;
3483 }
3484 #endif
3485 
3486 /*
3487  * Synchronize entity load avg of dequeued entity without locking
3488  * the previous rq.
3489  */
3490 void sync_entity_load_avg(struct sched_entity *se)
3491 {
3492 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
3493 	u64 last_update_time;
3494 
3495 	last_update_time = cfs_rq_last_update_time(cfs_rq);
3496 	__update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3497 }
3498 
3499 /*
3500  * Task first catches up with cfs_rq, and then subtract
3501  * itself from the cfs_rq (task must be off the queue now).
3502  */
3503 void remove_entity_load_avg(struct sched_entity *se)
3504 {
3505 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
3506 
3507 	/*
3508 	 * tasks cannot exit without having gone through wake_up_new_task() ->
3509 	 * post_init_entity_util_avg() which will have added things to the
3510 	 * cfs_rq, so we can remove unconditionally.
3511 	 *
3512 	 * Similarly for groups, they will have passed through
3513 	 * post_init_entity_util_avg() before unregister_sched_fair_group()
3514 	 * calls this.
3515 	 */
3516 
3517 	sync_entity_load_avg(se);
3518 	atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3519 	atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3520 }
3521 
3522 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3523 {
3524 	return cfs_rq->runnable_load_avg;
3525 }
3526 
3527 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3528 {
3529 	return cfs_rq->avg.load_avg;
3530 }
3531 
3532 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3533 
3534 #else /* CONFIG_SMP */
3535 
3536 static inline int
3537 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3538 {
3539 	return 0;
3540 }
3541 
3542 #define UPDATE_TG	0x0
3543 #define SKIP_AGE_LOAD	0x0
3544 
3545 static inline void update_load_avg(struct sched_entity *se, int not_used1)
3546 {
3547 	cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
3548 }
3549 
3550 static inline void
3551 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3552 static inline void
3553 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3554 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3555 
3556 static inline void
3557 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3558 static inline void
3559 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3560 
3561 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3562 {
3563 	return 0;
3564 }
3565 
3566 #endif /* CONFIG_SMP */
3567 
3568 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3569 {
3570 #ifdef CONFIG_SCHED_DEBUG
3571 	s64 d = se->vruntime - cfs_rq->min_vruntime;
3572 
3573 	if (d < 0)
3574 		d = -d;
3575 
3576 	if (d > 3*sysctl_sched_latency)
3577 		schedstat_inc(cfs_rq->nr_spread_over);
3578 #endif
3579 }
3580 
3581 static void
3582 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3583 {
3584 	u64 vruntime = cfs_rq->min_vruntime;
3585 
3586 	/*
3587 	 * The 'current' period is already promised to the current tasks,
3588 	 * however the extra weight of the new task will slow them down a
3589 	 * little, place the new task so that it fits in the slot that
3590 	 * stays open at the end.
3591 	 */
3592 	if (initial && sched_feat(START_DEBIT))
3593 		vruntime += sched_vslice(cfs_rq, se);
3594 
3595 	/* sleeps up to a single latency don't count. */
3596 	if (!initial) {
3597 		unsigned long thresh = sysctl_sched_latency;
3598 
3599 		/*
3600 		 * Halve their sleep time's effect, to allow
3601 		 * for a gentler effect of sleepers:
3602 		 */
3603 		if (sched_feat(GENTLE_FAIR_SLEEPERS))
3604 			thresh >>= 1;
3605 
3606 		vruntime -= thresh;
3607 	}
3608 
3609 	/* ensure we never gain time by being placed backwards. */
3610 	se->vruntime = max_vruntime(se->vruntime, vruntime);
3611 }
3612 
3613 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3614 
3615 static inline void check_schedstat_required(void)
3616 {
3617 #ifdef CONFIG_SCHEDSTATS
3618 	if (schedstat_enabled())
3619 		return;
3620 
3621 	/* Force schedstat enabled if a dependent tracepoint is active */
3622 	if (trace_sched_stat_wait_enabled()    ||
3623 			trace_sched_stat_sleep_enabled()   ||
3624 			trace_sched_stat_iowait_enabled()  ||
3625 			trace_sched_stat_blocked_enabled() ||
3626 			trace_sched_stat_runtime_enabled())  {
3627 		printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3628 			     "stat_blocked and stat_runtime require the "
3629 			     "kernel parameter schedstats=enable or "
3630 			     "kernel.sched_schedstats=1\n");
3631 	}
3632 #endif
3633 }
3634 
3635 
3636 /*
3637  * MIGRATION
3638  *
3639  *	dequeue
3640  *	  update_curr()
3641  *	    update_min_vruntime()
3642  *	  vruntime -= min_vruntime
3643  *
3644  *	enqueue
3645  *	  update_curr()
3646  *	    update_min_vruntime()
3647  *	  vruntime += min_vruntime
3648  *
3649  * this way the vruntime transition between RQs is done when both
3650  * min_vruntime are up-to-date.
3651  *
3652  * WAKEUP (remote)
3653  *
3654  *	->migrate_task_rq_fair() (p->state == TASK_WAKING)
3655  *	  vruntime -= min_vruntime
3656  *
3657  *	enqueue
3658  *	  update_curr()
3659  *	    update_min_vruntime()
3660  *	  vruntime += min_vruntime
3661  *
3662  * this way we don't have the most up-to-date min_vruntime on the originating
3663  * CPU and an up-to-date min_vruntime on the destination CPU.
3664  */
3665 
3666 static void
3667 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3668 {
3669 	bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3670 	bool curr = cfs_rq->curr == se;
3671 
3672 	/*
3673 	 * If we're the current task, we must renormalise before calling
3674 	 * update_curr().
3675 	 */
3676 	if (renorm && curr)
3677 		se->vruntime += cfs_rq->min_vruntime;
3678 
3679 	update_curr(cfs_rq);
3680 
3681 	/*
3682 	 * Otherwise, renormalise after, such that we're placed at the current
3683 	 * moment in time, instead of some random moment in the past. Being
3684 	 * placed in the past could significantly boost this task to the
3685 	 * fairness detriment of existing tasks.
3686 	 */
3687 	if (renorm && !curr)
3688 		se->vruntime += cfs_rq->min_vruntime;
3689 
3690 	/*
3691 	 * When enqueuing a sched_entity, we must:
3692 	 *   - Update loads to have both entity and cfs_rq synced with now.
3693 	 *   - Add its load to cfs_rq->runnable_avg
3694 	 *   - For group_entity, update its weight to reflect the new share of
3695 	 *     its group cfs_rq
3696 	 *   - Add its new weight to cfs_rq->load.weight
3697 	 */
3698 	update_load_avg(se, UPDATE_TG);
3699 	enqueue_entity_load_avg(cfs_rq, se);
3700 	update_cfs_shares(se);
3701 	account_entity_enqueue(cfs_rq, se);
3702 
3703 	if (flags & ENQUEUE_WAKEUP)
3704 		place_entity(cfs_rq, se, 0);
3705 
3706 	check_schedstat_required();
3707 	update_stats_enqueue(cfs_rq, se, flags);
3708 	check_spread(cfs_rq, se);
3709 	if (!curr)
3710 		__enqueue_entity(cfs_rq, se);
3711 	se->on_rq = 1;
3712 
3713 	if (cfs_rq->nr_running == 1) {
3714 		list_add_leaf_cfs_rq(cfs_rq);
3715 		check_enqueue_throttle(cfs_rq);
3716 	}
3717 }
3718 
3719 static void __clear_buddies_last(struct sched_entity *se)
3720 {
3721 	for_each_sched_entity(se) {
3722 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
3723 		if (cfs_rq->last != se)
3724 			break;
3725 
3726 		cfs_rq->last = NULL;
3727 	}
3728 }
3729 
3730 static void __clear_buddies_next(struct sched_entity *se)
3731 {
3732 	for_each_sched_entity(se) {
3733 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
3734 		if (cfs_rq->next != se)
3735 			break;
3736 
3737 		cfs_rq->next = NULL;
3738 	}
3739 }
3740 
3741 static void __clear_buddies_skip(struct sched_entity *se)
3742 {
3743 	for_each_sched_entity(se) {
3744 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
3745 		if (cfs_rq->skip != se)
3746 			break;
3747 
3748 		cfs_rq->skip = NULL;
3749 	}
3750 }
3751 
3752 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3753 {
3754 	if (cfs_rq->last == se)
3755 		__clear_buddies_last(se);
3756 
3757 	if (cfs_rq->next == se)
3758 		__clear_buddies_next(se);
3759 
3760 	if (cfs_rq->skip == se)
3761 		__clear_buddies_skip(se);
3762 }
3763 
3764 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3765 
3766 static void
3767 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3768 {
3769 	/*
3770 	 * Update run-time statistics of the 'current'.
3771 	 */
3772 	update_curr(cfs_rq);
3773 
3774 	/*
3775 	 * When dequeuing a sched_entity, we must:
3776 	 *   - Update loads to have both entity and cfs_rq synced with now.
3777 	 *   - Substract its load from the cfs_rq->runnable_avg.
3778 	 *   - Substract its previous weight from cfs_rq->load.weight.
3779 	 *   - For group entity, update its weight to reflect the new share
3780 	 *     of its group cfs_rq.
3781 	 */
3782 	update_load_avg(se, UPDATE_TG);
3783 	dequeue_entity_load_avg(cfs_rq, se);
3784 
3785 	update_stats_dequeue(cfs_rq, se, flags);
3786 
3787 	clear_buddies(cfs_rq, se);
3788 
3789 	if (se != cfs_rq->curr)
3790 		__dequeue_entity(cfs_rq, se);
3791 	se->on_rq = 0;
3792 	account_entity_dequeue(cfs_rq, se);
3793 
3794 	/*
3795 	 * Normalize after update_curr(); which will also have moved
3796 	 * min_vruntime if @se is the one holding it back. But before doing
3797 	 * update_min_vruntime() again, which will discount @se's position and
3798 	 * can move min_vruntime forward still more.
3799 	 */
3800 	if (!(flags & DEQUEUE_SLEEP))
3801 		se->vruntime -= cfs_rq->min_vruntime;
3802 
3803 	/* return excess runtime on last dequeue */
3804 	return_cfs_rq_runtime(cfs_rq);
3805 
3806 	update_cfs_shares(se);
3807 
3808 	/*
3809 	 * Now advance min_vruntime if @se was the entity holding it back,
3810 	 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3811 	 * put back on, and if we advance min_vruntime, we'll be placed back
3812 	 * further than we started -- ie. we'll be penalized.
3813 	 */
3814 	if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3815 		update_min_vruntime(cfs_rq);
3816 }
3817 
3818 /*
3819  * Preempt the current task with a newly woken task if needed:
3820  */
3821 static void
3822 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3823 {
3824 	unsigned long ideal_runtime, delta_exec;
3825 	struct sched_entity *se;
3826 	s64 delta;
3827 
3828 	ideal_runtime = sched_slice(cfs_rq, curr);
3829 	delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3830 	if (delta_exec > ideal_runtime) {
3831 		resched_curr(rq_of(cfs_rq));
3832 		/*
3833 		 * The current task ran long enough, ensure it doesn't get
3834 		 * re-elected due to buddy favours.
3835 		 */
3836 		clear_buddies(cfs_rq, curr);
3837 		return;
3838 	}
3839 
3840 	/*
3841 	 * Ensure that a task that missed wakeup preemption by a
3842 	 * narrow margin doesn't have to wait for a full slice.
3843 	 * This also mitigates buddy induced latencies under load.
3844 	 */
3845 	if (delta_exec < sysctl_sched_min_granularity)
3846 		return;
3847 
3848 	se = __pick_first_entity(cfs_rq);
3849 	delta = curr->vruntime - se->vruntime;
3850 
3851 	if (delta < 0)
3852 		return;
3853 
3854 	if (delta > ideal_runtime)
3855 		resched_curr(rq_of(cfs_rq));
3856 }
3857 
3858 static void
3859 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3860 {
3861 	/* 'current' is not kept within the tree. */
3862 	if (se->on_rq) {
3863 		/*
3864 		 * Any task has to be enqueued before it get to execute on
3865 		 * a CPU. So account for the time it spent waiting on the
3866 		 * runqueue.
3867 		 */
3868 		update_stats_wait_end(cfs_rq, se);
3869 		__dequeue_entity(cfs_rq, se);
3870 		update_load_avg(se, UPDATE_TG);
3871 	}
3872 
3873 	update_stats_curr_start(cfs_rq, se);
3874 	cfs_rq->curr = se;
3875 
3876 	/*
3877 	 * Track our maximum slice length, if the CPU's load is at
3878 	 * least twice that of our own weight (i.e. dont track it
3879 	 * when there are only lesser-weight tasks around):
3880 	 */
3881 	if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3882 		schedstat_set(se->statistics.slice_max,
3883 			max((u64)schedstat_val(se->statistics.slice_max),
3884 			    se->sum_exec_runtime - se->prev_sum_exec_runtime));
3885 	}
3886 
3887 	se->prev_sum_exec_runtime = se->sum_exec_runtime;
3888 }
3889 
3890 static int
3891 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3892 
3893 /*
3894  * Pick the next process, keeping these things in mind, in this order:
3895  * 1) keep things fair between processes/task groups
3896  * 2) pick the "next" process, since someone really wants that to run
3897  * 3) pick the "last" process, for cache locality
3898  * 4) do not run the "skip" process, if something else is available
3899  */
3900 static struct sched_entity *
3901 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3902 {
3903 	struct sched_entity *left = __pick_first_entity(cfs_rq);
3904 	struct sched_entity *se;
3905 
3906 	/*
3907 	 * If curr is set we have to see if its left of the leftmost entity
3908 	 * still in the tree, provided there was anything in the tree at all.
3909 	 */
3910 	if (!left || (curr && entity_before(curr, left)))
3911 		left = curr;
3912 
3913 	se = left; /* ideally we run the leftmost entity */
3914 
3915 	/*
3916 	 * Avoid running the skip buddy, if running something else can
3917 	 * be done without getting too unfair.
3918 	 */
3919 	if (cfs_rq->skip == se) {
3920 		struct sched_entity *second;
3921 
3922 		if (se == curr) {
3923 			second = __pick_first_entity(cfs_rq);
3924 		} else {
3925 			second = __pick_next_entity(se);
3926 			if (!second || (curr && entity_before(curr, second)))
3927 				second = curr;
3928 		}
3929 
3930 		if (second && wakeup_preempt_entity(second, left) < 1)
3931 			se = second;
3932 	}
3933 
3934 	/*
3935 	 * Prefer last buddy, try to return the CPU to a preempted task.
3936 	 */
3937 	if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3938 		se = cfs_rq->last;
3939 
3940 	/*
3941 	 * Someone really wants this to run. If it's not unfair, run it.
3942 	 */
3943 	if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3944 		se = cfs_rq->next;
3945 
3946 	clear_buddies(cfs_rq, se);
3947 
3948 	return se;
3949 }
3950 
3951 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3952 
3953 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3954 {
3955 	/*
3956 	 * If still on the runqueue then deactivate_task()
3957 	 * was not called and update_curr() has to be done:
3958 	 */
3959 	if (prev->on_rq)
3960 		update_curr(cfs_rq);
3961 
3962 	/* throttle cfs_rqs exceeding runtime */
3963 	check_cfs_rq_runtime(cfs_rq);
3964 
3965 	check_spread(cfs_rq, prev);
3966 
3967 	if (prev->on_rq) {
3968 		update_stats_wait_start(cfs_rq, prev);
3969 		/* Put 'current' back into the tree. */
3970 		__enqueue_entity(cfs_rq, prev);
3971 		/* in !on_rq case, update occurred at dequeue */
3972 		update_load_avg(prev, 0);
3973 	}
3974 	cfs_rq->curr = NULL;
3975 }
3976 
3977 static void
3978 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3979 {
3980 	/*
3981 	 * Update run-time statistics of the 'current'.
3982 	 */
3983 	update_curr(cfs_rq);
3984 
3985 	/*
3986 	 * Ensure that runnable average is periodically updated.
3987 	 */
3988 	update_load_avg(curr, UPDATE_TG);
3989 	update_cfs_shares(curr);
3990 
3991 #ifdef CONFIG_SCHED_HRTICK
3992 	/*
3993 	 * queued ticks are scheduled to match the slice, so don't bother
3994 	 * validating it and just reschedule.
3995 	 */
3996 	if (queued) {
3997 		resched_curr(rq_of(cfs_rq));
3998 		return;
3999 	}
4000 	/*
4001 	 * don't let the period tick interfere with the hrtick preemption
4002 	 */
4003 	if (!sched_feat(DOUBLE_TICK) &&
4004 			hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4005 		return;
4006 #endif
4007 
4008 	if (cfs_rq->nr_running > 1)
4009 		check_preempt_tick(cfs_rq, curr);
4010 }
4011 
4012 
4013 /**************************************************
4014  * CFS bandwidth control machinery
4015  */
4016 
4017 #ifdef CONFIG_CFS_BANDWIDTH
4018 
4019 #ifdef HAVE_JUMP_LABEL
4020 static struct static_key __cfs_bandwidth_used;
4021 
4022 static inline bool cfs_bandwidth_used(void)
4023 {
4024 	return static_key_false(&__cfs_bandwidth_used);
4025 }
4026 
4027 void cfs_bandwidth_usage_inc(void)
4028 {
4029 	static_key_slow_inc(&__cfs_bandwidth_used);
4030 }
4031 
4032 void cfs_bandwidth_usage_dec(void)
4033 {
4034 	static_key_slow_dec(&__cfs_bandwidth_used);
4035 }
4036 #else /* HAVE_JUMP_LABEL */
4037 static bool cfs_bandwidth_used(void)
4038 {
4039 	return true;
4040 }
4041 
4042 void cfs_bandwidth_usage_inc(void) {}
4043 void cfs_bandwidth_usage_dec(void) {}
4044 #endif /* HAVE_JUMP_LABEL */
4045 
4046 /*
4047  * default period for cfs group bandwidth.
4048  * default: 0.1s, units: nanoseconds
4049  */
4050 static inline u64 default_cfs_period(void)
4051 {
4052 	return 100000000ULL;
4053 }
4054 
4055 static inline u64 sched_cfs_bandwidth_slice(void)
4056 {
4057 	return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4058 }
4059 
4060 /*
4061  * Replenish runtime according to assigned quota and update expiration time.
4062  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4063  * additional synchronization around rq->lock.
4064  *
4065  * requires cfs_b->lock
4066  */
4067 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4068 {
4069 	u64 now;
4070 
4071 	if (cfs_b->quota == RUNTIME_INF)
4072 		return;
4073 
4074 	now = sched_clock_cpu(smp_processor_id());
4075 	cfs_b->runtime = cfs_b->quota;
4076 	cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4077 }
4078 
4079 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4080 {
4081 	return &tg->cfs_bandwidth;
4082 }
4083 
4084 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4085 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4086 {
4087 	if (unlikely(cfs_rq->throttle_count))
4088 		return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4089 
4090 	return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4091 }
4092 
4093 /* returns 0 on failure to allocate runtime */
4094 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4095 {
4096 	struct task_group *tg = cfs_rq->tg;
4097 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4098 	u64 amount = 0, min_amount, expires;
4099 
4100 	/* note: this is a positive sum as runtime_remaining <= 0 */
4101 	min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4102 
4103 	raw_spin_lock(&cfs_b->lock);
4104 	if (cfs_b->quota == RUNTIME_INF)
4105 		amount = min_amount;
4106 	else {
4107 		start_cfs_bandwidth(cfs_b);
4108 
4109 		if (cfs_b->runtime > 0) {
4110 			amount = min(cfs_b->runtime, min_amount);
4111 			cfs_b->runtime -= amount;
4112 			cfs_b->idle = 0;
4113 		}
4114 	}
4115 	expires = cfs_b->runtime_expires;
4116 	raw_spin_unlock(&cfs_b->lock);
4117 
4118 	cfs_rq->runtime_remaining += amount;
4119 	/*
4120 	 * we may have advanced our local expiration to account for allowed
4121 	 * spread between our sched_clock and the one on which runtime was
4122 	 * issued.
4123 	 */
4124 	if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4125 		cfs_rq->runtime_expires = expires;
4126 
4127 	return cfs_rq->runtime_remaining > 0;
4128 }
4129 
4130 /*
4131  * Note: This depends on the synchronization provided by sched_clock and the
4132  * fact that rq->clock snapshots this value.
4133  */
4134 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4135 {
4136 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4137 
4138 	/* if the deadline is ahead of our clock, nothing to do */
4139 	if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4140 		return;
4141 
4142 	if (cfs_rq->runtime_remaining < 0)
4143 		return;
4144 
4145 	/*
4146 	 * If the local deadline has passed we have to consider the
4147 	 * possibility that our sched_clock is 'fast' and the global deadline
4148 	 * has not truly expired.
4149 	 *
4150 	 * Fortunately we can check determine whether this the case by checking
4151 	 * whether the global deadline has advanced. It is valid to compare
4152 	 * cfs_b->runtime_expires without any locks since we only care about
4153 	 * exact equality, so a partial write will still work.
4154 	 */
4155 
4156 	if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4157 		/* extend local deadline, drift is bounded above by 2 ticks */
4158 		cfs_rq->runtime_expires += TICK_NSEC;
4159 	} else {
4160 		/* global deadline is ahead, expiration has passed */
4161 		cfs_rq->runtime_remaining = 0;
4162 	}
4163 }
4164 
4165 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4166 {
4167 	/* dock delta_exec before expiring quota (as it could span periods) */
4168 	cfs_rq->runtime_remaining -= delta_exec;
4169 	expire_cfs_rq_runtime(cfs_rq);
4170 
4171 	if (likely(cfs_rq->runtime_remaining > 0))
4172 		return;
4173 
4174 	/*
4175 	 * if we're unable to extend our runtime we resched so that the active
4176 	 * hierarchy can be throttled
4177 	 */
4178 	if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4179 		resched_curr(rq_of(cfs_rq));
4180 }
4181 
4182 static __always_inline
4183 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4184 {
4185 	if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4186 		return;
4187 
4188 	__account_cfs_rq_runtime(cfs_rq, delta_exec);
4189 }
4190 
4191 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4192 {
4193 	return cfs_bandwidth_used() && cfs_rq->throttled;
4194 }
4195 
4196 /* check whether cfs_rq, or any parent, is throttled */
4197 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4198 {
4199 	return cfs_bandwidth_used() && cfs_rq->throttle_count;
4200 }
4201 
4202 /*
4203  * Ensure that neither of the group entities corresponding to src_cpu or
4204  * dest_cpu are members of a throttled hierarchy when performing group
4205  * load-balance operations.
4206  */
4207 static inline int throttled_lb_pair(struct task_group *tg,
4208 				    int src_cpu, int dest_cpu)
4209 {
4210 	struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4211 
4212 	src_cfs_rq = tg->cfs_rq[src_cpu];
4213 	dest_cfs_rq = tg->cfs_rq[dest_cpu];
4214 
4215 	return throttled_hierarchy(src_cfs_rq) ||
4216 	       throttled_hierarchy(dest_cfs_rq);
4217 }
4218 
4219 /* updated child weight may affect parent so we have to do this bottom up */
4220 static int tg_unthrottle_up(struct task_group *tg, void *data)
4221 {
4222 	struct rq *rq = data;
4223 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4224 
4225 	cfs_rq->throttle_count--;
4226 	if (!cfs_rq->throttle_count) {
4227 		/* adjust cfs_rq_clock_task() */
4228 		cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4229 					     cfs_rq->throttled_clock_task;
4230 	}
4231 
4232 	return 0;
4233 }
4234 
4235 static int tg_throttle_down(struct task_group *tg, void *data)
4236 {
4237 	struct rq *rq = data;
4238 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4239 
4240 	/* group is entering throttled state, stop time */
4241 	if (!cfs_rq->throttle_count)
4242 		cfs_rq->throttled_clock_task = rq_clock_task(rq);
4243 	cfs_rq->throttle_count++;
4244 
4245 	return 0;
4246 }
4247 
4248 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4249 {
4250 	struct rq *rq = rq_of(cfs_rq);
4251 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4252 	struct sched_entity *se;
4253 	long task_delta, dequeue = 1;
4254 	bool empty;
4255 
4256 	se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4257 
4258 	/* freeze hierarchy runnable averages while throttled */
4259 	rcu_read_lock();
4260 	walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4261 	rcu_read_unlock();
4262 
4263 	task_delta = cfs_rq->h_nr_running;
4264 	for_each_sched_entity(se) {
4265 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4266 		/* throttled entity or throttle-on-deactivate */
4267 		if (!se->on_rq)
4268 			break;
4269 
4270 		if (dequeue)
4271 			dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4272 		qcfs_rq->h_nr_running -= task_delta;
4273 
4274 		if (qcfs_rq->load.weight)
4275 			dequeue = 0;
4276 	}
4277 
4278 	if (!se)
4279 		sub_nr_running(rq, task_delta);
4280 
4281 	cfs_rq->throttled = 1;
4282 	cfs_rq->throttled_clock = rq_clock(rq);
4283 	raw_spin_lock(&cfs_b->lock);
4284 	empty = list_empty(&cfs_b->throttled_cfs_rq);
4285 
4286 	/*
4287 	 * Add to the _head_ of the list, so that an already-started
4288 	 * distribute_cfs_runtime will not see us
4289 	 */
4290 	list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4291 
4292 	/*
4293 	 * If we're the first throttled task, make sure the bandwidth
4294 	 * timer is running.
4295 	 */
4296 	if (empty)
4297 		start_cfs_bandwidth(cfs_b);
4298 
4299 	raw_spin_unlock(&cfs_b->lock);
4300 }
4301 
4302 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4303 {
4304 	struct rq *rq = rq_of(cfs_rq);
4305 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4306 	struct sched_entity *se;
4307 	int enqueue = 1;
4308 	long task_delta;
4309 
4310 	se = cfs_rq->tg->se[cpu_of(rq)];
4311 
4312 	cfs_rq->throttled = 0;
4313 
4314 	update_rq_clock(rq);
4315 
4316 	raw_spin_lock(&cfs_b->lock);
4317 	cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4318 	list_del_rcu(&cfs_rq->throttled_list);
4319 	raw_spin_unlock(&cfs_b->lock);
4320 
4321 	/* update hierarchical throttle state */
4322 	walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4323 
4324 	if (!cfs_rq->load.weight)
4325 		return;
4326 
4327 	task_delta = cfs_rq->h_nr_running;
4328 	for_each_sched_entity(se) {
4329 		if (se->on_rq)
4330 			enqueue = 0;
4331 
4332 		cfs_rq = cfs_rq_of(se);
4333 		if (enqueue)
4334 			enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4335 		cfs_rq->h_nr_running += task_delta;
4336 
4337 		if (cfs_rq_throttled(cfs_rq))
4338 			break;
4339 	}
4340 
4341 	if (!se)
4342 		add_nr_running(rq, task_delta);
4343 
4344 	/* determine whether we need to wake up potentially idle cpu */
4345 	if (rq->curr == rq->idle && rq->cfs.nr_running)
4346 		resched_curr(rq);
4347 }
4348 
4349 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4350 		u64 remaining, u64 expires)
4351 {
4352 	struct cfs_rq *cfs_rq;
4353 	u64 runtime;
4354 	u64 starting_runtime = remaining;
4355 
4356 	rcu_read_lock();
4357 	list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4358 				throttled_list) {
4359 		struct rq *rq = rq_of(cfs_rq);
4360 		struct rq_flags rf;
4361 
4362 		rq_lock(rq, &rf);
4363 		if (!cfs_rq_throttled(cfs_rq))
4364 			goto next;
4365 
4366 		runtime = -cfs_rq->runtime_remaining + 1;
4367 		if (runtime > remaining)
4368 			runtime = remaining;
4369 		remaining -= runtime;
4370 
4371 		cfs_rq->runtime_remaining += runtime;
4372 		cfs_rq->runtime_expires = expires;
4373 
4374 		/* we check whether we're throttled above */
4375 		if (cfs_rq->runtime_remaining > 0)
4376 			unthrottle_cfs_rq(cfs_rq);
4377 
4378 next:
4379 		rq_unlock(rq, &rf);
4380 
4381 		if (!remaining)
4382 			break;
4383 	}
4384 	rcu_read_unlock();
4385 
4386 	return starting_runtime - remaining;
4387 }
4388 
4389 /*
4390  * Responsible for refilling a task_group's bandwidth and unthrottling its
4391  * cfs_rqs as appropriate. If there has been no activity within the last
4392  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4393  * used to track this state.
4394  */
4395 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4396 {
4397 	u64 runtime, runtime_expires;
4398 	int throttled;
4399 
4400 	/* no need to continue the timer with no bandwidth constraint */
4401 	if (cfs_b->quota == RUNTIME_INF)
4402 		goto out_deactivate;
4403 
4404 	throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4405 	cfs_b->nr_periods += overrun;
4406 
4407 	/*
4408 	 * idle depends on !throttled (for the case of a large deficit), and if
4409 	 * we're going inactive then everything else can be deferred
4410 	 */
4411 	if (cfs_b->idle && !throttled)
4412 		goto out_deactivate;
4413 
4414 	__refill_cfs_bandwidth_runtime(cfs_b);
4415 
4416 	if (!throttled) {
4417 		/* mark as potentially idle for the upcoming period */
4418 		cfs_b->idle = 1;
4419 		return 0;
4420 	}
4421 
4422 	/* account preceding periods in which throttling occurred */
4423 	cfs_b->nr_throttled += overrun;
4424 
4425 	runtime_expires = cfs_b->runtime_expires;
4426 
4427 	/*
4428 	 * This check is repeated as we are holding onto the new bandwidth while
4429 	 * we unthrottle. This can potentially race with an unthrottled group
4430 	 * trying to acquire new bandwidth from the global pool. This can result
4431 	 * in us over-using our runtime if it is all used during this loop, but
4432 	 * only by limited amounts in that extreme case.
4433 	 */
4434 	while (throttled && cfs_b->runtime > 0) {
4435 		runtime = cfs_b->runtime;
4436 		raw_spin_unlock(&cfs_b->lock);
4437 		/* we can't nest cfs_b->lock while distributing bandwidth */
4438 		runtime = distribute_cfs_runtime(cfs_b, runtime,
4439 						 runtime_expires);
4440 		raw_spin_lock(&cfs_b->lock);
4441 
4442 		throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4443 
4444 		cfs_b->runtime -= min(runtime, cfs_b->runtime);
4445 	}
4446 
4447 	/*
4448 	 * While we are ensured activity in the period following an
4449 	 * unthrottle, this also covers the case in which the new bandwidth is
4450 	 * insufficient to cover the existing bandwidth deficit.  (Forcing the
4451 	 * timer to remain active while there are any throttled entities.)
4452 	 */
4453 	cfs_b->idle = 0;
4454 
4455 	return 0;
4456 
4457 out_deactivate:
4458 	return 1;
4459 }
4460 
4461 /* a cfs_rq won't donate quota below this amount */
4462 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4463 /* minimum remaining period time to redistribute slack quota */
4464 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4465 /* how long we wait to gather additional slack before distributing */
4466 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4467 
4468 /*
4469  * Are we near the end of the current quota period?
4470  *
4471  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4472  * hrtimer base being cleared by hrtimer_start. In the case of
4473  * migrate_hrtimers, base is never cleared, so we are fine.
4474  */
4475 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4476 {
4477 	struct hrtimer *refresh_timer = &cfs_b->period_timer;
4478 	u64 remaining;
4479 
4480 	/* if the call-back is running a quota refresh is already occurring */
4481 	if (hrtimer_callback_running(refresh_timer))
4482 		return 1;
4483 
4484 	/* is a quota refresh about to occur? */
4485 	remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4486 	if (remaining < min_expire)
4487 		return 1;
4488 
4489 	return 0;
4490 }
4491 
4492 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4493 {
4494 	u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4495 
4496 	/* if there's a quota refresh soon don't bother with slack */
4497 	if (runtime_refresh_within(cfs_b, min_left))
4498 		return;
4499 
4500 	hrtimer_start(&cfs_b->slack_timer,
4501 			ns_to_ktime(cfs_bandwidth_slack_period),
4502 			HRTIMER_MODE_REL);
4503 }
4504 
4505 /* we know any runtime found here is valid as update_curr() precedes return */
4506 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4507 {
4508 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4509 	s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4510 
4511 	if (slack_runtime <= 0)
4512 		return;
4513 
4514 	raw_spin_lock(&cfs_b->lock);
4515 	if (cfs_b->quota != RUNTIME_INF &&
4516 	    cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4517 		cfs_b->runtime += slack_runtime;
4518 
4519 		/* we are under rq->lock, defer unthrottling using a timer */
4520 		if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4521 		    !list_empty(&cfs_b->throttled_cfs_rq))
4522 			start_cfs_slack_bandwidth(cfs_b);
4523 	}
4524 	raw_spin_unlock(&cfs_b->lock);
4525 
4526 	/* even if it's not valid for return we don't want to try again */
4527 	cfs_rq->runtime_remaining -= slack_runtime;
4528 }
4529 
4530 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4531 {
4532 	if (!cfs_bandwidth_used())
4533 		return;
4534 
4535 	if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4536 		return;
4537 
4538 	__return_cfs_rq_runtime(cfs_rq);
4539 }
4540 
4541 /*
4542  * This is done with a timer (instead of inline with bandwidth return) since
4543  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4544  */
4545 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4546 {
4547 	u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4548 	u64 expires;
4549 
4550 	/* confirm we're still not at a refresh boundary */
4551 	raw_spin_lock(&cfs_b->lock);
4552 	if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4553 		raw_spin_unlock(&cfs_b->lock);
4554 		return;
4555 	}
4556 
4557 	if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4558 		runtime = cfs_b->runtime;
4559 
4560 	expires = cfs_b->runtime_expires;
4561 	raw_spin_unlock(&cfs_b->lock);
4562 
4563 	if (!runtime)
4564 		return;
4565 
4566 	runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4567 
4568 	raw_spin_lock(&cfs_b->lock);
4569 	if (expires == cfs_b->runtime_expires)
4570 		cfs_b->runtime -= min(runtime, cfs_b->runtime);
4571 	raw_spin_unlock(&cfs_b->lock);
4572 }
4573 
4574 /*
4575  * When a group wakes up we want to make sure that its quota is not already
4576  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4577  * runtime as update_curr() throttling can not not trigger until it's on-rq.
4578  */
4579 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4580 {
4581 	if (!cfs_bandwidth_used())
4582 		return;
4583 
4584 	/* an active group must be handled by the update_curr()->put() path */
4585 	if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4586 		return;
4587 
4588 	/* ensure the group is not already throttled */
4589 	if (cfs_rq_throttled(cfs_rq))
4590 		return;
4591 
4592 	/* update runtime allocation */
4593 	account_cfs_rq_runtime(cfs_rq, 0);
4594 	if (cfs_rq->runtime_remaining <= 0)
4595 		throttle_cfs_rq(cfs_rq);
4596 }
4597 
4598 static void sync_throttle(struct task_group *tg, int cpu)
4599 {
4600 	struct cfs_rq *pcfs_rq, *cfs_rq;
4601 
4602 	if (!cfs_bandwidth_used())
4603 		return;
4604 
4605 	if (!tg->parent)
4606 		return;
4607 
4608 	cfs_rq = tg->cfs_rq[cpu];
4609 	pcfs_rq = tg->parent->cfs_rq[cpu];
4610 
4611 	cfs_rq->throttle_count = pcfs_rq->throttle_count;
4612 	cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4613 }
4614 
4615 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4616 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4617 {
4618 	if (!cfs_bandwidth_used())
4619 		return false;
4620 
4621 	if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4622 		return false;
4623 
4624 	/*
4625 	 * it's possible for a throttled entity to be forced into a running
4626 	 * state (e.g. set_curr_task), in this case we're finished.
4627 	 */
4628 	if (cfs_rq_throttled(cfs_rq))
4629 		return true;
4630 
4631 	throttle_cfs_rq(cfs_rq);
4632 	return true;
4633 }
4634 
4635 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4636 {
4637 	struct cfs_bandwidth *cfs_b =
4638 		container_of(timer, struct cfs_bandwidth, slack_timer);
4639 
4640 	do_sched_cfs_slack_timer(cfs_b);
4641 
4642 	return HRTIMER_NORESTART;
4643 }
4644 
4645 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4646 {
4647 	struct cfs_bandwidth *cfs_b =
4648 		container_of(timer, struct cfs_bandwidth, period_timer);
4649 	int overrun;
4650 	int idle = 0;
4651 
4652 	raw_spin_lock(&cfs_b->lock);
4653 	for (;;) {
4654 		overrun = hrtimer_forward_now(timer, cfs_b->period);
4655 		if (!overrun)
4656 			break;
4657 
4658 		idle = do_sched_cfs_period_timer(cfs_b, overrun);
4659 	}
4660 	if (idle)
4661 		cfs_b->period_active = 0;
4662 	raw_spin_unlock(&cfs_b->lock);
4663 
4664 	return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4665 }
4666 
4667 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4668 {
4669 	raw_spin_lock_init(&cfs_b->lock);
4670 	cfs_b->runtime = 0;
4671 	cfs_b->quota = RUNTIME_INF;
4672 	cfs_b->period = ns_to_ktime(default_cfs_period());
4673 
4674 	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4675 	hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4676 	cfs_b->period_timer.function = sched_cfs_period_timer;
4677 	hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4678 	cfs_b->slack_timer.function = sched_cfs_slack_timer;
4679 }
4680 
4681 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4682 {
4683 	cfs_rq->runtime_enabled = 0;
4684 	INIT_LIST_HEAD(&cfs_rq->throttled_list);
4685 }
4686 
4687 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4688 {
4689 	lockdep_assert_held(&cfs_b->lock);
4690 
4691 	if (!cfs_b->period_active) {
4692 		cfs_b->period_active = 1;
4693 		hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4694 		hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4695 	}
4696 }
4697 
4698 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4699 {
4700 	/* init_cfs_bandwidth() was not called */
4701 	if (!cfs_b->throttled_cfs_rq.next)
4702 		return;
4703 
4704 	hrtimer_cancel(&cfs_b->period_timer);
4705 	hrtimer_cancel(&cfs_b->slack_timer);
4706 }
4707 
4708 /*
4709  * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
4710  *
4711  * The race is harmless, since modifying bandwidth settings of unhooked group
4712  * bits doesn't do much.
4713  */
4714 
4715 /* cpu online calback */
4716 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4717 {
4718 	struct task_group *tg;
4719 
4720 	lockdep_assert_held(&rq->lock);
4721 
4722 	rcu_read_lock();
4723 	list_for_each_entry_rcu(tg, &task_groups, list) {
4724 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4725 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4726 
4727 		raw_spin_lock(&cfs_b->lock);
4728 		cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4729 		raw_spin_unlock(&cfs_b->lock);
4730 	}
4731 	rcu_read_unlock();
4732 }
4733 
4734 /* cpu offline callback */
4735 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4736 {
4737 	struct task_group *tg;
4738 
4739 	lockdep_assert_held(&rq->lock);
4740 
4741 	rcu_read_lock();
4742 	list_for_each_entry_rcu(tg, &task_groups, list) {
4743 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4744 
4745 		if (!cfs_rq->runtime_enabled)
4746 			continue;
4747 
4748 		/*
4749 		 * clock_task is not advancing so we just need to make sure
4750 		 * there's some valid quota amount
4751 		 */
4752 		cfs_rq->runtime_remaining = 1;
4753 		/*
4754 		 * Offline rq is schedulable till cpu is completely disabled
4755 		 * in take_cpu_down(), so we prevent new cfs throttling here.
4756 		 */
4757 		cfs_rq->runtime_enabled = 0;
4758 
4759 		if (cfs_rq_throttled(cfs_rq))
4760 			unthrottle_cfs_rq(cfs_rq);
4761 	}
4762 	rcu_read_unlock();
4763 }
4764 
4765 #else /* CONFIG_CFS_BANDWIDTH */
4766 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4767 {
4768 	return rq_clock_task(rq_of(cfs_rq));
4769 }
4770 
4771 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4772 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4773 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4774 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4775 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4776 
4777 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4778 {
4779 	return 0;
4780 }
4781 
4782 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4783 {
4784 	return 0;
4785 }
4786 
4787 static inline int throttled_lb_pair(struct task_group *tg,
4788 				    int src_cpu, int dest_cpu)
4789 {
4790 	return 0;
4791 }
4792 
4793 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4794 
4795 #ifdef CONFIG_FAIR_GROUP_SCHED
4796 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4797 #endif
4798 
4799 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4800 {
4801 	return NULL;
4802 }
4803 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4804 static inline void update_runtime_enabled(struct rq *rq) {}
4805 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4806 
4807 #endif /* CONFIG_CFS_BANDWIDTH */
4808 
4809 /**************************************************
4810  * CFS operations on tasks:
4811  */
4812 
4813 #ifdef CONFIG_SCHED_HRTICK
4814 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4815 {
4816 	struct sched_entity *se = &p->se;
4817 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4818 
4819 	SCHED_WARN_ON(task_rq(p) != rq);
4820 
4821 	if (rq->cfs.h_nr_running > 1) {
4822 		u64 slice = sched_slice(cfs_rq, se);
4823 		u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4824 		s64 delta = slice - ran;
4825 
4826 		if (delta < 0) {
4827 			if (rq->curr == p)
4828 				resched_curr(rq);
4829 			return;
4830 		}
4831 		hrtick_start(rq, delta);
4832 	}
4833 }
4834 
4835 /*
4836  * called from enqueue/dequeue and updates the hrtick when the
4837  * current task is from our class and nr_running is low enough
4838  * to matter.
4839  */
4840 static void hrtick_update(struct rq *rq)
4841 {
4842 	struct task_struct *curr = rq->curr;
4843 
4844 	if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4845 		return;
4846 
4847 	if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4848 		hrtick_start_fair(rq, curr);
4849 }
4850 #else /* !CONFIG_SCHED_HRTICK */
4851 static inline void
4852 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4853 {
4854 }
4855 
4856 static inline void hrtick_update(struct rq *rq)
4857 {
4858 }
4859 #endif
4860 
4861 /*
4862  * The enqueue_task method is called before nr_running is
4863  * increased. Here we update the fair scheduling stats and
4864  * then put the task into the rbtree:
4865  */
4866 static void
4867 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4868 {
4869 	struct cfs_rq *cfs_rq;
4870 	struct sched_entity *se = &p->se;
4871 
4872 	/*
4873 	 * If in_iowait is set, the code below may not trigger any cpufreq
4874 	 * utilization updates, so do it here explicitly with the IOWAIT flag
4875 	 * passed.
4876 	 */
4877 	if (p->in_iowait)
4878 		cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4879 
4880 	for_each_sched_entity(se) {
4881 		if (se->on_rq)
4882 			break;
4883 		cfs_rq = cfs_rq_of(se);
4884 		enqueue_entity(cfs_rq, se, flags);
4885 
4886 		/*
4887 		 * end evaluation on encountering a throttled cfs_rq
4888 		 *
4889 		 * note: in the case of encountering a throttled cfs_rq we will
4890 		 * post the final h_nr_running increment below.
4891 		 */
4892 		if (cfs_rq_throttled(cfs_rq))
4893 			break;
4894 		cfs_rq->h_nr_running++;
4895 
4896 		flags = ENQUEUE_WAKEUP;
4897 	}
4898 
4899 	for_each_sched_entity(se) {
4900 		cfs_rq = cfs_rq_of(se);
4901 		cfs_rq->h_nr_running++;
4902 
4903 		if (cfs_rq_throttled(cfs_rq))
4904 			break;
4905 
4906 		update_load_avg(se, UPDATE_TG);
4907 		update_cfs_shares(se);
4908 	}
4909 
4910 	if (!se)
4911 		add_nr_running(rq, 1);
4912 
4913 	hrtick_update(rq);
4914 }
4915 
4916 static void set_next_buddy(struct sched_entity *se);
4917 
4918 /*
4919  * The dequeue_task method is called before nr_running is
4920  * decreased. We remove the task from the rbtree and
4921  * update the fair scheduling stats:
4922  */
4923 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4924 {
4925 	struct cfs_rq *cfs_rq;
4926 	struct sched_entity *se = &p->se;
4927 	int task_sleep = flags & DEQUEUE_SLEEP;
4928 
4929 	for_each_sched_entity(se) {
4930 		cfs_rq = cfs_rq_of(se);
4931 		dequeue_entity(cfs_rq, se, flags);
4932 
4933 		/*
4934 		 * end evaluation on encountering a throttled cfs_rq
4935 		 *
4936 		 * note: in the case of encountering a throttled cfs_rq we will
4937 		 * post the final h_nr_running decrement below.
4938 		*/
4939 		if (cfs_rq_throttled(cfs_rq))
4940 			break;
4941 		cfs_rq->h_nr_running--;
4942 
4943 		/* Don't dequeue parent if it has other entities besides us */
4944 		if (cfs_rq->load.weight) {
4945 			/* Avoid re-evaluating load for this entity: */
4946 			se = parent_entity(se);
4947 			/*
4948 			 * Bias pick_next to pick a task from this cfs_rq, as
4949 			 * p is sleeping when it is within its sched_slice.
4950 			 */
4951 			if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4952 				set_next_buddy(se);
4953 			break;
4954 		}
4955 		flags |= DEQUEUE_SLEEP;
4956 	}
4957 
4958 	for_each_sched_entity(se) {
4959 		cfs_rq = cfs_rq_of(se);
4960 		cfs_rq->h_nr_running--;
4961 
4962 		if (cfs_rq_throttled(cfs_rq))
4963 			break;
4964 
4965 		update_load_avg(se, UPDATE_TG);
4966 		update_cfs_shares(se);
4967 	}
4968 
4969 	if (!se)
4970 		sub_nr_running(rq, 1);
4971 
4972 	hrtick_update(rq);
4973 }
4974 
4975 #ifdef CONFIG_SMP
4976 
4977 /* Working cpumask for: load_balance, load_balance_newidle. */
4978 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
4979 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
4980 
4981 #ifdef CONFIG_NO_HZ_COMMON
4982 /*
4983  * per rq 'load' arrray crap; XXX kill this.
4984  */
4985 
4986 /*
4987  * The exact cpuload calculated at every tick would be:
4988  *
4989  *   load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4990  *
4991  * If a cpu misses updates for n ticks (as it was idle) and update gets
4992  * called on the n+1-th tick when cpu may be busy, then we have:
4993  *
4994  *   load_n   = (1 - 1/2^i)^n * load_0
4995  *   load_n+1 = (1 - 1/2^i)   * load_n + (1/2^i) * cur_load
4996  *
4997  * decay_load_missed() below does efficient calculation of
4998  *
4999  *   load' = (1 - 1/2^i)^n * load
5000  *
5001  * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5002  * This allows us to precompute the above in said factors, thereby allowing the
5003  * reduction of an arbitrary n in O(log_2 n) steps. (See also
5004  * fixed_power_int())
5005  *
5006  * The calculation is approximated on a 128 point scale.
5007  */
5008 #define DEGRADE_SHIFT		7
5009 
5010 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5011 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5012 	{   0,   0,  0,  0,  0,  0, 0, 0 },
5013 	{  64,  32,  8,  0,  0,  0, 0, 0 },
5014 	{  96,  72, 40, 12,  1,  0, 0, 0 },
5015 	{ 112,  98, 75, 43, 15,  1, 0, 0 },
5016 	{ 120, 112, 98, 76, 45, 16, 2, 0 }
5017 };
5018 
5019 /*
5020  * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5021  * would be when CPU is idle and so we just decay the old load without
5022  * adding any new load.
5023  */
5024 static unsigned long
5025 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5026 {
5027 	int j = 0;
5028 
5029 	if (!missed_updates)
5030 		return load;
5031 
5032 	if (missed_updates >= degrade_zero_ticks[idx])
5033 		return 0;
5034 
5035 	if (idx == 1)
5036 		return load >> missed_updates;
5037 
5038 	while (missed_updates) {
5039 		if (missed_updates % 2)
5040 			load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5041 
5042 		missed_updates >>= 1;
5043 		j++;
5044 	}
5045 	return load;
5046 }
5047 #endif /* CONFIG_NO_HZ_COMMON */
5048 
5049 /**
5050  * __cpu_load_update - update the rq->cpu_load[] statistics
5051  * @this_rq: The rq to update statistics for
5052  * @this_load: The current load
5053  * @pending_updates: The number of missed updates
5054  *
5055  * Update rq->cpu_load[] statistics. This function is usually called every
5056  * scheduler tick (TICK_NSEC).
5057  *
5058  * This function computes a decaying average:
5059  *
5060  *   load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5061  *
5062  * Because of NOHZ it might not get called on every tick which gives need for
5063  * the @pending_updates argument.
5064  *
5065  *   load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5066  *             = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5067  *             = A * (A * load[i]_n-2 + B) + B
5068  *             = A * (A * (A * load[i]_n-3 + B) + B) + B
5069  *             = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5070  *             = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5071  *             = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5072  *             = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5073  *
5074  * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5075  * any change in load would have resulted in the tick being turned back on.
5076  *
5077  * For regular NOHZ, this reduces to:
5078  *
5079  *   load[i]_n = (1 - 1/2^i)^n * load[i]_0
5080  *
5081  * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5082  * term.
5083  */
5084 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5085 			    unsigned long pending_updates)
5086 {
5087 	unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5088 	int i, scale;
5089 
5090 	this_rq->nr_load_updates++;
5091 
5092 	/* Update our load: */
5093 	this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5094 	for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5095 		unsigned long old_load, new_load;
5096 
5097 		/* scale is effectively 1 << i now, and >> i divides by scale */
5098 
5099 		old_load = this_rq->cpu_load[i];
5100 #ifdef CONFIG_NO_HZ_COMMON
5101 		old_load = decay_load_missed(old_load, pending_updates - 1, i);
5102 		if (tickless_load) {
5103 			old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5104 			/*
5105 			 * old_load can never be a negative value because a
5106 			 * decayed tickless_load cannot be greater than the
5107 			 * original tickless_load.
5108 			 */
5109 			old_load += tickless_load;
5110 		}
5111 #endif
5112 		new_load = this_load;
5113 		/*
5114 		 * Round up the averaging division if load is increasing. This
5115 		 * prevents us from getting stuck on 9 if the load is 10, for
5116 		 * example.
5117 		 */
5118 		if (new_load > old_load)
5119 			new_load += scale - 1;
5120 
5121 		this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5122 	}
5123 
5124 	sched_avg_update(this_rq);
5125 }
5126 
5127 /* Used instead of source_load when we know the type == 0 */
5128 static unsigned long weighted_cpuload(const int cpu)
5129 {
5130 	return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
5131 }
5132 
5133 #ifdef CONFIG_NO_HZ_COMMON
5134 /*
5135  * There is no sane way to deal with nohz on smp when using jiffies because the
5136  * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5137  * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5138  *
5139  * Therefore we need to avoid the delta approach from the regular tick when
5140  * possible since that would seriously skew the load calculation. This is why we
5141  * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5142  * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5143  * loop exit, nohz_idle_balance, nohz full exit...)
5144  *
5145  * This means we might still be one tick off for nohz periods.
5146  */
5147 
5148 static void cpu_load_update_nohz(struct rq *this_rq,
5149 				 unsigned long curr_jiffies,
5150 				 unsigned long load)
5151 {
5152 	unsigned long pending_updates;
5153 
5154 	pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5155 	if (pending_updates) {
5156 		this_rq->last_load_update_tick = curr_jiffies;
5157 		/*
5158 		 * In the regular NOHZ case, we were idle, this means load 0.
5159 		 * In the NOHZ_FULL case, we were non-idle, we should consider
5160 		 * its weighted load.
5161 		 */
5162 		cpu_load_update(this_rq, load, pending_updates);
5163 	}
5164 }
5165 
5166 /*
5167  * Called from nohz_idle_balance() to update the load ratings before doing the
5168  * idle balance.
5169  */
5170 static void cpu_load_update_idle(struct rq *this_rq)
5171 {
5172 	/*
5173 	 * bail if there's load or we're actually up-to-date.
5174 	 */
5175 	if (weighted_cpuload(cpu_of(this_rq)))
5176 		return;
5177 
5178 	cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5179 }
5180 
5181 /*
5182  * Record CPU load on nohz entry so we know the tickless load to account
5183  * on nohz exit. cpu_load[0] happens then to be updated more frequently
5184  * than other cpu_load[idx] but it should be fine as cpu_load readers
5185  * shouldn't rely into synchronized cpu_load[*] updates.
5186  */
5187 void cpu_load_update_nohz_start(void)
5188 {
5189 	struct rq *this_rq = this_rq();
5190 
5191 	/*
5192 	 * This is all lockless but should be fine. If weighted_cpuload changes
5193 	 * concurrently we'll exit nohz. And cpu_load write can race with
5194 	 * cpu_load_update_idle() but both updater would be writing the same.
5195 	 */
5196 	this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
5197 }
5198 
5199 /*
5200  * Account the tickless load in the end of a nohz frame.
5201  */
5202 void cpu_load_update_nohz_stop(void)
5203 {
5204 	unsigned long curr_jiffies = READ_ONCE(jiffies);
5205 	struct rq *this_rq = this_rq();
5206 	unsigned long load;
5207 	struct rq_flags rf;
5208 
5209 	if (curr_jiffies == this_rq->last_load_update_tick)
5210 		return;
5211 
5212 	load = weighted_cpuload(cpu_of(this_rq));
5213 	rq_lock(this_rq, &rf);
5214 	update_rq_clock(this_rq);
5215 	cpu_load_update_nohz(this_rq, curr_jiffies, load);
5216 	rq_unlock(this_rq, &rf);
5217 }
5218 #else /* !CONFIG_NO_HZ_COMMON */
5219 static inline void cpu_load_update_nohz(struct rq *this_rq,
5220 					unsigned long curr_jiffies,
5221 					unsigned long load) { }
5222 #endif /* CONFIG_NO_HZ_COMMON */
5223 
5224 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5225 {
5226 #ifdef CONFIG_NO_HZ_COMMON
5227 	/* See the mess around cpu_load_update_nohz(). */
5228 	this_rq->last_load_update_tick = READ_ONCE(jiffies);
5229 #endif
5230 	cpu_load_update(this_rq, load, 1);
5231 }
5232 
5233 /*
5234  * Called from scheduler_tick()
5235  */
5236 void cpu_load_update_active(struct rq *this_rq)
5237 {
5238 	unsigned long load = weighted_cpuload(cpu_of(this_rq));
5239 
5240 	if (tick_nohz_tick_stopped())
5241 		cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5242 	else
5243 		cpu_load_update_periodic(this_rq, load);
5244 }
5245 
5246 /*
5247  * Return a low guess at the load of a migration-source cpu weighted
5248  * according to the scheduling class and "nice" value.
5249  *
5250  * We want to under-estimate the load of migration sources, to
5251  * balance conservatively.
5252  */
5253 static unsigned long source_load(int cpu, int type)
5254 {
5255 	struct rq *rq = cpu_rq(cpu);
5256 	unsigned long total = weighted_cpuload(cpu);
5257 
5258 	if (type == 0 || !sched_feat(LB_BIAS))
5259 		return total;
5260 
5261 	return min(rq->cpu_load[type-1], total);
5262 }
5263 
5264 /*
5265  * Return a high guess at the load of a migration-target cpu weighted
5266  * according to the scheduling class and "nice" value.
5267  */
5268 static unsigned long target_load(int cpu, int type)
5269 {
5270 	struct rq *rq = cpu_rq(cpu);
5271 	unsigned long total = weighted_cpuload(cpu);
5272 
5273 	if (type == 0 || !sched_feat(LB_BIAS))
5274 		return total;
5275 
5276 	return max(rq->cpu_load[type-1], total);
5277 }
5278 
5279 static unsigned long capacity_of(int cpu)
5280 {
5281 	return cpu_rq(cpu)->cpu_capacity;
5282 }
5283 
5284 static unsigned long capacity_orig_of(int cpu)
5285 {
5286 	return cpu_rq(cpu)->cpu_capacity_orig;
5287 }
5288 
5289 static unsigned long cpu_avg_load_per_task(int cpu)
5290 {
5291 	struct rq *rq = cpu_rq(cpu);
5292 	unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5293 	unsigned long load_avg = weighted_cpuload(cpu);
5294 
5295 	if (nr_running)
5296 		return load_avg / nr_running;
5297 
5298 	return 0;
5299 }
5300 
5301 static void record_wakee(struct task_struct *p)
5302 {
5303 	/*
5304 	 * Only decay a single time; tasks that have less then 1 wakeup per
5305 	 * jiffy will not have built up many flips.
5306 	 */
5307 	if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5308 		current->wakee_flips >>= 1;
5309 		current->wakee_flip_decay_ts = jiffies;
5310 	}
5311 
5312 	if (current->last_wakee != p) {
5313 		current->last_wakee = p;
5314 		current->wakee_flips++;
5315 	}
5316 }
5317 
5318 /*
5319  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5320  *
5321  * A waker of many should wake a different task than the one last awakened
5322  * at a frequency roughly N times higher than one of its wakees.
5323  *
5324  * In order to determine whether we should let the load spread vs consolidating
5325  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5326  * partner, and a factor of lls_size higher frequency in the other.
5327  *
5328  * With both conditions met, we can be relatively sure that the relationship is
5329  * non-monogamous, with partner count exceeding socket size.
5330  *
5331  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5332  * whatever is irrelevant, spread criteria is apparent partner count exceeds
5333  * socket size.
5334  */
5335 static int wake_wide(struct task_struct *p)
5336 {
5337 	unsigned int master = current->wakee_flips;
5338 	unsigned int slave = p->wakee_flips;
5339 	int factor = this_cpu_read(sd_llc_size);
5340 
5341 	if (master < slave)
5342 		swap(master, slave);
5343 	if (slave < factor || master < slave * factor)
5344 		return 0;
5345 	return 1;
5346 }
5347 
5348 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5349 		       int prev_cpu, int sync)
5350 {
5351 	int this_cpu = smp_processor_id();
5352 	bool affine = false;
5353 
5354 	/*
5355 	 * Common case: CPUs are in the same socket, and select_idle_sibling()
5356 	 * will do its thing regardless of what we return:
5357 	 */
5358 	if (cpus_share_cache(prev_cpu, this_cpu))
5359 		affine = true;
5360 	else
5361 		affine = numa_wake_affine(sd, p, this_cpu, prev_cpu, sync);
5362 
5363 	schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5364 	if (affine) {
5365 		schedstat_inc(sd->ttwu_move_affine);
5366 		schedstat_inc(p->se.statistics.nr_wakeups_affine);
5367 	}
5368 
5369 	return affine;
5370 }
5371 
5372 static inline int task_util(struct task_struct *p);
5373 static int cpu_util_wake(int cpu, struct task_struct *p);
5374 
5375 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5376 {
5377 	return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5378 }
5379 
5380 /*
5381  * find_idlest_group finds and returns the least busy CPU group within the
5382  * domain.
5383  */
5384 static struct sched_group *
5385 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5386 		  int this_cpu, int sd_flag)
5387 {
5388 	struct sched_group *idlest = NULL, *group = sd->groups;
5389 	struct sched_group *most_spare_sg = NULL;
5390 	unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5391 	unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
5392 	unsigned long most_spare = 0, this_spare = 0;
5393 	int load_idx = sd->forkexec_idx;
5394 	int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5395 	unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5396 				(sd->imbalance_pct-100) / 100;
5397 
5398 	if (sd_flag & SD_BALANCE_WAKE)
5399 		load_idx = sd->wake_idx;
5400 
5401 	do {
5402 		unsigned long load, avg_load, runnable_load;
5403 		unsigned long spare_cap, max_spare_cap;
5404 		int local_group;
5405 		int i;
5406 
5407 		/* Skip over this group if it has no CPUs allowed */
5408 		if (!cpumask_intersects(sched_group_span(group),
5409 					&p->cpus_allowed))
5410 			continue;
5411 
5412 		local_group = cpumask_test_cpu(this_cpu,
5413 					       sched_group_span(group));
5414 
5415 		/*
5416 		 * Tally up the load of all CPUs in the group and find
5417 		 * the group containing the CPU with most spare capacity.
5418 		 */
5419 		avg_load = 0;
5420 		runnable_load = 0;
5421 		max_spare_cap = 0;
5422 
5423 		for_each_cpu(i, sched_group_span(group)) {
5424 			/* Bias balancing toward cpus of our domain */
5425 			if (local_group)
5426 				load = source_load(i, load_idx);
5427 			else
5428 				load = target_load(i, load_idx);
5429 
5430 			runnable_load += load;
5431 
5432 			avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5433 
5434 			spare_cap = capacity_spare_wake(i, p);
5435 
5436 			if (spare_cap > max_spare_cap)
5437 				max_spare_cap = spare_cap;
5438 		}
5439 
5440 		/* Adjust by relative CPU capacity of the group */
5441 		avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5442 					group->sgc->capacity;
5443 		runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5444 					group->sgc->capacity;
5445 
5446 		if (local_group) {
5447 			this_runnable_load = runnable_load;
5448 			this_avg_load = avg_load;
5449 			this_spare = max_spare_cap;
5450 		} else {
5451 			if (min_runnable_load > (runnable_load + imbalance)) {
5452 				/*
5453 				 * The runnable load is significantly smaller
5454 				 * so we can pick this new cpu
5455 				 */
5456 				min_runnable_load = runnable_load;
5457 				min_avg_load = avg_load;
5458 				idlest = group;
5459 			} else if ((runnable_load < (min_runnable_load + imbalance)) &&
5460 				   (100*min_avg_load > imbalance_scale*avg_load)) {
5461 				/*
5462 				 * The runnable loads are close so take the
5463 				 * blocked load into account through avg_load.
5464 				 */
5465 				min_avg_load = avg_load;
5466 				idlest = group;
5467 			}
5468 
5469 			if (most_spare < max_spare_cap) {
5470 				most_spare = max_spare_cap;
5471 				most_spare_sg = group;
5472 			}
5473 		}
5474 	} while (group = group->next, group != sd->groups);
5475 
5476 	/*
5477 	 * The cross-over point between using spare capacity or least load
5478 	 * is too conservative for high utilization tasks on partially
5479 	 * utilized systems if we require spare_capacity > task_util(p),
5480 	 * so we allow for some task stuffing by using
5481 	 * spare_capacity > task_util(p)/2.
5482 	 *
5483 	 * Spare capacity can't be used for fork because the utilization has
5484 	 * not been set yet, we must first select a rq to compute the initial
5485 	 * utilization.
5486 	 */
5487 	if (sd_flag & SD_BALANCE_FORK)
5488 		goto skip_spare;
5489 
5490 	if (this_spare > task_util(p) / 2 &&
5491 	    imbalance_scale*this_spare > 100*most_spare)
5492 		return NULL;
5493 
5494 	if (most_spare > task_util(p) / 2)
5495 		return most_spare_sg;
5496 
5497 skip_spare:
5498 	if (!idlest)
5499 		return NULL;
5500 
5501 	if (min_runnable_load > (this_runnable_load + imbalance))
5502 		return NULL;
5503 
5504 	if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5505 	     (100*this_avg_load < imbalance_scale*min_avg_load))
5506 		return NULL;
5507 
5508 	return idlest;
5509 }
5510 
5511 /*
5512  * find_idlest_cpu - find the idlest cpu among the cpus in group.
5513  */
5514 static int
5515 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5516 {
5517 	unsigned long load, min_load = ULONG_MAX;
5518 	unsigned int min_exit_latency = UINT_MAX;
5519 	u64 latest_idle_timestamp = 0;
5520 	int least_loaded_cpu = this_cpu;
5521 	int shallowest_idle_cpu = -1;
5522 	int i;
5523 
5524 	/* Check if we have any choice: */
5525 	if (group->group_weight == 1)
5526 		return cpumask_first(sched_group_span(group));
5527 
5528 	/* Traverse only the allowed CPUs */
5529 	for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5530 		if (idle_cpu(i)) {
5531 			struct rq *rq = cpu_rq(i);
5532 			struct cpuidle_state *idle = idle_get_state(rq);
5533 			if (idle && idle->exit_latency < min_exit_latency) {
5534 				/*
5535 				 * We give priority to a CPU whose idle state
5536 				 * has the smallest exit latency irrespective
5537 				 * of any idle timestamp.
5538 				 */
5539 				min_exit_latency = idle->exit_latency;
5540 				latest_idle_timestamp = rq->idle_stamp;
5541 				shallowest_idle_cpu = i;
5542 			} else if ((!idle || idle->exit_latency == min_exit_latency) &&
5543 				   rq->idle_stamp > latest_idle_timestamp) {
5544 				/*
5545 				 * If equal or no active idle state, then
5546 				 * the most recently idled CPU might have
5547 				 * a warmer cache.
5548 				 */
5549 				latest_idle_timestamp = rq->idle_stamp;
5550 				shallowest_idle_cpu = i;
5551 			}
5552 		} else if (shallowest_idle_cpu == -1) {
5553 			load = weighted_cpuload(i);
5554 			if (load < min_load || (load == min_load && i == this_cpu)) {
5555 				min_load = load;
5556 				least_loaded_cpu = i;
5557 			}
5558 		}
5559 	}
5560 
5561 	return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5562 }
5563 
5564 #ifdef CONFIG_SCHED_SMT
5565 
5566 static inline void set_idle_cores(int cpu, int val)
5567 {
5568 	struct sched_domain_shared *sds;
5569 
5570 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5571 	if (sds)
5572 		WRITE_ONCE(sds->has_idle_cores, val);
5573 }
5574 
5575 static inline bool test_idle_cores(int cpu, bool def)
5576 {
5577 	struct sched_domain_shared *sds;
5578 
5579 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5580 	if (sds)
5581 		return READ_ONCE(sds->has_idle_cores);
5582 
5583 	return def;
5584 }
5585 
5586 /*
5587  * Scans the local SMT mask to see if the entire core is idle, and records this
5588  * information in sd_llc_shared->has_idle_cores.
5589  *
5590  * Since SMT siblings share all cache levels, inspecting this limited remote
5591  * state should be fairly cheap.
5592  */
5593 void __update_idle_core(struct rq *rq)
5594 {
5595 	int core = cpu_of(rq);
5596 	int cpu;
5597 
5598 	rcu_read_lock();
5599 	if (test_idle_cores(core, true))
5600 		goto unlock;
5601 
5602 	for_each_cpu(cpu, cpu_smt_mask(core)) {
5603 		if (cpu == core)
5604 			continue;
5605 
5606 		if (!idle_cpu(cpu))
5607 			goto unlock;
5608 	}
5609 
5610 	set_idle_cores(core, 1);
5611 unlock:
5612 	rcu_read_unlock();
5613 }
5614 
5615 /*
5616  * Scan the entire LLC domain for idle cores; this dynamically switches off if
5617  * there are no idle cores left in the system; tracked through
5618  * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5619  */
5620 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5621 {
5622 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5623 	int core, cpu;
5624 
5625 	if (!static_branch_likely(&sched_smt_present))
5626 		return -1;
5627 
5628 	if (!test_idle_cores(target, false))
5629 		return -1;
5630 
5631 	cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5632 
5633 	for_each_cpu_wrap(core, cpus, target) {
5634 		bool idle = true;
5635 
5636 		for_each_cpu(cpu, cpu_smt_mask(core)) {
5637 			cpumask_clear_cpu(cpu, cpus);
5638 			if (!idle_cpu(cpu))
5639 				idle = false;
5640 		}
5641 
5642 		if (idle)
5643 			return core;
5644 	}
5645 
5646 	/*
5647 	 * Failed to find an idle core; stop looking for one.
5648 	 */
5649 	set_idle_cores(target, 0);
5650 
5651 	return -1;
5652 }
5653 
5654 /*
5655  * Scan the local SMT mask for idle CPUs.
5656  */
5657 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5658 {
5659 	int cpu;
5660 
5661 	if (!static_branch_likely(&sched_smt_present))
5662 		return -1;
5663 
5664 	for_each_cpu(cpu, cpu_smt_mask(target)) {
5665 		if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5666 			continue;
5667 		if (idle_cpu(cpu))
5668 			return cpu;
5669 	}
5670 
5671 	return -1;
5672 }
5673 
5674 #else /* CONFIG_SCHED_SMT */
5675 
5676 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5677 {
5678 	return -1;
5679 }
5680 
5681 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5682 {
5683 	return -1;
5684 }
5685 
5686 #endif /* CONFIG_SCHED_SMT */
5687 
5688 /*
5689  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5690  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5691  * average idle time for this rq (as found in rq->avg_idle).
5692  */
5693 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5694 {
5695 	struct sched_domain *this_sd;
5696 	u64 avg_cost, avg_idle;
5697 	u64 time, cost;
5698 	s64 delta;
5699 	int cpu, nr = INT_MAX;
5700 
5701 	this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5702 	if (!this_sd)
5703 		return -1;
5704 
5705 	/*
5706 	 * Due to large variance we need a large fuzz factor; hackbench in
5707 	 * particularly is sensitive here.
5708 	 */
5709 	avg_idle = this_rq()->avg_idle / 512;
5710 	avg_cost = this_sd->avg_scan_cost + 1;
5711 
5712 	if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5713 		return -1;
5714 
5715 	if (sched_feat(SIS_PROP)) {
5716 		u64 span_avg = sd->span_weight * avg_idle;
5717 		if (span_avg > 4*avg_cost)
5718 			nr = div_u64(span_avg, avg_cost);
5719 		else
5720 			nr = 4;
5721 	}
5722 
5723 	time = local_clock();
5724 
5725 	for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5726 		if (!--nr)
5727 			return -1;
5728 		if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5729 			continue;
5730 		if (idle_cpu(cpu))
5731 			break;
5732 	}
5733 
5734 	time = local_clock() - time;
5735 	cost = this_sd->avg_scan_cost;
5736 	delta = (s64)(time - cost) / 8;
5737 	this_sd->avg_scan_cost += delta;
5738 
5739 	return cpu;
5740 }
5741 
5742 /*
5743  * Try and locate an idle core/thread in the LLC cache domain.
5744  */
5745 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5746 {
5747 	struct sched_domain *sd;
5748 	int i;
5749 
5750 	if (idle_cpu(target))
5751 		return target;
5752 
5753 	/*
5754 	 * If the previous cpu is cache affine and idle, don't be stupid.
5755 	 */
5756 	if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5757 		return prev;
5758 
5759 	sd = rcu_dereference(per_cpu(sd_llc, target));
5760 	if (!sd)
5761 		return target;
5762 
5763 	i = select_idle_core(p, sd, target);
5764 	if ((unsigned)i < nr_cpumask_bits)
5765 		return i;
5766 
5767 	i = select_idle_cpu(p, sd, target);
5768 	if ((unsigned)i < nr_cpumask_bits)
5769 		return i;
5770 
5771 	i = select_idle_smt(p, sd, target);
5772 	if ((unsigned)i < nr_cpumask_bits)
5773 		return i;
5774 
5775 	return target;
5776 }
5777 
5778 /*
5779  * cpu_util returns the amount of capacity of a CPU that is used by CFS
5780  * tasks. The unit of the return value must be the one of capacity so we can
5781  * compare the utilization with the capacity of the CPU that is available for
5782  * CFS task (ie cpu_capacity).
5783  *
5784  * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5785  * recent utilization of currently non-runnable tasks on a CPU. It represents
5786  * the amount of utilization of a CPU in the range [0..capacity_orig] where
5787  * capacity_orig is the cpu_capacity available at the highest frequency
5788  * (arch_scale_freq_capacity()).
5789  * The utilization of a CPU converges towards a sum equal to or less than the
5790  * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5791  * the running time on this CPU scaled by capacity_curr.
5792  *
5793  * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5794  * higher than capacity_orig because of unfortunate rounding in
5795  * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5796  * the average stabilizes with the new running time. We need to check that the
5797  * utilization stays within the range of [0..capacity_orig] and cap it if
5798  * necessary. Without utilization capping, a group could be seen as overloaded
5799  * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5800  * available capacity. We allow utilization to overshoot capacity_curr (but not
5801  * capacity_orig) as it useful for predicting the capacity required after task
5802  * migrations (scheduler-driven DVFS).
5803  */
5804 static int cpu_util(int cpu)
5805 {
5806 	unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5807 	unsigned long capacity = capacity_orig_of(cpu);
5808 
5809 	return (util >= capacity) ? capacity : util;
5810 }
5811 
5812 static inline int task_util(struct task_struct *p)
5813 {
5814 	return p->se.avg.util_avg;
5815 }
5816 
5817 /*
5818  * cpu_util_wake: Compute cpu utilization with any contributions from
5819  * the waking task p removed.
5820  */
5821 static int cpu_util_wake(int cpu, struct task_struct *p)
5822 {
5823 	unsigned long util, capacity;
5824 
5825 	/* Task has no contribution or is new */
5826 	if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5827 		return cpu_util(cpu);
5828 
5829 	capacity = capacity_orig_of(cpu);
5830 	util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5831 
5832 	return (util >= capacity) ? capacity : util;
5833 }
5834 
5835 /*
5836  * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5837  * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5838  *
5839  * In that case WAKE_AFFINE doesn't make sense and we'll let
5840  * BALANCE_WAKE sort things out.
5841  */
5842 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5843 {
5844 	long min_cap, max_cap;
5845 
5846 	min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5847 	max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5848 
5849 	/* Minimum capacity is close to max, no need to abort wake_affine */
5850 	if (max_cap - min_cap < max_cap >> 3)
5851 		return 0;
5852 
5853 	/* Bring task utilization in sync with prev_cpu */
5854 	sync_entity_load_avg(&p->se);
5855 
5856 	return min_cap * 1024 < task_util(p) * capacity_margin;
5857 }
5858 
5859 /*
5860  * select_task_rq_fair: Select target runqueue for the waking task in domains
5861  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5862  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5863  *
5864  * Balances load by selecting the idlest cpu in the idlest group, or under
5865  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5866  *
5867  * Returns the target cpu number.
5868  *
5869  * preempt must be disabled.
5870  */
5871 static int
5872 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5873 {
5874 	struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5875 	int cpu = smp_processor_id();
5876 	int new_cpu = prev_cpu;
5877 	int want_affine = 0;
5878 	int sync = wake_flags & WF_SYNC;
5879 
5880 	if (sd_flag & SD_BALANCE_WAKE) {
5881 		record_wakee(p);
5882 		want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
5883 			      && cpumask_test_cpu(cpu, &p->cpus_allowed);
5884 	}
5885 
5886 	rcu_read_lock();
5887 	for_each_domain(cpu, tmp) {
5888 		if (!(tmp->flags & SD_LOAD_BALANCE))
5889 			break;
5890 
5891 		/*
5892 		 * If both cpu and prev_cpu are part of this domain,
5893 		 * cpu is a valid SD_WAKE_AFFINE target.
5894 		 */
5895 		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5896 		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5897 			affine_sd = tmp;
5898 			break;
5899 		}
5900 
5901 		if (tmp->flags & sd_flag)
5902 			sd = tmp;
5903 		else if (!want_affine)
5904 			break;
5905 	}
5906 
5907 	if (affine_sd) {
5908 		sd = NULL; /* Prefer wake_affine over balance flags */
5909 		if (cpu == prev_cpu)
5910 			goto pick_cpu;
5911 
5912 		if (wake_affine(affine_sd, p, prev_cpu, sync))
5913 			new_cpu = cpu;
5914 	}
5915 
5916 	if (!sd) {
5917  pick_cpu:
5918 		if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5919 			new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
5920 
5921 	} else while (sd) {
5922 		struct sched_group *group;
5923 		int weight;
5924 
5925 		if (!(sd->flags & sd_flag)) {
5926 			sd = sd->child;
5927 			continue;
5928 		}
5929 
5930 		group = find_idlest_group(sd, p, cpu, sd_flag);
5931 		if (!group) {
5932 			sd = sd->child;
5933 			continue;
5934 		}
5935 
5936 		new_cpu = find_idlest_cpu(group, p, cpu);
5937 		if (new_cpu == -1 || new_cpu == cpu) {
5938 			/* Now try balancing at a lower domain level of cpu */
5939 			sd = sd->child;
5940 			continue;
5941 		}
5942 
5943 		/* Now try balancing at a lower domain level of new_cpu */
5944 		cpu = new_cpu;
5945 		weight = sd->span_weight;
5946 		sd = NULL;
5947 		for_each_domain(cpu, tmp) {
5948 			if (weight <= tmp->span_weight)
5949 				break;
5950 			if (tmp->flags & sd_flag)
5951 				sd = tmp;
5952 		}
5953 		/* while loop will break here if sd == NULL */
5954 	}
5955 	rcu_read_unlock();
5956 
5957 	return new_cpu;
5958 }
5959 
5960 /*
5961  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5962  * cfs_rq_of(p) references at time of call are still valid and identify the
5963  * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5964  */
5965 static void migrate_task_rq_fair(struct task_struct *p)
5966 {
5967 	/*
5968 	 * As blocked tasks retain absolute vruntime the migration needs to
5969 	 * deal with this by subtracting the old and adding the new
5970 	 * min_vruntime -- the latter is done by enqueue_entity() when placing
5971 	 * the task on the new runqueue.
5972 	 */
5973 	if (p->state == TASK_WAKING) {
5974 		struct sched_entity *se = &p->se;
5975 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
5976 		u64 min_vruntime;
5977 
5978 #ifndef CONFIG_64BIT
5979 		u64 min_vruntime_copy;
5980 
5981 		do {
5982 			min_vruntime_copy = cfs_rq->min_vruntime_copy;
5983 			smp_rmb();
5984 			min_vruntime = cfs_rq->min_vruntime;
5985 		} while (min_vruntime != min_vruntime_copy);
5986 #else
5987 		min_vruntime = cfs_rq->min_vruntime;
5988 #endif
5989 
5990 		se->vruntime -= min_vruntime;
5991 	}
5992 
5993 	/*
5994 	 * We are supposed to update the task to "current" time, then its up to date
5995 	 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5996 	 * what current time is, so simply throw away the out-of-date time. This
5997 	 * will result in the wakee task is less decayed, but giving the wakee more
5998 	 * load sounds not bad.
5999 	 */
6000 	remove_entity_load_avg(&p->se);
6001 
6002 	/* Tell new CPU we are migrated */
6003 	p->se.avg.last_update_time = 0;
6004 
6005 	/* We have migrated, no longer consider this task hot */
6006 	p->se.exec_start = 0;
6007 }
6008 
6009 static void task_dead_fair(struct task_struct *p)
6010 {
6011 	remove_entity_load_avg(&p->se);
6012 }
6013 #endif /* CONFIG_SMP */
6014 
6015 static unsigned long
6016 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6017 {
6018 	unsigned long gran = sysctl_sched_wakeup_granularity;
6019 
6020 	/*
6021 	 * Since its curr running now, convert the gran from real-time
6022 	 * to virtual-time in his units.
6023 	 *
6024 	 * By using 'se' instead of 'curr' we penalize light tasks, so
6025 	 * they get preempted easier. That is, if 'se' < 'curr' then
6026 	 * the resulting gran will be larger, therefore penalizing the
6027 	 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6028 	 * be smaller, again penalizing the lighter task.
6029 	 *
6030 	 * This is especially important for buddies when the leftmost
6031 	 * task is higher priority than the buddy.
6032 	 */
6033 	return calc_delta_fair(gran, se);
6034 }
6035 
6036 /*
6037  * Should 'se' preempt 'curr'.
6038  *
6039  *             |s1
6040  *        |s2
6041  *   |s3
6042  *         g
6043  *      |<--->|c
6044  *
6045  *  w(c, s1) = -1
6046  *  w(c, s2) =  0
6047  *  w(c, s3) =  1
6048  *
6049  */
6050 static int
6051 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6052 {
6053 	s64 gran, vdiff = curr->vruntime - se->vruntime;
6054 
6055 	if (vdiff <= 0)
6056 		return -1;
6057 
6058 	gran = wakeup_gran(curr, se);
6059 	if (vdiff > gran)
6060 		return 1;
6061 
6062 	return 0;
6063 }
6064 
6065 static void set_last_buddy(struct sched_entity *se)
6066 {
6067 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6068 		return;
6069 
6070 	for_each_sched_entity(se) {
6071 		if (SCHED_WARN_ON(!se->on_rq))
6072 			return;
6073 		cfs_rq_of(se)->last = se;
6074 	}
6075 }
6076 
6077 static void set_next_buddy(struct sched_entity *se)
6078 {
6079 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6080 		return;
6081 
6082 	for_each_sched_entity(se) {
6083 		if (SCHED_WARN_ON(!se->on_rq))
6084 			return;
6085 		cfs_rq_of(se)->next = se;
6086 	}
6087 }
6088 
6089 static void set_skip_buddy(struct sched_entity *se)
6090 {
6091 	for_each_sched_entity(se)
6092 		cfs_rq_of(se)->skip = se;
6093 }
6094 
6095 /*
6096  * Preempt the current task with a newly woken task if needed:
6097  */
6098 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6099 {
6100 	struct task_struct *curr = rq->curr;
6101 	struct sched_entity *se = &curr->se, *pse = &p->se;
6102 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6103 	int scale = cfs_rq->nr_running >= sched_nr_latency;
6104 	int next_buddy_marked = 0;
6105 
6106 	if (unlikely(se == pse))
6107 		return;
6108 
6109 	/*
6110 	 * This is possible from callers such as attach_tasks(), in which we
6111 	 * unconditionally check_prempt_curr() after an enqueue (which may have
6112 	 * lead to a throttle).  This both saves work and prevents false
6113 	 * next-buddy nomination below.
6114 	 */
6115 	if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6116 		return;
6117 
6118 	if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6119 		set_next_buddy(pse);
6120 		next_buddy_marked = 1;
6121 	}
6122 
6123 	/*
6124 	 * We can come here with TIF_NEED_RESCHED already set from new task
6125 	 * wake up path.
6126 	 *
6127 	 * Note: this also catches the edge-case of curr being in a throttled
6128 	 * group (e.g. via set_curr_task), since update_curr() (in the
6129 	 * enqueue of curr) will have resulted in resched being set.  This
6130 	 * prevents us from potentially nominating it as a false LAST_BUDDY
6131 	 * below.
6132 	 */
6133 	if (test_tsk_need_resched(curr))
6134 		return;
6135 
6136 	/* Idle tasks are by definition preempted by non-idle tasks. */
6137 	if (unlikely(curr->policy == SCHED_IDLE) &&
6138 	    likely(p->policy != SCHED_IDLE))
6139 		goto preempt;
6140 
6141 	/*
6142 	 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6143 	 * is driven by the tick):
6144 	 */
6145 	if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6146 		return;
6147 
6148 	find_matching_se(&se, &pse);
6149 	update_curr(cfs_rq_of(se));
6150 	BUG_ON(!pse);
6151 	if (wakeup_preempt_entity(se, pse) == 1) {
6152 		/*
6153 		 * Bias pick_next to pick the sched entity that is
6154 		 * triggering this preemption.
6155 		 */
6156 		if (!next_buddy_marked)
6157 			set_next_buddy(pse);
6158 		goto preempt;
6159 	}
6160 
6161 	return;
6162 
6163 preempt:
6164 	resched_curr(rq);
6165 	/*
6166 	 * Only set the backward buddy when the current task is still
6167 	 * on the rq. This can happen when a wakeup gets interleaved
6168 	 * with schedule on the ->pre_schedule() or idle_balance()
6169 	 * point, either of which can * drop the rq lock.
6170 	 *
6171 	 * Also, during early boot the idle thread is in the fair class,
6172 	 * for obvious reasons its a bad idea to schedule back to it.
6173 	 */
6174 	if (unlikely(!se->on_rq || curr == rq->idle))
6175 		return;
6176 
6177 	if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6178 		set_last_buddy(se);
6179 }
6180 
6181 static struct task_struct *
6182 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6183 {
6184 	struct cfs_rq *cfs_rq = &rq->cfs;
6185 	struct sched_entity *se;
6186 	struct task_struct *p;
6187 	int new_tasks;
6188 
6189 again:
6190 #ifdef CONFIG_FAIR_GROUP_SCHED
6191 	if (!cfs_rq->nr_running)
6192 		goto idle;
6193 
6194 	if (prev->sched_class != &fair_sched_class)
6195 		goto simple;
6196 
6197 	/*
6198 	 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6199 	 * likely that a next task is from the same cgroup as the current.
6200 	 *
6201 	 * Therefore attempt to avoid putting and setting the entire cgroup
6202 	 * hierarchy, only change the part that actually changes.
6203 	 */
6204 
6205 	do {
6206 		struct sched_entity *curr = cfs_rq->curr;
6207 
6208 		/*
6209 		 * Since we got here without doing put_prev_entity() we also
6210 		 * have to consider cfs_rq->curr. If it is still a runnable
6211 		 * entity, update_curr() will update its vruntime, otherwise
6212 		 * forget we've ever seen it.
6213 		 */
6214 		if (curr) {
6215 			if (curr->on_rq)
6216 				update_curr(cfs_rq);
6217 			else
6218 				curr = NULL;
6219 
6220 			/*
6221 			 * This call to check_cfs_rq_runtime() will do the
6222 			 * throttle and dequeue its entity in the parent(s).
6223 			 * Therefore the 'simple' nr_running test will indeed
6224 			 * be correct.
6225 			 */
6226 			if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6227 				goto simple;
6228 		}
6229 
6230 		se = pick_next_entity(cfs_rq, curr);
6231 		cfs_rq = group_cfs_rq(se);
6232 	} while (cfs_rq);
6233 
6234 	p = task_of(se);
6235 
6236 	/*
6237 	 * Since we haven't yet done put_prev_entity and if the selected task
6238 	 * is a different task than we started out with, try and touch the
6239 	 * least amount of cfs_rqs.
6240 	 */
6241 	if (prev != p) {
6242 		struct sched_entity *pse = &prev->se;
6243 
6244 		while (!(cfs_rq = is_same_group(se, pse))) {
6245 			int se_depth = se->depth;
6246 			int pse_depth = pse->depth;
6247 
6248 			if (se_depth <= pse_depth) {
6249 				put_prev_entity(cfs_rq_of(pse), pse);
6250 				pse = parent_entity(pse);
6251 			}
6252 			if (se_depth >= pse_depth) {
6253 				set_next_entity(cfs_rq_of(se), se);
6254 				se = parent_entity(se);
6255 			}
6256 		}
6257 
6258 		put_prev_entity(cfs_rq, pse);
6259 		set_next_entity(cfs_rq, se);
6260 	}
6261 
6262 	if (hrtick_enabled(rq))
6263 		hrtick_start_fair(rq, p);
6264 
6265 	return p;
6266 simple:
6267 	cfs_rq = &rq->cfs;
6268 #endif
6269 
6270 	if (!cfs_rq->nr_running)
6271 		goto idle;
6272 
6273 	put_prev_task(rq, prev);
6274 
6275 	do {
6276 		se = pick_next_entity(cfs_rq, NULL);
6277 		set_next_entity(cfs_rq, se);
6278 		cfs_rq = group_cfs_rq(se);
6279 	} while (cfs_rq);
6280 
6281 	p = task_of(se);
6282 
6283 	if (hrtick_enabled(rq))
6284 		hrtick_start_fair(rq, p);
6285 
6286 	return p;
6287 
6288 idle:
6289 	new_tasks = idle_balance(rq, rf);
6290 
6291 	/*
6292 	 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6293 	 * possible for any higher priority task to appear. In that case we
6294 	 * must re-start the pick_next_entity() loop.
6295 	 */
6296 	if (new_tasks < 0)
6297 		return RETRY_TASK;
6298 
6299 	if (new_tasks > 0)
6300 		goto again;
6301 
6302 	return NULL;
6303 }
6304 
6305 /*
6306  * Account for a descheduled task:
6307  */
6308 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6309 {
6310 	struct sched_entity *se = &prev->se;
6311 	struct cfs_rq *cfs_rq;
6312 
6313 	for_each_sched_entity(se) {
6314 		cfs_rq = cfs_rq_of(se);
6315 		put_prev_entity(cfs_rq, se);
6316 	}
6317 }
6318 
6319 /*
6320  * sched_yield() is very simple
6321  *
6322  * The magic of dealing with the ->skip buddy is in pick_next_entity.
6323  */
6324 static void yield_task_fair(struct rq *rq)
6325 {
6326 	struct task_struct *curr = rq->curr;
6327 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6328 	struct sched_entity *se = &curr->se;
6329 
6330 	/*
6331 	 * Are we the only task in the tree?
6332 	 */
6333 	if (unlikely(rq->nr_running == 1))
6334 		return;
6335 
6336 	clear_buddies(cfs_rq, se);
6337 
6338 	if (curr->policy != SCHED_BATCH) {
6339 		update_rq_clock(rq);
6340 		/*
6341 		 * Update run-time statistics of the 'current'.
6342 		 */
6343 		update_curr(cfs_rq);
6344 		/*
6345 		 * Tell update_rq_clock() that we've just updated,
6346 		 * so we don't do microscopic update in schedule()
6347 		 * and double the fastpath cost.
6348 		 */
6349 		rq_clock_skip_update(rq, true);
6350 	}
6351 
6352 	set_skip_buddy(se);
6353 }
6354 
6355 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6356 {
6357 	struct sched_entity *se = &p->se;
6358 
6359 	/* throttled hierarchies are not runnable */
6360 	if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6361 		return false;
6362 
6363 	/* Tell the scheduler that we'd really like pse to run next. */
6364 	set_next_buddy(se);
6365 
6366 	yield_task_fair(rq);
6367 
6368 	return true;
6369 }
6370 
6371 #ifdef CONFIG_SMP
6372 /**************************************************
6373  * Fair scheduling class load-balancing methods.
6374  *
6375  * BASICS
6376  *
6377  * The purpose of load-balancing is to achieve the same basic fairness the
6378  * per-cpu scheduler provides, namely provide a proportional amount of compute
6379  * time to each task. This is expressed in the following equation:
6380  *
6381  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
6382  *
6383  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6384  * W_i,0 is defined as:
6385  *
6386  *   W_i,0 = \Sum_j w_i,j                                             (2)
6387  *
6388  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6389  * is derived from the nice value as per sched_prio_to_weight[].
6390  *
6391  * The weight average is an exponential decay average of the instantaneous
6392  * weight:
6393  *
6394  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
6395  *
6396  * C_i is the compute capacity of cpu i, typically it is the
6397  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6398  * can also include other factors [XXX].
6399  *
6400  * To achieve this balance we define a measure of imbalance which follows
6401  * directly from (1):
6402  *
6403  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
6404  *
6405  * We them move tasks around to minimize the imbalance. In the continuous
6406  * function space it is obvious this converges, in the discrete case we get
6407  * a few fun cases generally called infeasible weight scenarios.
6408  *
6409  * [XXX expand on:
6410  *     - infeasible weights;
6411  *     - local vs global optima in the discrete case. ]
6412  *
6413  *
6414  * SCHED DOMAINS
6415  *
6416  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6417  * for all i,j solution, we create a tree of cpus that follows the hardware
6418  * topology where each level pairs two lower groups (or better). This results
6419  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6420  * tree to only the first of the previous level and we decrease the frequency
6421  * of load-balance at each level inv. proportional to the number of cpus in
6422  * the groups.
6423  *
6424  * This yields:
6425  *
6426  *     log_2 n     1     n
6427  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
6428  *     i = 0      2^i   2^i
6429  *                               `- size of each group
6430  *         |         |     `- number of cpus doing load-balance
6431  *         |         `- freq
6432  *         `- sum over all levels
6433  *
6434  * Coupled with a limit on how many tasks we can migrate every balance pass,
6435  * this makes (5) the runtime complexity of the balancer.
6436  *
6437  * An important property here is that each CPU is still (indirectly) connected
6438  * to every other cpu in at most O(log n) steps:
6439  *
6440  * The adjacency matrix of the resulting graph is given by:
6441  *
6442  *             log_2 n
6443  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
6444  *             k = 0
6445  *
6446  * And you'll find that:
6447  *
6448  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
6449  *
6450  * Showing there's indeed a path between every cpu in at most O(log n) steps.
6451  * The task movement gives a factor of O(m), giving a convergence complexity
6452  * of:
6453  *
6454  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
6455  *
6456  *
6457  * WORK CONSERVING
6458  *
6459  * In order to avoid CPUs going idle while there's still work to do, new idle
6460  * balancing is more aggressive and has the newly idle cpu iterate up the domain
6461  * tree itself instead of relying on other CPUs to bring it work.
6462  *
6463  * This adds some complexity to both (5) and (8) but it reduces the total idle
6464  * time.
6465  *
6466  * [XXX more?]
6467  *
6468  *
6469  * CGROUPS
6470  *
6471  * Cgroups make a horror show out of (2), instead of a simple sum we get:
6472  *
6473  *                                s_k,i
6474  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
6475  *                                 S_k
6476  *
6477  * Where
6478  *
6479  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
6480  *
6481  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6482  *
6483  * The big problem is S_k, its a global sum needed to compute a local (W_i)
6484  * property.
6485  *
6486  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6487  *      rewrite all of this once again.]
6488  */
6489 
6490 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6491 
6492 enum fbq_type { regular, remote, all };
6493 
6494 #define LBF_ALL_PINNED	0x01
6495 #define LBF_NEED_BREAK	0x02
6496 #define LBF_DST_PINNED  0x04
6497 #define LBF_SOME_PINNED	0x08
6498 
6499 struct lb_env {
6500 	struct sched_domain	*sd;
6501 
6502 	struct rq		*src_rq;
6503 	int			src_cpu;
6504 
6505 	int			dst_cpu;
6506 	struct rq		*dst_rq;
6507 
6508 	struct cpumask		*dst_grpmask;
6509 	int			new_dst_cpu;
6510 	enum cpu_idle_type	idle;
6511 	long			imbalance;
6512 	/* The set of CPUs under consideration for load-balancing */
6513 	struct cpumask		*cpus;
6514 
6515 	unsigned int		flags;
6516 
6517 	unsigned int		loop;
6518 	unsigned int		loop_break;
6519 	unsigned int		loop_max;
6520 
6521 	enum fbq_type		fbq_type;
6522 	struct list_head	tasks;
6523 };
6524 
6525 /*
6526  * Is this task likely cache-hot:
6527  */
6528 static int task_hot(struct task_struct *p, struct lb_env *env)
6529 {
6530 	s64 delta;
6531 
6532 	lockdep_assert_held(&env->src_rq->lock);
6533 
6534 	if (p->sched_class != &fair_sched_class)
6535 		return 0;
6536 
6537 	if (unlikely(p->policy == SCHED_IDLE))
6538 		return 0;
6539 
6540 	/*
6541 	 * Buddy candidates are cache hot:
6542 	 */
6543 	if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6544 			(&p->se == cfs_rq_of(&p->se)->next ||
6545 			 &p->se == cfs_rq_of(&p->se)->last))
6546 		return 1;
6547 
6548 	if (sysctl_sched_migration_cost == -1)
6549 		return 1;
6550 	if (sysctl_sched_migration_cost == 0)
6551 		return 0;
6552 
6553 	delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6554 
6555 	return delta < (s64)sysctl_sched_migration_cost;
6556 }
6557 
6558 #ifdef CONFIG_NUMA_BALANCING
6559 /*
6560  * Returns 1, if task migration degrades locality
6561  * Returns 0, if task migration improves locality i.e migration preferred.
6562  * Returns -1, if task migration is not affected by locality.
6563  */
6564 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6565 {
6566 	struct numa_group *numa_group = rcu_dereference(p->numa_group);
6567 	unsigned long src_faults, dst_faults;
6568 	int src_nid, dst_nid;
6569 
6570 	if (!static_branch_likely(&sched_numa_balancing))
6571 		return -1;
6572 
6573 	if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6574 		return -1;
6575 
6576 	src_nid = cpu_to_node(env->src_cpu);
6577 	dst_nid = cpu_to_node(env->dst_cpu);
6578 
6579 	if (src_nid == dst_nid)
6580 		return -1;
6581 
6582 	/* Migrating away from the preferred node is always bad. */
6583 	if (src_nid == p->numa_preferred_nid) {
6584 		if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6585 			return 1;
6586 		else
6587 			return -1;
6588 	}
6589 
6590 	/* Encourage migration to the preferred node. */
6591 	if (dst_nid == p->numa_preferred_nid)
6592 		return 0;
6593 
6594 	/* Leaving a core idle is often worse than degrading locality. */
6595 	if (env->idle != CPU_NOT_IDLE)
6596 		return -1;
6597 
6598 	if (numa_group) {
6599 		src_faults = group_faults(p, src_nid);
6600 		dst_faults = group_faults(p, dst_nid);
6601 	} else {
6602 		src_faults = task_faults(p, src_nid);
6603 		dst_faults = task_faults(p, dst_nid);
6604 	}
6605 
6606 	return dst_faults < src_faults;
6607 }
6608 
6609 #else
6610 static inline int migrate_degrades_locality(struct task_struct *p,
6611 					     struct lb_env *env)
6612 {
6613 	return -1;
6614 }
6615 #endif
6616 
6617 /*
6618  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6619  */
6620 static
6621 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6622 {
6623 	int tsk_cache_hot;
6624 
6625 	lockdep_assert_held(&env->src_rq->lock);
6626 
6627 	/*
6628 	 * We do not migrate tasks that are:
6629 	 * 1) throttled_lb_pair, or
6630 	 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6631 	 * 3) running (obviously), or
6632 	 * 4) are cache-hot on their current CPU.
6633 	 */
6634 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6635 		return 0;
6636 
6637 	if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
6638 		int cpu;
6639 
6640 		schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6641 
6642 		env->flags |= LBF_SOME_PINNED;
6643 
6644 		/*
6645 		 * Remember if this task can be migrated to any other cpu in
6646 		 * our sched_group. We may want to revisit it if we couldn't
6647 		 * meet load balance goals by pulling other tasks on src_cpu.
6648 		 *
6649 		 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
6650 		 * already computed one in current iteration.
6651 		 */
6652 		if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
6653 			return 0;
6654 
6655 		/* Prevent to re-select dst_cpu via env's cpus */
6656 		for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6657 			if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
6658 				env->flags |= LBF_DST_PINNED;
6659 				env->new_dst_cpu = cpu;
6660 				break;
6661 			}
6662 		}
6663 
6664 		return 0;
6665 	}
6666 
6667 	/* Record that we found atleast one task that could run on dst_cpu */
6668 	env->flags &= ~LBF_ALL_PINNED;
6669 
6670 	if (task_running(env->src_rq, p)) {
6671 		schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6672 		return 0;
6673 	}
6674 
6675 	/*
6676 	 * Aggressive migration if:
6677 	 * 1) destination numa is preferred
6678 	 * 2) task is cache cold, or
6679 	 * 3) too many balance attempts have failed.
6680 	 */
6681 	tsk_cache_hot = migrate_degrades_locality(p, env);
6682 	if (tsk_cache_hot == -1)
6683 		tsk_cache_hot = task_hot(p, env);
6684 
6685 	if (tsk_cache_hot <= 0 ||
6686 	    env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6687 		if (tsk_cache_hot == 1) {
6688 			schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6689 			schedstat_inc(p->se.statistics.nr_forced_migrations);
6690 		}
6691 		return 1;
6692 	}
6693 
6694 	schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6695 	return 0;
6696 }
6697 
6698 /*
6699  * detach_task() -- detach the task for the migration specified in env
6700  */
6701 static void detach_task(struct task_struct *p, struct lb_env *env)
6702 {
6703 	lockdep_assert_held(&env->src_rq->lock);
6704 
6705 	p->on_rq = TASK_ON_RQ_MIGRATING;
6706 	deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
6707 	set_task_cpu(p, env->dst_cpu);
6708 }
6709 
6710 /*
6711  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6712  * part of active balancing operations within "domain".
6713  *
6714  * Returns a task if successful and NULL otherwise.
6715  */
6716 static struct task_struct *detach_one_task(struct lb_env *env)
6717 {
6718 	struct task_struct *p, *n;
6719 
6720 	lockdep_assert_held(&env->src_rq->lock);
6721 
6722 	list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6723 		if (!can_migrate_task(p, env))
6724 			continue;
6725 
6726 		detach_task(p, env);
6727 
6728 		/*
6729 		 * Right now, this is only the second place where
6730 		 * lb_gained[env->idle] is updated (other is detach_tasks)
6731 		 * so we can safely collect stats here rather than
6732 		 * inside detach_tasks().
6733 		 */
6734 		schedstat_inc(env->sd->lb_gained[env->idle]);
6735 		return p;
6736 	}
6737 	return NULL;
6738 }
6739 
6740 static const unsigned int sched_nr_migrate_break = 32;
6741 
6742 /*
6743  * detach_tasks() -- tries to detach up to imbalance weighted load from
6744  * busiest_rq, as part of a balancing operation within domain "sd".
6745  *
6746  * Returns number of detached tasks if successful and 0 otherwise.
6747  */
6748 static int detach_tasks(struct lb_env *env)
6749 {
6750 	struct list_head *tasks = &env->src_rq->cfs_tasks;
6751 	struct task_struct *p;
6752 	unsigned long load;
6753 	int detached = 0;
6754 
6755 	lockdep_assert_held(&env->src_rq->lock);
6756 
6757 	if (env->imbalance <= 0)
6758 		return 0;
6759 
6760 	while (!list_empty(tasks)) {
6761 		/*
6762 		 * We don't want to steal all, otherwise we may be treated likewise,
6763 		 * which could at worst lead to a livelock crash.
6764 		 */
6765 		if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6766 			break;
6767 
6768 		p = list_first_entry(tasks, struct task_struct, se.group_node);
6769 
6770 		env->loop++;
6771 		/* We've more or less seen every task there is, call it quits */
6772 		if (env->loop > env->loop_max)
6773 			break;
6774 
6775 		/* take a breather every nr_migrate tasks */
6776 		if (env->loop > env->loop_break) {
6777 			env->loop_break += sched_nr_migrate_break;
6778 			env->flags |= LBF_NEED_BREAK;
6779 			break;
6780 		}
6781 
6782 		if (!can_migrate_task(p, env))
6783 			goto next;
6784 
6785 		load = task_h_load(p);
6786 
6787 		if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6788 			goto next;
6789 
6790 		if ((load / 2) > env->imbalance)
6791 			goto next;
6792 
6793 		detach_task(p, env);
6794 		list_add(&p->se.group_node, &env->tasks);
6795 
6796 		detached++;
6797 		env->imbalance -= load;
6798 
6799 #ifdef CONFIG_PREEMPT
6800 		/*
6801 		 * NEWIDLE balancing is a source of latency, so preemptible
6802 		 * kernels will stop after the first task is detached to minimize
6803 		 * the critical section.
6804 		 */
6805 		if (env->idle == CPU_NEWLY_IDLE)
6806 			break;
6807 #endif
6808 
6809 		/*
6810 		 * We only want to steal up to the prescribed amount of
6811 		 * weighted load.
6812 		 */
6813 		if (env->imbalance <= 0)
6814 			break;
6815 
6816 		continue;
6817 next:
6818 		list_move_tail(&p->se.group_node, tasks);
6819 	}
6820 
6821 	/*
6822 	 * Right now, this is one of only two places we collect this stat
6823 	 * so we can safely collect detach_one_task() stats here rather
6824 	 * than inside detach_one_task().
6825 	 */
6826 	schedstat_add(env->sd->lb_gained[env->idle], detached);
6827 
6828 	return detached;
6829 }
6830 
6831 /*
6832  * attach_task() -- attach the task detached by detach_task() to its new rq.
6833  */
6834 static void attach_task(struct rq *rq, struct task_struct *p)
6835 {
6836 	lockdep_assert_held(&rq->lock);
6837 
6838 	BUG_ON(task_rq(p) != rq);
6839 	activate_task(rq, p, ENQUEUE_NOCLOCK);
6840 	p->on_rq = TASK_ON_RQ_QUEUED;
6841 	check_preempt_curr(rq, p, 0);
6842 }
6843 
6844 /*
6845  * attach_one_task() -- attaches the task returned from detach_one_task() to
6846  * its new rq.
6847  */
6848 static void attach_one_task(struct rq *rq, struct task_struct *p)
6849 {
6850 	struct rq_flags rf;
6851 
6852 	rq_lock(rq, &rf);
6853 	update_rq_clock(rq);
6854 	attach_task(rq, p);
6855 	rq_unlock(rq, &rf);
6856 }
6857 
6858 /*
6859  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6860  * new rq.
6861  */
6862 static void attach_tasks(struct lb_env *env)
6863 {
6864 	struct list_head *tasks = &env->tasks;
6865 	struct task_struct *p;
6866 	struct rq_flags rf;
6867 
6868 	rq_lock(env->dst_rq, &rf);
6869 	update_rq_clock(env->dst_rq);
6870 
6871 	while (!list_empty(tasks)) {
6872 		p = list_first_entry(tasks, struct task_struct, se.group_node);
6873 		list_del_init(&p->se.group_node);
6874 
6875 		attach_task(env->dst_rq, p);
6876 	}
6877 
6878 	rq_unlock(env->dst_rq, &rf);
6879 }
6880 
6881 #ifdef CONFIG_FAIR_GROUP_SCHED
6882 
6883 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
6884 {
6885 	if (cfs_rq->load.weight)
6886 		return false;
6887 
6888 	if (cfs_rq->avg.load_sum)
6889 		return false;
6890 
6891 	if (cfs_rq->avg.util_sum)
6892 		return false;
6893 
6894 	if (cfs_rq->runnable_load_sum)
6895 		return false;
6896 
6897 	return true;
6898 }
6899 
6900 static void update_blocked_averages(int cpu)
6901 {
6902 	struct rq *rq = cpu_rq(cpu);
6903 	struct cfs_rq *cfs_rq, *pos;
6904 	struct rq_flags rf;
6905 
6906 	rq_lock_irqsave(rq, &rf);
6907 	update_rq_clock(rq);
6908 
6909 	/*
6910 	 * Iterates the task_group tree in a bottom up fashion, see
6911 	 * list_add_leaf_cfs_rq() for details.
6912 	 */
6913 	for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
6914 		struct sched_entity *se;
6915 
6916 		/* throttled entities do not contribute to load */
6917 		if (throttled_hierarchy(cfs_rq))
6918 			continue;
6919 
6920 		if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6921 			update_tg_load_avg(cfs_rq, 0);
6922 
6923 		/* Propagate pending load changes to the parent, if any: */
6924 		se = cfs_rq->tg->se[cpu];
6925 		if (se && !skip_blocked_update(se))
6926 			update_load_avg(se, 0);
6927 
6928 		/*
6929 		 * There can be a lot of idle CPU cgroups.  Don't let fully
6930 		 * decayed cfs_rqs linger on the list.
6931 		 */
6932 		if (cfs_rq_is_decayed(cfs_rq))
6933 			list_del_leaf_cfs_rq(cfs_rq);
6934 	}
6935 	rq_unlock_irqrestore(rq, &rf);
6936 }
6937 
6938 /*
6939  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6940  * This needs to be done in a top-down fashion because the load of a child
6941  * group is a fraction of its parents load.
6942  */
6943 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6944 {
6945 	struct rq *rq = rq_of(cfs_rq);
6946 	struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6947 	unsigned long now = jiffies;
6948 	unsigned long load;
6949 
6950 	if (cfs_rq->last_h_load_update == now)
6951 		return;
6952 
6953 	cfs_rq->h_load_next = NULL;
6954 	for_each_sched_entity(se) {
6955 		cfs_rq = cfs_rq_of(se);
6956 		cfs_rq->h_load_next = se;
6957 		if (cfs_rq->last_h_load_update == now)
6958 			break;
6959 	}
6960 
6961 	if (!se) {
6962 		cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6963 		cfs_rq->last_h_load_update = now;
6964 	}
6965 
6966 	while ((se = cfs_rq->h_load_next) != NULL) {
6967 		load = cfs_rq->h_load;
6968 		load = div64_ul(load * se->avg.load_avg,
6969 			cfs_rq_load_avg(cfs_rq) + 1);
6970 		cfs_rq = group_cfs_rq(se);
6971 		cfs_rq->h_load = load;
6972 		cfs_rq->last_h_load_update = now;
6973 	}
6974 }
6975 
6976 static unsigned long task_h_load(struct task_struct *p)
6977 {
6978 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
6979 
6980 	update_cfs_rq_h_load(cfs_rq);
6981 	return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6982 			cfs_rq_load_avg(cfs_rq) + 1);
6983 }
6984 #else
6985 static inline void update_blocked_averages(int cpu)
6986 {
6987 	struct rq *rq = cpu_rq(cpu);
6988 	struct cfs_rq *cfs_rq = &rq->cfs;
6989 	struct rq_flags rf;
6990 
6991 	rq_lock_irqsave(rq, &rf);
6992 	update_rq_clock(rq);
6993 	update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6994 	rq_unlock_irqrestore(rq, &rf);
6995 }
6996 
6997 static unsigned long task_h_load(struct task_struct *p)
6998 {
6999 	return p->se.avg.load_avg;
7000 }
7001 #endif
7002 
7003 /********** Helpers for find_busiest_group ************************/
7004 
7005 enum group_type {
7006 	group_other = 0,
7007 	group_imbalanced,
7008 	group_overloaded,
7009 };
7010 
7011 /*
7012  * sg_lb_stats - stats of a sched_group required for load_balancing
7013  */
7014 struct sg_lb_stats {
7015 	unsigned long avg_load; /*Avg load across the CPUs of the group */
7016 	unsigned long group_load; /* Total load over the CPUs of the group */
7017 	unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7018 	unsigned long load_per_task;
7019 	unsigned long group_capacity;
7020 	unsigned long group_util; /* Total utilization of the group */
7021 	unsigned int sum_nr_running; /* Nr tasks running in the group */
7022 	unsigned int idle_cpus;
7023 	unsigned int group_weight;
7024 	enum group_type group_type;
7025 	int group_no_capacity;
7026 #ifdef CONFIG_NUMA_BALANCING
7027 	unsigned int nr_numa_running;
7028 	unsigned int nr_preferred_running;
7029 #endif
7030 };
7031 
7032 /*
7033  * sd_lb_stats - Structure to store the statistics of a sched_domain
7034  *		 during load balancing.
7035  */
7036 struct sd_lb_stats {
7037 	struct sched_group *busiest;	/* Busiest group in this sd */
7038 	struct sched_group *local;	/* Local group in this sd */
7039 	unsigned long total_load;	/* Total load of all groups in sd */
7040 	unsigned long total_capacity;	/* Total capacity of all groups in sd */
7041 	unsigned long avg_load;	/* Average load across all groups in sd */
7042 
7043 	struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7044 	struct sg_lb_stats local_stat;	/* Statistics of the local group */
7045 };
7046 
7047 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7048 {
7049 	/*
7050 	 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7051 	 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7052 	 * We must however clear busiest_stat::avg_load because
7053 	 * update_sd_pick_busiest() reads this before assignment.
7054 	 */
7055 	*sds = (struct sd_lb_stats){
7056 		.busiest = NULL,
7057 		.local = NULL,
7058 		.total_load = 0UL,
7059 		.total_capacity = 0UL,
7060 		.busiest_stat = {
7061 			.avg_load = 0UL,
7062 			.sum_nr_running = 0,
7063 			.group_type = group_other,
7064 		},
7065 	};
7066 }
7067 
7068 /**
7069  * get_sd_load_idx - Obtain the load index for a given sched domain.
7070  * @sd: The sched_domain whose load_idx is to be obtained.
7071  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7072  *
7073  * Return: The load index.
7074  */
7075 static inline int get_sd_load_idx(struct sched_domain *sd,
7076 					enum cpu_idle_type idle)
7077 {
7078 	int load_idx;
7079 
7080 	switch (idle) {
7081 	case CPU_NOT_IDLE:
7082 		load_idx = sd->busy_idx;
7083 		break;
7084 
7085 	case CPU_NEWLY_IDLE:
7086 		load_idx = sd->newidle_idx;
7087 		break;
7088 	default:
7089 		load_idx = sd->idle_idx;
7090 		break;
7091 	}
7092 
7093 	return load_idx;
7094 }
7095 
7096 static unsigned long scale_rt_capacity(int cpu)
7097 {
7098 	struct rq *rq = cpu_rq(cpu);
7099 	u64 total, used, age_stamp, avg;
7100 	s64 delta;
7101 
7102 	/*
7103 	 * Since we're reading these variables without serialization make sure
7104 	 * we read them once before doing sanity checks on them.
7105 	 */
7106 	age_stamp = READ_ONCE(rq->age_stamp);
7107 	avg = READ_ONCE(rq->rt_avg);
7108 	delta = __rq_clock_broken(rq) - age_stamp;
7109 
7110 	if (unlikely(delta < 0))
7111 		delta = 0;
7112 
7113 	total = sched_avg_period() + delta;
7114 
7115 	used = div_u64(avg, total);
7116 
7117 	if (likely(used < SCHED_CAPACITY_SCALE))
7118 		return SCHED_CAPACITY_SCALE - used;
7119 
7120 	return 1;
7121 }
7122 
7123 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7124 {
7125 	unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7126 	struct sched_group *sdg = sd->groups;
7127 
7128 	cpu_rq(cpu)->cpu_capacity_orig = capacity;
7129 
7130 	capacity *= scale_rt_capacity(cpu);
7131 	capacity >>= SCHED_CAPACITY_SHIFT;
7132 
7133 	if (!capacity)
7134 		capacity = 1;
7135 
7136 	cpu_rq(cpu)->cpu_capacity = capacity;
7137 	sdg->sgc->capacity = capacity;
7138 	sdg->sgc->min_capacity = capacity;
7139 }
7140 
7141 void update_group_capacity(struct sched_domain *sd, int cpu)
7142 {
7143 	struct sched_domain *child = sd->child;
7144 	struct sched_group *group, *sdg = sd->groups;
7145 	unsigned long capacity, min_capacity;
7146 	unsigned long interval;
7147 
7148 	interval = msecs_to_jiffies(sd->balance_interval);
7149 	interval = clamp(interval, 1UL, max_load_balance_interval);
7150 	sdg->sgc->next_update = jiffies + interval;
7151 
7152 	if (!child) {
7153 		update_cpu_capacity(sd, cpu);
7154 		return;
7155 	}
7156 
7157 	capacity = 0;
7158 	min_capacity = ULONG_MAX;
7159 
7160 	if (child->flags & SD_OVERLAP) {
7161 		/*
7162 		 * SD_OVERLAP domains cannot assume that child groups
7163 		 * span the current group.
7164 		 */
7165 
7166 		for_each_cpu(cpu, sched_group_span(sdg)) {
7167 			struct sched_group_capacity *sgc;
7168 			struct rq *rq = cpu_rq(cpu);
7169 
7170 			/*
7171 			 * build_sched_domains() -> init_sched_groups_capacity()
7172 			 * gets here before we've attached the domains to the
7173 			 * runqueues.
7174 			 *
7175 			 * Use capacity_of(), which is set irrespective of domains
7176 			 * in update_cpu_capacity().
7177 			 *
7178 			 * This avoids capacity from being 0 and
7179 			 * causing divide-by-zero issues on boot.
7180 			 */
7181 			if (unlikely(!rq->sd)) {
7182 				capacity += capacity_of(cpu);
7183 			} else {
7184 				sgc = rq->sd->groups->sgc;
7185 				capacity += sgc->capacity;
7186 			}
7187 
7188 			min_capacity = min(capacity, min_capacity);
7189 		}
7190 	} else  {
7191 		/*
7192 		 * !SD_OVERLAP domains can assume that child groups
7193 		 * span the current group.
7194 		 */
7195 
7196 		group = child->groups;
7197 		do {
7198 			struct sched_group_capacity *sgc = group->sgc;
7199 
7200 			capacity += sgc->capacity;
7201 			min_capacity = min(sgc->min_capacity, min_capacity);
7202 			group = group->next;
7203 		} while (group != child->groups);
7204 	}
7205 
7206 	sdg->sgc->capacity = capacity;
7207 	sdg->sgc->min_capacity = min_capacity;
7208 }
7209 
7210 /*
7211  * Check whether the capacity of the rq has been noticeably reduced by side
7212  * activity. The imbalance_pct is used for the threshold.
7213  * Return true is the capacity is reduced
7214  */
7215 static inline int
7216 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7217 {
7218 	return ((rq->cpu_capacity * sd->imbalance_pct) <
7219 				(rq->cpu_capacity_orig * 100));
7220 }
7221 
7222 /*
7223  * Group imbalance indicates (and tries to solve) the problem where balancing
7224  * groups is inadequate due to ->cpus_allowed constraints.
7225  *
7226  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7227  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7228  * Something like:
7229  *
7230  *	{ 0 1 2 3 } { 4 5 6 7 }
7231  *	        *     * * *
7232  *
7233  * If we were to balance group-wise we'd place two tasks in the first group and
7234  * two tasks in the second group. Clearly this is undesired as it will overload
7235  * cpu 3 and leave one of the cpus in the second group unused.
7236  *
7237  * The current solution to this issue is detecting the skew in the first group
7238  * by noticing the lower domain failed to reach balance and had difficulty
7239  * moving tasks due to affinity constraints.
7240  *
7241  * When this is so detected; this group becomes a candidate for busiest; see
7242  * update_sd_pick_busiest(). And calculate_imbalance() and
7243  * find_busiest_group() avoid some of the usual balance conditions to allow it
7244  * to create an effective group imbalance.
7245  *
7246  * This is a somewhat tricky proposition since the next run might not find the
7247  * group imbalance and decide the groups need to be balanced again. A most
7248  * subtle and fragile situation.
7249  */
7250 
7251 static inline int sg_imbalanced(struct sched_group *group)
7252 {
7253 	return group->sgc->imbalance;
7254 }
7255 
7256 /*
7257  * group_has_capacity returns true if the group has spare capacity that could
7258  * be used by some tasks.
7259  * We consider that a group has spare capacity if the  * number of task is
7260  * smaller than the number of CPUs or if the utilization is lower than the
7261  * available capacity for CFS tasks.
7262  * For the latter, we use a threshold to stabilize the state, to take into
7263  * account the variance of the tasks' load and to return true if the available
7264  * capacity in meaningful for the load balancer.
7265  * As an example, an available capacity of 1% can appear but it doesn't make
7266  * any benefit for the load balance.
7267  */
7268 static inline bool
7269 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7270 {
7271 	if (sgs->sum_nr_running < sgs->group_weight)
7272 		return true;
7273 
7274 	if ((sgs->group_capacity * 100) >
7275 			(sgs->group_util * env->sd->imbalance_pct))
7276 		return true;
7277 
7278 	return false;
7279 }
7280 
7281 /*
7282  *  group_is_overloaded returns true if the group has more tasks than it can
7283  *  handle.
7284  *  group_is_overloaded is not equals to !group_has_capacity because a group
7285  *  with the exact right number of tasks, has no more spare capacity but is not
7286  *  overloaded so both group_has_capacity and group_is_overloaded return
7287  *  false.
7288  */
7289 static inline bool
7290 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7291 {
7292 	if (sgs->sum_nr_running <= sgs->group_weight)
7293 		return false;
7294 
7295 	if ((sgs->group_capacity * 100) <
7296 			(sgs->group_util * env->sd->imbalance_pct))
7297 		return true;
7298 
7299 	return false;
7300 }
7301 
7302 /*
7303  * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7304  * per-CPU capacity than sched_group ref.
7305  */
7306 static inline bool
7307 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7308 {
7309 	return sg->sgc->min_capacity * capacity_margin <
7310 						ref->sgc->min_capacity * 1024;
7311 }
7312 
7313 static inline enum
7314 group_type group_classify(struct sched_group *group,
7315 			  struct sg_lb_stats *sgs)
7316 {
7317 	if (sgs->group_no_capacity)
7318 		return group_overloaded;
7319 
7320 	if (sg_imbalanced(group))
7321 		return group_imbalanced;
7322 
7323 	return group_other;
7324 }
7325 
7326 /**
7327  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7328  * @env: The load balancing environment.
7329  * @group: sched_group whose statistics are to be updated.
7330  * @load_idx: Load index of sched_domain of this_cpu for load calc.
7331  * @local_group: Does group contain this_cpu.
7332  * @sgs: variable to hold the statistics for this group.
7333  * @overload: Indicate more than one runnable task for any CPU.
7334  */
7335 static inline void update_sg_lb_stats(struct lb_env *env,
7336 			struct sched_group *group, int load_idx,
7337 			int local_group, struct sg_lb_stats *sgs,
7338 			bool *overload)
7339 {
7340 	unsigned long load;
7341 	int i, nr_running;
7342 
7343 	memset(sgs, 0, sizeof(*sgs));
7344 
7345 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7346 		struct rq *rq = cpu_rq(i);
7347 
7348 		/* Bias balancing toward cpus of our domain */
7349 		if (local_group)
7350 			load = target_load(i, load_idx);
7351 		else
7352 			load = source_load(i, load_idx);
7353 
7354 		sgs->group_load += load;
7355 		sgs->group_util += cpu_util(i);
7356 		sgs->sum_nr_running += rq->cfs.h_nr_running;
7357 
7358 		nr_running = rq->nr_running;
7359 		if (nr_running > 1)
7360 			*overload = true;
7361 
7362 #ifdef CONFIG_NUMA_BALANCING
7363 		sgs->nr_numa_running += rq->nr_numa_running;
7364 		sgs->nr_preferred_running += rq->nr_preferred_running;
7365 #endif
7366 		sgs->sum_weighted_load += weighted_cpuload(i);
7367 		/*
7368 		 * No need to call idle_cpu() if nr_running is not 0
7369 		 */
7370 		if (!nr_running && idle_cpu(i))
7371 			sgs->idle_cpus++;
7372 	}
7373 
7374 	/* Adjust by relative CPU capacity of the group */
7375 	sgs->group_capacity = group->sgc->capacity;
7376 	sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7377 
7378 	if (sgs->sum_nr_running)
7379 		sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7380 
7381 	sgs->group_weight = group->group_weight;
7382 
7383 	sgs->group_no_capacity = group_is_overloaded(env, sgs);
7384 	sgs->group_type = group_classify(group, sgs);
7385 }
7386 
7387 /**
7388  * update_sd_pick_busiest - return 1 on busiest group
7389  * @env: The load balancing environment.
7390  * @sds: sched_domain statistics
7391  * @sg: sched_group candidate to be checked for being the busiest
7392  * @sgs: sched_group statistics
7393  *
7394  * Determine if @sg is a busier group than the previously selected
7395  * busiest group.
7396  *
7397  * Return: %true if @sg is a busier group than the previously selected
7398  * busiest group. %false otherwise.
7399  */
7400 static bool update_sd_pick_busiest(struct lb_env *env,
7401 				   struct sd_lb_stats *sds,
7402 				   struct sched_group *sg,
7403 				   struct sg_lb_stats *sgs)
7404 {
7405 	struct sg_lb_stats *busiest = &sds->busiest_stat;
7406 
7407 	if (sgs->group_type > busiest->group_type)
7408 		return true;
7409 
7410 	if (sgs->group_type < busiest->group_type)
7411 		return false;
7412 
7413 	if (sgs->avg_load <= busiest->avg_load)
7414 		return false;
7415 
7416 	if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7417 		goto asym_packing;
7418 
7419 	/*
7420 	 * Candidate sg has no more than one task per CPU and
7421 	 * has higher per-CPU capacity. Migrating tasks to less
7422 	 * capable CPUs may harm throughput. Maximize throughput,
7423 	 * power/energy consequences are not considered.
7424 	 */
7425 	if (sgs->sum_nr_running <= sgs->group_weight &&
7426 	    group_smaller_cpu_capacity(sds->local, sg))
7427 		return false;
7428 
7429 asym_packing:
7430 	/* This is the busiest node in its class. */
7431 	if (!(env->sd->flags & SD_ASYM_PACKING))
7432 		return true;
7433 
7434 	/* No ASYM_PACKING if target cpu is already busy */
7435 	if (env->idle == CPU_NOT_IDLE)
7436 		return true;
7437 	/*
7438 	 * ASYM_PACKING needs to move all the work to the highest
7439 	 * prority CPUs in the group, therefore mark all groups
7440 	 * of lower priority than ourself as busy.
7441 	 */
7442 	if (sgs->sum_nr_running &&
7443 	    sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7444 		if (!sds->busiest)
7445 			return true;
7446 
7447 		/* Prefer to move from lowest priority cpu's work */
7448 		if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7449 				      sg->asym_prefer_cpu))
7450 			return true;
7451 	}
7452 
7453 	return false;
7454 }
7455 
7456 #ifdef CONFIG_NUMA_BALANCING
7457 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7458 {
7459 	if (sgs->sum_nr_running > sgs->nr_numa_running)
7460 		return regular;
7461 	if (sgs->sum_nr_running > sgs->nr_preferred_running)
7462 		return remote;
7463 	return all;
7464 }
7465 
7466 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7467 {
7468 	if (rq->nr_running > rq->nr_numa_running)
7469 		return regular;
7470 	if (rq->nr_running > rq->nr_preferred_running)
7471 		return remote;
7472 	return all;
7473 }
7474 #else
7475 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7476 {
7477 	return all;
7478 }
7479 
7480 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7481 {
7482 	return regular;
7483 }
7484 #endif /* CONFIG_NUMA_BALANCING */
7485 
7486 /**
7487  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7488  * @env: The load balancing environment.
7489  * @sds: variable to hold the statistics for this sched_domain.
7490  */
7491 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7492 {
7493 	struct sched_domain *child = env->sd->child;
7494 	struct sched_group *sg = env->sd->groups;
7495 	struct sg_lb_stats *local = &sds->local_stat;
7496 	struct sg_lb_stats tmp_sgs;
7497 	int load_idx, prefer_sibling = 0;
7498 	bool overload = false;
7499 
7500 	if (child && child->flags & SD_PREFER_SIBLING)
7501 		prefer_sibling = 1;
7502 
7503 	load_idx = get_sd_load_idx(env->sd, env->idle);
7504 
7505 	do {
7506 		struct sg_lb_stats *sgs = &tmp_sgs;
7507 		int local_group;
7508 
7509 		local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7510 		if (local_group) {
7511 			sds->local = sg;
7512 			sgs = local;
7513 
7514 			if (env->idle != CPU_NEWLY_IDLE ||
7515 			    time_after_eq(jiffies, sg->sgc->next_update))
7516 				update_group_capacity(env->sd, env->dst_cpu);
7517 		}
7518 
7519 		update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7520 						&overload);
7521 
7522 		if (local_group)
7523 			goto next_group;
7524 
7525 		/*
7526 		 * In case the child domain prefers tasks go to siblings
7527 		 * first, lower the sg capacity so that we'll try
7528 		 * and move all the excess tasks away. We lower the capacity
7529 		 * of a group only if the local group has the capacity to fit
7530 		 * these excess tasks. The extra check prevents the case where
7531 		 * you always pull from the heaviest group when it is already
7532 		 * under-utilized (possible with a large weight task outweighs
7533 		 * the tasks on the system).
7534 		 */
7535 		if (prefer_sibling && sds->local &&
7536 		    group_has_capacity(env, local) &&
7537 		    (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7538 			sgs->group_no_capacity = 1;
7539 			sgs->group_type = group_classify(sg, sgs);
7540 		}
7541 
7542 		if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7543 			sds->busiest = sg;
7544 			sds->busiest_stat = *sgs;
7545 		}
7546 
7547 next_group:
7548 		/* Now, start updating sd_lb_stats */
7549 		sds->total_load += sgs->group_load;
7550 		sds->total_capacity += sgs->group_capacity;
7551 
7552 		sg = sg->next;
7553 	} while (sg != env->sd->groups);
7554 
7555 	if (env->sd->flags & SD_NUMA)
7556 		env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7557 
7558 	if (!env->sd->parent) {
7559 		/* update overload indicator if we are at root domain */
7560 		if (env->dst_rq->rd->overload != overload)
7561 			env->dst_rq->rd->overload = overload;
7562 	}
7563 
7564 }
7565 
7566 /**
7567  * check_asym_packing - Check to see if the group is packed into the
7568  *			sched domain.
7569  *
7570  * This is primarily intended to used at the sibling level.  Some
7571  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
7572  * case of POWER7, it can move to lower SMT modes only when higher
7573  * threads are idle.  When in lower SMT modes, the threads will
7574  * perform better since they share less core resources.  Hence when we
7575  * have idle threads, we want them to be the higher ones.
7576  *
7577  * This packing function is run on idle threads.  It checks to see if
7578  * the busiest CPU in this domain (core in the P7 case) has a higher
7579  * CPU number than the packing function is being run on.  Here we are
7580  * assuming lower CPU number will be equivalent to lower a SMT thread
7581  * number.
7582  *
7583  * Return: 1 when packing is required and a task should be moved to
7584  * this CPU.  The amount of the imbalance is returned in *imbalance.
7585  *
7586  * @env: The load balancing environment.
7587  * @sds: Statistics of the sched_domain which is to be packed
7588  */
7589 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7590 {
7591 	int busiest_cpu;
7592 
7593 	if (!(env->sd->flags & SD_ASYM_PACKING))
7594 		return 0;
7595 
7596 	if (env->idle == CPU_NOT_IDLE)
7597 		return 0;
7598 
7599 	if (!sds->busiest)
7600 		return 0;
7601 
7602 	busiest_cpu = sds->busiest->asym_prefer_cpu;
7603 	if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
7604 		return 0;
7605 
7606 	env->imbalance = DIV_ROUND_CLOSEST(
7607 		sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7608 		SCHED_CAPACITY_SCALE);
7609 
7610 	return 1;
7611 }
7612 
7613 /**
7614  * fix_small_imbalance - Calculate the minor imbalance that exists
7615  *			amongst the groups of a sched_domain, during
7616  *			load balancing.
7617  * @env: The load balancing environment.
7618  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7619  */
7620 static inline
7621 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7622 {
7623 	unsigned long tmp, capa_now = 0, capa_move = 0;
7624 	unsigned int imbn = 2;
7625 	unsigned long scaled_busy_load_per_task;
7626 	struct sg_lb_stats *local, *busiest;
7627 
7628 	local = &sds->local_stat;
7629 	busiest = &sds->busiest_stat;
7630 
7631 	if (!local->sum_nr_running)
7632 		local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7633 	else if (busiest->load_per_task > local->load_per_task)
7634 		imbn = 1;
7635 
7636 	scaled_busy_load_per_task =
7637 		(busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7638 		busiest->group_capacity;
7639 
7640 	if (busiest->avg_load + scaled_busy_load_per_task >=
7641 	    local->avg_load + (scaled_busy_load_per_task * imbn)) {
7642 		env->imbalance = busiest->load_per_task;
7643 		return;
7644 	}
7645 
7646 	/*
7647 	 * OK, we don't have enough imbalance to justify moving tasks,
7648 	 * however we may be able to increase total CPU capacity used by
7649 	 * moving them.
7650 	 */
7651 
7652 	capa_now += busiest->group_capacity *
7653 			min(busiest->load_per_task, busiest->avg_load);
7654 	capa_now += local->group_capacity *
7655 			min(local->load_per_task, local->avg_load);
7656 	capa_now /= SCHED_CAPACITY_SCALE;
7657 
7658 	/* Amount of load we'd subtract */
7659 	if (busiest->avg_load > scaled_busy_load_per_task) {
7660 		capa_move += busiest->group_capacity *
7661 			    min(busiest->load_per_task,
7662 				busiest->avg_load - scaled_busy_load_per_task);
7663 	}
7664 
7665 	/* Amount of load we'd add */
7666 	if (busiest->avg_load * busiest->group_capacity <
7667 	    busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7668 		tmp = (busiest->avg_load * busiest->group_capacity) /
7669 		      local->group_capacity;
7670 	} else {
7671 		tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7672 		      local->group_capacity;
7673 	}
7674 	capa_move += local->group_capacity *
7675 		    min(local->load_per_task, local->avg_load + tmp);
7676 	capa_move /= SCHED_CAPACITY_SCALE;
7677 
7678 	/* Move if we gain throughput */
7679 	if (capa_move > capa_now)
7680 		env->imbalance = busiest->load_per_task;
7681 }
7682 
7683 /**
7684  * calculate_imbalance - Calculate the amount of imbalance present within the
7685  *			 groups of a given sched_domain during load balance.
7686  * @env: load balance environment
7687  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7688  */
7689 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7690 {
7691 	unsigned long max_pull, load_above_capacity = ~0UL;
7692 	struct sg_lb_stats *local, *busiest;
7693 
7694 	local = &sds->local_stat;
7695 	busiest = &sds->busiest_stat;
7696 
7697 	if (busiest->group_type == group_imbalanced) {
7698 		/*
7699 		 * In the group_imb case we cannot rely on group-wide averages
7700 		 * to ensure cpu-load equilibrium, look at wider averages. XXX
7701 		 */
7702 		busiest->load_per_task =
7703 			min(busiest->load_per_task, sds->avg_load);
7704 	}
7705 
7706 	/*
7707 	 * Avg load of busiest sg can be less and avg load of local sg can
7708 	 * be greater than avg load across all sgs of sd because avg load
7709 	 * factors in sg capacity and sgs with smaller group_type are
7710 	 * skipped when updating the busiest sg:
7711 	 */
7712 	if (busiest->avg_load <= sds->avg_load ||
7713 	    local->avg_load >= sds->avg_load) {
7714 		env->imbalance = 0;
7715 		return fix_small_imbalance(env, sds);
7716 	}
7717 
7718 	/*
7719 	 * If there aren't any idle cpus, avoid creating some.
7720 	 */
7721 	if (busiest->group_type == group_overloaded &&
7722 	    local->group_type   == group_overloaded) {
7723 		load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7724 		if (load_above_capacity > busiest->group_capacity) {
7725 			load_above_capacity -= busiest->group_capacity;
7726 			load_above_capacity *= scale_load_down(NICE_0_LOAD);
7727 			load_above_capacity /= busiest->group_capacity;
7728 		} else
7729 			load_above_capacity = ~0UL;
7730 	}
7731 
7732 	/*
7733 	 * We're trying to get all the cpus to the average_load, so we don't
7734 	 * want to push ourselves above the average load, nor do we wish to
7735 	 * reduce the max loaded cpu below the average load. At the same time,
7736 	 * we also don't want to reduce the group load below the group
7737 	 * capacity. Thus we look for the minimum possible imbalance.
7738 	 */
7739 	max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7740 
7741 	/* How much load to actually move to equalise the imbalance */
7742 	env->imbalance = min(
7743 		max_pull * busiest->group_capacity,
7744 		(sds->avg_load - local->avg_load) * local->group_capacity
7745 	) / SCHED_CAPACITY_SCALE;
7746 
7747 	/*
7748 	 * if *imbalance is less than the average load per runnable task
7749 	 * there is no guarantee that any tasks will be moved so we'll have
7750 	 * a think about bumping its value to force at least one task to be
7751 	 * moved
7752 	 */
7753 	if (env->imbalance < busiest->load_per_task)
7754 		return fix_small_imbalance(env, sds);
7755 }
7756 
7757 /******* find_busiest_group() helpers end here *********************/
7758 
7759 /**
7760  * find_busiest_group - Returns the busiest group within the sched_domain
7761  * if there is an imbalance.
7762  *
7763  * Also calculates the amount of weighted load which should be moved
7764  * to restore balance.
7765  *
7766  * @env: The load balancing environment.
7767  *
7768  * Return:	- The busiest group if imbalance exists.
7769  */
7770 static struct sched_group *find_busiest_group(struct lb_env *env)
7771 {
7772 	struct sg_lb_stats *local, *busiest;
7773 	struct sd_lb_stats sds;
7774 
7775 	init_sd_lb_stats(&sds);
7776 
7777 	/*
7778 	 * Compute the various statistics relavent for load balancing at
7779 	 * this level.
7780 	 */
7781 	update_sd_lb_stats(env, &sds);
7782 	local = &sds.local_stat;
7783 	busiest = &sds.busiest_stat;
7784 
7785 	/* ASYM feature bypasses nice load balance check */
7786 	if (check_asym_packing(env, &sds))
7787 		return sds.busiest;
7788 
7789 	/* There is no busy sibling group to pull tasks from */
7790 	if (!sds.busiest || busiest->sum_nr_running == 0)
7791 		goto out_balanced;
7792 
7793 	sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7794 						/ sds.total_capacity;
7795 
7796 	/*
7797 	 * If the busiest group is imbalanced the below checks don't
7798 	 * work because they assume all things are equal, which typically
7799 	 * isn't true due to cpus_allowed constraints and the like.
7800 	 */
7801 	if (busiest->group_type == group_imbalanced)
7802 		goto force_balance;
7803 
7804 	/* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7805 	if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7806 	    busiest->group_no_capacity)
7807 		goto force_balance;
7808 
7809 	/*
7810 	 * If the local group is busier than the selected busiest group
7811 	 * don't try and pull any tasks.
7812 	 */
7813 	if (local->avg_load >= busiest->avg_load)
7814 		goto out_balanced;
7815 
7816 	/*
7817 	 * Don't pull any tasks if this group is already above the domain
7818 	 * average load.
7819 	 */
7820 	if (local->avg_load >= sds.avg_load)
7821 		goto out_balanced;
7822 
7823 	if (env->idle == CPU_IDLE) {
7824 		/*
7825 		 * This cpu is idle. If the busiest group is not overloaded
7826 		 * and there is no imbalance between this and busiest group
7827 		 * wrt idle cpus, it is balanced. The imbalance becomes
7828 		 * significant if the diff is greater than 1 otherwise we
7829 		 * might end up to just move the imbalance on another group
7830 		 */
7831 		if ((busiest->group_type != group_overloaded) &&
7832 				(local->idle_cpus <= (busiest->idle_cpus + 1)))
7833 			goto out_balanced;
7834 	} else {
7835 		/*
7836 		 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7837 		 * imbalance_pct to be conservative.
7838 		 */
7839 		if (100 * busiest->avg_load <=
7840 				env->sd->imbalance_pct * local->avg_load)
7841 			goto out_balanced;
7842 	}
7843 
7844 force_balance:
7845 	/* Looks like there is an imbalance. Compute it */
7846 	calculate_imbalance(env, &sds);
7847 	return sds.busiest;
7848 
7849 out_balanced:
7850 	env->imbalance = 0;
7851 	return NULL;
7852 }
7853 
7854 /*
7855  * find_busiest_queue - find the busiest runqueue among the cpus in group.
7856  */
7857 static struct rq *find_busiest_queue(struct lb_env *env,
7858 				     struct sched_group *group)
7859 {
7860 	struct rq *busiest = NULL, *rq;
7861 	unsigned long busiest_load = 0, busiest_capacity = 1;
7862 	int i;
7863 
7864 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7865 		unsigned long capacity, wl;
7866 		enum fbq_type rt;
7867 
7868 		rq = cpu_rq(i);
7869 		rt = fbq_classify_rq(rq);
7870 
7871 		/*
7872 		 * We classify groups/runqueues into three groups:
7873 		 *  - regular: there are !numa tasks
7874 		 *  - remote:  there are numa tasks that run on the 'wrong' node
7875 		 *  - all:     there is no distinction
7876 		 *
7877 		 * In order to avoid migrating ideally placed numa tasks,
7878 		 * ignore those when there's better options.
7879 		 *
7880 		 * If we ignore the actual busiest queue to migrate another
7881 		 * task, the next balance pass can still reduce the busiest
7882 		 * queue by moving tasks around inside the node.
7883 		 *
7884 		 * If we cannot move enough load due to this classification
7885 		 * the next pass will adjust the group classification and
7886 		 * allow migration of more tasks.
7887 		 *
7888 		 * Both cases only affect the total convergence complexity.
7889 		 */
7890 		if (rt > env->fbq_type)
7891 			continue;
7892 
7893 		capacity = capacity_of(i);
7894 
7895 		wl = weighted_cpuload(i);
7896 
7897 		/*
7898 		 * When comparing with imbalance, use weighted_cpuload()
7899 		 * which is not scaled with the cpu capacity.
7900 		 */
7901 
7902 		if (rq->nr_running == 1 && wl > env->imbalance &&
7903 		    !check_cpu_capacity(rq, env->sd))
7904 			continue;
7905 
7906 		/*
7907 		 * For the load comparisons with the other cpu's, consider
7908 		 * the weighted_cpuload() scaled with the cpu capacity, so
7909 		 * that the load can be moved away from the cpu that is
7910 		 * potentially running at a lower capacity.
7911 		 *
7912 		 * Thus we're looking for max(wl_i / capacity_i), crosswise
7913 		 * multiplication to rid ourselves of the division works out
7914 		 * to: wl_i * capacity_j > wl_j * capacity_i;  where j is
7915 		 * our previous maximum.
7916 		 */
7917 		if (wl * busiest_capacity > busiest_load * capacity) {
7918 			busiest_load = wl;
7919 			busiest_capacity = capacity;
7920 			busiest = rq;
7921 		}
7922 	}
7923 
7924 	return busiest;
7925 }
7926 
7927 /*
7928  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7929  * so long as it is large enough.
7930  */
7931 #define MAX_PINNED_INTERVAL	512
7932 
7933 static int need_active_balance(struct lb_env *env)
7934 {
7935 	struct sched_domain *sd = env->sd;
7936 
7937 	if (env->idle == CPU_NEWLY_IDLE) {
7938 
7939 		/*
7940 		 * ASYM_PACKING needs to force migrate tasks from busy but
7941 		 * lower priority CPUs in order to pack all tasks in the
7942 		 * highest priority CPUs.
7943 		 */
7944 		if ((sd->flags & SD_ASYM_PACKING) &&
7945 		    sched_asym_prefer(env->dst_cpu, env->src_cpu))
7946 			return 1;
7947 	}
7948 
7949 	/*
7950 	 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7951 	 * It's worth migrating the task if the src_cpu's capacity is reduced
7952 	 * because of other sched_class or IRQs if more capacity stays
7953 	 * available on dst_cpu.
7954 	 */
7955 	if ((env->idle != CPU_NOT_IDLE) &&
7956 	    (env->src_rq->cfs.h_nr_running == 1)) {
7957 		if ((check_cpu_capacity(env->src_rq, sd)) &&
7958 		    (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7959 			return 1;
7960 	}
7961 
7962 	return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7963 }
7964 
7965 static int active_load_balance_cpu_stop(void *data);
7966 
7967 static int should_we_balance(struct lb_env *env)
7968 {
7969 	struct sched_group *sg = env->sd->groups;
7970 	int cpu, balance_cpu = -1;
7971 
7972 	/*
7973 	 * In the newly idle case, we will allow all the cpu's
7974 	 * to do the newly idle load balance.
7975 	 */
7976 	if (env->idle == CPU_NEWLY_IDLE)
7977 		return 1;
7978 
7979 	/* Try to find first idle cpu */
7980 	for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
7981 		if (!idle_cpu(cpu))
7982 			continue;
7983 
7984 		balance_cpu = cpu;
7985 		break;
7986 	}
7987 
7988 	if (balance_cpu == -1)
7989 		balance_cpu = group_balance_cpu(sg);
7990 
7991 	/*
7992 	 * First idle cpu or the first cpu(busiest) in this sched group
7993 	 * is eligible for doing load balancing at this and above domains.
7994 	 */
7995 	return balance_cpu == env->dst_cpu;
7996 }
7997 
7998 /*
7999  * Check this_cpu to ensure it is balanced within domain. Attempt to move
8000  * tasks if there is an imbalance.
8001  */
8002 static int load_balance(int this_cpu, struct rq *this_rq,
8003 			struct sched_domain *sd, enum cpu_idle_type idle,
8004 			int *continue_balancing)
8005 {
8006 	int ld_moved, cur_ld_moved, active_balance = 0;
8007 	struct sched_domain *sd_parent = sd->parent;
8008 	struct sched_group *group;
8009 	struct rq *busiest;
8010 	struct rq_flags rf;
8011 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8012 
8013 	struct lb_env env = {
8014 		.sd		= sd,
8015 		.dst_cpu	= this_cpu,
8016 		.dst_rq		= this_rq,
8017 		.dst_grpmask    = sched_group_span(sd->groups),
8018 		.idle		= idle,
8019 		.loop_break	= sched_nr_migrate_break,
8020 		.cpus		= cpus,
8021 		.fbq_type	= all,
8022 		.tasks		= LIST_HEAD_INIT(env.tasks),
8023 	};
8024 
8025 	cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8026 
8027 	schedstat_inc(sd->lb_count[idle]);
8028 
8029 redo:
8030 	if (!should_we_balance(&env)) {
8031 		*continue_balancing = 0;
8032 		goto out_balanced;
8033 	}
8034 
8035 	group = find_busiest_group(&env);
8036 	if (!group) {
8037 		schedstat_inc(sd->lb_nobusyg[idle]);
8038 		goto out_balanced;
8039 	}
8040 
8041 	busiest = find_busiest_queue(&env, group);
8042 	if (!busiest) {
8043 		schedstat_inc(sd->lb_nobusyq[idle]);
8044 		goto out_balanced;
8045 	}
8046 
8047 	BUG_ON(busiest == env.dst_rq);
8048 
8049 	schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8050 
8051 	env.src_cpu = busiest->cpu;
8052 	env.src_rq = busiest;
8053 
8054 	ld_moved = 0;
8055 	if (busiest->nr_running > 1) {
8056 		/*
8057 		 * Attempt to move tasks. If find_busiest_group has found
8058 		 * an imbalance but busiest->nr_running <= 1, the group is
8059 		 * still unbalanced. ld_moved simply stays zero, so it is
8060 		 * correctly treated as an imbalance.
8061 		 */
8062 		env.flags |= LBF_ALL_PINNED;
8063 		env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
8064 
8065 more_balance:
8066 		rq_lock_irqsave(busiest, &rf);
8067 		update_rq_clock(busiest);
8068 
8069 		/*
8070 		 * cur_ld_moved - load moved in current iteration
8071 		 * ld_moved     - cumulative load moved across iterations
8072 		 */
8073 		cur_ld_moved = detach_tasks(&env);
8074 
8075 		/*
8076 		 * We've detached some tasks from busiest_rq. Every
8077 		 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8078 		 * unlock busiest->lock, and we are able to be sure
8079 		 * that nobody can manipulate the tasks in parallel.
8080 		 * See task_rq_lock() family for the details.
8081 		 */
8082 
8083 		rq_unlock(busiest, &rf);
8084 
8085 		if (cur_ld_moved) {
8086 			attach_tasks(&env);
8087 			ld_moved += cur_ld_moved;
8088 		}
8089 
8090 		local_irq_restore(rf.flags);
8091 
8092 		if (env.flags & LBF_NEED_BREAK) {
8093 			env.flags &= ~LBF_NEED_BREAK;
8094 			goto more_balance;
8095 		}
8096 
8097 		/*
8098 		 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8099 		 * us and move them to an alternate dst_cpu in our sched_group
8100 		 * where they can run. The upper limit on how many times we
8101 		 * iterate on same src_cpu is dependent on number of cpus in our
8102 		 * sched_group.
8103 		 *
8104 		 * This changes load balance semantics a bit on who can move
8105 		 * load to a given_cpu. In addition to the given_cpu itself
8106 		 * (or a ilb_cpu acting on its behalf where given_cpu is
8107 		 * nohz-idle), we now have balance_cpu in a position to move
8108 		 * load to given_cpu. In rare situations, this may cause
8109 		 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8110 		 * _independently_ and at _same_ time to move some load to
8111 		 * given_cpu) causing exceess load to be moved to given_cpu.
8112 		 * This however should not happen so much in practice and
8113 		 * moreover subsequent load balance cycles should correct the
8114 		 * excess load moved.
8115 		 */
8116 		if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8117 
8118 			/* Prevent to re-select dst_cpu via env's cpus */
8119 			cpumask_clear_cpu(env.dst_cpu, env.cpus);
8120 
8121 			env.dst_rq	 = cpu_rq(env.new_dst_cpu);
8122 			env.dst_cpu	 = env.new_dst_cpu;
8123 			env.flags	&= ~LBF_DST_PINNED;
8124 			env.loop	 = 0;
8125 			env.loop_break	 = sched_nr_migrate_break;
8126 
8127 			/*
8128 			 * Go back to "more_balance" rather than "redo" since we
8129 			 * need to continue with same src_cpu.
8130 			 */
8131 			goto more_balance;
8132 		}
8133 
8134 		/*
8135 		 * We failed to reach balance because of affinity.
8136 		 */
8137 		if (sd_parent) {
8138 			int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8139 
8140 			if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8141 				*group_imbalance = 1;
8142 		}
8143 
8144 		/* All tasks on this runqueue were pinned by CPU affinity */
8145 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
8146 			cpumask_clear_cpu(cpu_of(busiest), cpus);
8147 			/*
8148 			 * Attempting to continue load balancing at the current
8149 			 * sched_domain level only makes sense if there are
8150 			 * active CPUs remaining as possible busiest CPUs to
8151 			 * pull load from which are not contained within the
8152 			 * destination group that is receiving any migrated
8153 			 * load.
8154 			 */
8155 			if (!cpumask_subset(cpus, env.dst_grpmask)) {
8156 				env.loop = 0;
8157 				env.loop_break = sched_nr_migrate_break;
8158 				goto redo;
8159 			}
8160 			goto out_all_pinned;
8161 		}
8162 	}
8163 
8164 	if (!ld_moved) {
8165 		schedstat_inc(sd->lb_failed[idle]);
8166 		/*
8167 		 * Increment the failure counter only on periodic balance.
8168 		 * We do not want newidle balance, which can be very
8169 		 * frequent, pollute the failure counter causing
8170 		 * excessive cache_hot migrations and active balances.
8171 		 */
8172 		if (idle != CPU_NEWLY_IDLE)
8173 			sd->nr_balance_failed++;
8174 
8175 		if (need_active_balance(&env)) {
8176 			unsigned long flags;
8177 
8178 			raw_spin_lock_irqsave(&busiest->lock, flags);
8179 
8180 			/* don't kick the active_load_balance_cpu_stop,
8181 			 * if the curr task on busiest cpu can't be
8182 			 * moved to this_cpu
8183 			 */
8184 			if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8185 				raw_spin_unlock_irqrestore(&busiest->lock,
8186 							    flags);
8187 				env.flags |= LBF_ALL_PINNED;
8188 				goto out_one_pinned;
8189 			}
8190 
8191 			/*
8192 			 * ->active_balance synchronizes accesses to
8193 			 * ->active_balance_work.  Once set, it's cleared
8194 			 * only after active load balance is finished.
8195 			 */
8196 			if (!busiest->active_balance) {
8197 				busiest->active_balance = 1;
8198 				busiest->push_cpu = this_cpu;
8199 				active_balance = 1;
8200 			}
8201 			raw_spin_unlock_irqrestore(&busiest->lock, flags);
8202 
8203 			if (active_balance) {
8204 				stop_one_cpu_nowait(cpu_of(busiest),
8205 					active_load_balance_cpu_stop, busiest,
8206 					&busiest->active_balance_work);
8207 			}
8208 
8209 			/* We've kicked active balancing, force task migration. */
8210 			sd->nr_balance_failed = sd->cache_nice_tries+1;
8211 		}
8212 	} else
8213 		sd->nr_balance_failed = 0;
8214 
8215 	if (likely(!active_balance)) {
8216 		/* We were unbalanced, so reset the balancing interval */
8217 		sd->balance_interval = sd->min_interval;
8218 	} else {
8219 		/*
8220 		 * If we've begun active balancing, start to back off. This
8221 		 * case may not be covered by the all_pinned logic if there
8222 		 * is only 1 task on the busy runqueue (because we don't call
8223 		 * detach_tasks).
8224 		 */
8225 		if (sd->balance_interval < sd->max_interval)
8226 			sd->balance_interval *= 2;
8227 	}
8228 
8229 	goto out;
8230 
8231 out_balanced:
8232 	/*
8233 	 * We reach balance although we may have faced some affinity
8234 	 * constraints. Clear the imbalance flag if it was set.
8235 	 */
8236 	if (sd_parent) {
8237 		int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8238 
8239 		if (*group_imbalance)
8240 			*group_imbalance = 0;
8241 	}
8242 
8243 out_all_pinned:
8244 	/*
8245 	 * We reach balance because all tasks are pinned at this level so
8246 	 * we can't migrate them. Let the imbalance flag set so parent level
8247 	 * can try to migrate them.
8248 	 */
8249 	schedstat_inc(sd->lb_balanced[idle]);
8250 
8251 	sd->nr_balance_failed = 0;
8252 
8253 out_one_pinned:
8254 	/* tune up the balancing interval */
8255 	if (((env.flags & LBF_ALL_PINNED) &&
8256 			sd->balance_interval < MAX_PINNED_INTERVAL) ||
8257 			(sd->balance_interval < sd->max_interval))
8258 		sd->balance_interval *= 2;
8259 
8260 	ld_moved = 0;
8261 out:
8262 	return ld_moved;
8263 }
8264 
8265 static inline unsigned long
8266 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8267 {
8268 	unsigned long interval = sd->balance_interval;
8269 
8270 	if (cpu_busy)
8271 		interval *= sd->busy_factor;
8272 
8273 	/* scale ms to jiffies */
8274 	interval = msecs_to_jiffies(interval);
8275 	interval = clamp(interval, 1UL, max_load_balance_interval);
8276 
8277 	return interval;
8278 }
8279 
8280 static inline void
8281 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8282 {
8283 	unsigned long interval, next;
8284 
8285 	/* used by idle balance, so cpu_busy = 0 */
8286 	interval = get_sd_balance_interval(sd, 0);
8287 	next = sd->last_balance + interval;
8288 
8289 	if (time_after(*next_balance, next))
8290 		*next_balance = next;
8291 }
8292 
8293 /*
8294  * idle_balance is called by schedule() if this_cpu is about to become
8295  * idle. Attempts to pull tasks from other CPUs.
8296  */
8297 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8298 {
8299 	unsigned long next_balance = jiffies + HZ;
8300 	int this_cpu = this_rq->cpu;
8301 	struct sched_domain *sd;
8302 	int pulled_task = 0;
8303 	u64 curr_cost = 0;
8304 
8305 	/*
8306 	 * We must set idle_stamp _before_ calling idle_balance(), such that we
8307 	 * measure the duration of idle_balance() as idle time.
8308 	 */
8309 	this_rq->idle_stamp = rq_clock(this_rq);
8310 
8311 	/*
8312 	 * This is OK, because current is on_cpu, which avoids it being picked
8313 	 * for load-balance and preemption/IRQs are still disabled avoiding
8314 	 * further scheduler activity on it and we're being very careful to
8315 	 * re-start the picking loop.
8316 	 */
8317 	rq_unpin_lock(this_rq, rf);
8318 
8319 	if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8320 	    !this_rq->rd->overload) {
8321 		rcu_read_lock();
8322 		sd = rcu_dereference_check_sched_domain(this_rq->sd);
8323 		if (sd)
8324 			update_next_balance(sd, &next_balance);
8325 		rcu_read_unlock();
8326 
8327 		goto out;
8328 	}
8329 
8330 	raw_spin_unlock(&this_rq->lock);
8331 
8332 	update_blocked_averages(this_cpu);
8333 	rcu_read_lock();
8334 	for_each_domain(this_cpu, sd) {
8335 		int continue_balancing = 1;
8336 		u64 t0, domain_cost;
8337 
8338 		if (!(sd->flags & SD_LOAD_BALANCE))
8339 			continue;
8340 
8341 		if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8342 			update_next_balance(sd, &next_balance);
8343 			break;
8344 		}
8345 
8346 		if (sd->flags & SD_BALANCE_NEWIDLE) {
8347 			t0 = sched_clock_cpu(this_cpu);
8348 
8349 			pulled_task = load_balance(this_cpu, this_rq,
8350 						   sd, CPU_NEWLY_IDLE,
8351 						   &continue_balancing);
8352 
8353 			domain_cost = sched_clock_cpu(this_cpu) - t0;
8354 			if (domain_cost > sd->max_newidle_lb_cost)
8355 				sd->max_newidle_lb_cost = domain_cost;
8356 
8357 			curr_cost += domain_cost;
8358 		}
8359 
8360 		update_next_balance(sd, &next_balance);
8361 
8362 		/*
8363 		 * Stop searching for tasks to pull if there are
8364 		 * now runnable tasks on this rq.
8365 		 */
8366 		if (pulled_task || this_rq->nr_running > 0)
8367 			break;
8368 	}
8369 	rcu_read_unlock();
8370 
8371 	raw_spin_lock(&this_rq->lock);
8372 
8373 	if (curr_cost > this_rq->max_idle_balance_cost)
8374 		this_rq->max_idle_balance_cost = curr_cost;
8375 
8376 	/*
8377 	 * While browsing the domains, we released the rq lock, a task could
8378 	 * have been enqueued in the meantime. Since we're not going idle,
8379 	 * pretend we pulled a task.
8380 	 */
8381 	if (this_rq->cfs.h_nr_running && !pulled_task)
8382 		pulled_task = 1;
8383 
8384 out:
8385 	/* Move the next balance forward */
8386 	if (time_after(this_rq->next_balance, next_balance))
8387 		this_rq->next_balance = next_balance;
8388 
8389 	/* Is there a task of a high priority class? */
8390 	if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8391 		pulled_task = -1;
8392 
8393 	if (pulled_task)
8394 		this_rq->idle_stamp = 0;
8395 
8396 	rq_repin_lock(this_rq, rf);
8397 
8398 	return pulled_task;
8399 }
8400 
8401 /*
8402  * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8403  * running tasks off the busiest CPU onto idle CPUs. It requires at
8404  * least 1 task to be running on each physical CPU where possible, and
8405  * avoids physical / logical imbalances.
8406  */
8407 static int active_load_balance_cpu_stop(void *data)
8408 {
8409 	struct rq *busiest_rq = data;
8410 	int busiest_cpu = cpu_of(busiest_rq);
8411 	int target_cpu = busiest_rq->push_cpu;
8412 	struct rq *target_rq = cpu_rq(target_cpu);
8413 	struct sched_domain *sd;
8414 	struct task_struct *p = NULL;
8415 	struct rq_flags rf;
8416 
8417 	rq_lock_irq(busiest_rq, &rf);
8418 
8419 	/* make sure the requested cpu hasn't gone down in the meantime */
8420 	if (unlikely(busiest_cpu != smp_processor_id() ||
8421 		     !busiest_rq->active_balance))
8422 		goto out_unlock;
8423 
8424 	/* Is there any task to move? */
8425 	if (busiest_rq->nr_running <= 1)
8426 		goto out_unlock;
8427 
8428 	/*
8429 	 * This condition is "impossible", if it occurs
8430 	 * we need to fix it. Originally reported by
8431 	 * Bjorn Helgaas on a 128-cpu setup.
8432 	 */
8433 	BUG_ON(busiest_rq == target_rq);
8434 
8435 	/* Search for an sd spanning us and the target CPU. */
8436 	rcu_read_lock();
8437 	for_each_domain(target_cpu, sd) {
8438 		if ((sd->flags & SD_LOAD_BALANCE) &&
8439 		    cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8440 				break;
8441 	}
8442 
8443 	if (likely(sd)) {
8444 		struct lb_env env = {
8445 			.sd		= sd,
8446 			.dst_cpu	= target_cpu,
8447 			.dst_rq		= target_rq,
8448 			.src_cpu	= busiest_rq->cpu,
8449 			.src_rq		= busiest_rq,
8450 			.idle		= CPU_IDLE,
8451 			/*
8452 			 * can_migrate_task() doesn't need to compute new_dst_cpu
8453 			 * for active balancing. Since we have CPU_IDLE, but no
8454 			 * @dst_grpmask we need to make that test go away with lying
8455 			 * about DST_PINNED.
8456 			 */
8457 			.flags		= LBF_DST_PINNED,
8458 		};
8459 
8460 		schedstat_inc(sd->alb_count);
8461 		update_rq_clock(busiest_rq);
8462 
8463 		p = detach_one_task(&env);
8464 		if (p) {
8465 			schedstat_inc(sd->alb_pushed);
8466 			/* Active balancing done, reset the failure counter. */
8467 			sd->nr_balance_failed = 0;
8468 		} else {
8469 			schedstat_inc(sd->alb_failed);
8470 		}
8471 	}
8472 	rcu_read_unlock();
8473 out_unlock:
8474 	busiest_rq->active_balance = 0;
8475 	rq_unlock(busiest_rq, &rf);
8476 
8477 	if (p)
8478 		attach_one_task(target_rq, p);
8479 
8480 	local_irq_enable();
8481 
8482 	return 0;
8483 }
8484 
8485 static inline int on_null_domain(struct rq *rq)
8486 {
8487 	return unlikely(!rcu_dereference_sched(rq->sd));
8488 }
8489 
8490 #ifdef CONFIG_NO_HZ_COMMON
8491 /*
8492  * idle load balancing details
8493  * - When one of the busy CPUs notice that there may be an idle rebalancing
8494  *   needed, they will kick the idle load balancer, which then does idle
8495  *   load balancing for all the idle CPUs.
8496  */
8497 static struct {
8498 	cpumask_var_t idle_cpus_mask;
8499 	atomic_t nr_cpus;
8500 	unsigned long next_balance;     /* in jiffy units */
8501 } nohz ____cacheline_aligned;
8502 
8503 static inline int find_new_ilb(void)
8504 {
8505 	int ilb = cpumask_first(nohz.idle_cpus_mask);
8506 
8507 	if (ilb < nr_cpu_ids && idle_cpu(ilb))
8508 		return ilb;
8509 
8510 	return nr_cpu_ids;
8511 }
8512 
8513 /*
8514  * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8515  * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8516  * CPU (if there is one).
8517  */
8518 static void nohz_balancer_kick(void)
8519 {
8520 	int ilb_cpu;
8521 
8522 	nohz.next_balance++;
8523 
8524 	ilb_cpu = find_new_ilb();
8525 
8526 	if (ilb_cpu >= nr_cpu_ids)
8527 		return;
8528 
8529 	if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8530 		return;
8531 	/*
8532 	 * Use smp_send_reschedule() instead of resched_cpu().
8533 	 * This way we generate a sched IPI on the target cpu which
8534 	 * is idle. And the softirq performing nohz idle load balance
8535 	 * will be run before returning from the IPI.
8536 	 */
8537 	smp_send_reschedule(ilb_cpu);
8538 	return;
8539 }
8540 
8541 void nohz_balance_exit_idle(unsigned int cpu)
8542 {
8543 	if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8544 		/*
8545 		 * Completely isolated CPUs don't ever set, so we must test.
8546 		 */
8547 		if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8548 			cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8549 			atomic_dec(&nohz.nr_cpus);
8550 		}
8551 		clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8552 	}
8553 }
8554 
8555 static inline void set_cpu_sd_state_busy(void)
8556 {
8557 	struct sched_domain *sd;
8558 	int cpu = smp_processor_id();
8559 
8560 	rcu_read_lock();
8561 	sd = rcu_dereference(per_cpu(sd_llc, cpu));
8562 
8563 	if (!sd || !sd->nohz_idle)
8564 		goto unlock;
8565 	sd->nohz_idle = 0;
8566 
8567 	atomic_inc(&sd->shared->nr_busy_cpus);
8568 unlock:
8569 	rcu_read_unlock();
8570 }
8571 
8572 void set_cpu_sd_state_idle(void)
8573 {
8574 	struct sched_domain *sd;
8575 	int cpu = smp_processor_id();
8576 
8577 	rcu_read_lock();
8578 	sd = rcu_dereference(per_cpu(sd_llc, cpu));
8579 
8580 	if (!sd || sd->nohz_idle)
8581 		goto unlock;
8582 	sd->nohz_idle = 1;
8583 
8584 	atomic_dec(&sd->shared->nr_busy_cpus);
8585 unlock:
8586 	rcu_read_unlock();
8587 }
8588 
8589 /*
8590  * This routine will record that the cpu is going idle with tick stopped.
8591  * This info will be used in performing idle load balancing in the future.
8592  */
8593 void nohz_balance_enter_idle(int cpu)
8594 {
8595 	/*
8596 	 * If this cpu is going down, then nothing needs to be done.
8597 	 */
8598 	if (!cpu_active(cpu))
8599 		return;
8600 
8601 	/* Spare idle load balancing on CPUs that don't want to be disturbed: */
8602 	if (!is_housekeeping_cpu(cpu))
8603 		return;
8604 
8605 	if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8606 		return;
8607 
8608 	/*
8609 	 * If we're a completely isolated CPU, we don't play.
8610 	 */
8611 	if (on_null_domain(cpu_rq(cpu)))
8612 		return;
8613 
8614 	cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8615 	atomic_inc(&nohz.nr_cpus);
8616 	set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8617 }
8618 #endif
8619 
8620 static DEFINE_SPINLOCK(balancing);
8621 
8622 /*
8623  * Scale the max load_balance interval with the number of CPUs in the system.
8624  * This trades load-balance latency on larger machines for less cross talk.
8625  */
8626 void update_max_interval(void)
8627 {
8628 	max_load_balance_interval = HZ*num_online_cpus()/10;
8629 }
8630 
8631 /*
8632  * It checks each scheduling domain to see if it is due to be balanced,
8633  * and initiates a balancing operation if so.
8634  *
8635  * Balancing parameters are set up in init_sched_domains.
8636  */
8637 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8638 {
8639 	int continue_balancing = 1;
8640 	int cpu = rq->cpu;
8641 	unsigned long interval;
8642 	struct sched_domain *sd;
8643 	/* Earliest time when we have to do rebalance again */
8644 	unsigned long next_balance = jiffies + 60*HZ;
8645 	int update_next_balance = 0;
8646 	int need_serialize, need_decay = 0;
8647 	u64 max_cost = 0;
8648 
8649 	update_blocked_averages(cpu);
8650 
8651 	rcu_read_lock();
8652 	for_each_domain(cpu, sd) {
8653 		/*
8654 		 * Decay the newidle max times here because this is a regular
8655 		 * visit to all the domains. Decay ~1% per second.
8656 		 */
8657 		if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8658 			sd->max_newidle_lb_cost =
8659 				(sd->max_newidle_lb_cost * 253) / 256;
8660 			sd->next_decay_max_lb_cost = jiffies + HZ;
8661 			need_decay = 1;
8662 		}
8663 		max_cost += sd->max_newidle_lb_cost;
8664 
8665 		if (!(sd->flags & SD_LOAD_BALANCE))
8666 			continue;
8667 
8668 		/*
8669 		 * Stop the load balance at this level. There is another
8670 		 * CPU in our sched group which is doing load balancing more
8671 		 * actively.
8672 		 */
8673 		if (!continue_balancing) {
8674 			if (need_decay)
8675 				continue;
8676 			break;
8677 		}
8678 
8679 		interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8680 
8681 		need_serialize = sd->flags & SD_SERIALIZE;
8682 		if (need_serialize) {
8683 			if (!spin_trylock(&balancing))
8684 				goto out;
8685 		}
8686 
8687 		if (time_after_eq(jiffies, sd->last_balance + interval)) {
8688 			if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8689 				/*
8690 				 * The LBF_DST_PINNED logic could have changed
8691 				 * env->dst_cpu, so we can't know our idle
8692 				 * state even if we migrated tasks. Update it.
8693 				 */
8694 				idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8695 			}
8696 			sd->last_balance = jiffies;
8697 			interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8698 		}
8699 		if (need_serialize)
8700 			spin_unlock(&balancing);
8701 out:
8702 		if (time_after(next_balance, sd->last_balance + interval)) {
8703 			next_balance = sd->last_balance + interval;
8704 			update_next_balance = 1;
8705 		}
8706 	}
8707 	if (need_decay) {
8708 		/*
8709 		 * Ensure the rq-wide value also decays but keep it at a
8710 		 * reasonable floor to avoid funnies with rq->avg_idle.
8711 		 */
8712 		rq->max_idle_balance_cost =
8713 			max((u64)sysctl_sched_migration_cost, max_cost);
8714 	}
8715 	rcu_read_unlock();
8716 
8717 	/*
8718 	 * next_balance will be updated only when there is a need.
8719 	 * When the cpu is attached to null domain for ex, it will not be
8720 	 * updated.
8721 	 */
8722 	if (likely(update_next_balance)) {
8723 		rq->next_balance = next_balance;
8724 
8725 #ifdef CONFIG_NO_HZ_COMMON
8726 		/*
8727 		 * If this CPU has been elected to perform the nohz idle
8728 		 * balance. Other idle CPUs have already rebalanced with
8729 		 * nohz_idle_balance() and nohz.next_balance has been
8730 		 * updated accordingly. This CPU is now running the idle load
8731 		 * balance for itself and we need to update the
8732 		 * nohz.next_balance accordingly.
8733 		 */
8734 		if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8735 			nohz.next_balance = rq->next_balance;
8736 #endif
8737 	}
8738 }
8739 
8740 #ifdef CONFIG_NO_HZ_COMMON
8741 /*
8742  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8743  * rebalancing for all the cpus for whom scheduler ticks are stopped.
8744  */
8745 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8746 {
8747 	int this_cpu = this_rq->cpu;
8748 	struct rq *rq;
8749 	int balance_cpu;
8750 	/* Earliest time when we have to do rebalance again */
8751 	unsigned long next_balance = jiffies + 60*HZ;
8752 	int update_next_balance = 0;
8753 
8754 	if (idle != CPU_IDLE ||
8755 	    !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8756 		goto end;
8757 
8758 	for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8759 		if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8760 			continue;
8761 
8762 		/*
8763 		 * If this cpu gets work to do, stop the load balancing
8764 		 * work being done for other cpus. Next load
8765 		 * balancing owner will pick it up.
8766 		 */
8767 		if (need_resched())
8768 			break;
8769 
8770 		rq = cpu_rq(balance_cpu);
8771 
8772 		/*
8773 		 * If time for next balance is due,
8774 		 * do the balance.
8775 		 */
8776 		if (time_after_eq(jiffies, rq->next_balance)) {
8777 			struct rq_flags rf;
8778 
8779 			rq_lock_irq(rq, &rf);
8780 			update_rq_clock(rq);
8781 			cpu_load_update_idle(rq);
8782 			rq_unlock_irq(rq, &rf);
8783 
8784 			rebalance_domains(rq, CPU_IDLE);
8785 		}
8786 
8787 		if (time_after(next_balance, rq->next_balance)) {
8788 			next_balance = rq->next_balance;
8789 			update_next_balance = 1;
8790 		}
8791 	}
8792 
8793 	/*
8794 	 * next_balance will be updated only when there is a need.
8795 	 * When the CPU is attached to null domain for ex, it will not be
8796 	 * updated.
8797 	 */
8798 	if (likely(update_next_balance))
8799 		nohz.next_balance = next_balance;
8800 end:
8801 	clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8802 }
8803 
8804 /*
8805  * Current heuristic for kicking the idle load balancer in the presence
8806  * of an idle cpu in the system.
8807  *   - This rq has more than one task.
8808  *   - This rq has at least one CFS task and the capacity of the CPU is
8809  *     significantly reduced because of RT tasks or IRQs.
8810  *   - At parent of LLC scheduler domain level, this cpu's scheduler group has
8811  *     multiple busy cpu.
8812  *   - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8813  *     domain span are idle.
8814  */
8815 static inline bool nohz_kick_needed(struct rq *rq)
8816 {
8817 	unsigned long now = jiffies;
8818 	struct sched_domain_shared *sds;
8819 	struct sched_domain *sd;
8820 	int nr_busy, i, cpu = rq->cpu;
8821 	bool kick = false;
8822 
8823 	if (unlikely(rq->idle_balance))
8824 		return false;
8825 
8826        /*
8827 	* We may be recently in ticked or tickless idle mode. At the first
8828 	* busy tick after returning from idle, we will update the busy stats.
8829 	*/
8830 	set_cpu_sd_state_busy();
8831 	nohz_balance_exit_idle(cpu);
8832 
8833 	/*
8834 	 * None are in tickless mode and hence no need for NOHZ idle load
8835 	 * balancing.
8836 	 */
8837 	if (likely(!atomic_read(&nohz.nr_cpus)))
8838 		return false;
8839 
8840 	if (time_before(now, nohz.next_balance))
8841 		return false;
8842 
8843 	if (rq->nr_running >= 2)
8844 		return true;
8845 
8846 	rcu_read_lock();
8847 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
8848 	if (sds) {
8849 		/*
8850 		 * XXX: write a coherent comment on why we do this.
8851 		 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8852 		 */
8853 		nr_busy = atomic_read(&sds->nr_busy_cpus);
8854 		if (nr_busy > 1) {
8855 			kick = true;
8856 			goto unlock;
8857 		}
8858 
8859 	}
8860 
8861 	sd = rcu_dereference(rq->sd);
8862 	if (sd) {
8863 		if ((rq->cfs.h_nr_running >= 1) &&
8864 				check_cpu_capacity(rq, sd)) {
8865 			kick = true;
8866 			goto unlock;
8867 		}
8868 	}
8869 
8870 	sd = rcu_dereference(per_cpu(sd_asym, cpu));
8871 	if (sd) {
8872 		for_each_cpu(i, sched_domain_span(sd)) {
8873 			if (i == cpu ||
8874 			    !cpumask_test_cpu(i, nohz.idle_cpus_mask))
8875 				continue;
8876 
8877 			if (sched_asym_prefer(i, cpu)) {
8878 				kick = true;
8879 				goto unlock;
8880 			}
8881 		}
8882 	}
8883 unlock:
8884 	rcu_read_unlock();
8885 	return kick;
8886 }
8887 #else
8888 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8889 #endif
8890 
8891 /*
8892  * run_rebalance_domains is triggered when needed from the scheduler tick.
8893  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8894  */
8895 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
8896 {
8897 	struct rq *this_rq = this_rq();
8898 	enum cpu_idle_type idle = this_rq->idle_balance ?
8899 						CPU_IDLE : CPU_NOT_IDLE;
8900 
8901 	/*
8902 	 * If this cpu has a pending nohz_balance_kick, then do the
8903 	 * balancing on behalf of the other idle cpus whose ticks are
8904 	 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8905 	 * give the idle cpus a chance to load balance. Else we may
8906 	 * load balance only within the local sched_domain hierarchy
8907 	 * and abort nohz_idle_balance altogether if we pull some load.
8908 	 */
8909 	nohz_idle_balance(this_rq, idle);
8910 	rebalance_domains(this_rq, idle);
8911 }
8912 
8913 /*
8914  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8915  */
8916 void trigger_load_balance(struct rq *rq)
8917 {
8918 	/* Don't need to rebalance while attached to NULL domain */
8919 	if (unlikely(on_null_domain(rq)))
8920 		return;
8921 
8922 	if (time_after_eq(jiffies, rq->next_balance))
8923 		raise_softirq(SCHED_SOFTIRQ);
8924 #ifdef CONFIG_NO_HZ_COMMON
8925 	if (nohz_kick_needed(rq))
8926 		nohz_balancer_kick();
8927 #endif
8928 }
8929 
8930 static void rq_online_fair(struct rq *rq)
8931 {
8932 	update_sysctl();
8933 
8934 	update_runtime_enabled(rq);
8935 }
8936 
8937 static void rq_offline_fair(struct rq *rq)
8938 {
8939 	update_sysctl();
8940 
8941 	/* Ensure any throttled groups are reachable by pick_next_task */
8942 	unthrottle_offline_cfs_rqs(rq);
8943 }
8944 
8945 #endif /* CONFIG_SMP */
8946 
8947 /*
8948  * scheduler tick hitting a task of our scheduling class:
8949  */
8950 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8951 {
8952 	struct cfs_rq *cfs_rq;
8953 	struct sched_entity *se = &curr->se;
8954 
8955 	for_each_sched_entity(se) {
8956 		cfs_rq = cfs_rq_of(se);
8957 		entity_tick(cfs_rq, se, queued);
8958 	}
8959 
8960 	if (static_branch_unlikely(&sched_numa_balancing))
8961 		task_tick_numa(rq, curr);
8962 }
8963 
8964 /*
8965  * called on fork with the child task as argument from the parent's context
8966  *  - child not yet on the tasklist
8967  *  - preemption disabled
8968  */
8969 static void task_fork_fair(struct task_struct *p)
8970 {
8971 	struct cfs_rq *cfs_rq;
8972 	struct sched_entity *se = &p->se, *curr;
8973 	struct rq *rq = this_rq();
8974 	struct rq_flags rf;
8975 
8976 	rq_lock(rq, &rf);
8977 	update_rq_clock(rq);
8978 
8979 	cfs_rq = task_cfs_rq(current);
8980 	curr = cfs_rq->curr;
8981 	if (curr) {
8982 		update_curr(cfs_rq);
8983 		se->vruntime = curr->vruntime;
8984 	}
8985 	place_entity(cfs_rq, se, 1);
8986 
8987 	if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8988 		/*
8989 		 * Upon rescheduling, sched_class::put_prev_task() will place
8990 		 * 'current' within the tree based on its new key value.
8991 		 */
8992 		swap(curr->vruntime, se->vruntime);
8993 		resched_curr(rq);
8994 	}
8995 
8996 	se->vruntime -= cfs_rq->min_vruntime;
8997 	rq_unlock(rq, &rf);
8998 }
8999 
9000 /*
9001  * Priority of the task has changed. Check to see if we preempt
9002  * the current task.
9003  */
9004 static void
9005 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9006 {
9007 	if (!task_on_rq_queued(p))
9008 		return;
9009 
9010 	/*
9011 	 * Reschedule if we are currently running on this runqueue and
9012 	 * our priority decreased, or if we are not currently running on
9013 	 * this runqueue and our priority is higher than the current's
9014 	 */
9015 	if (rq->curr == p) {
9016 		if (p->prio > oldprio)
9017 			resched_curr(rq);
9018 	} else
9019 		check_preempt_curr(rq, p, 0);
9020 }
9021 
9022 static inline bool vruntime_normalized(struct task_struct *p)
9023 {
9024 	struct sched_entity *se = &p->se;
9025 
9026 	/*
9027 	 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9028 	 * the dequeue_entity(.flags=0) will already have normalized the
9029 	 * vruntime.
9030 	 */
9031 	if (p->on_rq)
9032 		return true;
9033 
9034 	/*
9035 	 * When !on_rq, vruntime of the task has usually NOT been normalized.
9036 	 * But there are some cases where it has already been normalized:
9037 	 *
9038 	 * - A forked child which is waiting for being woken up by
9039 	 *   wake_up_new_task().
9040 	 * - A task which has been woken up by try_to_wake_up() and
9041 	 *   waiting for actually being woken up by sched_ttwu_pending().
9042 	 */
9043 	if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9044 		return true;
9045 
9046 	return false;
9047 }
9048 
9049 #ifdef CONFIG_FAIR_GROUP_SCHED
9050 /*
9051  * Propagate the changes of the sched_entity across the tg tree to make it
9052  * visible to the root
9053  */
9054 static void propagate_entity_cfs_rq(struct sched_entity *se)
9055 {
9056 	struct cfs_rq *cfs_rq;
9057 
9058 	/* Start to propagate at parent */
9059 	se = se->parent;
9060 
9061 	for_each_sched_entity(se) {
9062 		cfs_rq = cfs_rq_of(se);
9063 
9064 		if (cfs_rq_throttled(cfs_rq))
9065 			break;
9066 
9067 		update_load_avg(se, UPDATE_TG);
9068 	}
9069 }
9070 #else
9071 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9072 #endif
9073 
9074 static void detach_entity_cfs_rq(struct sched_entity *se)
9075 {
9076 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
9077 
9078 	/* Catch up with the cfs_rq and remove our load when we leave */
9079 	update_load_avg(se, 0);
9080 	detach_entity_load_avg(cfs_rq, se);
9081 	update_tg_load_avg(cfs_rq, false);
9082 	propagate_entity_cfs_rq(se);
9083 }
9084 
9085 static void attach_entity_cfs_rq(struct sched_entity *se)
9086 {
9087 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
9088 
9089 #ifdef CONFIG_FAIR_GROUP_SCHED
9090 	/*
9091 	 * Since the real-depth could have been changed (only FAIR
9092 	 * class maintain depth value), reset depth properly.
9093 	 */
9094 	se->depth = se->parent ? se->parent->depth + 1 : 0;
9095 #endif
9096 
9097 	/* Synchronize entity with its cfs_rq */
9098 	update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9099 	attach_entity_load_avg(cfs_rq, se);
9100 	update_tg_load_avg(cfs_rq, false);
9101 	propagate_entity_cfs_rq(se);
9102 }
9103 
9104 static void detach_task_cfs_rq(struct task_struct *p)
9105 {
9106 	struct sched_entity *se = &p->se;
9107 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
9108 
9109 	if (!vruntime_normalized(p)) {
9110 		/*
9111 		 * Fix up our vruntime so that the current sleep doesn't
9112 		 * cause 'unlimited' sleep bonus.
9113 		 */
9114 		place_entity(cfs_rq, se, 0);
9115 		se->vruntime -= cfs_rq->min_vruntime;
9116 	}
9117 
9118 	detach_entity_cfs_rq(se);
9119 }
9120 
9121 static void attach_task_cfs_rq(struct task_struct *p)
9122 {
9123 	struct sched_entity *se = &p->se;
9124 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
9125 
9126 	attach_entity_cfs_rq(se);
9127 
9128 	if (!vruntime_normalized(p))
9129 		se->vruntime += cfs_rq->min_vruntime;
9130 }
9131 
9132 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9133 {
9134 	detach_task_cfs_rq(p);
9135 }
9136 
9137 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9138 {
9139 	attach_task_cfs_rq(p);
9140 
9141 	if (task_on_rq_queued(p)) {
9142 		/*
9143 		 * We were most likely switched from sched_rt, so
9144 		 * kick off the schedule if running, otherwise just see
9145 		 * if we can still preempt the current task.
9146 		 */
9147 		if (rq->curr == p)
9148 			resched_curr(rq);
9149 		else
9150 			check_preempt_curr(rq, p, 0);
9151 	}
9152 }
9153 
9154 /* Account for a task changing its policy or group.
9155  *
9156  * This routine is mostly called to set cfs_rq->curr field when a task
9157  * migrates between groups/classes.
9158  */
9159 static void set_curr_task_fair(struct rq *rq)
9160 {
9161 	struct sched_entity *se = &rq->curr->se;
9162 
9163 	for_each_sched_entity(se) {
9164 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
9165 
9166 		set_next_entity(cfs_rq, se);
9167 		/* ensure bandwidth has been allocated on our new cfs_rq */
9168 		account_cfs_rq_runtime(cfs_rq, 0);
9169 	}
9170 }
9171 
9172 void init_cfs_rq(struct cfs_rq *cfs_rq)
9173 {
9174 	cfs_rq->tasks_timeline = RB_ROOT;
9175 	cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9176 #ifndef CONFIG_64BIT
9177 	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9178 #endif
9179 #ifdef CONFIG_SMP
9180 #ifdef CONFIG_FAIR_GROUP_SCHED
9181 	cfs_rq->propagate_avg = 0;
9182 #endif
9183 	atomic_long_set(&cfs_rq->removed_load_avg, 0);
9184 	atomic_long_set(&cfs_rq->removed_util_avg, 0);
9185 #endif
9186 }
9187 
9188 #ifdef CONFIG_FAIR_GROUP_SCHED
9189 static void task_set_group_fair(struct task_struct *p)
9190 {
9191 	struct sched_entity *se = &p->se;
9192 
9193 	set_task_rq(p, task_cpu(p));
9194 	se->depth = se->parent ? se->parent->depth + 1 : 0;
9195 }
9196 
9197 static void task_move_group_fair(struct task_struct *p)
9198 {
9199 	detach_task_cfs_rq(p);
9200 	set_task_rq(p, task_cpu(p));
9201 
9202 #ifdef CONFIG_SMP
9203 	/* Tell se's cfs_rq has been changed -- migrated */
9204 	p->se.avg.last_update_time = 0;
9205 #endif
9206 	attach_task_cfs_rq(p);
9207 }
9208 
9209 static void task_change_group_fair(struct task_struct *p, int type)
9210 {
9211 	switch (type) {
9212 	case TASK_SET_GROUP:
9213 		task_set_group_fair(p);
9214 		break;
9215 
9216 	case TASK_MOVE_GROUP:
9217 		task_move_group_fair(p);
9218 		break;
9219 	}
9220 }
9221 
9222 void free_fair_sched_group(struct task_group *tg)
9223 {
9224 	int i;
9225 
9226 	destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9227 
9228 	for_each_possible_cpu(i) {
9229 		if (tg->cfs_rq)
9230 			kfree(tg->cfs_rq[i]);
9231 		if (tg->se)
9232 			kfree(tg->se[i]);
9233 	}
9234 
9235 	kfree(tg->cfs_rq);
9236 	kfree(tg->se);
9237 }
9238 
9239 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9240 {
9241 	struct sched_entity *se;
9242 	struct cfs_rq *cfs_rq;
9243 	int i;
9244 
9245 	tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9246 	if (!tg->cfs_rq)
9247 		goto err;
9248 	tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9249 	if (!tg->se)
9250 		goto err;
9251 
9252 	tg->shares = NICE_0_LOAD;
9253 
9254 	init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9255 
9256 	for_each_possible_cpu(i) {
9257 		cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9258 				      GFP_KERNEL, cpu_to_node(i));
9259 		if (!cfs_rq)
9260 			goto err;
9261 
9262 		se = kzalloc_node(sizeof(struct sched_entity),
9263 				  GFP_KERNEL, cpu_to_node(i));
9264 		if (!se)
9265 			goto err_free_rq;
9266 
9267 		init_cfs_rq(cfs_rq);
9268 		init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9269 		init_entity_runnable_average(se);
9270 	}
9271 
9272 	return 1;
9273 
9274 err_free_rq:
9275 	kfree(cfs_rq);
9276 err:
9277 	return 0;
9278 }
9279 
9280 void online_fair_sched_group(struct task_group *tg)
9281 {
9282 	struct sched_entity *se;
9283 	struct rq *rq;
9284 	int i;
9285 
9286 	for_each_possible_cpu(i) {
9287 		rq = cpu_rq(i);
9288 		se = tg->se[i];
9289 
9290 		raw_spin_lock_irq(&rq->lock);
9291 		update_rq_clock(rq);
9292 		attach_entity_cfs_rq(se);
9293 		sync_throttle(tg, i);
9294 		raw_spin_unlock_irq(&rq->lock);
9295 	}
9296 }
9297 
9298 void unregister_fair_sched_group(struct task_group *tg)
9299 {
9300 	unsigned long flags;
9301 	struct rq *rq;
9302 	int cpu;
9303 
9304 	for_each_possible_cpu(cpu) {
9305 		if (tg->se[cpu])
9306 			remove_entity_load_avg(tg->se[cpu]);
9307 
9308 		/*
9309 		 * Only empty task groups can be destroyed; so we can speculatively
9310 		 * check on_list without danger of it being re-added.
9311 		 */
9312 		if (!tg->cfs_rq[cpu]->on_list)
9313 			continue;
9314 
9315 		rq = cpu_rq(cpu);
9316 
9317 		raw_spin_lock_irqsave(&rq->lock, flags);
9318 		list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9319 		raw_spin_unlock_irqrestore(&rq->lock, flags);
9320 	}
9321 }
9322 
9323 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9324 			struct sched_entity *se, int cpu,
9325 			struct sched_entity *parent)
9326 {
9327 	struct rq *rq = cpu_rq(cpu);
9328 
9329 	cfs_rq->tg = tg;
9330 	cfs_rq->rq = rq;
9331 	init_cfs_rq_runtime(cfs_rq);
9332 
9333 	tg->cfs_rq[cpu] = cfs_rq;
9334 	tg->se[cpu] = se;
9335 
9336 	/* se could be NULL for root_task_group */
9337 	if (!se)
9338 		return;
9339 
9340 	if (!parent) {
9341 		se->cfs_rq = &rq->cfs;
9342 		se->depth = 0;
9343 	} else {
9344 		se->cfs_rq = parent->my_q;
9345 		se->depth = parent->depth + 1;
9346 	}
9347 
9348 	se->my_q = cfs_rq;
9349 	/* guarantee group entities always have weight */
9350 	update_load_set(&se->load, NICE_0_LOAD);
9351 	se->parent = parent;
9352 }
9353 
9354 static DEFINE_MUTEX(shares_mutex);
9355 
9356 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9357 {
9358 	int i;
9359 
9360 	/*
9361 	 * We can't change the weight of the root cgroup.
9362 	 */
9363 	if (!tg->se[0])
9364 		return -EINVAL;
9365 
9366 	shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9367 
9368 	mutex_lock(&shares_mutex);
9369 	if (tg->shares == shares)
9370 		goto done;
9371 
9372 	tg->shares = shares;
9373 	for_each_possible_cpu(i) {
9374 		struct rq *rq = cpu_rq(i);
9375 		struct sched_entity *se = tg->se[i];
9376 		struct rq_flags rf;
9377 
9378 		/* Propagate contribution to hierarchy */
9379 		rq_lock_irqsave(rq, &rf);
9380 		update_rq_clock(rq);
9381 		for_each_sched_entity(se) {
9382 			update_load_avg(se, UPDATE_TG);
9383 			update_cfs_shares(se);
9384 		}
9385 		rq_unlock_irqrestore(rq, &rf);
9386 	}
9387 
9388 done:
9389 	mutex_unlock(&shares_mutex);
9390 	return 0;
9391 }
9392 #else /* CONFIG_FAIR_GROUP_SCHED */
9393 
9394 void free_fair_sched_group(struct task_group *tg) { }
9395 
9396 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9397 {
9398 	return 1;
9399 }
9400 
9401 void online_fair_sched_group(struct task_group *tg) { }
9402 
9403 void unregister_fair_sched_group(struct task_group *tg) { }
9404 
9405 #endif /* CONFIG_FAIR_GROUP_SCHED */
9406 
9407 
9408 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9409 {
9410 	struct sched_entity *se = &task->se;
9411 	unsigned int rr_interval = 0;
9412 
9413 	/*
9414 	 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9415 	 * idle runqueue:
9416 	 */
9417 	if (rq->cfs.load.weight)
9418 		rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9419 
9420 	return rr_interval;
9421 }
9422 
9423 /*
9424  * All the scheduling class methods:
9425  */
9426 const struct sched_class fair_sched_class = {
9427 	.next			= &idle_sched_class,
9428 	.enqueue_task		= enqueue_task_fair,
9429 	.dequeue_task		= dequeue_task_fair,
9430 	.yield_task		= yield_task_fair,
9431 	.yield_to_task		= yield_to_task_fair,
9432 
9433 	.check_preempt_curr	= check_preempt_wakeup,
9434 
9435 	.pick_next_task		= pick_next_task_fair,
9436 	.put_prev_task		= put_prev_task_fair,
9437 
9438 #ifdef CONFIG_SMP
9439 	.select_task_rq		= select_task_rq_fair,
9440 	.migrate_task_rq	= migrate_task_rq_fair,
9441 
9442 	.rq_online		= rq_online_fair,
9443 	.rq_offline		= rq_offline_fair,
9444 
9445 	.task_dead		= task_dead_fair,
9446 	.set_cpus_allowed	= set_cpus_allowed_common,
9447 #endif
9448 
9449 	.set_curr_task          = set_curr_task_fair,
9450 	.task_tick		= task_tick_fair,
9451 	.task_fork		= task_fork_fair,
9452 
9453 	.prio_changed		= prio_changed_fair,
9454 	.switched_from		= switched_from_fair,
9455 	.switched_to		= switched_to_fair,
9456 
9457 	.get_rr_interval	= get_rr_interval_fair,
9458 
9459 	.update_curr		= update_curr_fair,
9460 
9461 #ifdef CONFIG_FAIR_GROUP_SCHED
9462 	.task_change_group	= task_change_group_fair,
9463 #endif
9464 };
9465 
9466 #ifdef CONFIG_SCHED_DEBUG
9467 void print_cfs_stats(struct seq_file *m, int cpu)
9468 {
9469 	struct cfs_rq *cfs_rq, *pos;
9470 
9471 	rcu_read_lock();
9472 	for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
9473 		print_cfs_rq(m, cpu, cfs_rq);
9474 	rcu_read_unlock();
9475 }
9476 
9477 #ifdef CONFIG_NUMA_BALANCING
9478 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9479 {
9480 	int node;
9481 	unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9482 
9483 	for_each_online_node(node) {
9484 		if (p->numa_faults) {
9485 			tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9486 			tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9487 		}
9488 		if (p->numa_group) {
9489 			gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9490 			gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9491 		}
9492 		print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9493 	}
9494 }
9495 #endif /* CONFIG_NUMA_BALANCING */
9496 #endif /* CONFIG_SCHED_DEBUG */
9497 
9498 __init void init_sched_fair_class(void)
9499 {
9500 #ifdef CONFIG_SMP
9501 	open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9502 
9503 #ifdef CONFIG_NO_HZ_COMMON
9504 	nohz.next_balance = jiffies;
9505 	zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9506 #endif
9507 #endif /* SMP */
9508 
9509 }
9510