xref: /openbmc/linux/kernel/sched/fair.c (revision ff6defa6)
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 <pzijlstr@redhat.com>
21  */
22 
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
33 
34 #include <trace/events/sched.h>
35 
36 #include "sched.h"
37 
38 /*
39  * Targeted preemption latency for CPU-bound tasks:
40  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41  *
42  * NOTE: this latency value is not the same as the concept of
43  * 'timeslice length' - timeslices in CFS are of variable length
44  * and have no persistent notion like in traditional, time-slice
45  * based scheduling concepts.
46  *
47  * (to see the precise effective timeslice length of your workload,
48  *  run vmstat and monitor the context-switches (cs) field)
49  */
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
52 
53 /*
54  * The initial- and re-scaling of tunables is configurable
55  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
56  *
57  * Options are:
58  * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59  * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60  * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61  */
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 	= SCHED_TUNABLESCALING_LOG;
64 
65 /*
66  * Minimal preemption granularity for CPU-bound tasks:
67  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68  */
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
71 
72 /*
73  * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74  */
75 static unsigned int sched_nr_latency = 8;
76 
77 /*
78  * After fork, child runs first. If set to 0 (default) then
79  * parent will (try to) run first.
80  */
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
82 
83 /*
84  * SCHED_OTHER wake-up granularity.
85  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86  *
87  * This option delays the preemption effects of decoupled workloads
88  * and reduces their over-scheduling. Synchronous workloads will still
89  * have immediate wakeup/sleep latencies.
90  */
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
95 
96 /*
97  * The exponential sliding  window over which load is averaged for shares
98  * distribution.
99  * (default: 10msec)
100  */
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 
103 #ifdef CONFIG_CFS_BANDWIDTH
104 /*
105  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106  * each time a cfs_rq requests quota.
107  *
108  * Note: in the case that the slice exceeds the runtime remaining (either due
109  * to consumption or the quota being specified to be smaller than the slice)
110  * we will always only issue the remaining available time.
111  *
112  * default: 5 msec, units: microseconds
113   */
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
115 #endif
116 
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
118 {
119 	lw->weight += inc;
120 	lw->inv_weight = 0;
121 }
122 
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
124 {
125 	lw->weight -= dec;
126 	lw->inv_weight = 0;
127 }
128 
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
130 {
131 	lw->weight = w;
132 	lw->inv_weight = 0;
133 }
134 
135 /*
136  * Increase the granularity value when there are more CPUs,
137  * because with more CPUs the 'effective latency' as visible
138  * to users decreases. But the relationship is not linear,
139  * so pick a second-best guess by going with the log2 of the
140  * number of CPUs.
141  *
142  * This idea comes from the SD scheduler of Con Kolivas:
143  */
144 static int get_update_sysctl_factor(void)
145 {
146 	unsigned int cpus = min_t(int, num_online_cpus(), 8);
147 	unsigned int factor;
148 
149 	switch (sysctl_sched_tunable_scaling) {
150 	case SCHED_TUNABLESCALING_NONE:
151 		factor = 1;
152 		break;
153 	case SCHED_TUNABLESCALING_LINEAR:
154 		factor = cpus;
155 		break;
156 	case SCHED_TUNABLESCALING_LOG:
157 	default:
158 		factor = 1 + ilog2(cpus);
159 		break;
160 	}
161 
162 	return factor;
163 }
164 
165 static void update_sysctl(void)
166 {
167 	unsigned int factor = get_update_sysctl_factor();
168 
169 #define SET_SYSCTL(name) \
170 	(sysctl_##name = (factor) * normalized_sysctl_##name)
171 	SET_SYSCTL(sched_min_granularity);
172 	SET_SYSCTL(sched_latency);
173 	SET_SYSCTL(sched_wakeup_granularity);
174 #undef SET_SYSCTL
175 }
176 
177 void sched_init_granularity(void)
178 {
179 	update_sysctl();
180 }
181 
182 #define WMULT_CONST	(~0U)
183 #define WMULT_SHIFT	32
184 
185 static void __update_inv_weight(struct load_weight *lw)
186 {
187 	unsigned long w;
188 
189 	if (likely(lw->inv_weight))
190 		return;
191 
192 	w = scale_load_down(lw->weight);
193 
194 	if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
195 		lw->inv_weight = 1;
196 	else if (unlikely(!w))
197 		lw->inv_weight = WMULT_CONST;
198 	else
199 		lw->inv_weight = WMULT_CONST / w;
200 }
201 
202 /*
203  * delta_exec * weight / lw.weight
204  *   OR
205  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206  *
207  * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208  * we're guaranteed shift stays positive because inv_weight is guaranteed to
209  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210  *
211  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212  * weight/lw.weight <= 1, and therefore our shift will also be positive.
213  */
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
215 {
216 	u64 fact = scale_load_down(weight);
217 	int shift = WMULT_SHIFT;
218 
219 	__update_inv_weight(lw);
220 
221 	if (unlikely(fact >> 32)) {
222 		while (fact >> 32) {
223 			fact >>= 1;
224 			shift--;
225 		}
226 	}
227 
228 	/* hint to use a 32x32->64 mul */
229 	fact = (u64)(u32)fact * lw->inv_weight;
230 
231 	while (fact >> 32) {
232 		fact >>= 1;
233 		shift--;
234 	}
235 
236 	return mul_u64_u32_shr(delta_exec, fact, shift);
237 }
238 
239 
240 const struct sched_class fair_sched_class;
241 
242 /**************************************************************
243  * CFS operations on generic schedulable entities:
244  */
245 
246 #ifdef CONFIG_FAIR_GROUP_SCHED
247 
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
250 {
251 	return cfs_rq->rq;
252 }
253 
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se)	(!se->my_q)
256 
257 static inline struct task_struct *task_of(struct sched_entity *se)
258 {
259 #ifdef CONFIG_SCHED_DEBUG
260 	WARN_ON_ONCE(!entity_is_task(se));
261 #endif
262 	return container_of(se, struct task_struct, se);
263 }
264 
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 		for (; se; se = se->parent)
268 
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
270 {
271 	return p->se.cfs_rq;
272 }
273 
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
276 {
277 	return se->cfs_rq;
278 }
279 
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
282 {
283 	return grp->my_q;
284 }
285 
286 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
287 				       int force_update);
288 
289 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 {
291 	if (!cfs_rq->on_list) {
292 		/*
293 		 * Ensure we either appear before our parent (if already
294 		 * enqueued) or force our parent to appear after us when it is
295 		 * enqueued.  The fact that we always enqueue bottom-up
296 		 * reduces this to two cases.
297 		 */
298 		if (cfs_rq->tg->parent &&
299 		    cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
300 			list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
301 				&rq_of(cfs_rq)->leaf_cfs_rq_list);
302 		} else {
303 			list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
304 				&rq_of(cfs_rq)->leaf_cfs_rq_list);
305 		}
306 
307 		cfs_rq->on_list = 1;
308 		/* We should have no load, but we need to update last_decay. */
309 		update_cfs_rq_blocked_load(cfs_rq, 0);
310 	}
311 }
312 
313 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
314 {
315 	if (cfs_rq->on_list) {
316 		list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
317 		cfs_rq->on_list = 0;
318 	}
319 }
320 
321 /* Iterate thr' all leaf cfs_rq's on a runqueue */
322 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
323 	list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 
325 /* Do the two (enqueued) entities belong to the same group ? */
326 static inline struct cfs_rq *
327 is_same_group(struct sched_entity *se, struct sched_entity *pse)
328 {
329 	if (se->cfs_rq == pse->cfs_rq)
330 		return se->cfs_rq;
331 
332 	return NULL;
333 }
334 
335 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 {
337 	return se->parent;
338 }
339 
340 static void
341 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
342 {
343 	int se_depth, pse_depth;
344 
345 	/*
346 	 * preemption test can be made between sibling entities who are in the
347 	 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
348 	 * both tasks until we find their ancestors who are siblings of common
349 	 * parent.
350 	 */
351 
352 	/* First walk up until both entities are at same depth */
353 	se_depth = (*se)->depth;
354 	pse_depth = (*pse)->depth;
355 
356 	while (se_depth > pse_depth) {
357 		se_depth--;
358 		*se = parent_entity(*se);
359 	}
360 
361 	while (pse_depth > se_depth) {
362 		pse_depth--;
363 		*pse = parent_entity(*pse);
364 	}
365 
366 	while (!is_same_group(*se, *pse)) {
367 		*se = parent_entity(*se);
368 		*pse = parent_entity(*pse);
369 	}
370 }
371 
372 #else	/* !CONFIG_FAIR_GROUP_SCHED */
373 
374 static inline struct task_struct *task_of(struct sched_entity *se)
375 {
376 	return container_of(se, struct task_struct, se);
377 }
378 
379 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
380 {
381 	return container_of(cfs_rq, struct rq, cfs);
382 }
383 
384 #define entity_is_task(se)	1
385 
386 #define for_each_sched_entity(se) \
387 		for (; se; se = NULL)
388 
389 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
390 {
391 	return &task_rq(p)->cfs;
392 }
393 
394 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
395 {
396 	struct task_struct *p = task_of(se);
397 	struct rq *rq = task_rq(p);
398 
399 	return &rq->cfs;
400 }
401 
402 /* runqueue "owned" by this group */
403 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
404 {
405 	return NULL;
406 }
407 
408 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
409 {
410 }
411 
412 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
413 {
414 }
415 
416 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
417 		for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 
419 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 {
421 	return NULL;
422 }
423 
424 static inline void
425 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
426 {
427 }
428 
429 #endif	/* CONFIG_FAIR_GROUP_SCHED */
430 
431 static __always_inline
432 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
433 
434 /**************************************************************
435  * Scheduling class tree data structure manipulation methods:
436  */
437 
438 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
439 {
440 	s64 delta = (s64)(vruntime - max_vruntime);
441 	if (delta > 0)
442 		max_vruntime = vruntime;
443 
444 	return max_vruntime;
445 }
446 
447 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
448 {
449 	s64 delta = (s64)(vruntime - min_vruntime);
450 	if (delta < 0)
451 		min_vruntime = vruntime;
452 
453 	return min_vruntime;
454 }
455 
456 static inline int entity_before(struct sched_entity *a,
457 				struct sched_entity *b)
458 {
459 	return (s64)(a->vruntime - b->vruntime) < 0;
460 }
461 
462 static void update_min_vruntime(struct cfs_rq *cfs_rq)
463 {
464 	u64 vruntime = cfs_rq->min_vruntime;
465 
466 	if (cfs_rq->curr)
467 		vruntime = cfs_rq->curr->vruntime;
468 
469 	if (cfs_rq->rb_leftmost) {
470 		struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
471 						   struct sched_entity,
472 						   run_node);
473 
474 		if (!cfs_rq->curr)
475 			vruntime = se->vruntime;
476 		else
477 			vruntime = min_vruntime(vruntime, se->vruntime);
478 	}
479 
480 	/* ensure we never gain time by being placed backwards. */
481 	cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
482 #ifndef CONFIG_64BIT
483 	smp_wmb();
484 	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
485 #endif
486 }
487 
488 /*
489  * Enqueue an entity into the rb-tree:
490  */
491 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
492 {
493 	struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
494 	struct rb_node *parent = NULL;
495 	struct sched_entity *entry;
496 	int leftmost = 1;
497 
498 	/*
499 	 * Find the right place in the rbtree:
500 	 */
501 	while (*link) {
502 		parent = *link;
503 		entry = rb_entry(parent, struct sched_entity, run_node);
504 		/*
505 		 * We dont care about collisions. Nodes with
506 		 * the same key stay together.
507 		 */
508 		if (entity_before(se, entry)) {
509 			link = &parent->rb_left;
510 		} else {
511 			link = &parent->rb_right;
512 			leftmost = 0;
513 		}
514 	}
515 
516 	/*
517 	 * Maintain a cache of leftmost tree entries (it is frequently
518 	 * used):
519 	 */
520 	if (leftmost)
521 		cfs_rq->rb_leftmost = &se->run_node;
522 
523 	rb_link_node(&se->run_node, parent, link);
524 	rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
525 }
526 
527 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
528 {
529 	if (cfs_rq->rb_leftmost == &se->run_node) {
530 		struct rb_node *next_node;
531 
532 		next_node = rb_next(&se->run_node);
533 		cfs_rq->rb_leftmost = next_node;
534 	}
535 
536 	rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
537 }
538 
539 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
540 {
541 	struct rb_node *left = cfs_rq->rb_leftmost;
542 
543 	if (!left)
544 		return NULL;
545 
546 	return rb_entry(left, struct sched_entity, run_node);
547 }
548 
549 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
550 {
551 	struct rb_node *next = rb_next(&se->run_node);
552 
553 	if (!next)
554 		return NULL;
555 
556 	return rb_entry(next, struct sched_entity, run_node);
557 }
558 
559 #ifdef CONFIG_SCHED_DEBUG
560 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
561 {
562 	struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
563 
564 	if (!last)
565 		return NULL;
566 
567 	return rb_entry(last, struct sched_entity, run_node);
568 }
569 
570 /**************************************************************
571  * Scheduling class statistics methods:
572  */
573 
574 int sched_proc_update_handler(struct ctl_table *table, int write,
575 		void __user *buffer, size_t *lenp,
576 		loff_t *ppos)
577 {
578 	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
579 	int factor = get_update_sysctl_factor();
580 
581 	if (ret || !write)
582 		return ret;
583 
584 	sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
585 					sysctl_sched_min_granularity);
586 
587 #define WRT_SYSCTL(name) \
588 	(normalized_sysctl_##name = sysctl_##name / (factor))
589 	WRT_SYSCTL(sched_min_granularity);
590 	WRT_SYSCTL(sched_latency);
591 	WRT_SYSCTL(sched_wakeup_granularity);
592 #undef WRT_SYSCTL
593 
594 	return 0;
595 }
596 #endif
597 
598 /*
599  * delta /= w
600  */
601 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
602 {
603 	if (unlikely(se->load.weight != NICE_0_LOAD))
604 		delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 
606 	return delta;
607 }
608 
609 /*
610  * The idea is to set a period in which each task runs once.
611  *
612  * When there are too many tasks (sched_nr_latency) we have to stretch
613  * this period because otherwise the slices get too small.
614  *
615  * p = (nr <= nl) ? l : l*nr/nl
616  */
617 static u64 __sched_period(unsigned long nr_running)
618 {
619 	u64 period = sysctl_sched_latency;
620 	unsigned long nr_latency = sched_nr_latency;
621 
622 	if (unlikely(nr_running > nr_latency)) {
623 		period = sysctl_sched_min_granularity;
624 		period *= nr_running;
625 	}
626 
627 	return period;
628 }
629 
630 /*
631  * We calculate the wall-time slice from the period by taking a part
632  * proportional to the weight.
633  *
634  * s = p*P[w/rw]
635  */
636 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
637 {
638 	u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
639 
640 	for_each_sched_entity(se) {
641 		struct load_weight *load;
642 		struct load_weight lw;
643 
644 		cfs_rq = cfs_rq_of(se);
645 		load = &cfs_rq->load;
646 
647 		if (unlikely(!se->on_rq)) {
648 			lw = cfs_rq->load;
649 
650 			update_load_add(&lw, se->load.weight);
651 			load = &lw;
652 		}
653 		slice = __calc_delta(slice, se->load.weight, load);
654 	}
655 	return slice;
656 }
657 
658 /*
659  * We calculate the vruntime slice of a to-be-inserted task.
660  *
661  * vs = s/w
662  */
663 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
664 {
665 	return calc_delta_fair(sched_slice(cfs_rq, se), se);
666 }
667 
668 #ifdef CONFIG_SMP
669 static int select_idle_sibling(struct task_struct *p, int cpu);
670 static unsigned long task_h_load(struct task_struct *p);
671 
672 static inline void __update_task_entity_contrib(struct sched_entity *se);
673 
674 /* Give new task start runnable values to heavy its load in infant time */
675 void init_task_runnable_average(struct task_struct *p)
676 {
677 	u32 slice;
678 
679 	p->se.avg.decay_count = 0;
680 	slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
681 	p->se.avg.runnable_avg_sum = slice;
682 	p->se.avg.runnable_avg_period = slice;
683 	__update_task_entity_contrib(&p->se);
684 }
685 #else
686 void init_task_runnable_average(struct task_struct *p)
687 {
688 }
689 #endif
690 
691 /*
692  * Update the current task's runtime statistics.
693  */
694 static void update_curr(struct cfs_rq *cfs_rq)
695 {
696 	struct sched_entity *curr = cfs_rq->curr;
697 	u64 now = rq_clock_task(rq_of(cfs_rq));
698 	u64 delta_exec;
699 
700 	if (unlikely(!curr))
701 		return;
702 
703 	delta_exec = now - curr->exec_start;
704 	if (unlikely((s64)delta_exec <= 0))
705 		return;
706 
707 	curr->exec_start = now;
708 
709 	schedstat_set(curr->statistics.exec_max,
710 		      max(delta_exec, curr->statistics.exec_max));
711 
712 	curr->sum_exec_runtime += delta_exec;
713 	schedstat_add(cfs_rq, exec_clock, delta_exec);
714 
715 	curr->vruntime += calc_delta_fair(delta_exec, curr);
716 	update_min_vruntime(cfs_rq);
717 
718 	if (entity_is_task(curr)) {
719 		struct task_struct *curtask = task_of(curr);
720 
721 		trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
722 		cpuacct_charge(curtask, delta_exec);
723 		account_group_exec_runtime(curtask, delta_exec);
724 	}
725 
726 	account_cfs_rq_runtime(cfs_rq, delta_exec);
727 }
728 
729 static void update_curr_fair(struct rq *rq)
730 {
731 	update_curr(cfs_rq_of(&rq->curr->se));
732 }
733 
734 static inline void
735 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
736 {
737 	schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
738 }
739 
740 /*
741  * Task is being enqueued - update stats:
742  */
743 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
744 {
745 	/*
746 	 * Are we enqueueing a waiting task? (for current tasks
747 	 * a dequeue/enqueue event is a NOP)
748 	 */
749 	if (se != cfs_rq->curr)
750 		update_stats_wait_start(cfs_rq, se);
751 }
752 
753 static void
754 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
755 {
756 	schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
757 			rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
758 	schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
759 	schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
760 			rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
761 #ifdef CONFIG_SCHEDSTATS
762 	if (entity_is_task(se)) {
763 		trace_sched_stat_wait(task_of(se),
764 			rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
765 	}
766 #endif
767 	schedstat_set(se->statistics.wait_start, 0);
768 }
769 
770 static inline void
771 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
772 {
773 	/*
774 	 * Mark the end of the wait period if dequeueing a
775 	 * waiting task:
776 	 */
777 	if (se != cfs_rq->curr)
778 		update_stats_wait_end(cfs_rq, se);
779 }
780 
781 /*
782  * We are picking a new current task - update its stats:
783  */
784 static inline void
785 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
786 {
787 	/*
788 	 * We are starting a new run period:
789 	 */
790 	se->exec_start = rq_clock_task(rq_of(cfs_rq));
791 }
792 
793 /**************************************************
794  * Scheduling class queueing methods:
795  */
796 
797 #ifdef CONFIG_NUMA_BALANCING
798 /*
799  * Approximate time to scan a full NUMA task in ms. The task scan period is
800  * calculated based on the tasks virtual memory size and
801  * numa_balancing_scan_size.
802  */
803 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
804 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
805 
806 /* Portion of address space to scan in MB */
807 unsigned int sysctl_numa_balancing_scan_size = 256;
808 
809 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
810 unsigned int sysctl_numa_balancing_scan_delay = 1000;
811 
812 static unsigned int task_nr_scan_windows(struct task_struct *p)
813 {
814 	unsigned long rss = 0;
815 	unsigned long nr_scan_pages;
816 
817 	/*
818 	 * Calculations based on RSS as non-present and empty pages are skipped
819 	 * by the PTE scanner and NUMA hinting faults should be trapped based
820 	 * on resident pages
821 	 */
822 	nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
823 	rss = get_mm_rss(p->mm);
824 	if (!rss)
825 		rss = nr_scan_pages;
826 
827 	rss = round_up(rss, nr_scan_pages);
828 	return rss / nr_scan_pages;
829 }
830 
831 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
832 #define MAX_SCAN_WINDOW 2560
833 
834 static unsigned int task_scan_min(struct task_struct *p)
835 {
836 	unsigned int scan_size = ACCESS_ONCE(sysctl_numa_balancing_scan_size);
837 	unsigned int scan, floor;
838 	unsigned int windows = 1;
839 
840 	if (scan_size < MAX_SCAN_WINDOW)
841 		windows = MAX_SCAN_WINDOW / scan_size;
842 	floor = 1000 / windows;
843 
844 	scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
845 	return max_t(unsigned int, floor, scan);
846 }
847 
848 static unsigned int task_scan_max(struct task_struct *p)
849 {
850 	unsigned int smin = task_scan_min(p);
851 	unsigned int smax;
852 
853 	/* Watch for min being lower than max due to floor calculations */
854 	smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
855 	return max(smin, smax);
856 }
857 
858 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
859 {
860 	rq->nr_numa_running += (p->numa_preferred_nid != -1);
861 	rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
862 }
863 
864 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
865 {
866 	rq->nr_numa_running -= (p->numa_preferred_nid != -1);
867 	rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
868 }
869 
870 struct numa_group {
871 	atomic_t refcount;
872 
873 	spinlock_t lock; /* nr_tasks, tasks */
874 	int nr_tasks;
875 	pid_t gid;
876 
877 	struct rcu_head rcu;
878 	nodemask_t active_nodes;
879 	unsigned long total_faults;
880 	/*
881 	 * Faults_cpu is used to decide whether memory should move
882 	 * towards the CPU. As a consequence, these stats are weighted
883 	 * more by CPU use than by memory faults.
884 	 */
885 	unsigned long *faults_cpu;
886 	unsigned long faults[0];
887 };
888 
889 /* Shared or private faults. */
890 #define NR_NUMA_HINT_FAULT_TYPES 2
891 
892 /* Memory and CPU locality */
893 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
894 
895 /* Averaged statistics, and temporary buffers. */
896 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
897 
898 pid_t task_numa_group_id(struct task_struct *p)
899 {
900 	return p->numa_group ? p->numa_group->gid : 0;
901 }
902 
903 /*
904  * The averaged statistics, shared & private, memory & cpu,
905  * occupy the first half of the array. The second half of the
906  * array is for current counters, which are averaged into the
907  * first set by task_numa_placement.
908  */
909 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
910 {
911 	return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
912 }
913 
914 static inline unsigned long task_faults(struct task_struct *p, int nid)
915 {
916 	if (!p->numa_faults)
917 		return 0;
918 
919 	return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
920 		p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
921 }
922 
923 static inline unsigned long group_faults(struct task_struct *p, int nid)
924 {
925 	if (!p->numa_group)
926 		return 0;
927 
928 	return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
929 		p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
930 }
931 
932 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
933 {
934 	return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
935 		group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
936 }
937 
938 /* Handle placement on systems where not all nodes are directly connected. */
939 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
940 					int maxdist, bool task)
941 {
942 	unsigned long score = 0;
943 	int node;
944 
945 	/*
946 	 * All nodes are directly connected, and the same distance
947 	 * from each other. No need for fancy placement algorithms.
948 	 */
949 	if (sched_numa_topology_type == NUMA_DIRECT)
950 		return 0;
951 
952 	/*
953 	 * This code is called for each node, introducing N^2 complexity,
954 	 * which should be ok given the number of nodes rarely exceeds 8.
955 	 */
956 	for_each_online_node(node) {
957 		unsigned long faults;
958 		int dist = node_distance(nid, node);
959 
960 		/*
961 		 * The furthest away nodes in the system are not interesting
962 		 * for placement; nid was already counted.
963 		 */
964 		if (dist == sched_max_numa_distance || node == nid)
965 			continue;
966 
967 		/*
968 		 * On systems with a backplane NUMA topology, compare groups
969 		 * of nodes, and move tasks towards the group with the most
970 		 * memory accesses. When comparing two nodes at distance
971 		 * "hoplimit", only nodes closer by than "hoplimit" are part
972 		 * of each group. Skip other nodes.
973 		 */
974 		if (sched_numa_topology_type == NUMA_BACKPLANE &&
975 					dist > maxdist)
976 			continue;
977 
978 		/* Add up the faults from nearby nodes. */
979 		if (task)
980 			faults = task_faults(p, node);
981 		else
982 			faults = group_faults(p, node);
983 
984 		/*
985 		 * On systems with a glueless mesh NUMA topology, there are
986 		 * no fixed "groups of nodes". Instead, nodes that are not
987 		 * directly connected bounce traffic through intermediate
988 		 * nodes; a numa_group can occupy any set of nodes.
989 		 * The further away a node is, the less the faults count.
990 		 * This seems to result in good task placement.
991 		 */
992 		if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
993 			faults *= (sched_max_numa_distance - dist);
994 			faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
995 		}
996 
997 		score += faults;
998 	}
999 
1000 	return score;
1001 }
1002 
1003 /*
1004  * These return the fraction of accesses done by a particular task, or
1005  * task group, on a particular numa node.  The group weight is given a
1006  * larger multiplier, in order to group tasks together that are almost
1007  * evenly spread out between numa nodes.
1008  */
1009 static inline unsigned long task_weight(struct task_struct *p, int nid,
1010 					int dist)
1011 {
1012 	unsigned long faults, total_faults;
1013 
1014 	if (!p->numa_faults)
1015 		return 0;
1016 
1017 	total_faults = p->total_numa_faults;
1018 
1019 	if (!total_faults)
1020 		return 0;
1021 
1022 	faults = task_faults(p, nid);
1023 	faults += score_nearby_nodes(p, nid, dist, true);
1024 
1025 	return 1000 * faults / total_faults;
1026 }
1027 
1028 static inline unsigned long group_weight(struct task_struct *p, int nid,
1029 					 int dist)
1030 {
1031 	unsigned long faults, total_faults;
1032 
1033 	if (!p->numa_group)
1034 		return 0;
1035 
1036 	total_faults = p->numa_group->total_faults;
1037 
1038 	if (!total_faults)
1039 		return 0;
1040 
1041 	faults = group_faults(p, nid);
1042 	faults += score_nearby_nodes(p, nid, dist, false);
1043 
1044 	return 1000 * faults / total_faults;
1045 }
1046 
1047 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1048 				int src_nid, int dst_cpu)
1049 {
1050 	struct numa_group *ng = p->numa_group;
1051 	int dst_nid = cpu_to_node(dst_cpu);
1052 	int last_cpupid, this_cpupid;
1053 
1054 	this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1055 
1056 	/*
1057 	 * Multi-stage node selection is used in conjunction with a periodic
1058 	 * migration fault to build a temporal task<->page relation. By using
1059 	 * a two-stage filter we remove short/unlikely relations.
1060 	 *
1061 	 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1062 	 * a task's usage of a particular page (n_p) per total usage of this
1063 	 * page (n_t) (in a given time-span) to a probability.
1064 	 *
1065 	 * Our periodic faults will sample this probability and getting the
1066 	 * same result twice in a row, given these samples are fully
1067 	 * independent, is then given by P(n)^2, provided our sample period
1068 	 * is sufficiently short compared to the usage pattern.
1069 	 *
1070 	 * This quadric squishes small probabilities, making it less likely we
1071 	 * act on an unlikely task<->page relation.
1072 	 */
1073 	last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1074 	if (!cpupid_pid_unset(last_cpupid) &&
1075 				cpupid_to_nid(last_cpupid) != dst_nid)
1076 		return false;
1077 
1078 	/* Always allow migrate on private faults */
1079 	if (cpupid_match_pid(p, last_cpupid))
1080 		return true;
1081 
1082 	/* A shared fault, but p->numa_group has not been set up yet. */
1083 	if (!ng)
1084 		return true;
1085 
1086 	/*
1087 	 * Do not migrate if the destination is not a node that
1088 	 * is actively used by this numa group.
1089 	 */
1090 	if (!node_isset(dst_nid, ng->active_nodes))
1091 		return false;
1092 
1093 	/*
1094 	 * Source is a node that is not actively used by this
1095 	 * numa group, while the destination is. Migrate.
1096 	 */
1097 	if (!node_isset(src_nid, ng->active_nodes))
1098 		return true;
1099 
1100 	/*
1101 	 * Both source and destination are nodes in active
1102 	 * use by this numa group. Maximize memory bandwidth
1103 	 * by migrating from more heavily used groups, to less
1104 	 * heavily used ones, spreading the load around.
1105 	 * Use a 1/4 hysteresis to avoid spurious page movement.
1106 	 */
1107 	return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1108 }
1109 
1110 static unsigned long weighted_cpuload(const int cpu);
1111 static unsigned long source_load(int cpu, int type);
1112 static unsigned long target_load(int cpu, int type);
1113 static unsigned long capacity_of(int cpu);
1114 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1115 
1116 /* Cached statistics for all CPUs within a node */
1117 struct numa_stats {
1118 	unsigned long nr_running;
1119 	unsigned long load;
1120 
1121 	/* Total compute capacity of CPUs on a node */
1122 	unsigned long compute_capacity;
1123 
1124 	/* Approximate capacity in terms of runnable tasks on a node */
1125 	unsigned long task_capacity;
1126 	int has_free_capacity;
1127 };
1128 
1129 /*
1130  * XXX borrowed from update_sg_lb_stats
1131  */
1132 static void update_numa_stats(struct numa_stats *ns, int nid)
1133 {
1134 	int smt, cpu, cpus = 0;
1135 	unsigned long capacity;
1136 
1137 	memset(ns, 0, sizeof(*ns));
1138 	for_each_cpu(cpu, cpumask_of_node(nid)) {
1139 		struct rq *rq = cpu_rq(cpu);
1140 
1141 		ns->nr_running += rq->nr_running;
1142 		ns->load += weighted_cpuload(cpu);
1143 		ns->compute_capacity += capacity_of(cpu);
1144 
1145 		cpus++;
1146 	}
1147 
1148 	/*
1149 	 * If we raced with hotplug and there are no CPUs left in our mask
1150 	 * the @ns structure is NULL'ed and task_numa_compare() will
1151 	 * not find this node attractive.
1152 	 *
1153 	 * We'll either bail at !has_free_capacity, or we'll detect a huge
1154 	 * imbalance and bail there.
1155 	 */
1156 	if (!cpus)
1157 		return;
1158 
1159 	/* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1160 	smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1161 	capacity = cpus / smt; /* cores */
1162 
1163 	ns->task_capacity = min_t(unsigned, capacity,
1164 		DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1165 	ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1166 }
1167 
1168 struct task_numa_env {
1169 	struct task_struct *p;
1170 
1171 	int src_cpu, src_nid;
1172 	int dst_cpu, dst_nid;
1173 
1174 	struct numa_stats src_stats, dst_stats;
1175 
1176 	int imbalance_pct;
1177 	int dist;
1178 
1179 	struct task_struct *best_task;
1180 	long best_imp;
1181 	int best_cpu;
1182 };
1183 
1184 static void task_numa_assign(struct task_numa_env *env,
1185 			     struct task_struct *p, long imp)
1186 {
1187 	if (env->best_task)
1188 		put_task_struct(env->best_task);
1189 	if (p)
1190 		get_task_struct(p);
1191 
1192 	env->best_task = p;
1193 	env->best_imp = imp;
1194 	env->best_cpu = env->dst_cpu;
1195 }
1196 
1197 static bool load_too_imbalanced(long src_load, long dst_load,
1198 				struct task_numa_env *env)
1199 {
1200 	long imb, old_imb;
1201 	long orig_src_load, orig_dst_load;
1202 	long src_capacity, dst_capacity;
1203 
1204 	/*
1205 	 * The load is corrected for the CPU capacity available on each node.
1206 	 *
1207 	 * src_load        dst_load
1208 	 * ------------ vs ---------
1209 	 * src_capacity    dst_capacity
1210 	 */
1211 	src_capacity = env->src_stats.compute_capacity;
1212 	dst_capacity = env->dst_stats.compute_capacity;
1213 
1214 	/* We care about the slope of the imbalance, not the direction. */
1215 	if (dst_load < src_load)
1216 		swap(dst_load, src_load);
1217 
1218 	/* Is the difference below the threshold? */
1219 	imb = dst_load * src_capacity * 100 -
1220 	      src_load * dst_capacity * env->imbalance_pct;
1221 	if (imb <= 0)
1222 		return false;
1223 
1224 	/*
1225 	 * The imbalance is above the allowed threshold.
1226 	 * Compare it with the old imbalance.
1227 	 */
1228 	orig_src_load = env->src_stats.load;
1229 	orig_dst_load = env->dst_stats.load;
1230 
1231 	if (orig_dst_load < orig_src_load)
1232 		swap(orig_dst_load, orig_src_load);
1233 
1234 	old_imb = orig_dst_load * src_capacity * 100 -
1235 		  orig_src_load * dst_capacity * env->imbalance_pct;
1236 
1237 	/* Would this change make things worse? */
1238 	return (imb > old_imb);
1239 }
1240 
1241 /*
1242  * This checks if the overall compute and NUMA accesses of the system would
1243  * be improved if the source tasks was migrated to the target dst_cpu taking
1244  * into account that it might be best if task running on the dst_cpu should
1245  * be exchanged with the source task
1246  */
1247 static void task_numa_compare(struct task_numa_env *env,
1248 			      long taskimp, long groupimp)
1249 {
1250 	struct rq *src_rq = cpu_rq(env->src_cpu);
1251 	struct rq *dst_rq = cpu_rq(env->dst_cpu);
1252 	struct task_struct *cur;
1253 	long src_load, dst_load;
1254 	long load;
1255 	long imp = env->p->numa_group ? groupimp : taskimp;
1256 	long moveimp = imp;
1257 	int dist = env->dist;
1258 
1259 	rcu_read_lock();
1260 
1261 	raw_spin_lock_irq(&dst_rq->lock);
1262 	cur = dst_rq->curr;
1263 	/*
1264 	 * No need to move the exiting task, and this ensures that ->curr
1265 	 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1266 	 * is safe under RCU read lock.
1267 	 * Note that rcu_read_lock() itself can't protect from the final
1268 	 * put_task_struct() after the last schedule().
1269 	 */
1270 	if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1271 		cur = NULL;
1272 	raw_spin_unlock_irq(&dst_rq->lock);
1273 
1274 	/*
1275 	 * Because we have preemption enabled we can get migrated around and
1276 	 * end try selecting ourselves (current == env->p) as a swap candidate.
1277 	 */
1278 	if (cur == env->p)
1279 		goto unlock;
1280 
1281 	/*
1282 	 * "imp" is the fault differential for the source task between the
1283 	 * source and destination node. Calculate the total differential for
1284 	 * the source task and potential destination task. The more negative
1285 	 * the value is, the more rmeote accesses that would be expected to
1286 	 * be incurred if the tasks were swapped.
1287 	 */
1288 	if (cur) {
1289 		/* Skip this swap candidate if cannot move to the source cpu */
1290 		if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1291 			goto unlock;
1292 
1293 		/*
1294 		 * If dst and source tasks are in the same NUMA group, or not
1295 		 * in any group then look only at task weights.
1296 		 */
1297 		if (cur->numa_group == env->p->numa_group) {
1298 			imp = taskimp + task_weight(cur, env->src_nid, dist) -
1299 			      task_weight(cur, env->dst_nid, dist);
1300 			/*
1301 			 * Add some hysteresis to prevent swapping the
1302 			 * tasks within a group over tiny differences.
1303 			 */
1304 			if (cur->numa_group)
1305 				imp -= imp/16;
1306 		} else {
1307 			/*
1308 			 * Compare the group weights. If a task is all by
1309 			 * itself (not part of a group), use the task weight
1310 			 * instead.
1311 			 */
1312 			if (cur->numa_group)
1313 				imp += group_weight(cur, env->src_nid, dist) -
1314 				       group_weight(cur, env->dst_nid, dist);
1315 			else
1316 				imp += task_weight(cur, env->src_nid, dist) -
1317 				       task_weight(cur, env->dst_nid, dist);
1318 		}
1319 	}
1320 
1321 	if (imp <= env->best_imp && moveimp <= env->best_imp)
1322 		goto unlock;
1323 
1324 	if (!cur) {
1325 		/* Is there capacity at our destination? */
1326 		if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1327 		    !env->dst_stats.has_free_capacity)
1328 			goto unlock;
1329 
1330 		goto balance;
1331 	}
1332 
1333 	/* Balance doesn't matter much if we're running a task per cpu */
1334 	if (imp > env->best_imp && src_rq->nr_running == 1 &&
1335 			dst_rq->nr_running == 1)
1336 		goto assign;
1337 
1338 	/*
1339 	 * In the overloaded case, try and keep the load balanced.
1340 	 */
1341 balance:
1342 	load = task_h_load(env->p);
1343 	dst_load = env->dst_stats.load + load;
1344 	src_load = env->src_stats.load - load;
1345 
1346 	if (moveimp > imp && moveimp > env->best_imp) {
1347 		/*
1348 		 * If the improvement from just moving env->p direction is
1349 		 * better than swapping tasks around, check if a move is
1350 		 * possible. Store a slightly smaller score than moveimp,
1351 		 * so an actually idle CPU will win.
1352 		 */
1353 		if (!load_too_imbalanced(src_load, dst_load, env)) {
1354 			imp = moveimp - 1;
1355 			cur = NULL;
1356 			goto assign;
1357 		}
1358 	}
1359 
1360 	if (imp <= env->best_imp)
1361 		goto unlock;
1362 
1363 	if (cur) {
1364 		load = task_h_load(cur);
1365 		dst_load -= load;
1366 		src_load += load;
1367 	}
1368 
1369 	if (load_too_imbalanced(src_load, dst_load, env))
1370 		goto unlock;
1371 
1372 	/*
1373 	 * One idle CPU per node is evaluated for a task numa move.
1374 	 * Call select_idle_sibling to maybe find a better one.
1375 	 */
1376 	if (!cur)
1377 		env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1378 
1379 assign:
1380 	task_numa_assign(env, cur, imp);
1381 unlock:
1382 	rcu_read_unlock();
1383 }
1384 
1385 static void task_numa_find_cpu(struct task_numa_env *env,
1386 				long taskimp, long groupimp)
1387 {
1388 	int cpu;
1389 
1390 	for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1391 		/* Skip this CPU if the source task cannot migrate */
1392 		if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1393 			continue;
1394 
1395 		env->dst_cpu = cpu;
1396 		task_numa_compare(env, taskimp, groupimp);
1397 	}
1398 }
1399 
1400 static int task_numa_migrate(struct task_struct *p)
1401 {
1402 	struct task_numa_env env = {
1403 		.p = p,
1404 
1405 		.src_cpu = task_cpu(p),
1406 		.src_nid = task_node(p),
1407 
1408 		.imbalance_pct = 112,
1409 
1410 		.best_task = NULL,
1411 		.best_imp = 0,
1412 		.best_cpu = -1
1413 	};
1414 	struct sched_domain *sd;
1415 	unsigned long taskweight, groupweight;
1416 	int nid, ret, dist;
1417 	long taskimp, groupimp;
1418 
1419 	/*
1420 	 * Pick the lowest SD_NUMA domain, as that would have the smallest
1421 	 * imbalance and would be the first to start moving tasks about.
1422 	 *
1423 	 * And we want to avoid any moving of tasks about, as that would create
1424 	 * random movement of tasks -- counter the numa conditions we're trying
1425 	 * to satisfy here.
1426 	 */
1427 	rcu_read_lock();
1428 	sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1429 	if (sd)
1430 		env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1431 	rcu_read_unlock();
1432 
1433 	/*
1434 	 * Cpusets can break the scheduler domain tree into smaller
1435 	 * balance domains, some of which do not cross NUMA boundaries.
1436 	 * Tasks that are "trapped" in such domains cannot be migrated
1437 	 * elsewhere, so there is no point in (re)trying.
1438 	 */
1439 	if (unlikely(!sd)) {
1440 		p->numa_preferred_nid = task_node(p);
1441 		return -EINVAL;
1442 	}
1443 
1444 	env.dst_nid = p->numa_preferred_nid;
1445 	dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1446 	taskweight = task_weight(p, env.src_nid, dist);
1447 	groupweight = group_weight(p, env.src_nid, dist);
1448 	update_numa_stats(&env.src_stats, env.src_nid);
1449 	taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1450 	groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1451 	update_numa_stats(&env.dst_stats, env.dst_nid);
1452 
1453 	/* Try to find a spot on the preferred nid. */
1454 	task_numa_find_cpu(&env, taskimp, groupimp);
1455 
1456 	/*
1457 	 * Look at other nodes in these cases:
1458 	 * - there is no space available on the preferred_nid
1459 	 * - the task is part of a numa_group that is interleaved across
1460 	 *   multiple NUMA nodes; in order to better consolidate the group,
1461 	 *   we need to check other locations.
1462 	 */
1463 	if (env.best_cpu == -1 || (p->numa_group &&
1464 			nodes_weight(p->numa_group->active_nodes) > 1)) {
1465 		for_each_online_node(nid) {
1466 			if (nid == env.src_nid || nid == p->numa_preferred_nid)
1467 				continue;
1468 
1469 			dist = node_distance(env.src_nid, env.dst_nid);
1470 			if (sched_numa_topology_type == NUMA_BACKPLANE &&
1471 						dist != env.dist) {
1472 				taskweight = task_weight(p, env.src_nid, dist);
1473 				groupweight = group_weight(p, env.src_nid, dist);
1474 			}
1475 
1476 			/* Only consider nodes where both task and groups benefit */
1477 			taskimp = task_weight(p, nid, dist) - taskweight;
1478 			groupimp = group_weight(p, nid, dist) - groupweight;
1479 			if (taskimp < 0 && groupimp < 0)
1480 				continue;
1481 
1482 			env.dist = dist;
1483 			env.dst_nid = nid;
1484 			update_numa_stats(&env.dst_stats, env.dst_nid);
1485 			task_numa_find_cpu(&env, taskimp, groupimp);
1486 		}
1487 	}
1488 
1489 	/*
1490 	 * If the task is part of a workload that spans multiple NUMA nodes,
1491 	 * and is migrating into one of the workload's active nodes, remember
1492 	 * this node as the task's preferred numa node, so the workload can
1493 	 * settle down.
1494 	 * A task that migrated to a second choice node will be better off
1495 	 * trying for a better one later. Do not set the preferred node here.
1496 	 */
1497 	if (p->numa_group) {
1498 		if (env.best_cpu == -1)
1499 			nid = env.src_nid;
1500 		else
1501 			nid = env.dst_nid;
1502 
1503 		if (node_isset(nid, p->numa_group->active_nodes))
1504 			sched_setnuma(p, env.dst_nid);
1505 	}
1506 
1507 	/* No better CPU than the current one was found. */
1508 	if (env.best_cpu == -1)
1509 		return -EAGAIN;
1510 
1511 	/*
1512 	 * Reset the scan period if the task is being rescheduled on an
1513 	 * alternative node to recheck if the tasks is now properly placed.
1514 	 */
1515 	p->numa_scan_period = task_scan_min(p);
1516 
1517 	if (env.best_task == NULL) {
1518 		ret = migrate_task_to(p, env.best_cpu);
1519 		if (ret != 0)
1520 			trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1521 		return ret;
1522 	}
1523 
1524 	ret = migrate_swap(p, env.best_task);
1525 	if (ret != 0)
1526 		trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1527 	put_task_struct(env.best_task);
1528 	return ret;
1529 }
1530 
1531 /* Attempt to migrate a task to a CPU on the preferred node. */
1532 static void numa_migrate_preferred(struct task_struct *p)
1533 {
1534 	unsigned long interval = HZ;
1535 
1536 	/* This task has no NUMA fault statistics yet */
1537 	if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1538 		return;
1539 
1540 	/* Periodically retry migrating the task to the preferred node */
1541 	interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1542 	p->numa_migrate_retry = jiffies + interval;
1543 
1544 	/* Success if task is already running on preferred CPU */
1545 	if (task_node(p) == p->numa_preferred_nid)
1546 		return;
1547 
1548 	/* Otherwise, try migrate to a CPU on the preferred node */
1549 	task_numa_migrate(p);
1550 }
1551 
1552 /*
1553  * Find the nodes on which the workload is actively running. We do this by
1554  * tracking the nodes from which NUMA hinting faults are triggered. This can
1555  * be different from the set of nodes where the workload's memory is currently
1556  * located.
1557  *
1558  * The bitmask is used to make smarter decisions on when to do NUMA page
1559  * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1560  * are added when they cause over 6/16 of the maximum number of faults, but
1561  * only removed when they drop below 3/16.
1562  */
1563 static void update_numa_active_node_mask(struct numa_group *numa_group)
1564 {
1565 	unsigned long faults, max_faults = 0;
1566 	int nid;
1567 
1568 	for_each_online_node(nid) {
1569 		faults = group_faults_cpu(numa_group, nid);
1570 		if (faults > max_faults)
1571 			max_faults = faults;
1572 	}
1573 
1574 	for_each_online_node(nid) {
1575 		faults = group_faults_cpu(numa_group, nid);
1576 		if (!node_isset(nid, numa_group->active_nodes)) {
1577 			if (faults > max_faults * 6 / 16)
1578 				node_set(nid, numa_group->active_nodes);
1579 		} else if (faults < max_faults * 3 / 16)
1580 			node_clear(nid, numa_group->active_nodes);
1581 	}
1582 }
1583 
1584 /*
1585  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1586  * increments. The more local the fault statistics are, the higher the scan
1587  * period will be for the next scan window. If local/(local+remote) ratio is
1588  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1589  * the scan period will decrease. Aim for 70% local accesses.
1590  */
1591 #define NUMA_PERIOD_SLOTS 10
1592 #define NUMA_PERIOD_THRESHOLD 7
1593 
1594 /*
1595  * Increase the scan period (slow down scanning) if the majority of
1596  * our memory is already on our local node, or if the majority of
1597  * the page accesses are shared with other processes.
1598  * Otherwise, decrease the scan period.
1599  */
1600 static void update_task_scan_period(struct task_struct *p,
1601 			unsigned long shared, unsigned long private)
1602 {
1603 	unsigned int period_slot;
1604 	int ratio;
1605 	int diff;
1606 
1607 	unsigned long remote = p->numa_faults_locality[0];
1608 	unsigned long local = p->numa_faults_locality[1];
1609 
1610 	/*
1611 	 * If there were no record hinting faults then either the task is
1612 	 * completely idle or all activity is areas that are not of interest
1613 	 * to automatic numa balancing. Scan slower
1614 	 */
1615 	if (local + shared == 0) {
1616 		p->numa_scan_period = min(p->numa_scan_period_max,
1617 			p->numa_scan_period << 1);
1618 
1619 		p->mm->numa_next_scan = jiffies +
1620 			msecs_to_jiffies(p->numa_scan_period);
1621 
1622 		return;
1623 	}
1624 
1625 	/*
1626 	 * Prepare to scale scan period relative to the current period.
1627 	 *	 == NUMA_PERIOD_THRESHOLD scan period stays the same
1628 	 *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1629 	 *	 >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1630 	 */
1631 	period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1632 	ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1633 	if (ratio >= NUMA_PERIOD_THRESHOLD) {
1634 		int slot = ratio - NUMA_PERIOD_THRESHOLD;
1635 		if (!slot)
1636 			slot = 1;
1637 		diff = slot * period_slot;
1638 	} else {
1639 		diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1640 
1641 		/*
1642 		 * Scale scan rate increases based on sharing. There is an
1643 		 * inverse relationship between the degree of sharing and
1644 		 * the adjustment made to the scanning period. Broadly
1645 		 * speaking the intent is that there is little point
1646 		 * scanning faster if shared accesses dominate as it may
1647 		 * simply bounce migrations uselessly
1648 		 */
1649 		ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1650 		diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1651 	}
1652 
1653 	p->numa_scan_period = clamp(p->numa_scan_period + diff,
1654 			task_scan_min(p), task_scan_max(p));
1655 	memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1656 }
1657 
1658 /*
1659  * Get the fraction of time the task has been running since the last
1660  * NUMA placement cycle. The scheduler keeps similar statistics, but
1661  * decays those on a 32ms period, which is orders of magnitude off
1662  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1663  * stats only if the task is so new there are no NUMA statistics yet.
1664  */
1665 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1666 {
1667 	u64 runtime, delta, now;
1668 	/* Use the start of this time slice to avoid calculations. */
1669 	now = p->se.exec_start;
1670 	runtime = p->se.sum_exec_runtime;
1671 
1672 	if (p->last_task_numa_placement) {
1673 		delta = runtime - p->last_sum_exec_runtime;
1674 		*period = now - p->last_task_numa_placement;
1675 	} else {
1676 		delta = p->se.avg.runnable_avg_sum;
1677 		*period = p->se.avg.runnable_avg_period;
1678 	}
1679 
1680 	p->last_sum_exec_runtime = runtime;
1681 	p->last_task_numa_placement = now;
1682 
1683 	return delta;
1684 }
1685 
1686 /*
1687  * Determine the preferred nid for a task in a numa_group. This needs to
1688  * be done in a way that produces consistent results with group_weight,
1689  * otherwise workloads might not converge.
1690  */
1691 static int preferred_group_nid(struct task_struct *p, int nid)
1692 {
1693 	nodemask_t nodes;
1694 	int dist;
1695 
1696 	/* Direct connections between all NUMA nodes. */
1697 	if (sched_numa_topology_type == NUMA_DIRECT)
1698 		return nid;
1699 
1700 	/*
1701 	 * On a system with glueless mesh NUMA topology, group_weight
1702 	 * scores nodes according to the number of NUMA hinting faults on
1703 	 * both the node itself, and on nearby nodes.
1704 	 */
1705 	if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1706 		unsigned long score, max_score = 0;
1707 		int node, max_node = nid;
1708 
1709 		dist = sched_max_numa_distance;
1710 
1711 		for_each_online_node(node) {
1712 			score = group_weight(p, node, dist);
1713 			if (score > max_score) {
1714 				max_score = score;
1715 				max_node = node;
1716 			}
1717 		}
1718 		return max_node;
1719 	}
1720 
1721 	/*
1722 	 * Finding the preferred nid in a system with NUMA backplane
1723 	 * interconnect topology is more involved. The goal is to locate
1724 	 * tasks from numa_groups near each other in the system, and
1725 	 * untangle workloads from different sides of the system. This requires
1726 	 * searching down the hierarchy of node groups, recursively searching
1727 	 * inside the highest scoring group of nodes. The nodemask tricks
1728 	 * keep the complexity of the search down.
1729 	 */
1730 	nodes = node_online_map;
1731 	for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1732 		unsigned long max_faults = 0;
1733 		nodemask_t max_group;
1734 		int a, b;
1735 
1736 		/* Are there nodes at this distance from each other? */
1737 		if (!find_numa_distance(dist))
1738 			continue;
1739 
1740 		for_each_node_mask(a, nodes) {
1741 			unsigned long faults = 0;
1742 			nodemask_t this_group;
1743 			nodes_clear(this_group);
1744 
1745 			/* Sum group's NUMA faults; includes a==b case. */
1746 			for_each_node_mask(b, nodes) {
1747 				if (node_distance(a, b) < dist) {
1748 					faults += group_faults(p, b);
1749 					node_set(b, this_group);
1750 					node_clear(b, nodes);
1751 				}
1752 			}
1753 
1754 			/* Remember the top group. */
1755 			if (faults > max_faults) {
1756 				max_faults = faults;
1757 				max_group = this_group;
1758 				/*
1759 				 * subtle: at the smallest distance there is
1760 				 * just one node left in each "group", the
1761 				 * winner is the preferred nid.
1762 				 */
1763 				nid = a;
1764 			}
1765 		}
1766 		/* Next round, evaluate the nodes within max_group. */
1767 		nodes = max_group;
1768 	}
1769 	return nid;
1770 }
1771 
1772 static void task_numa_placement(struct task_struct *p)
1773 {
1774 	int seq, nid, max_nid = -1, max_group_nid = -1;
1775 	unsigned long max_faults = 0, max_group_faults = 0;
1776 	unsigned long fault_types[2] = { 0, 0 };
1777 	unsigned long total_faults;
1778 	u64 runtime, period;
1779 	spinlock_t *group_lock = NULL;
1780 
1781 	seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1782 	if (p->numa_scan_seq == seq)
1783 		return;
1784 	p->numa_scan_seq = seq;
1785 	p->numa_scan_period_max = task_scan_max(p);
1786 
1787 	total_faults = p->numa_faults_locality[0] +
1788 		       p->numa_faults_locality[1];
1789 	runtime = numa_get_avg_runtime(p, &period);
1790 
1791 	/* If the task is part of a group prevent parallel updates to group stats */
1792 	if (p->numa_group) {
1793 		group_lock = &p->numa_group->lock;
1794 		spin_lock_irq(group_lock);
1795 	}
1796 
1797 	/* Find the node with the highest number of faults */
1798 	for_each_online_node(nid) {
1799 		/* Keep track of the offsets in numa_faults array */
1800 		int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1801 		unsigned long faults = 0, group_faults = 0;
1802 		int priv;
1803 
1804 		for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1805 			long diff, f_diff, f_weight;
1806 
1807 			mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1808 			membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1809 			cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1810 			cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1811 
1812 			/* Decay existing window, copy faults since last scan */
1813 			diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1814 			fault_types[priv] += p->numa_faults[membuf_idx];
1815 			p->numa_faults[membuf_idx] = 0;
1816 
1817 			/*
1818 			 * Normalize the faults_from, so all tasks in a group
1819 			 * count according to CPU use, instead of by the raw
1820 			 * number of faults. Tasks with little runtime have
1821 			 * little over-all impact on throughput, and thus their
1822 			 * faults are less important.
1823 			 */
1824 			f_weight = div64_u64(runtime << 16, period + 1);
1825 			f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1826 				   (total_faults + 1);
1827 			f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1828 			p->numa_faults[cpubuf_idx] = 0;
1829 
1830 			p->numa_faults[mem_idx] += diff;
1831 			p->numa_faults[cpu_idx] += f_diff;
1832 			faults += p->numa_faults[mem_idx];
1833 			p->total_numa_faults += diff;
1834 			if (p->numa_group) {
1835 				/*
1836 				 * safe because we can only change our own group
1837 				 *
1838 				 * mem_idx represents the offset for a given
1839 				 * nid and priv in a specific region because it
1840 				 * is at the beginning of the numa_faults array.
1841 				 */
1842 				p->numa_group->faults[mem_idx] += diff;
1843 				p->numa_group->faults_cpu[mem_idx] += f_diff;
1844 				p->numa_group->total_faults += diff;
1845 				group_faults += p->numa_group->faults[mem_idx];
1846 			}
1847 		}
1848 
1849 		if (faults > max_faults) {
1850 			max_faults = faults;
1851 			max_nid = nid;
1852 		}
1853 
1854 		if (group_faults > max_group_faults) {
1855 			max_group_faults = group_faults;
1856 			max_group_nid = nid;
1857 		}
1858 	}
1859 
1860 	update_task_scan_period(p, fault_types[0], fault_types[1]);
1861 
1862 	if (p->numa_group) {
1863 		update_numa_active_node_mask(p->numa_group);
1864 		spin_unlock_irq(group_lock);
1865 		max_nid = preferred_group_nid(p, max_group_nid);
1866 	}
1867 
1868 	if (max_faults) {
1869 		/* Set the new preferred node */
1870 		if (max_nid != p->numa_preferred_nid)
1871 			sched_setnuma(p, max_nid);
1872 
1873 		if (task_node(p) != p->numa_preferred_nid)
1874 			numa_migrate_preferred(p);
1875 	}
1876 }
1877 
1878 static inline int get_numa_group(struct numa_group *grp)
1879 {
1880 	return atomic_inc_not_zero(&grp->refcount);
1881 }
1882 
1883 static inline void put_numa_group(struct numa_group *grp)
1884 {
1885 	if (atomic_dec_and_test(&grp->refcount))
1886 		kfree_rcu(grp, rcu);
1887 }
1888 
1889 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1890 			int *priv)
1891 {
1892 	struct numa_group *grp, *my_grp;
1893 	struct task_struct *tsk;
1894 	bool join = false;
1895 	int cpu = cpupid_to_cpu(cpupid);
1896 	int i;
1897 
1898 	if (unlikely(!p->numa_group)) {
1899 		unsigned int size = sizeof(struct numa_group) +
1900 				    4*nr_node_ids*sizeof(unsigned long);
1901 
1902 		grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1903 		if (!grp)
1904 			return;
1905 
1906 		atomic_set(&grp->refcount, 1);
1907 		spin_lock_init(&grp->lock);
1908 		grp->gid = p->pid;
1909 		/* Second half of the array tracks nids where faults happen */
1910 		grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1911 						nr_node_ids;
1912 
1913 		node_set(task_node(current), grp->active_nodes);
1914 
1915 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1916 			grp->faults[i] = p->numa_faults[i];
1917 
1918 		grp->total_faults = p->total_numa_faults;
1919 
1920 		grp->nr_tasks++;
1921 		rcu_assign_pointer(p->numa_group, grp);
1922 	}
1923 
1924 	rcu_read_lock();
1925 	tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1926 
1927 	if (!cpupid_match_pid(tsk, cpupid))
1928 		goto no_join;
1929 
1930 	grp = rcu_dereference(tsk->numa_group);
1931 	if (!grp)
1932 		goto no_join;
1933 
1934 	my_grp = p->numa_group;
1935 	if (grp == my_grp)
1936 		goto no_join;
1937 
1938 	/*
1939 	 * Only join the other group if its bigger; if we're the bigger group,
1940 	 * the other task will join us.
1941 	 */
1942 	if (my_grp->nr_tasks > grp->nr_tasks)
1943 		goto no_join;
1944 
1945 	/*
1946 	 * Tie-break on the grp address.
1947 	 */
1948 	if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1949 		goto no_join;
1950 
1951 	/* Always join threads in the same process. */
1952 	if (tsk->mm == current->mm)
1953 		join = true;
1954 
1955 	/* Simple filter to avoid false positives due to PID collisions */
1956 	if (flags & TNF_SHARED)
1957 		join = true;
1958 
1959 	/* Update priv based on whether false sharing was detected */
1960 	*priv = !join;
1961 
1962 	if (join && !get_numa_group(grp))
1963 		goto no_join;
1964 
1965 	rcu_read_unlock();
1966 
1967 	if (!join)
1968 		return;
1969 
1970 	BUG_ON(irqs_disabled());
1971 	double_lock_irq(&my_grp->lock, &grp->lock);
1972 
1973 	for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1974 		my_grp->faults[i] -= p->numa_faults[i];
1975 		grp->faults[i] += p->numa_faults[i];
1976 	}
1977 	my_grp->total_faults -= p->total_numa_faults;
1978 	grp->total_faults += p->total_numa_faults;
1979 
1980 	my_grp->nr_tasks--;
1981 	grp->nr_tasks++;
1982 
1983 	spin_unlock(&my_grp->lock);
1984 	spin_unlock_irq(&grp->lock);
1985 
1986 	rcu_assign_pointer(p->numa_group, grp);
1987 
1988 	put_numa_group(my_grp);
1989 	return;
1990 
1991 no_join:
1992 	rcu_read_unlock();
1993 	return;
1994 }
1995 
1996 void task_numa_free(struct task_struct *p)
1997 {
1998 	struct numa_group *grp = p->numa_group;
1999 	void *numa_faults = p->numa_faults;
2000 	unsigned long flags;
2001 	int i;
2002 
2003 	if (grp) {
2004 		spin_lock_irqsave(&grp->lock, flags);
2005 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2006 			grp->faults[i] -= p->numa_faults[i];
2007 		grp->total_faults -= p->total_numa_faults;
2008 
2009 		grp->nr_tasks--;
2010 		spin_unlock_irqrestore(&grp->lock, flags);
2011 		RCU_INIT_POINTER(p->numa_group, NULL);
2012 		put_numa_group(grp);
2013 	}
2014 
2015 	p->numa_faults = NULL;
2016 	kfree(numa_faults);
2017 }
2018 
2019 /*
2020  * Got a PROT_NONE fault for a page on @node.
2021  */
2022 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2023 {
2024 	struct task_struct *p = current;
2025 	bool migrated = flags & TNF_MIGRATED;
2026 	int cpu_node = task_node(current);
2027 	int local = !!(flags & TNF_FAULT_LOCAL);
2028 	int priv;
2029 
2030 	if (!numabalancing_enabled)
2031 		return;
2032 
2033 	/* for example, ksmd faulting in a user's mm */
2034 	if (!p->mm)
2035 		return;
2036 
2037 	/* Allocate buffer to track faults on a per-node basis */
2038 	if (unlikely(!p->numa_faults)) {
2039 		int size = sizeof(*p->numa_faults) *
2040 			   NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2041 
2042 		p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2043 		if (!p->numa_faults)
2044 			return;
2045 
2046 		p->total_numa_faults = 0;
2047 		memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2048 	}
2049 
2050 	/*
2051 	 * First accesses are treated as private, otherwise consider accesses
2052 	 * to be private if the accessing pid has not changed
2053 	 */
2054 	if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2055 		priv = 1;
2056 	} else {
2057 		priv = cpupid_match_pid(p, last_cpupid);
2058 		if (!priv && !(flags & TNF_NO_GROUP))
2059 			task_numa_group(p, last_cpupid, flags, &priv);
2060 	}
2061 
2062 	/*
2063 	 * If a workload spans multiple NUMA nodes, a shared fault that
2064 	 * occurs wholly within the set of nodes that the workload is
2065 	 * actively using should be counted as local. This allows the
2066 	 * scan rate to slow down when a workload has settled down.
2067 	 */
2068 	if (!priv && !local && p->numa_group &&
2069 			node_isset(cpu_node, p->numa_group->active_nodes) &&
2070 			node_isset(mem_node, p->numa_group->active_nodes))
2071 		local = 1;
2072 
2073 	task_numa_placement(p);
2074 
2075 	/*
2076 	 * Retry task to preferred node migration periodically, in case it
2077 	 * case it previously failed, or the scheduler moved us.
2078 	 */
2079 	if (time_after(jiffies, p->numa_migrate_retry))
2080 		numa_migrate_preferred(p);
2081 
2082 	if (migrated)
2083 		p->numa_pages_migrated += pages;
2084 
2085 	p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2086 	p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2087 	p->numa_faults_locality[local] += pages;
2088 }
2089 
2090 static void reset_ptenuma_scan(struct task_struct *p)
2091 {
2092 	ACCESS_ONCE(p->mm->numa_scan_seq)++;
2093 	p->mm->numa_scan_offset = 0;
2094 }
2095 
2096 /*
2097  * The expensive part of numa migration is done from task_work context.
2098  * Triggered from task_tick_numa().
2099  */
2100 void task_numa_work(struct callback_head *work)
2101 {
2102 	unsigned long migrate, next_scan, now = jiffies;
2103 	struct task_struct *p = current;
2104 	struct mm_struct *mm = p->mm;
2105 	struct vm_area_struct *vma;
2106 	unsigned long start, end;
2107 	unsigned long nr_pte_updates = 0;
2108 	long pages;
2109 
2110 	WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2111 
2112 	work->next = work; /* protect against double add */
2113 	/*
2114 	 * Who cares about NUMA placement when they're dying.
2115 	 *
2116 	 * NOTE: make sure not to dereference p->mm before this check,
2117 	 * exit_task_work() happens _after_ exit_mm() so we could be called
2118 	 * without p->mm even though we still had it when we enqueued this
2119 	 * work.
2120 	 */
2121 	if (p->flags & PF_EXITING)
2122 		return;
2123 
2124 	if (!mm->numa_next_scan) {
2125 		mm->numa_next_scan = now +
2126 			msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2127 	}
2128 
2129 	/*
2130 	 * Enforce maximal scan/migration frequency..
2131 	 */
2132 	migrate = mm->numa_next_scan;
2133 	if (time_before(now, migrate))
2134 		return;
2135 
2136 	if (p->numa_scan_period == 0) {
2137 		p->numa_scan_period_max = task_scan_max(p);
2138 		p->numa_scan_period = task_scan_min(p);
2139 	}
2140 
2141 	next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2142 	if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2143 		return;
2144 
2145 	/*
2146 	 * Delay this task enough that another task of this mm will likely win
2147 	 * the next time around.
2148 	 */
2149 	p->node_stamp += 2 * TICK_NSEC;
2150 
2151 	start = mm->numa_scan_offset;
2152 	pages = sysctl_numa_balancing_scan_size;
2153 	pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2154 	if (!pages)
2155 		return;
2156 
2157 	down_read(&mm->mmap_sem);
2158 	vma = find_vma(mm, start);
2159 	if (!vma) {
2160 		reset_ptenuma_scan(p);
2161 		start = 0;
2162 		vma = mm->mmap;
2163 	}
2164 	for (; vma; vma = vma->vm_next) {
2165 		if (!vma_migratable(vma) || !vma_policy_mof(vma))
2166 			continue;
2167 
2168 		/*
2169 		 * Shared library pages mapped by multiple processes are not
2170 		 * migrated as it is expected they are cache replicated. Avoid
2171 		 * hinting faults in read-only file-backed mappings or the vdso
2172 		 * as migrating the pages will be of marginal benefit.
2173 		 */
2174 		if (!vma->vm_mm ||
2175 		    (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2176 			continue;
2177 
2178 		/*
2179 		 * Skip inaccessible VMAs to avoid any confusion between
2180 		 * PROT_NONE and NUMA hinting ptes
2181 		 */
2182 		if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2183 			continue;
2184 
2185 		do {
2186 			start = max(start, vma->vm_start);
2187 			end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2188 			end = min(end, vma->vm_end);
2189 			nr_pte_updates += change_prot_numa(vma, start, end);
2190 
2191 			/*
2192 			 * Scan sysctl_numa_balancing_scan_size but ensure that
2193 			 * at least one PTE is updated so that unused virtual
2194 			 * address space is quickly skipped.
2195 			 */
2196 			if (nr_pte_updates)
2197 				pages -= (end - start) >> PAGE_SHIFT;
2198 
2199 			start = end;
2200 			if (pages <= 0)
2201 				goto out;
2202 
2203 			cond_resched();
2204 		} while (end != vma->vm_end);
2205 	}
2206 
2207 out:
2208 	/*
2209 	 * It is possible to reach the end of the VMA list but the last few
2210 	 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2211 	 * would find the !migratable VMA on the next scan but not reset the
2212 	 * scanner to the start so check it now.
2213 	 */
2214 	if (vma)
2215 		mm->numa_scan_offset = start;
2216 	else
2217 		reset_ptenuma_scan(p);
2218 	up_read(&mm->mmap_sem);
2219 }
2220 
2221 /*
2222  * Drive the periodic memory faults..
2223  */
2224 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2225 {
2226 	struct callback_head *work = &curr->numa_work;
2227 	u64 period, now;
2228 
2229 	/*
2230 	 * We don't care about NUMA placement if we don't have memory.
2231 	 */
2232 	if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2233 		return;
2234 
2235 	/*
2236 	 * Using runtime rather than walltime has the dual advantage that
2237 	 * we (mostly) drive the selection from busy threads and that the
2238 	 * task needs to have done some actual work before we bother with
2239 	 * NUMA placement.
2240 	 */
2241 	now = curr->se.sum_exec_runtime;
2242 	period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2243 
2244 	if (now - curr->node_stamp > period) {
2245 		if (!curr->node_stamp)
2246 			curr->numa_scan_period = task_scan_min(curr);
2247 		curr->node_stamp += period;
2248 
2249 		if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2250 			init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2251 			task_work_add(curr, work, true);
2252 		}
2253 	}
2254 }
2255 #else
2256 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2257 {
2258 }
2259 
2260 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2261 {
2262 }
2263 
2264 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2265 {
2266 }
2267 #endif /* CONFIG_NUMA_BALANCING */
2268 
2269 static void
2270 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2271 {
2272 	update_load_add(&cfs_rq->load, se->load.weight);
2273 	if (!parent_entity(se))
2274 		update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2275 #ifdef CONFIG_SMP
2276 	if (entity_is_task(se)) {
2277 		struct rq *rq = rq_of(cfs_rq);
2278 
2279 		account_numa_enqueue(rq, task_of(se));
2280 		list_add(&se->group_node, &rq->cfs_tasks);
2281 	}
2282 #endif
2283 	cfs_rq->nr_running++;
2284 }
2285 
2286 static void
2287 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2288 {
2289 	update_load_sub(&cfs_rq->load, se->load.weight);
2290 	if (!parent_entity(se))
2291 		update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2292 	if (entity_is_task(se)) {
2293 		account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2294 		list_del_init(&se->group_node);
2295 	}
2296 	cfs_rq->nr_running--;
2297 }
2298 
2299 #ifdef CONFIG_FAIR_GROUP_SCHED
2300 # ifdef CONFIG_SMP
2301 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2302 {
2303 	long tg_weight;
2304 
2305 	/*
2306 	 * Use this CPU's actual weight instead of the last load_contribution
2307 	 * to gain a more accurate current total weight. See
2308 	 * update_cfs_rq_load_contribution().
2309 	 */
2310 	tg_weight = atomic_long_read(&tg->load_avg);
2311 	tg_weight -= cfs_rq->tg_load_contrib;
2312 	tg_weight += cfs_rq->load.weight;
2313 
2314 	return tg_weight;
2315 }
2316 
2317 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2318 {
2319 	long tg_weight, load, shares;
2320 
2321 	tg_weight = calc_tg_weight(tg, cfs_rq);
2322 	load = cfs_rq->load.weight;
2323 
2324 	shares = (tg->shares * load);
2325 	if (tg_weight)
2326 		shares /= tg_weight;
2327 
2328 	if (shares < MIN_SHARES)
2329 		shares = MIN_SHARES;
2330 	if (shares > tg->shares)
2331 		shares = tg->shares;
2332 
2333 	return shares;
2334 }
2335 # else /* CONFIG_SMP */
2336 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2337 {
2338 	return tg->shares;
2339 }
2340 # endif /* CONFIG_SMP */
2341 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2342 			    unsigned long weight)
2343 {
2344 	if (se->on_rq) {
2345 		/* commit outstanding execution time */
2346 		if (cfs_rq->curr == se)
2347 			update_curr(cfs_rq);
2348 		account_entity_dequeue(cfs_rq, se);
2349 	}
2350 
2351 	update_load_set(&se->load, weight);
2352 
2353 	if (se->on_rq)
2354 		account_entity_enqueue(cfs_rq, se);
2355 }
2356 
2357 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2358 
2359 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2360 {
2361 	struct task_group *tg;
2362 	struct sched_entity *se;
2363 	long shares;
2364 
2365 	tg = cfs_rq->tg;
2366 	se = tg->se[cpu_of(rq_of(cfs_rq))];
2367 	if (!se || throttled_hierarchy(cfs_rq))
2368 		return;
2369 #ifndef CONFIG_SMP
2370 	if (likely(se->load.weight == tg->shares))
2371 		return;
2372 #endif
2373 	shares = calc_cfs_shares(cfs_rq, tg);
2374 
2375 	reweight_entity(cfs_rq_of(se), se, shares);
2376 }
2377 #else /* CONFIG_FAIR_GROUP_SCHED */
2378 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2379 {
2380 }
2381 #endif /* CONFIG_FAIR_GROUP_SCHED */
2382 
2383 #ifdef CONFIG_SMP
2384 /*
2385  * We choose a half-life close to 1 scheduling period.
2386  * Note: The tables below are dependent on this value.
2387  */
2388 #define LOAD_AVG_PERIOD 32
2389 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2390 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2391 
2392 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2393 static const u32 runnable_avg_yN_inv[] = {
2394 	0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2395 	0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2396 	0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2397 	0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2398 	0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2399 	0x85aac367, 0x82cd8698,
2400 };
2401 
2402 /*
2403  * Precomputed \Sum y^k { 1<=k<=n }.  These are floor(true_value) to prevent
2404  * over-estimates when re-combining.
2405  */
2406 static const u32 runnable_avg_yN_sum[] = {
2407 	    0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2408 	 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2409 	17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2410 };
2411 
2412 /*
2413  * Approximate:
2414  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2415  */
2416 static __always_inline u64 decay_load(u64 val, u64 n)
2417 {
2418 	unsigned int local_n;
2419 
2420 	if (!n)
2421 		return val;
2422 	else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2423 		return 0;
2424 
2425 	/* after bounds checking we can collapse to 32-bit */
2426 	local_n = n;
2427 
2428 	/*
2429 	 * As y^PERIOD = 1/2, we can combine
2430 	 *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2431 	 * With a look-up table which covers y^n (n<PERIOD)
2432 	 *
2433 	 * To achieve constant time decay_load.
2434 	 */
2435 	if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2436 		val >>= local_n / LOAD_AVG_PERIOD;
2437 		local_n %= LOAD_AVG_PERIOD;
2438 	}
2439 
2440 	val *= runnable_avg_yN_inv[local_n];
2441 	/* We don't use SRR here since we always want to round down. */
2442 	return val >> 32;
2443 }
2444 
2445 /*
2446  * For updates fully spanning n periods, the contribution to runnable
2447  * average will be: \Sum 1024*y^n
2448  *
2449  * We can compute this reasonably efficiently by combining:
2450  *   y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for  n <PERIOD}
2451  */
2452 static u32 __compute_runnable_contrib(u64 n)
2453 {
2454 	u32 contrib = 0;
2455 
2456 	if (likely(n <= LOAD_AVG_PERIOD))
2457 		return runnable_avg_yN_sum[n];
2458 	else if (unlikely(n >= LOAD_AVG_MAX_N))
2459 		return LOAD_AVG_MAX;
2460 
2461 	/* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2462 	do {
2463 		contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2464 		contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2465 
2466 		n -= LOAD_AVG_PERIOD;
2467 	} while (n > LOAD_AVG_PERIOD);
2468 
2469 	contrib = decay_load(contrib, n);
2470 	return contrib + runnable_avg_yN_sum[n];
2471 }
2472 
2473 /*
2474  * We can represent the historical contribution to runnable average as the
2475  * coefficients of a geometric series.  To do this we sub-divide our runnable
2476  * history into segments of approximately 1ms (1024us); label the segment that
2477  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2478  *
2479  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2480  *      p0            p1           p2
2481  *     (now)       (~1ms ago)  (~2ms ago)
2482  *
2483  * Let u_i denote the fraction of p_i that the entity was runnable.
2484  *
2485  * We then designate the fractions u_i as our co-efficients, yielding the
2486  * following representation of historical load:
2487  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2488  *
2489  * We choose y based on the with of a reasonably scheduling period, fixing:
2490  *   y^32 = 0.5
2491  *
2492  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2493  * approximately half as much as the contribution to load within the last ms
2494  * (u_0).
2495  *
2496  * When a period "rolls over" and we have new u_0`, multiplying the previous
2497  * sum again by y is sufficient to update:
2498  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2499  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2500  */
2501 static __always_inline int __update_entity_runnable_avg(u64 now,
2502 							struct sched_avg *sa,
2503 							int runnable)
2504 {
2505 	u64 delta, periods;
2506 	u32 runnable_contrib;
2507 	int delta_w, decayed = 0;
2508 
2509 	delta = now - sa->last_runnable_update;
2510 	/*
2511 	 * This should only happen when time goes backwards, which it
2512 	 * unfortunately does during sched clock init when we swap over to TSC.
2513 	 */
2514 	if ((s64)delta < 0) {
2515 		sa->last_runnable_update = now;
2516 		return 0;
2517 	}
2518 
2519 	/*
2520 	 * Use 1024ns as the unit of measurement since it's a reasonable
2521 	 * approximation of 1us and fast to compute.
2522 	 */
2523 	delta >>= 10;
2524 	if (!delta)
2525 		return 0;
2526 	sa->last_runnable_update = now;
2527 
2528 	/* delta_w is the amount already accumulated against our next period */
2529 	delta_w = sa->runnable_avg_period % 1024;
2530 	if (delta + delta_w >= 1024) {
2531 		/* period roll-over */
2532 		decayed = 1;
2533 
2534 		/*
2535 		 * Now that we know we're crossing a period boundary, figure
2536 		 * out how much from delta we need to complete the current
2537 		 * period and accrue it.
2538 		 */
2539 		delta_w = 1024 - delta_w;
2540 		if (runnable)
2541 			sa->runnable_avg_sum += delta_w;
2542 		sa->runnable_avg_period += delta_w;
2543 
2544 		delta -= delta_w;
2545 
2546 		/* Figure out how many additional periods this update spans */
2547 		periods = delta / 1024;
2548 		delta %= 1024;
2549 
2550 		sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2551 						  periods + 1);
2552 		sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2553 						     periods + 1);
2554 
2555 		/* Efficiently calculate \sum (1..n_period) 1024*y^i */
2556 		runnable_contrib = __compute_runnable_contrib(periods);
2557 		if (runnable)
2558 			sa->runnable_avg_sum += runnable_contrib;
2559 		sa->runnable_avg_period += runnable_contrib;
2560 	}
2561 
2562 	/* Remainder of delta accrued against u_0` */
2563 	if (runnable)
2564 		sa->runnable_avg_sum += delta;
2565 	sa->runnable_avg_period += delta;
2566 
2567 	return decayed;
2568 }
2569 
2570 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2571 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2572 {
2573 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
2574 	u64 decays = atomic64_read(&cfs_rq->decay_counter);
2575 
2576 	decays -= se->avg.decay_count;
2577 	if (!decays)
2578 		return 0;
2579 
2580 	se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2581 	se->avg.decay_count = 0;
2582 
2583 	return decays;
2584 }
2585 
2586 #ifdef CONFIG_FAIR_GROUP_SCHED
2587 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2588 						 int force_update)
2589 {
2590 	struct task_group *tg = cfs_rq->tg;
2591 	long tg_contrib;
2592 
2593 	tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2594 	tg_contrib -= cfs_rq->tg_load_contrib;
2595 
2596 	if (!tg_contrib)
2597 		return;
2598 
2599 	if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2600 		atomic_long_add(tg_contrib, &tg->load_avg);
2601 		cfs_rq->tg_load_contrib += tg_contrib;
2602 	}
2603 }
2604 
2605 /*
2606  * Aggregate cfs_rq runnable averages into an equivalent task_group
2607  * representation for computing load contributions.
2608  */
2609 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2610 						  struct cfs_rq *cfs_rq)
2611 {
2612 	struct task_group *tg = cfs_rq->tg;
2613 	long contrib;
2614 
2615 	/* The fraction of a cpu used by this cfs_rq */
2616 	contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2617 			  sa->runnable_avg_period + 1);
2618 	contrib -= cfs_rq->tg_runnable_contrib;
2619 
2620 	if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2621 		atomic_add(contrib, &tg->runnable_avg);
2622 		cfs_rq->tg_runnable_contrib += contrib;
2623 	}
2624 }
2625 
2626 static inline void __update_group_entity_contrib(struct sched_entity *se)
2627 {
2628 	struct cfs_rq *cfs_rq = group_cfs_rq(se);
2629 	struct task_group *tg = cfs_rq->tg;
2630 	int runnable_avg;
2631 
2632 	u64 contrib;
2633 
2634 	contrib = cfs_rq->tg_load_contrib * tg->shares;
2635 	se->avg.load_avg_contrib = div_u64(contrib,
2636 				     atomic_long_read(&tg->load_avg) + 1);
2637 
2638 	/*
2639 	 * For group entities we need to compute a correction term in the case
2640 	 * that they are consuming <1 cpu so that we would contribute the same
2641 	 * load as a task of equal weight.
2642 	 *
2643 	 * Explicitly co-ordinating this measurement would be expensive, but
2644 	 * fortunately the sum of each cpus contribution forms a usable
2645 	 * lower-bound on the true value.
2646 	 *
2647 	 * Consider the aggregate of 2 contributions.  Either they are disjoint
2648 	 * (and the sum represents true value) or they are disjoint and we are
2649 	 * understating by the aggregate of their overlap.
2650 	 *
2651 	 * Extending this to N cpus, for a given overlap, the maximum amount we
2652 	 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2653 	 * cpus that overlap for this interval and w_i is the interval width.
2654 	 *
2655 	 * On a small machine; the first term is well-bounded which bounds the
2656 	 * total error since w_i is a subset of the period.  Whereas on a
2657 	 * larger machine, while this first term can be larger, if w_i is the
2658 	 * of consequential size guaranteed to see n_i*w_i quickly converge to
2659 	 * our upper bound of 1-cpu.
2660 	 */
2661 	runnable_avg = atomic_read(&tg->runnable_avg);
2662 	if (runnable_avg < NICE_0_LOAD) {
2663 		se->avg.load_avg_contrib *= runnable_avg;
2664 		se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2665 	}
2666 }
2667 
2668 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2669 {
2670 	__update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2671 	__update_tg_runnable_avg(&rq->avg, &rq->cfs);
2672 }
2673 #else /* CONFIG_FAIR_GROUP_SCHED */
2674 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2675 						 int force_update) {}
2676 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2677 						  struct cfs_rq *cfs_rq) {}
2678 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2679 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2680 #endif /* CONFIG_FAIR_GROUP_SCHED */
2681 
2682 static inline void __update_task_entity_contrib(struct sched_entity *se)
2683 {
2684 	u32 contrib;
2685 
2686 	/* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2687 	contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2688 	contrib /= (se->avg.runnable_avg_period + 1);
2689 	se->avg.load_avg_contrib = scale_load(contrib);
2690 }
2691 
2692 /* Compute the current contribution to load_avg by se, return any delta */
2693 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2694 {
2695 	long old_contrib = se->avg.load_avg_contrib;
2696 
2697 	if (entity_is_task(se)) {
2698 		__update_task_entity_contrib(se);
2699 	} else {
2700 		__update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2701 		__update_group_entity_contrib(se);
2702 	}
2703 
2704 	return se->avg.load_avg_contrib - old_contrib;
2705 }
2706 
2707 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2708 						 long load_contrib)
2709 {
2710 	if (likely(load_contrib < cfs_rq->blocked_load_avg))
2711 		cfs_rq->blocked_load_avg -= load_contrib;
2712 	else
2713 		cfs_rq->blocked_load_avg = 0;
2714 }
2715 
2716 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2717 
2718 /* Update a sched_entity's runnable average */
2719 static inline void update_entity_load_avg(struct sched_entity *se,
2720 					  int update_cfs_rq)
2721 {
2722 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
2723 	long contrib_delta;
2724 	u64 now;
2725 
2726 	/*
2727 	 * For a group entity we need to use their owned cfs_rq_clock_task() in
2728 	 * case they are the parent of a throttled hierarchy.
2729 	 */
2730 	if (entity_is_task(se))
2731 		now = cfs_rq_clock_task(cfs_rq);
2732 	else
2733 		now = cfs_rq_clock_task(group_cfs_rq(se));
2734 
2735 	if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2736 		return;
2737 
2738 	contrib_delta = __update_entity_load_avg_contrib(se);
2739 
2740 	if (!update_cfs_rq)
2741 		return;
2742 
2743 	if (se->on_rq)
2744 		cfs_rq->runnable_load_avg += contrib_delta;
2745 	else
2746 		subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2747 }
2748 
2749 /*
2750  * Decay the load contributed by all blocked children and account this so that
2751  * their contribution may appropriately discounted when they wake up.
2752  */
2753 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2754 {
2755 	u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2756 	u64 decays;
2757 
2758 	decays = now - cfs_rq->last_decay;
2759 	if (!decays && !force_update)
2760 		return;
2761 
2762 	if (atomic_long_read(&cfs_rq->removed_load)) {
2763 		unsigned long removed_load;
2764 		removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2765 		subtract_blocked_load_contrib(cfs_rq, removed_load);
2766 	}
2767 
2768 	if (decays) {
2769 		cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2770 						      decays);
2771 		atomic64_add(decays, &cfs_rq->decay_counter);
2772 		cfs_rq->last_decay = now;
2773 	}
2774 
2775 	__update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2776 }
2777 
2778 /* Add the load generated by se into cfs_rq's child load-average */
2779 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2780 						  struct sched_entity *se,
2781 						  int wakeup)
2782 {
2783 	/*
2784 	 * We track migrations using entity decay_count <= 0, on a wake-up
2785 	 * migration we use a negative decay count to track the remote decays
2786 	 * accumulated while sleeping.
2787 	 *
2788 	 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2789 	 * are seen by enqueue_entity_load_avg() as a migration with an already
2790 	 * constructed load_avg_contrib.
2791 	 */
2792 	if (unlikely(se->avg.decay_count <= 0)) {
2793 		se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2794 		if (se->avg.decay_count) {
2795 			/*
2796 			 * In a wake-up migration we have to approximate the
2797 			 * time sleeping.  This is because we can't synchronize
2798 			 * clock_task between the two cpus, and it is not
2799 			 * guaranteed to be read-safe.  Instead, we can
2800 			 * approximate this using our carried decays, which are
2801 			 * explicitly atomically readable.
2802 			 */
2803 			se->avg.last_runnable_update -= (-se->avg.decay_count)
2804 							<< 20;
2805 			update_entity_load_avg(se, 0);
2806 			/* Indicate that we're now synchronized and on-rq */
2807 			se->avg.decay_count = 0;
2808 		}
2809 		wakeup = 0;
2810 	} else {
2811 		__synchronize_entity_decay(se);
2812 	}
2813 
2814 	/* migrated tasks did not contribute to our blocked load */
2815 	if (wakeup) {
2816 		subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2817 		update_entity_load_avg(se, 0);
2818 	}
2819 
2820 	cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2821 	/* we force update consideration on load-balancer moves */
2822 	update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2823 }
2824 
2825 /*
2826  * Remove se's load from this cfs_rq child load-average, if the entity is
2827  * transitioning to a blocked state we track its projected decay using
2828  * blocked_load_avg.
2829  */
2830 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2831 						  struct sched_entity *se,
2832 						  int sleep)
2833 {
2834 	update_entity_load_avg(se, 1);
2835 	/* we force update consideration on load-balancer moves */
2836 	update_cfs_rq_blocked_load(cfs_rq, !sleep);
2837 
2838 	cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2839 	if (sleep) {
2840 		cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2841 		se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2842 	} /* migrations, e.g. sleep=0 leave decay_count == 0 */
2843 }
2844 
2845 /*
2846  * Update the rq's load with the elapsed running time before entering
2847  * idle. if the last scheduled task is not a CFS task, idle_enter will
2848  * be the only way to update the runnable statistic.
2849  */
2850 void idle_enter_fair(struct rq *this_rq)
2851 {
2852 	update_rq_runnable_avg(this_rq, 1);
2853 }
2854 
2855 /*
2856  * Update the rq's load with the elapsed idle time before a task is
2857  * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2858  * be the only way to update the runnable statistic.
2859  */
2860 void idle_exit_fair(struct rq *this_rq)
2861 {
2862 	update_rq_runnable_avg(this_rq, 0);
2863 }
2864 
2865 static int idle_balance(struct rq *this_rq);
2866 
2867 #else /* CONFIG_SMP */
2868 
2869 static inline void update_entity_load_avg(struct sched_entity *se,
2870 					  int update_cfs_rq) {}
2871 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2872 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2873 					   struct sched_entity *se,
2874 					   int wakeup) {}
2875 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2876 					   struct sched_entity *se,
2877 					   int sleep) {}
2878 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2879 					      int force_update) {}
2880 
2881 static inline int idle_balance(struct rq *rq)
2882 {
2883 	return 0;
2884 }
2885 
2886 #endif /* CONFIG_SMP */
2887 
2888 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2889 {
2890 #ifdef CONFIG_SCHEDSTATS
2891 	struct task_struct *tsk = NULL;
2892 
2893 	if (entity_is_task(se))
2894 		tsk = task_of(se);
2895 
2896 	if (se->statistics.sleep_start) {
2897 		u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2898 
2899 		if ((s64)delta < 0)
2900 			delta = 0;
2901 
2902 		if (unlikely(delta > se->statistics.sleep_max))
2903 			se->statistics.sleep_max = delta;
2904 
2905 		se->statistics.sleep_start = 0;
2906 		se->statistics.sum_sleep_runtime += delta;
2907 
2908 		if (tsk) {
2909 			account_scheduler_latency(tsk, delta >> 10, 1);
2910 			trace_sched_stat_sleep(tsk, delta);
2911 		}
2912 	}
2913 	if (se->statistics.block_start) {
2914 		u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2915 
2916 		if ((s64)delta < 0)
2917 			delta = 0;
2918 
2919 		if (unlikely(delta > se->statistics.block_max))
2920 			se->statistics.block_max = delta;
2921 
2922 		se->statistics.block_start = 0;
2923 		se->statistics.sum_sleep_runtime += delta;
2924 
2925 		if (tsk) {
2926 			if (tsk->in_iowait) {
2927 				se->statistics.iowait_sum += delta;
2928 				se->statistics.iowait_count++;
2929 				trace_sched_stat_iowait(tsk, delta);
2930 			}
2931 
2932 			trace_sched_stat_blocked(tsk, delta);
2933 
2934 			/*
2935 			 * Blocking time is in units of nanosecs, so shift by
2936 			 * 20 to get a milliseconds-range estimation of the
2937 			 * amount of time that the task spent sleeping:
2938 			 */
2939 			if (unlikely(prof_on == SLEEP_PROFILING)) {
2940 				profile_hits(SLEEP_PROFILING,
2941 						(void *)get_wchan(tsk),
2942 						delta >> 20);
2943 			}
2944 			account_scheduler_latency(tsk, delta >> 10, 0);
2945 		}
2946 	}
2947 #endif
2948 }
2949 
2950 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2951 {
2952 #ifdef CONFIG_SCHED_DEBUG
2953 	s64 d = se->vruntime - cfs_rq->min_vruntime;
2954 
2955 	if (d < 0)
2956 		d = -d;
2957 
2958 	if (d > 3*sysctl_sched_latency)
2959 		schedstat_inc(cfs_rq, nr_spread_over);
2960 #endif
2961 }
2962 
2963 static void
2964 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2965 {
2966 	u64 vruntime = cfs_rq->min_vruntime;
2967 
2968 	/*
2969 	 * The 'current' period is already promised to the current tasks,
2970 	 * however the extra weight of the new task will slow them down a
2971 	 * little, place the new task so that it fits in the slot that
2972 	 * stays open at the end.
2973 	 */
2974 	if (initial && sched_feat(START_DEBIT))
2975 		vruntime += sched_vslice(cfs_rq, se);
2976 
2977 	/* sleeps up to a single latency don't count. */
2978 	if (!initial) {
2979 		unsigned long thresh = sysctl_sched_latency;
2980 
2981 		/*
2982 		 * Halve their sleep time's effect, to allow
2983 		 * for a gentler effect of sleepers:
2984 		 */
2985 		if (sched_feat(GENTLE_FAIR_SLEEPERS))
2986 			thresh >>= 1;
2987 
2988 		vruntime -= thresh;
2989 	}
2990 
2991 	/* ensure we never gain time by being placed backwards. */
2992 	se->vruntime = max_vruntime(se->vruntime, vruntime);
2993 }
2994 
2995 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2996 
2997 static void
2998 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2999 {
3000 	/*
3001 	 * Update the normalized vruntime before updating min_vruntime
3002 	 * through calling update_curr().
3003 	 */
3004 	if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3005 		se->vruntime += cfs_rq->min_vruntime;
3006 
3007 	/*
3008 	 * Update run-time statistics of the 'current'.
3009 	 */
3010 	update_curr(cfs_rq);
3011 	enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
3012 	account_entity_enqueue(cfs_rq, se);
3013 	update_cfs_shares(cfs_rq);
3014 
3015 	if (flags & ENQUEUE_WAKEUP) {
3016 		place_entity(cfs_rq, se, 0);
3017 		enqueue_sleeper(cfs_rq, se);
3018 	}
3019 
3020 	update_stats_enqueue(cfs_rq, se);
3021 	check_spread(cfs_rq, se);
3022 	if (se != cfs_rq->curr)
3023 		__enqueue_entity(cfs_rq, se);
3024 	se->on_rq = 1;
3025 
3026 	if (cfs_rq->nr_running == 1) {
3027 		list_add_leaf_cfs_rq(cfs_rq);
3028 		check_enqueue_throttle(cfs_rq);
3029 	}
3030 }
3031 
3032 static void __clear_buddies_last(struct sched_entity *se)
3033 {
3034 	for_each_sched_entity(se) {
3035 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
3036 		if (cfs_rq->last != se)
3037 			break;
3038 
3039 		cfs_rq->last = NULL;
3040 	}
3041 }
3042 
3043 static void __clear_buddies_next(struct sched_entity *se)
3044 {
3045 	for_each_sched_entity(se) {
3046 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
3047 		if (cfs_rq->next != se)
3048 			break;
3049 
3050 		cfs_rq->next = NULL;
3051 	}
3052 }
3053 
3054 static void __clear_buddies_skip(struct sched_entity *se)
3055 {
3056 	for_each_sched_entity(se) {
3057 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
3058 		if (cfs_rq->skip != se)
3059 			break;
3060 
3061 		cfs_rq->skip = NULL;
3062 	}
3063 }
3064 
3065 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3066 {
3067 	if (cfs_rq->last == se)
3068 		__clear_buddies_last(se);
3069 
3070 	if (cfs_rq->next == se)
3071 		__clear_buddies_next(se);
3072 
3073 	if (cfs_rq->skip == se)
3074 		__clear_buddies_skip(se);
3075 }
3076 
3077 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3078 
3079 static void
3080 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3081 {
3082 	/*
3083 	 * Update run-time statistics of the 'current'.
3084 	 */
3085 	update_curr(cfs_rq);
3086 	dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
3087 
3088 	update_stats_dequeue(cfs_rq, se);
3089 	if (flags & DEQUEUE_SLEEP) {
3090 #ifdef CONFIG_SCHEDSTATS
3091 		if (entity_is_task(se)) {
3092 			struct task_struct *tsk = task_of(se);
3093 
3094 			if (tsk->state & TASK_INTERRUPTIBLE)
3095 				se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3096 			if (tsk->state & TASK_UNINTERRUPTIBLE)
3097 				se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3098 		}
3099 #endif
3100 	}
3101 
3102 	clear_buddies(cfs_rq, se);
3103 
3104 	if (se != cfs_rq->curr)
3105 		__dequeue_entity(cfs_rq, se);
3106 	se->on_rq = 0;
3107 	account_entity_dequeue(cfs_rq, se);
3108 
3109 	/*
3110 	 * Normalize the entity after updating the min_vruntime because the
3111 	 * update can refer to the ->curr item and we need to reflect this
3112 	 * movement in our normalized position.
3113 	 */
3114 	if (!(flags & DEQUEUE_SLEEP))
3115 		se->vruntime -= cfs_rq->min_vruntime;
3116 
3117 	/* return excess runtime on last dequeue */
3118 	return_cfs_rq_runtime(cfs_rq);
3119 
3120 	update_min_vruntime(cfs_rq);
3121 	update_cfs_shares(cfs_rq);
3122 }
3123 
3124 /*
3125  * Preempt the current task with a newly woken task if needed:
3126  */
3127 static void
3128 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3129 {
3130 	unsigned long ideal_runtime, delta_exec;
3131 	struct sched_entity *se;
3132 	s64 delta;
3133 
3134 	ideal_runtime = sched_slice(cfs_rq, curr);
3135 	delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3136 	if (delta_exec > ideal_runtime) {
3137 		resched_curr(rq_of(cfs_rq));
3138 		/*
3139 		 * The current task ran long enough, ensure it doesn't get
3140 		 * re-elected due to buddy favours.
3141 		 */
3142 		clear_buddies(cfs_rq, curr);
3143 		return;
3144 	}
3145 
3146 	/*
3147 	 * Ensure that a task that missed wakeup preemption by a
3148 	 * narrow margin doesn't have to wait for a full slice.
3149 	 * This also mitigates buddy induced latencies under load.
3150 	 */
3151 	if (delta_exec < sysctl_sched_min_granularity)
3152 		return;
3153 
3154 	se = __pick_first_entity(cfs_rq);
3155 	delta = curr->vruntime - se->vruntime;
3156 
3157 	if (delta < 0)
3158 		return;
3159 
3160 	if (delta > ideal_runtime)
3161 		resched_curr(rq_of(cfs_rq));
3162 }
3163 
3164 static void
3165 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3166 {
3167 	/* 'current' is not kept within the tree. */
3168 	if (se->on_rq) {
3169 		/*
3170 		 * Any task has to be enqueued before it get to execute on
3171 		 * a CPU. So account for the time it spent waiting on the
3172 		 * runqueue.
3173 		 */
3174 		update_stats_wait_end(cfs_rq, se);
3175 		__dequeue_entity(cfs_rq, se);
3176 	}
3177 
3178 	update_stats_curr_start(cfs_rq, se);
3179 	cfs_rq->curr = se;
3180 #ifdef CONFIG_SCHEDSTATS
3181 	/*
3182 	 * Track our maximum slice length, if the CPU's load is at
3183 	 * least twice that of our own weight (i.e. dont track it
3184 	 * when there are only lesser-weight tasks around):
3185 	 */
3186 	if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3187 		se->statistics.slice_max = max(se->statistics.slice_max,
3188 			se->sum_exec_runtime - se->prev_sum_exec_runtime);
3189 	}
3190 #endif
3191 	se->prev_sum_exec_runtime = se->sum_exec_runtime;
3192 }
3193 
3194 static int
3195 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3196 
3197 /*
3198  * Pick the next process, keeping these things in mind, in this order:
3199  * 1) keep things fair between processes/task groups
3200  * 2) pick the "next" process, since someone really wants that to run
3201  * 3) pick the "last" process, for cache locality
3202  * 4) do not run the "skip" process, if something else is available
3203  */
3204 static struct sched_entity *
3205 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3206 {
3207 	struct sched_entity *left = __pick_first_entity(cfs_rq);
3208 	struct sched_entity *se;
3209 
3210 	/*
3211 	 * If curr is set we have to see if its left of the leftmost entity
3212 	 * still in the tree, provided there was anything in the tree at all.
3213 	 */
3214 	if (!left || (curr && entity_before(curr, left)))
3215 		left = curr;
3216 
3217 	se = left; /* ideally we run the leftmost entity */
3218 
3219 	/*
3220 	 * Avoid running the skip buddy, if running something else can
3221 	 * be done without getting too unfair.
3222 	 */
3223 	if (cfs_rq->skip == se) {
3224 		struct sched_entity *second;
3225 
3226 		if (se == curr) {
3227 			second = __pick_first_entity(cfs_rq);
3228 		} else {
3229 			second = __pick_next_entity(se);
3230 			if (!second || (curr && entity_before(curr, second)))
3231 				second = curr;
3232 		}
3233 
3234 		if (second && wakeup_preempt_entity(second, left) < 1)
3235 			se = second;
3236 	}
3237 
3238 	/*
3239 	 * Prefer last buddy, try to return the CPU to a preempted task.
3240 	 */
3241 	if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3242 		se = cfs_rq->last;
3243 
3244 	/*
3245 	 * Someone really wants this to run. If it's not unfair, run it.
3246 	 */
3247 	if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3248 		se = cfs_rq->next;
3249 
3250 	clear_buddies(cfs_rq, se);
3251 
3252 	return se;
3253 }
3254 
3255 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3256 
3257 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3258 {
3259 	/*
3260 	 * If still on the runqueue then deactivate_task()
3261 	 * was not called and update_curr() has to be done:
3262 	 */
3263 	if (prev->on_rq)
3264 		update_curr(cfs_rq);
3265 
3266 	/* throttle cfs_rqs exceeding runtime */
3267 	check_cfs_rq_runtime(cfs_rq);
3268 
3269 	check_spread(cfs_rq, prev);
3270 	if (prev->on_rq) {
3271 		update_stats_wait_start(cfs_rq, prev);
3272 		/* Put 'current' back into the tree. */
3273 		__enqueue_entity(cfs_rq, prev);
3274 		/* in !on_rq case, update occurred at dequeue */
3275 		update_entity_load_avg(prev, 1);
3276 	}
3277 	cfs_rq->curr = NULL;
3278 }
3279 
3280 static void
3281 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3282 {
3283 	/*
3284 	 * Update run-time statistics of the 'current'.
3285 	 */
3286 	update_curr(cfs_rq);
3287 
3288 	/*
3289 	 * Ensure that runnable average is periodically updated.
3290 	 */
3291 	update_entity_load_avg(curr, 1);
3292 	update_cfs_rq_blocked_load(cfs_rq, 1);
3293 	update_cfs_shares(cfs_rq);
3294 
3295 #ifdef CONFIG_SCHED_HRTICK
3296 	/*
3297 	 * queued ticks are scheduled to match the slice, so don't bother
3298 	 * validating it and just reschedule.
3299 	 */
3300 	if (queued) {
3301 		resched_curr(rq_of(cfs_rq));
3302 		return;
3303 	}
3304 	/*
3305 	 * don't let the period tick interfere with the hrtick preemption
3306 	 */
3307 	if (!sched_feat(DOUBLE_TICK) &&
3308 			hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3309 		return;
3310 #endif
3311 
3312 	if (cfs_rq->nr_running > 1)
3313 		check_preempt_tick(cfs_rq, curr);
3314 }
3315 
3316 
3317 /**************************************************
3318  * CFS bandwidth control machinery
3319  */
3320 
3321 #ifdef CONFIG_CFS_BANDWIDTH
3322 
3323 #ifdef HAVE_JUMP_LABEL
3324 static struct static_key __cfs_bandwidth_used;
3325 
3326 static inline bool cfs_bandwidth_used(void)
3327 {
3328 	return static_key_false(&__cfs_bandwidth_used);
3329 }
3330 
3331 void cfs_bandwidth_usage_inc(void)
3332 {
3333 	static_key_slow_inc(&__cfs_bandwidth_used);
3334 }
3335 
3336 void cfs_bandwidth_usage_dec(void)
3337 {
3338 	static_key_slow_dec(&__cfs_bandwidth_used);
3339 }
3340 #else /* HAVE_JUMP_LABEL */
3341 static bool cfs_bandwidth_used(void)
3342 {
3343 	return true;
3344 }
3345 
3346 void cfs_bandwidth_usage_inc(void) {}
3347 void cfs_bandwidth_usage_dec(void) {}
3348 #endif /* HAVE_JUMP_LABEL */
3349 
3350 /*
3351  * default period for cfs group bandwidth.
3352  * default: 0.1s, units: nanoseconds
3353  */
3354 static inline u64 default_cfs_period(void)
3355 {
3356 	return 100000000ULL;
3357 }
3358 
3359 static inline u64 sched_cfs_bandwidth_slice(void)
3360 {
3361 	return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3362 }
3363 
3364 /*
3365  * Replenish runtime according to assigned quota and update expiration time.
3366  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3367  * additional synchronization around rq->lock.
3368  *
3369  * requires cfs_b->lock
3370  */
3371 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3372 {
3373 	u64 now;
3374 
3375 	if (cfs_b->quota == RUNTIME_INF)
3376 		return;
3377 
3378 	now = sched_clock_cpu(smp_processor_id());
3379 	cfs_b->runtime = cfs_b->quota;
3380 	cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3381 }
3382 
3383 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3384 {
3385 	return &tg->cfs_bandwidth;
3386 }
3387 
3388 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3389 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3390 {
3391 	if (unlikely(cfs_rq->throttle_count))
3392 		return cfs_rq->throttled_clock_task;
3393 
3394 	return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3395 }
3396 
3397 /* returns 0 on failure to allocate runtime */
3398 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3399 {
3400 	struct task_group *tg = cfs_rq->tg;
3401 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3402 	u64 amount = 0, min_amount, expires;
3403 
3404 	/* note: this is a positive sum as runtime_remaining <= 0 */
3405 	min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3406 
3407 	raw_spin_lock(&cfs_b->lock);
3408 	if (cfs_b->quota == RUNTIME_INF)
3409 		amount = min_amount;
3410 	else {
3411 		/*
3412 		 * If the bandwidth pool has become inactive, then at least one
3413 		 * period must have elapsed since the last consumption.
3414 		 * Refresh the global state and ensure bandwidth timer becomes
3415 		 * active.
3416 		 */
3417 		if (!cfs_b->timer_active) {
3418 			__refill_cfs_bandwidth_runtime(cfs_b);
3419 			__start_cfs_bandwidth(cfs_b, false);
3420 		}
3421 
3422 		if (cfs_b->runtime > 0) {
3423 			amount = min(cfs_b->runtime, min_amount);
3424 			cfs_b->runtime -= amount;
3425 			cfs_b->idle = 0;
3426 		}
3427 	}
3428 	expires = cfs_b->runtime_expires;
3429 	raw_spin_unlock(&cfs_b->lock);
3430 
3431 	cfs_rq->runtime_remaining += amount;
3432 	/*
3433 	 * we may have advanced our local expiration to account for allowed
3434 	 * spread between our sched_clock and the one on which runtime was
3435 	 * issued.
3436 	 */
3437 	if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3438 		cfs_rq->runtime_expires = expires;
3439 
3440 	return cfs_rq->runtime_remaining > 0;
3441 }
3442 
3443 /*
3444  * Note: This depends on the synchronization provided by sched_clock and the
3445  * fact that rq->clock snapshots this value.
3446  */
3447 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3448 {
3449 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3450 
3451 	/* if the deadline is ahead of our clock, nothing to do */
3452 	if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3453 		return;
3454 
3455 	if (cfs_rq->runtime_remaining < 0)
3456 		return;
3457 
3458 	/*
3459 	 * If the local deadline has passed we have to consider the
3460 	 * possibility that our sched_clock is 'fast' and the global deadline
3461 	 * has not truly expired.
3462 	 *
3463 	 * Fortunately we can check determine whether this the case by checking
3464 	 * whether the global deadline has advanced. It is valid to compare
3465 	 * cfs_b->runtime_expires without any locks since we only care about
3466 	 * exact equality, so a partial write will still work.
3467 	 */
3468 
3469 	if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3470 		/* extend local deadline, drift is bounded above by 2 ticks */
3471 		cfs_rq->runtime_expires += TICK_NSEC;
3472 	} else {
3473 		/* global deadline is ahead, expiration has passed */
3474 		cfs_rq->runtime_remaining = 0;
3475 	}
3476 }
3477 
3478 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3479 {
3480 	/* dock delta_exec before expiring quota (as it could span periods) */
3481 	cfs_rq->runtime_remaining -= delta_exec;
3482 	expire_cfs_rq_runtime(cfs_rq);
3483 
3484 	if (likely(cfs_rq->runtime_remaining > 0))
3485 		return;
3486 
3487 	/*
3488 	 * if we're unable to extend our runtime we resched so that the active
3489 	 * hierarchy can be throttled
3490 	 */
3491 	if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3492 		resched_curr(rq_of(cfs_rq));
3493 }
3494 
3495 static __always_inline
3496 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3497 {
3498 	if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3499 		return;
3500 
3501 	__account_cfs_rq_runtime(cfs_rq, delta_exec);
3502 }
3503 
3504 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3505 {
3506 	return cfs_bandwidth_used() && cfs_rq->throttled;
3507 }
3508 
3509 /* check whether cfs_rq, or any parent, is throttled */
3510 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3511 {
3512 	return cfs_bandwidth_used() && cfs_rq->throttle_count;
3513 }
3514 
3515 /*
3516  * Ensure that neither of the group entities corresponding to src_cpu or
3517  * dest_cpu are members of a throttled hierarchy when performing group
3518  * load-balance operations.
3519  */
3520 static inline int throttled_lb_pair(struct task_group *tg,
3521 				    int src_cpu, int dest_cpu)
3522 {
3523 	struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3524 
3525 	src_cfs_rq = tg->cfs_rq[src_cpu];
3526 	dest_cfs_rq = tg->cfs_rq[dest_cpu];
3527 
3528 	return throttled_hierarchy(src_cfs_rq) ||
3529 	       throttled_hierarchy(dest_cfs_rq);
3530 }
3531 
3532 /* updated child weight may affect parent so we have to do this bottom up */
3533 static int tg_unthrottle_up(struct task_group *tg, void *data)
3534 {
3535 	struct rq *rq = data;
3536 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3537 
3538 	cfs_rq->throttle_count--;
3539 #ifdef CONFIG_SMP
3540 	if (!cfs_rq->throttle_count) {
3541 		/* adjust cfs_rq_clock_task() */
3542 		cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3543 					     cfs_rq->throttled_clock_task;
3544 	}
3545 #endif
3546 
3547 	return 0;
3548 }
3549 
3550 static int tg_throttle_down(struct task_group *tg, void *data)
3551 {
3552 	struct rq *rq = data;
3553 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3554 
3555 	/* group is entering throttled state, stop time */
3556 	if (!cfs_rq->throttle_count)
3557 		cfs_rq->throttled_clock_task = rq_clock_task(rq);
3558 	cfs_rq->throttle_count++;
3559 
3560 	return 0;
3561 }
3562 
3563 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3564 {
3565 	struct rq *rq = rq_of(cfs_rq);
3566 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3567 	struct sched_entity *se;
3568 	long task_delta, dequeue = 1;
3569 
3570 	se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3571 
3572 	/* freeze hierarchy runnable averages while throttled */
3573 	rcu_read_lock();
3574 	walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3575 	rcu_read_unlock();
3576 
3577 	task_delta = cfs_rq->h_nr_running;
3578 	for_each_sched_entity(se) {
3579 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3580 		/* throttled entity or throttle-on-deactivate */
3581 		if (!se->on_rq)
3582 			break;
3583 
3584 		if (dequeue)
3585 			dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3586 		qcfs_rq->h_nr_running -= task_delta;
3587 
3588 		if (qcfs_rq->load.weight)
3589 			dequeue = 0;
3590 	}
3591 
3592 	if (!se)
3593 		sub_nr_running(rq, task_delta);
3594 
3595 	cfs_rq->throttled = 1;
3596 	cfs_rq->throttled_clock = rq_clock(rq);
3597 	raw_spin_lock(&cfs_b->lock);
3598 	/*
3599 	 * Add to the _head_ of the list, so that an already-started
3600 	 * distribute_cfs_runtime will not see us
3601 	 */
3602 	list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3603 	if (!cfs_b->timer_active)
3604 		__start_cfs_bandwidth(cfs_b, false);
3605 	raw_spin_unlock(&cfs_b->lock);
3606 }
3607 
3608 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3609 {
3610 	struct rq *rq = rq_of(cfs_rq);
3611 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3612 	struct sched_entity *se;
3613 	int enqueue = 1;
3614 	long task_delta;
3615 
3616 	se = cfs_rq->tg->se[cpu_of(rq)];
3617 
3618 	cfs_rq->throttled = 0;
3619 
3620 	update_rq_clock(rq);
3621 
3622 	raw_spin_lock(&cfs_b->lock);
3623 	cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3624 	list_del_rcu(&cfs_rq->throttled_list);
3625 	raw_spin_unlock(&cfs_b->lock);
3626 
3627 	/* update hierarchical throttle state */
3628 	walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3629 
3630 	if (!cfs_rq->load.weight)
3631 		return;
3632 
3633 	task_delta = cfs_rq->h_nr_running;
3634 	for_each_sched_entity(se) {
3635 		if (se->on_rq)
3636 			enqueue = 0;
3637 
3638 		cfs_rq = cfs_rq_of(se);
3639 		if (enqueue)
3640 			enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3641 		cfs_rq->h_nr_running += task_delta;
3642 
3643 		if (cfs_rq_throttled(cfs_rq))
3644 			break;
3645 	}
3646 
3647 	if (!se)
3648 		add_nr_running(rq, task_delta);
3649 
3650 	/* determine whether we need to wake up potentially idle cpu */
3651 	if (rq->curr == rq->idle && rq->cfs.nr_running)
3652 		resched_curr(rq);
3653 }
3654 
3655 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3656 		u64 remaining, u64 expires)
3657 {
3658 	struct cfs_rq *cfs_rq;
3659 	u64 runtime;
3660 	u64 starting_runtime = remaining;
3661 
3662 	rcu_read_lock();
3663 	list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3664 				throttled_list) {
3665 		struct rq *rq = rq_of(cfs_rq);
3666 
3667 		raw_spin_lock(&rq->lock);
3668 		if (!cfs_rq_throttled(cfs_rq))
3669 			goto next;
3670 
3671 		runtime = -cfs_rq->runtime_remaining + 1;
3672 		if (runtime > remaining)
3673 			runtime = remaining;
3674 		remaining -= runtime;
3675 
3676 		cfs_rq->runtime_remaining += runtime;
3677 		cfs_rq->runtime_expires = expires;
3678 
3679 		/* we check whether we're throttled above */
3680 		if (cfs_rq->runtime_remaining > 0)
3681 			unthrottle_cfs_rq(cfs_rq);
3682 
3683 next:
3684 		raw_spin_unlock(&rq->lock);
3685 
3686 		if (!remaining)
3687 			break;
3688 	}
3689 	rcu_read_unlock();
3690 
3691 	return starting_runtime - remaining;
3692 }
3693 
3694 /*
3695  * Responsible for refilling a task_group's bandwidth and unthrottling its
3696  * cfs_rqs as appropriate. If there has been no activity within the last
3697  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3698  * used to track this state.
3699  */
3700 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3701 {
3702 	u64 runtime, runtime_expires;
3703 	int throttled;
3704 
3705 	/* no need to continue the timer with no bandwidth constraint */
3706 	if (cfs_b->quota == RUNTIME_INF)
3707 		goto out_deactivate;
3708 
3709 	throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3710 	cfs_b->nr_periods += overrun;
3711 
3712 	/*
3713 	 * idle depends on !throttled (for the case of a large deficit), and if
3714 	 * we're going inactive then everything else can be deferred
3715 	 */
3716 	if (cfs_b->idle && !throttled)
3717 		goto out_deactivate;
3718 
3719 	/*
3720 	 * if we have relooped after returning idle once, we need to update our
3721 	 * status as actually running, so that other cpus doing
3722 	 * __start_cfs_bandwidth will stop trying to cancel us.
3723 	 */
3724 	cfs_b->timer_active = 1;
3725 
3726 	__refill_cfs_bandwidth_runtime(cfs_b);
3727 
3728 	if (!throttled) {
3729 		/* mark as potentially idle for the upcoming period */
3730 		cfs_b->idle = 1;
3731 		return 0;
3732 	}
3733 
3734 	/* account preceding periods in which throttling occurred */
3735 	cfs_b->nr_throttled += overrun;
3736 
3737 	runtime_expires = cfs_b->runtime_expires;
3738 
3739 	/*
3740 	 * This check is repeated as we are holding onto the new bandwidth while
3741 	 * we unthrottle. This can potentially race with an unthrottled group
3742 	 * trying to acquire new bandwidth from the global pool. This can result
3743 	 * in us over-using our runtime if it is all used during this loop, but
3744 	 * only by limited amounts in that extreme case.
3745 	 */
3746 	while (throttled && cfs_b->runtime > 0) {
3747 		runtime = cfs_b->runtime;
3748 		raw_spin_unlock(&cfs_b->lock);
3749 		/* we can't nest cfs_b->lock while distributing bandwidth */
3750 		runtime = distribute_cfs_runtime(cfs_b, runtime,
3751 						 runtime_expires);
3752 		raw_spin_lock(&cfs_b->lock);
3753 
3754 		throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3755 
3756 		cfs_b->runtime -= min(runtime, cfs_b->runtime);
3757 	}
3758 
3759 	/*
3760 	 * While we are ensured activity in the period following an
3761 	 * unthrottle, this also covers the case in which the new bandwidth is
3762 	 * insufficient to cover the existing bandwidth deficit.  (Forcing the
3763 	 * timer to remain active while there are any throttled entities.)
3764 	 */
3765 	cfs_b->idle = 0;
3766 
3767 	return 0;
3768 
3769 out_deactivate:
3770 	cfs_b->timer_active = 0;
3771 	return 1;
3772 }
3773 
3774 /* a cfs_rq won't donate quota below this amount */
3775 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3776 /* minimum remaining period time to redistribute slack quota */
3777 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3778 /* how long we wait to gather additional slack before distributing */
3779 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3780 
3781 /*
3782  * Are we near the end of the current quota period?
3783  *
3784  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3785  * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3786  * migrate_hrtimers, base is never cleared, so we are fine.
3787  */
3788 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3789 {
3790 	struct hrtimer *refresh_timer = &cfs_b->period_timer;
3791 	u64 remaining;
3792 
3793 	/* if the call-back is running a quota refresh is already occurring */
3794 	if (hrtimer_callback_running(refresh_timer))
3795 		return 1;
3796 
3797 	/* is a quota refresh about to occur? */
3798 	remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3799 	if (remaining < min_expire)
3800 		return 1;
3801 
3802 	return 0;
3803 }
3804 
3805 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3806 {
3807 	u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3808 
3809 	/* if there's a quota refresh soon don't bother with slack */
3810 	if (runtime_refresh_within(cfs_b, min_left))
3811 		return;
3812 
3813 	start_bandwidth_timer(&cfs_b->slack_timer,
3814 				ns_to_ktime(cfs_bandwidth_slack_period));
3815 }
3816 
3817 /* we know any runtime found here is valid as update_curr() precedes return */
3818 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3819 {
3820 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3821 	s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3822 
3823 	if (slack_runtime <= 0)
3824 		return;
3825 
3826 	raw_spin_lock(&cfs_b->lock);
3827 	if (cfs_b->quota != RUNTIME_INF &&
3828 	    cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3829 		cfs_b->runtime += slack_runtime;
3830 
3831 		/* we are under rq->lock, defer unthrottling using a timer */
3832 		if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3833 		    !list_empty(&cfs_b->throttled_cfs_rq))
3834 			start_cfs_slack_bandwidth(cfs_b);
3835 	}
3836 	raw_spin_unlock(&cfs_b->lock);
3837 
3838 	/* even if it's not valid for return we don't want to try again */
3839 	cfs_rq->runtime_remaining -= slack_runtime;
3840 }
3841 
3842 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3843 {
3844 	if (!cfs_bandwidth_used())
3845 		return;
3846 
3847 	if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3848 		return;
3849 
3850 	__return_cfs_rq_runtime(cfs_rq);
3851 }
3852 
3853 /*
3854  * This is done with a timer (instead of inline with bandwidth return) since
3855  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3856  */
3857 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3858 {
3859 	u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3860 	u64 expires;
3861 
3862 	/* confirm we're still not at a refresh boundary */
3863 	raw_spin_lock(&cfs_b->lock);
3864 	if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3865 		raw_spin_unlock(&cfs_b->lock);
3866 		return;
3867 	}
3868 
3869 	if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3870 		runtime = cfs_b->runtime;
3871 
3872 	expires = cfs_b->runtime_expires;
3873 	raw_spin_unlock(&cfs_b->lock);
3874 
3875 	if (!runtime)
3876 		return;
3877 
3878 	runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3879 
3880 	raw_spin_lock(&cfs_b->lock);
3881 	if (expires == cfs_b->runtime_expires)
3882 		cfs_b->runtime -= min(runtime, cfs_b->runtime);
3883 	raw_spin_unlock(&cfs_b->lock);
3884 }
3885 
3886 /*
3887  * When a group wakes up we want to make sure that its quota is not already
3888  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3889  * runtime as update_curr() throttling can not not trigger until it's on-rq.
3890  */
3891 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3892 {
3893 	if (!cfs_bandwidth_used())
3894 		return;
3895 
3896 	/* an active group must be handled by the update_curr()->put() path */
3897 	if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3898 		return;
3899 
3900 	/* ensure the group is not already throttled */
3901 	if (cfs_rq_throttled(cfs_rq))
3902 		return;
3903 
3904 	/* update runtime allocation */
3905 	account_cfs_rq_runtime(cfs_rq, 0);
3906 	if (cfs_rq->runtime_remaining <= 0)
3907 		throttle_cfs_rq(cfs_rq);
3908 }
3909 
3910 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3911 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3912 {
3913 	if (!cfs_bandwidth_used())
3914 		return false;
3915 
3916 	if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3917 		return false;
3918 
3919 	/*
3920 	 * it's possible for a throttled entity to be forced into a running
3921 	 * state (e.g. set_curr_task), in this case we're finished.
3922 	 */
3923 	if (cfs_rq_throttled(cfs_rq))
3924 		return true;
3925 
3926 	throttle_cfs_rq(cfs_rq);
3927 	return true;
3928 }
3929 
3930 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3931 {
3932 	struct cfs_bandwidth *cfs_b =
3933 		container_of(timer, struct cfs_bandwidth, slack_timer);
3934 	do_sched_cfs_slack_timer(cfs_b);
3935 
3936 	return HRTIMER_NORESTART;
3937 }
3938 
3939 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3940 {
3941 	struct cfs_bandwidth *cfs_b =
3942 		container_of(timer, struct cfs_bandwidth, period_timer);
3943 	ktime_t now;
3944 	int overrun;
3945 	int idle = 0;
3946 
3947 	raw_spin_lock(&cfs_b->lock);
3948 	for (;;) {
3949 		now = hrtimer_cb_get_time(timer);
3950 		overrun = hrtimer_forward(timer, now, cfs_b->period);
3951 
3952 		if (!overrun)
3953 			break;
3954 
3955 		idle = do_sched_cfs_period_timer(cfs_b, overrun);
3956 	}
3957 	raw_spin_unlock(&cfs_b->lock);
3958 
3959 	return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3960 }
3961 
3962 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3963 {
3964 	raw_spin_lock_init(&cfs_b->lock);
3965 	cfs_b->runtime = 0;
3966 	cfs_b->quota = RUNTIME_INF;
3967 	cfs_b->period = ns_to_ktime(default_cfs_period());
3968 
3969 	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3970 	hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3971 	cfs_b->period_timer.function = sched_cfs_period_timer;
3972 	hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3973 	cfs_b->slack_timer.function = sched_cfs_slack_timer;
3974 }
3975 
3976 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3977 {
3978 	cfs_rq->runtime_enabled = 0;
3979 	INIT_LIST_HEAD(&cfs_rq->throttled_list);
3980 }
3981 
3982 /* requires cfs_b->lock, may release to reprogram timer */
3983 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
3984 {
3985 	/*
3986 	 * The timer may be active because we're trying to set a new bandwidth
3987 	 * period or because we're racing with the tear-down path
3988 	 * (timer_active==0 becomes visible before the hrtimer call-back
3989 	 * terminates).  In either case we ensure that it's re-programmed
3990 	 */
3991 	while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3992 	       hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3993 		/* bounce the lock to allow do_sched_cfs_period_timer to run */
3994 		raw_spin_unlock(&cfs_b->lock);
3995 		cpu_relax();
3996 		raw_spin_lock(&cfs_b->lock);
3997 		/* if someone else restarted the timer then we're done */
3998 		if (!force && cfs_b->timer_active)
3999 			return;
4000 	}
4001 
4002 	cfs_b->timer_active = 1;
4003 	start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
4004 }
4005 
4006 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4007 {
4008 	hrtimer_cancel(&cfs_b->period_timer);
4009 	hrtimer_cancel(&cfs_b->slack_timer);
4010 }
4011 
4012 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4013 {
4014 	struct cfs_rq *cfs_rq;
4015 
4016 	for_each_leaf_cfs_rq(rq, cfs_rq) {
4017 		struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4018 
4019 		raw_spin_lock(&cfs_b->lock);
4020 		cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4021 		raw_spin_unlock(&cfs_b->lock);
4022 	}
4023 }
4024 
4025 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4026 {
4027 	struct cfs_rq *cfs_rq;
4028 
4029 	for_each_leaf_cfs_rq(rq, cfs_rq) {
4030 		if (!cfs_rq->runtime_enabled)
4031 			continue;
4032 
4033 		/*
4034 		 * clock_task is not advancing so we just need to make sure
4035 		 * there's some valid quota amount
4036 		 */
4037 		cfs_rq->runtime_remaining = 1;
4038 		/*
4039 		 * Offline rq is schedulable till cpu is completely disabled
4040 		 * in take_cpu_down(), so we prevent new cfs throttling here.
4041 		 */
4042 		cfs_rq->runtime_enabled = 0;
4043 
4044 		if (cfs_rq_throttled(cfs_rq))
4045 			unthrottle_cfs_rq(cfs_rq);
4046 	}
4047 }
4048 
4049 #else /* CONFIG_CFS_BANDWIDTH */
4050 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4051 {
4052 	return rq_clock_task(rq_of(cfs_rq));
4053 }
4054 
4055 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4056 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4057 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4058 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4059 
4060 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4061 {
4062 	return 0;
4063 }
4064 
4065 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4066 {
4067 	return 0;
4068 }
4069 
4070 static inline int throttled_lb_pair(struct task_group *tg,
4071 				    int src_cpu, int dest_cpu)
4072 {
4073 	return 0;
4074 }
4075 
4076 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4077 
4078 #ifdef CONFIG_FAIR_GROUP_SCHED
4079 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4080 #endif
4081 
4082 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4083 {
4084 	return NULL;
4085 }
4086 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4087 static inline void update_runtime_enabled(struct rq *rq) {}
4088 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4089 
4090 #endif /* CONFIG_CFS_BANDWIDTH */
4091 
4092 /**************************************************
4093  * CFS operations on tasks:
4094  */
4095 
4096 #ifdef CONFIG_SCHED_HRTICK
4097 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4098 {
4099 	struct sched_entity *se = &p->se;
4100 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4101 
4102 	WARN_ON(task_rq(p) != rq);
4103 
4104 	if (cfs_rq->nr_running > 1) {
4105 		u64 slice = sched_slice(cfs_rq, se);
4106 		u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4107 		s64 delta = slice - ran;
4108 
4109 		if (delta < 0) {
4110 			if (rq->curr == p)
4111 				resched_curr(rq);
4112 			return;
4113 		}
4114 		hrtick_start(rq, delta);
4115 	}
4116 }
4117 
4118 /*
4119  * called from enqueue/dequeue and updates the hrtick when the
4120  * current task is from our class and nr_running is low enough
4121  * to matter.
4122  */
4123 static void hrtick_update(struct rq *rq)
4124 {
4125 	struct task_struct *curr = rq->curr;
4126 
4127 	if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4128 		return;
4129 
4130 	if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4131 		hrtick_start_fair(rq, curr);
4132 }
4133 #else /* !CONFIG_SCHED_HRTICK */
4134 static inline void
4135 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4136 {
4137 }
4138 
4139 static inline void hrtick_update(struct rq *rq)
4140 {
4141 }
4142 #endif
4143 
4144 /*
4145  * The enqueue_task method is called before nr_running is
4146  * increased. Here we update the fair scheduling stats and
4147  * then put the task into the rbtree:
4148  */
4149 static void
4150 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4151 {
4152 	struct cfs_rq *cfs_rq;
4153 	struct sched_entity *se = &p->se;
4154 
4155 	for_each_sched_entity(se) {
4156 		if (se->on_rq)
4157 			break;
4158 		cfs_rq = cfs_rq_of(se);
4159 		enqueue_entity(cfs_rq, se, flags);
4160 
4161 		/*
4162 		 * end evaluation on encountering a throttled cfs_rq
4163 		 *
4164 		 * note: in the case of encountering a throttled cfs_rq we will
4165 		 * post the final h_nr_running increment below.
4166 		*/
4167 		if (cfs_rq_throttled(cfs_rq))
4168 			break;
4169 		cfs_rq->h_nr_running++;
4170 
4171 		flags = ENQUEUE_WAKEUP;
4172 	}
4173 
4174 	for_each_sched_entity(se) {
4175 		cfs_rq = cfs_rq_of(se);
4176 		cfs_rq->h_nr_running++;
4177 
4178 		if (cfs_rq_throttled(cfs_rq))
4179 			break;
4180 
4181 		update_cfs_shares(cfs_rq);
4182 		update_entity_load_avg(se, 1);
4183 	}
4184 
4185 	if (!se) {
4186 		update_rq_runnable_avg(rq, rq->nr_running);
4187 		add_nr_running(rq, 1);
4188 	}
4189 	hrtick_update(rq);
4190 }
4191 
4192 static void set_next_buddy(struct sched_entity *se);
4193 
4194 /*
4195  * The dequeue_task method is called before nr_running is
4196  * decreased. We remove the task from the rbtree and
4197  * update the fair scheduling stats:
4198  */
4199 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4200 {
4201 	struct cfs_rq *cfs_rq;
4202 	struct sched_entity *se = &p->se;
4203 	int task_sleep = flags & DEQUEUE_SLEEP;
4204 
4205 	for_each_sched_entity(se) {
4206 		cfs_rq = cfs_rq_of(se);
4207 		dequeue_entity(cfs_rq, se, flags);
4208 
4209 		/*
4210 		 * end evaluation on encountering a throttled cfs_rq
4211 		 *
4212 		 * note: in the case of encountering a throttled cfs_rq we will
4213 		 * post the final h_nr_running decrement below.
4214 		*/
4215 		if (cfs_rq_throttled(cfs_rq))
4216 			break;
4217 		cfs_rq->h_nr_running--;
4218 
4219 		/* Don't dequeue parent if it has other entities besides us */
4220 		if (cfs_rq->load.weight) {
4221 			/*
4222 			 * Bias pick_next to pick a task from this cfs_rq, as
4223 			 * p is sleeping when it is within its sched_slice.
4224 			 */
4225 			if (task_sleep && parent_entity(se))
4226 				set_next_buddy(parent_entity(se));
4227 
4228 			/* avoid re-evaluating load for this entity */
4229 			se = parent_entity(se);
4230 			break;
4231 		}
4232 		flags |= DEQUEUE_SLEEP;
4233 	}
4234 
4235 	for_each_sched_entity(se) {
4236 		cfs_rq = cfs_rq_of(se);
4237 		cfs_rq->h_nr_running--;
4238 
4239 		if (cfs_rq_throttled(cfs_rq))
4240 			break;
4241 
4242 		update_cfs_shares(cfs_rq);
4243 		update_entity_load_avg(se, 1);
4244 	}
4245 
4246 	if (!se) {
4247 		sub_nr_running(rq, 1);
4248 		update_rq_runnable_avg(rq, 1);
4249 	}
4250 	hrtick_update(rq);
4251 }
4252 
4253 #ifdef CONFIG_SMP
4254 /* Used instead of source_load when we know the type == 0 */
4255 static unsigned long weighted_cpuload(const int cpu)
4256 {
4257 	return cpu_rq(cpu)->cfs.runnable_load_avg;
4258 }
4259 
4260 /*
4261  * Return a low guess at the load of a migration-source cpu weighted
4262  * according to the scheduling class and "nice" value.
4263  *
4264  * We want to under-estimate the load of migration sources, to
4265  * balance conservatively.
4266  */
4267 static unsigned long source_load(int cpu, int type)
4268 {
4269 	struct rq *rq = cpu_rq(cpu);
4270 	unsigned long total = weighted_cpuload(cpu);
4271 
4272 	if (type == 0 || !sched_feat(LB_BIAS))
4273 		return total;
4274 
4275 	return min(rq->cpu_load[type-1], total);
4276 }
4277 
4278 /*
4279  * Return a high guess at the load of a migration-target cpu weighted
4280  * according to the scheduling class and "nice" value.
4281  */
4282 static unsigned long target_load(int cpu, int type)
4283 {
4284 	struct rq *rq = cpu_rq(cpu);
4285 	unsigned long total = weighted_cpuload(cpu);
4286 
4287 	if (type == 0 || !sched_feat(LB_BIAS))
4288 		return total;
4289 
4290 	return max(rq->cpu_load[type-1], total);
4291 }
4292 
4293 static unsigned long capacity_of(int cpu)
4294 {
4295 	return cpu_rq(cpu)->cpu_capacity;
4296 }
4297 
4298 static unsigned long cpu_avg_load_per_task(int cpu)
4299 {
4300 	struct rq *rq = cpu_rq(cpu);
4301 	unsigned long nr_running = ACCESS_ONCE(rq->cfs.h_nr_running);
4302 	unsigned long load_avg = rq->cfs.runnable_load_avg;
4303 
4304 	if (nr_running)
4305 		return load_avg / nr_running;
4306 
4307 	return 0;
4308 }
4309 
4310 static void record_wakee(struct task_struct *p)
4311 {
4312 	/*
4313 	 * Rough decay (wiping) for cost saving, don't worry
4314 	 * about the boundary, really active task won't care
4315 	 * about the loss.
4316 	 */
4317 	if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4318 		current->wakee_flips >>= 1;
4319 		current->wakee_flip_decay_ts = jiffies;
4320 	}
4321 
4322 	if (current->last_wakee != p) {
4323 		current->last_wakee = p;
4324 		current->wakee_flips++;
4325 	}
4326 }
4327 
4328 static void task_waking_fair(struct task_struct *p)
4329 {
4330 	struct sched_entity *se = &p->se;
4331 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4332 	u64 min_vruntime;
4333 
4334 #ifndef CONFIG_64BIT
4335 	u64 min_vruntime_copy;
4336 
4337 	do {
4338 		min_vruntime_copy = cfs_rq->min_vruntime_copy;
4339 		smp_rmb();
4340 		min_vruntime = cfs_rq->min_vruntime;
4341 	} while (min_vruntime != min_vruntime_copy);
4342 #else
4343 	min_vruntime = cfs_rq->min_vruntime;
4344 #endif
4345 
4346 	se->vruntime -= min_vruntime;
4347 	record_wakee(p);
4348 }
4349 
4350 #ifdef CONFIG_FAIR_GROUP_SCHED
4351 /*
4352  * effective_load() calculates the load change as seen from the root_task_group
4353  *
4354  * Adding load to a group doesn't make a group heavier, but can cause movement
4355  * of group shares between cpus. Assuming the shares were perfectly aligned one
4356  * can calculate the shift in shares.
4357  *
4358  * Calculate the effective load difference if @wl is added (subtracted) to @tg
4359  * on this @cpu and results in a total addition (subtraction) of @wg to the
4360  * total group weight.
4361  *
4362  * Given a runqueue weight distribution (rw_i) we can compute a shares
4363  * distribution (s_i) using:
4364  *
4365  *   s_i = rw_i / \Sum rw_j						(1)
4366  *
4367  * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4368  * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4369  * shares distribution (s_i):
4370  *
4371  *   rw_i = {   2,   4,   1,   0 }
4372  *   s_i  = { 2/7, 4/7, 1/7,   0 }
4373  *
4374  * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4375  * task used to run on and the CPU the waker is running on), we need to
4376  * compute the effect of waking a task on either CPU and, in case of a sync
4377  * wakeup, compute the effect of the current task going to sleep.
4378  *
4379  * So for a change of @wl to the local @cpu with an overall group weight change
4380  * of @wl we can compute the new shares distribution (s'_i) using:
4381  *
4382  *   s'_i = (rw_i + @wl) / (@wg + \Sum rw_j)				(2)
4383  *
4384  * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4385  * differences in waking a task to CPU 0. The additional task changes the
4386  * weight and shares distributions like:
4387  *
4388  *   rw'_i = {   3,   4,   1,   0 }
4389  *   s'_i  = { 3/8, 4/8, 1/8,   0 }
4390  *
4391  * We can then compute the difference in effective weight by using:
4392  *
4393  *   dw_i = S * (s'_i - s_i)						(3)
4394  *
4395  * Where 'S' is the group weight as seen by its parent.
4396  *
4397  * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4398  * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4399  * 4/7) times the weight of the group.
4400  */
4401 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4402 {
4403 	struct sched_entity *se = tg->se[cpu];
4404 
4405 	if (!tg->parent)	/* the trivial, non-cgroup case */
4406 		return wl;
4407 
4408 	for_each_sched_entity(se) {
4409 		long w, W;
4410 
4411 		tg = se->my_q->tg;
4412 
4413 		/*
4414 		 * W = @wg + \Sum rw_j
4415 		 */
4416 		W = wg + calc_tg_weight(tg, se->my_q);
4417 
4418 		/*
4419 		 * w = rw_i + @wl
4420 		 */
4421 		w = se->my_q->load.weight + wl;
4422 
4423 		/*
4424 		 * wl = S * s'_i; see (2)
4425 		 */
4426 		if (W > 0 && w < W)
4427 			wl = (w * tg->shares) / W;
4428 		else
4429 			wl = tg->shares;
4430 
4431 		/*
4432 		 * Per the above, wl is the new se->load.weight value; since
4433 		 * those are clipped to [MIN_SHARES, ...) do so now. See
4434 		 * calc_cfs_shares().
4435 		 */
4436 		if (wl < MIN_SHARES)
4437 			wl = MIN_SHARES;
4438 
4439 		/*
4440 		 * wl = dw_i = S * (s'_i - s_i); see (3)
4441 		 */
4442 		wl -= se->load.weight;
4443 
4444 		/*
4445 		 * Recursively apply this logic to all parent groups to compute
4446 		 * the final effective load change on the root group. Since
4447 		 * only the @tg group gets extra weight, all parent groups can
4448 		 * only redistribute existing shares. @wl is the shift in shares
4449 		 * resulting from this level per the above.
4450 		 */
4451 		wg = 0;
4452 	}
4453 
4454 	return wl;
4455 }
4456 #else
4457 
4458 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4459 {
4460 	return wl;
4461 }
4462 
4463 #endif
4464 
4465 static int wake_wide(struct task_struct *p)
4466 {
4467 	int factor = this_cpu_read(sd_llc_size);
4468 
4469 	/*
4470 	 * Yeah, it's the switching-frequency, could means many wakee or
4471 	 * rapidly switch, use factor here will just help to automatically
4472 	 * adjust the loose-degree, so bigger node will lead to more pull.
4473 	 */
4474 	if (p->wakee_flips > factor) {
4475 		/*
4476 		 * wakee is somewhat hot, it needs certain amount of cpu
4477 		 * resource, so if waker is far more hot, prefer to leave
4478 		 * it alone.
4479 		 */
4480 		if (current->wakee_flips > (factor * p->wakee_flips))
4481 			return 1;
4482 	}
4483 
4484 	return 0;
4485 }
4486 
4487 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4488 {
4489 	s64 this_load, load;
4490 	s64 this_eff_load, prev_eff_load;
4491 	int idx, this_cpu, prev_cpu;
4492 	struct task_group *tg;
4493 	unsigned long weight;
4494 	int balanced;
4495 
4496 	/*
4497 	 * If we wake multiple tasks be careful to not bounce
4498 	 * ourselves around too much.
4499 	 */
4500 	if (wake_wide(p))
4501 		return 0;
4502 
4503 	idx	  = sd->wake_idx;
4504 	this_cpu  = smp_processor_id();
4505 	prev_cpu  = task_cpu(p);
4506 	load	  = source_load(prev_cpu, idx);
4507 	this_load = target_load(this_cpu, idx);
4508 
4509 	/*
4510 	 * If sync wakeup then subtract the (maximum possible)
4511 	 * effect of the currently running task from the load
4512 	 * of the current CPU:
4513 	 */
4514 	if (sync) {
4515 		tg = task_group(current);
4516 		weight = current->se.load.weight;
4517 
4518 		this_load += effective_load(tg, this_cpu, -weight, -weight);
4519 		load += effective_load(tg, prev_cpu, 0, -weight);
4520 	}
4521 
4522 	tg = task_group(p);
4523 	weight = p->se.load.weight;
4524 
4525 	/*
4526 	 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4527 	 * due to the sync cause above having dropped this_load to 0, we'll
4528 	 * always have an imbalance, but there's really nothing you can do
4529 	 * about that, so that's good too.
4530 	 *
4531 	 * Otherwise check if either cpus are near enough in load to allow this
4532 	 * task to be woken on this_cpu.
4533 	 */
4534 	this_eff_load = 100;
4535 	this_eff_load *= capacity_of(prev_cpu);
4536 
4537 	prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4538 	prev_eff_load *= capacity_of(this_cpu);
4539 
4540 	if (this_load > 0) {
4541 		this_eff_load *= this_load +
4542 			effective_load(tg, this_cpu, weight, weight);
4543 
4544 		prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4545 	}
4546 
4547 	balanced = this_eff_load <= prev_eff_load;
4548 
4549 	schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4550 
4551 	if (!balanced)
4552 		return 0;
4553 
4554 	schedstat_inc(sd, ttwu_move_affine);
4555 	schedstat_inc(p, se.statistics.nr_wakeups_affine);
4556 
4557 	return 1;
4558 }
4559 
4560 /*
4561  * find_idlest_group finds and returns the least busy CPU group within the
4562  * domain.
4563  */
4564 static struct sched_group *
4565 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4566 		  int this_cpu, int sd_flag)
4567 {
4568 	struct sched_group *idlest = NULL, *group = sd->groups;
4569 	unsigned long min_load = ULONG_MAX, this_load = 0;
4570 	int load_idx = sd->forkexec_idx;
4571 	int imbalance = 100 + (sd->imbalance_pct-100)/2;
4572 
4573 	if (sd_flag & SD_BALANCE_WAKE)
4574 		load_idx = sd->wake_idx;
4575 
4576 	do {
4577 		unsigned long load, avg_load;
4578 		int local_group;
4579 		int i;
4580 
4581 		/* Skip over this group if it has no CPUs allowed */
4582 		if (!cpumask_intersects(sched_group_cpus(group),
4583 					tsk_cpus_allowed(p)))
4584 			continue;
4585 
4586 		local_group = cpumask_test_cpu(this_cpu,
4587 					       sched_group_cpus(group));
4588 
4589 		/* Tally up the load of all CPUs in the group */
4590 		avg_load = 0;
4591 
4592 		for_each_cpu(i, sched_group_cpus(group)) {
4593 			/* Bias balancing toward cpus of our domain */
4594 			if (local_group)
4595 				load = source_load(i, load_idx);
4596 			else
4597 				load = target_load(i, load_idx);
4598 
4599 			avg_load += load;
4600 		}
4601 
4602 		/* Adjust by relative CPU capacity of the group */
4603 		avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4604 
4605 		if (local_group) {
4606 			this_load = avg_load;
4607 		} else if (avg_load < min_load) {
4608 			min_load = avg_load;
4609 			idlest = group;
4610 		}
4611 	} while (group = group->next, group != sd->groups);
4612 
4613 	if (!idlest || 100*this_load < imbalance*min_load)
4614 		return NULL;
4615 	return idlest;
4616 }
4617 
4618 /*
4619  * find_idlest_cpu - find the idlest cpu among the cpus in group.
4620  */
4621 static int
4622 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4623 {
4624 	unsigned long load, min_load = ULONG_MAX;
4625 	unsigned int min_exit_latency = UINT_MAX;
4626 	u64 latest_idle_timestamp = 0;
4627 	int least_loaded_cpu = this_cpu;
4628 	int shallowest_idle_cpu = -1;
4629 	int i;
4630 
4631 	/* Traverse only the allowed CPUs */
4632 	for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4633 		if (idle_cpu(i)) {
4634 			struct rq *rq = cpu_rq(i);
4635 			struct cpuidle_state *idle = idle_get_state(rq);
4636 			if (idle && idle->exit_latency < min_exit_latency) {
4637 				/*
4638 				 * We give priority to a CPU whose idle state
4639 				 * has the smallest exit latency irrespective
4640 				 * of any idle timestamp.
4641 				 */
4642 				min_exit_latency = idle->exit_latency;
4643 				latest_idle_timestamp = rq->idle_stamp;
4644 				shallowest_idle_cpu = i;
4645 			} else if ((!idle || idle->exit_latency == min_exit_latency) &&
4646 				   rq->idle_stamp > latest_idle_timestamp) {
4647 				/*
4648 				 * If equal or no active idle state, then
4649 				 * the most recently idled CPU might have
4650 				 * a warmer cache.
4651 				 */
4652 				latest_idle_timestamp = rq->idle_stamp;
4653 				shallowest_idle_cpu = i;
4654 			}
4655 		} else if (shallowest_idle_cpu == -1) {
4656 			load = weighted_cpuload(i);
4657 			if (load < min_load || (load == min_load && i == this_cpu)) {
4658 				min_load = load;
4659 				least_loaded_cpu = i;
4660 			}
4661 		}
4662 	}
4663 
4664 	return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4665 }
4666 
4667 /*
4668  * Try and locate an idle CPU in the sched_domain.
4669  */
4670 static int select_idle_sibling(struct task_struct *p, int target)
4671 {
4672 	struct sched_domain *sd;
4673 	struct sched_group *sg;
4674 	int i = task_cpu(p);
4675 
4676 	if (idle_cpu(target))
4677 		return target;
4678 
4679 	/*
4680 	 * If the prevous cpu is cache affine and idle, don't be stupid.
4681 	 */
4682 	if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4683 		return i;
4684 
4685 	/*
4686 	 * Otherwise, iterate the domains and find an elegible idle cpu.
4687 	 */
4688 	sd = rcu_dereference(per_cpu(sd_llc, target));
4689 	for_each_lower_domain(sd) {
4690 		sg = sd->groups;
4691 		do {
4692 			if (!cpumask_intersects(sched_group_cpus(sg),
4693 						tsk_cpus_allowed(p)))
4694 				goto next;
4695 
4696 			for_each_cpu(i, sched_group_cpus(sg)) {
4697 				if (i == target || !idle_cpu(i))
4698 					goto next;
4699 			}
4700 
4701 			target = cpumask_first_and(sched_group_cpus(sg),
4702 					tsk_cpus_allowed(p));
4703 			goto done;
4704 next:
4705 			sg = sg->next;
4706 		} while (sg != sd->groups);
4707 	}
4708 done:
4709 	return target;
4710 }
4711 
4712 /*
4713  * select_task_rq_fair: Select target runqueue for the waking task in domains
4714  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4715  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4716  *
4717  * Balances load by selecting the idlest cpu in the idlest group, or under
4718  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4719  *
4720  * Returns the target cpu number.
4721  *
4722  * preempt must be disabled.
4723  */
4724 static int
4725 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4726 {
4727 	struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4728 	int cpu = smp_processor_id();
4729 	int new_cpu = cpu;
4730 	int want_affine = 0;
4731 	int sync = wake_flags & WF_SYNC;
4732 
4733 	if (sd_flag & SD_BALANCE_WAKE)
4734 		want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4735 
4736 	rcu_read_lock();
4737 	for_each_domain(cpu, tmp) {
4738 		if (!(tmp->flags & SD_LOAD_BALANCE))
4739 			continue;
4740 
4741 		/*
4742 		 * If both cpu and prev_cpu are part of this domain,
4743 		 * cpu is a valid SD_WAKE_AFFINE target.
4744 		 */
4745 		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4746 		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4747 			affine_sd = tmp;
4748 			break;
4749 		}
4750 
4751 		if (tmp->flags & sd_flag)
4752 			sd = tmp;
4753 	}
4754 
4755 	if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4756 		prev_cpu = cpu;
4757 
4758 	if (sd_flag & SD_BALANCE_WAKE) {
4759 		new_cpu = select_idle_sibling(p, prev_cpu);
4760 		goto unlock;
4761 	}
4762 
4763 	while (sd) {
4764 		struct sched_group *group;
4765 		int weight;
4766 
4767 		if (!(sd->flags & sd_flag)) {
4768 			sd = sd->child;
4769 			continue;
4770 		}
4771 
4772 		group = find_idlest_group(sd, p, cpu, sd_flag);
4773 		if (!group) {
4774 			sd = sd->child;
4775 			continue;
4776 		}
4777 
4778 		new_cpu = find_idlest_cpu(group, p, cpu);
4779 		if (new_cpu == -1 || new_cpu == cpu) {
4780 			/* Now try balancing at a lower domain level of cpu */
4781 			sd = sd->child;
4782 			continue;
4783 		}
4784 
4785 		/* Now try balancing at a lower domain level of new_cpu */
4786 		cpu = new_cpu;
4787 		weight = sd->span_weight;
4788 		sd = NULL;
4789 		for_each_domain(cpu, tmp) {
4790 			if (weight <= tmp->span_weight)
4791 				break;
4792 			if (tmp->flags & sd_flag)
4793 				sd = tmp;
4794 		}
4795 		/* while loop will break here if sd == NULL */
4796 	}
4797 unlock:
4798 	rcu_read_unlock();
4799 
4800 	return new_cpu;
4801 }
4802 
4803 /*
4804  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4805  * cfs_rq_of(p) references at time of call are still valid and identify the
4806  * previous cpu.  However, the caller only guarantees p->pi_lock is held; no
4807  * other assumptions, including the state of rq->lock, should be made.
4808  */
4809 static void
4810 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4811 {
4812 	struct sched_entity *se = &p->se;
4813 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4814 
4815 	/*
4816 	 * Load tracking: accumulate removed load so that it can be processed
4817 	 * when we next update owning cfs_rq under rq->lock.  Tasks contribute
4818 	 * to blocked load iff they have a positive decay-count.  It can never
4819 	 * be negative here since on-rq tasks have decay-count == 0.
4820 	 */
4821 	if (se->avg.decay_count) {
4822 		se->avg.decay_count = -__synchronize_entity_decay(se);
4823 		atomic_long_add(se->avg.load_avg_contrib,
4824 						&cfs_rq->removed_load);
4825 	}
4826 
4827 	/* We have migrated, no longer consider this task hot */
4828 	se->exec_start = 0;
4829 }
4830 #endif /* CONFIG_SMP */
4831 
4832 static unsigned long
4833 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4834 {
4835 	unsigned long gran = sysctl_sched_wakeup_granularity;
4836 
4837 	/*
4838 	 * Since its curr running now, convert the gran from real-time
4839 	 * to virtual-time in his units.
4840 	 *
4841 	 * By using 'se' instead of 'curr' we penalize light tasks, so
4842 	 * they get preempted easier. That is, if 'se' < 'curr' then
4843 	 * the resulting gran will be larger, therefore penalizing the
4844 	 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4845 	 * be smaller, again penalizing the lighter task.
4846 	 *
4847 	 * This is especially important for buddies when the leftmost
4848 	 * task is higher priority than the buddy.
4849 	 */
4850 	return calc_delta_fair(gran, se);
4851 }
4852 
4853 /*
4854  * Should 'se' preempt 'curr'.
4855  *
4856  *             |s1
4857  *        |s2
4858  *   |s3
4859  *         g
4860  *      |<--->|c
4861  *
4862  *  w(c, s1) = -1
4863  *  w(c, s2) =  0
4864  *  w(c, s3) =  1
4865  *
4866  */
4867 static int
4868 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4869 {
4870 	s64 gran, vdiff = curr->vruntime - se->vruntime;
4871 
4872 	if (vdiff <= 0)
4873 		return -1;
4874 
4875 	gran = wakeup_gran(curr, se);
4876 	if (vdiff > gran)
4877 		return 1;
4878 
4879 	return 0;
4880 }
4881 
4882 static void set_last_buddy(struct sched_entity *se)
4883 {
4884 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4885 		return;
4886 
4887 	for_each_sched_entity(se)
4888 		cfs_rq_of(se)->last = se;
4889 }
4890 
4891 static void set_next_buddy(struct sched_entity *se)
4892 {
4893 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4894 		return;
4895 
4896 	for_each_sched_entity(se)
4897 		cfs_rq_of(se)->next = se;
4898 }
4899 
4900 static void set_skip_buddy(struct sched_entity *se)
4901 {
4902 	for_each_sched_entity(se)
4903 		cfs_rq_of(se)->skip = se;
4904 }
4905 
4906 /*
4907  * Preempt the current task with a newly woken task if needed:
4908  */
4909 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4910 {
4911 	struct task_struct *curr = rq->curr;
4912 	struct sched_entity *se = &curr->se, *pse = &p->se;
4913 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4914 	int scale = cfs_rq->nr_running >= sched_nr_latency;
4915 	int next_buddy_marked = 0;
4916 
4917 	if (unlikely(se == pse))
4918 		return;
4919 
4920 	/*
4921 	 * This is possible from callers such as attach_tasks(), in which we
4922 	 * unconditionally check_prempt_curr() after an enqueue (which may have
4923 	 * lead to a throttle).  This both saves work and prevents false
4924 	 * next-buddy nomination below.
4925 	 */
4926 	if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4927 		return;
4928 
4929 	if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4930 		set_next_buddy(pse);
4931 		next_buddy_marked = 1;
4932 	}
4933 
4934 	/*
4935 	 * We can come here with TIF_NEED_RESCHED already set from new task
4936 	 * wake up path.
4937 	 *
4938 	 * Note: this also catches the edge-case of curr being in a throttled
4939 	 * group (e.g. via set_curr_task), since update_curr() (in the
4940 	 * enqueue of curr) will have resulted in resched being set.  This
4941 	 * prevents us from potentially nominating it as a false LAST_BUDDY
4942 	 * below.
4943 	 */
4944 	if (test_tsk_need_resched(curr))
4945 		return;
4946 
4947 	/* Idle tasks are by definition preempted by non-idle tasks. */
4948 	if (unlikely(curr->policy == SCHED_IDLE) &&
4949 	    likely(p->policy != SCHED_IDLE))
4950 		goto preempt;
4951 
4952 	/*
4953 	 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4954 	 * is driven by the tick):
4955 	 */
4956 	if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4957 		return;
4958 
4959 	find_matching_se(&se, &pse);
4960 	update_curr(cfs_rq_of(se));
4961 	BUG_ON(!pse);
4962 	if (wakeup_preempt_entity(se, pse) == 1) {
4963 		/*
4964 		 * Bias pick_next to pick the sched entity that is
4965 		 * triggering this preemption.
4966 		 */
4967 		if (!next_buddy_marked)
4968 			set_next_buddy(pse);
4969 		goto preempt;
4970 	}
4971 
4972 	return;
4973 
4974 preempt:
4975 	resched_curr(rq);
4976 	/*
4977 	 * Only set the backward buddy when the current task is still
4978 	 * on the rq. This can happen when a wakeup gets interleaved
4979 	 * with schedule on the ->pre_schedule() or idle_balance()
4980 	 * point, either of which can * drop the rq lock.
4981 	 *
4982 	 * Also, during early boot the idle thread is in the fair class,
4983 	 * for obvious reasons its a bad idea to schedule back to it.
4984 	 */
4985 	if (unlikely(!se->on_rq || curr == rq->idle))
4986 		return;
4987 
4988 	if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4989 		set_last_buddy(se);
4990 }
4991 
4992 static struct task_struct *
4993 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4994 {
4995 	struct cfs_rq *cfs_rq = &rq->cfs;
4996 	struct sched_entity *se;
4997 	struct task_struct *p;
4998 	int new_tasks;
4999 
5000 again:
5001 #ifdef CONFIG_FAIR_GROUP_SCHED
5002 	if (!cfs_rq->nr_running)
5003 		goto idle;
5004 
5005 	if (prev->sched_class != &fair_sched_class)
5006 		goto simple;
5007 
5008 	/*
5009 	 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5010 	 * likely that a next task is from the same cgroup as the current.
5011 	 *
5012 	 * Therefore attempt to avoid putting and setting the entire cgroup
5013 	 * hierarchy, only change the part that actually changes.
5014 	 */
5015 
5016 	do {
5017 		struct sched_entity *curr = cfs_rq->curr;
5018 
5019 		/*
5020 		 * Since we got here without doing put_prev_entity() we also
5021 		 * have to consider cfs_rq->curr. If it is still a runnable
5022 		 * entity, update_curr() will update its vruntime, otherwise
5023 		 * forget we've ever seen it.
5024 		 */
5025 		if (curr && curr->on_rq)
5026 			update_curr(cfs_rq);
5027 		else
5028 			curr = NULL;
5029 
5030 		/*
5031 		 * This call to check_cfs_rq_runtime() will do the throttle and
5032 		 * dequeue its entity in the parent(s). Therefore the 'simple'
5033 		 * nr_running test will indeed be correct.
5034 		 */
5035 		if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5036 			goto simple;
5037 
5038 		se = pick_next_entity(cfs_rq, curr);
5039 		cfs_rq = group_cfs_rq(se);
5040 	} while (cfs_rq);
5041 
5042 	p = task_of(se);
5043 
5044 	/*
5045 	 * Since we haven't yet done put_prev_entity and if the selected task
5046 	 * is a different task than we started out with, try and touch the
5047 	 * least amount of cfs_rqs.
5048 	 */
5049 	if (prev != p) {
5050 		struct sched_entity *pse = &prev->se;
5051 
5052 		while (!(cfs_rq = is_same_group(se, pse))) {
5053 			int se_depth = se->depth;
5054 			int pse_depth = pse->depth;
5055 
5056 			if (se_depth <= pse_depth) {
5057 				put_prev_entity(cfs_rq_of(pse), pse);
5058 				pse = parent_entity(pse);
5059 			}
5060 			if (se_depth >= pse_depth) {
5061 				set_next_entity(cfs_rq_of(se), se);
5062 				se = parent_entity(se);
5063 			}
5064 		}
5065 
5066 		put_prev_entity(cfs_rq, pse);
5067 		set_next_entity(cfs_rq, se);
5068 	}
5069 
5070 	if (hrtick_enabled(rq))
5071 		hrtick_start_fair(rq, p);
5072 
5073 	return p;
5074 simple:
5075 	cfs_rq = &rq->cfs;
5076 #endif
5077 
5078 	if (!cfs_rq->nr_running)
5079 		goto idle;
5080 
5081 	put_prev_task(rq, prev);
5082 
5083 	do {
5084 		se = pick_next_entity(cfs_rq, NULL);
5085 		set_next_entity(cfs_rq, se);
5086 		cfs_rq = group_cfs_rq(se);
5087 	} while (cfs_rq);
5088 
5089 	p = task_of(se);
5090 
5091 	if (hrtick_enabled(rq))
5092 		hrtick_start_fair(rq, p);
5093 
5094 	return p;
5095 
5096 idle:
5097 	new_tasks = idle_balance(rq);
5098 	/*
5099 	 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5100 	 * possible for any higher priority task to appear. In that case we
5101 	 * must re-start the pick_next_entity() loop.
5102 	 */
5103 	if (new_tasks < 0)
5104 		return RETRY_TASK;
5105 
5106 	if (new_tasks > 0)
5107 		goto again;
5108 
5109 	return NULL;
5110 }
5111 
5112 /*
5113  * Account for a descheduled task:
5114  */
5115 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5116 {
5117 	struct sched_entity *se = &prev->se;
5118 	struct cfs_rq *cfs_rq;
5119 
5120 	for_each_sched_entity(se) {
5121 		cfs_rq = cfs_rq_of(se);
5122 		put_prev_entity(cfs_rq, se);
5123 	}
5124 }
5125 
5126 /*
5127  * sched_yield() is very simple
5128  *
5129  * The magic of dealing with the ->skip buddy is in pick_next_entity.
5130  */
5131 static void yield_task_fair(struct rq *rq)
5132 {
5133 	struct task_struct *curr = rq->curr;
5134 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5135 	struct sched_entity *se = &curr->se;
5136 
5137 	/*
5138 	 * Are we the only task in the tree?
5139 	 */
5140 	if (unlikely(rq->nr_running == 1))
5141 		return;
5142 
5143 	clear_buddies(cfs_rq, se);
5144 
5145 	if (curr->policy != SCHED_BATCH) {
5146 		update_rq_clock(rq);
5147 		/*
5148 		 * Update run-time statistics of the 'current'.
5149 		 */
5150 		update_curr(cfs_rq);
5151 		/*
5152 		 * Tell update_rq_clock() that we've just updated,
5153 		 * so we don't do microscopic update in schedule()
5154 		 * and double the fastpath cost.
5155 		 */
5156 		 rq->skip_clock_update = 1;
5157 	}
5158 
5159 	set_skip_buddy(se);
5160 }
5161 
5162 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5163 {
5164 	struct sched_entity *se = &p->se;
5165 
5166 	/* throttled hierarchies are not runnable */
5167 	if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5168 		return false;
5169 
5170 	/* Tell the scheduler that we'd really like pse to run next. */
5171 	set_next_buddy(se);
5172 
5173 	yield_task_fair(rq);
5174 
5175 	return true;
5176 }
5177 
5178 #ifdef CONFIG_SMP
5179 /**************************************************
5180  * Fair scheduling class load-balancing methods.
5181  *
5182  * BASICS
5183  *
5184  * The purpose of load-balancing is to achieve the same basic fairness the
5185  * per-cpu scheduler provides, namely provide a proportional amount of compute
5186  * time to each task. This is expressed in the following equation:
5187  *
5188  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
5189  *
5190  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5191  * W_i,0 is defined as:
5192  *
5193  *   W_i,0 = \Sum_j w_i,j                                             (2)
5194  *
5195  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5196  * is derived from the nice value as per prio_to_weight[].
5197  *
5198  * The weight average is an exponential decay average of the instantaneous
5199  * weight:
5200  *
5201  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
5202  *
5203  * C_i is the compute capacity of cpu i, typically it is the
5204  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5205  * can also include other factors [XXX].
5206  *
5207  * To achieve this balance we define a measure of imbalance which follows
5208  * directly from (1):
5209  *
5210  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
5211  *
5212  * We them move tasks around to minimize the imbalance. In the continuous
5213  * function space it is obvious this converges, in the discrete case we get
5214  * a few fun cases generally called infeasible weight scenarios.
5215  *
5216  * [XXX expand on:
5217  *     - infeasible weights;
5218  *     - local vs global optima in the discrete case. ]
5219  *
5220  *
5221  * SCHED DOMAINS
5222  *
5223  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5224  * for all i,j solution, we create a tree of cpus that follows the hardware
5225  * topology where each level pairs two lower groups (or better). This results
5226  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5227  * tree to only the first of the previous level and we decrease the frequency
5228  * of load-balance at each level inv. proportional to the number of cpus in
5229  * the groups.
5230  *
5231  * This yields:
5232  *
5233  *     log_2 n     1     n
5234  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
5235  *     i = 0      2^i   2^i
5236  *                               `- size of each group
5237  *         |         |     `- number of cpus doing load-balance
5238  *         |         `- freq
5239  *         `- sum over all levels
5240  *
5241  * Coupled with a limit on how many tasks we can migrate every balance pass,
5242  * this makes (5) the runtime complexity of the balancer.
5243  *
5244  * An important property here is that each CPU is still (indirectly) connected
5245  * to every other cpu in at most O(log n) steps:
5246  *
5247  * The adjacency matrix of the resulting graph is given by:
5248  *
5249  *             log_2 n
5250  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
5251  *             k = 0
5252  *
5253  * And you'll find that:
5254  *
5255  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
5256  *
5257  * Showing there's indeed a path between every cpu in at most O(log n) steps.
5258  * The task movement gives a factor of O(m), giving a convergence complexity
5259  * of:
5260  *
5261  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
5262  *
5263  *
5264  * WORK CONSERVING
5265  *
5266  * In order to avoid CPUs going idle while there's still work to do, new idle
5267  * balancing is more aggressive and has the newly idle cpu iterate up the domain
5268  * tree itself instead of relying on other CPUs to bring it work.
5269  *
5270  * This adds some complexity to both (5) and (8) but it reduces the total idle
5271  * time.
5272  *
5273  * [XXX more?]
5274  *
5275  *
5276  * CGROUPS
5277  *
5278  * Cgroups make a horror show out of (2), instead of a simple sum we get:
5279  *
5280  *                                s_k,i
5281  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
5282  *                                 S_k
5283  *
5284  * Where
5285  *
5286  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
5287  *
5288  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5289  *
5290  * The big problem is S_k, its a global sum needed to compute a local (W_i)
5291  * property.
5292  *
5293  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5294  *      rewrite all of this once again.]
5295  */
5296 
5297 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5298 
5299 enum fbq_type { regular, remote, all };
5300 
5301 #define LBF_ALL_PINNED	0x01
5302 #define LBF_NEED_BREAK	0x02
5303 #define LBF_DST_PINNED  0x04
5304 #define LBF_SOME_PINNED	0x08
5305 
5306 struct lb_env {
5307 	struct sched_domain	*sd;
5308 
5309 	struct rq		*src_rq;
5310 	int			src_cpu;
5311 
5312 	int			dst_cpu;
5313 	struct rq		*dst_rq;
5314 
5315 	struct cpumask		*dst_grpmask;
5316 	int			new_dst_cpu;
5317 	enum cpu_idle_type	idle;
5318 	long			imbalance;
5319 	/* The set of CPUs under consideration for load-balancing */
5320 	struct cpumask		*cpus;
5321 
5322 	unsigned int		flags;
5323 
5324 	unsigned int		loop;
5325 	unsigned int		loop_break;
5326 	unsigned int		loop_max;
5327 
5328 	enum fbq_type		fbq_type;
5329 	struct list_head	tasks;
5330 };
5331 
5332 /*
5333  * Is this task likely cache-hot:
5334  */
5335 static int task_hot(struct task_struct *p, struct lb_env *env)
5336 {
5337 	s64 delta;
5338 
5339 	lockdep_assert_held(&env->src_rq->lock);
5340 
5341 	if (p->sched_class != &fair_sched_class)
5342 		return 0;
5343 
5344 	if (unlikely(p->policy == SCHED_IDLE))
5345 		return 0;
5346 
5347 	/*
5348 	 * Buddy candidates are cache hot:
5349 	 */
5350 	if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5351 			(&p->se == cfs_rq_of(&p->se)->next ||
5352 			 &p->se == cfs_rq_of(&p->se)->last))
5353 		return 1;
5354 
5355 	if (sysctl_sched_migration_cost == -1)
5356 		return 1;
5357 	if (sysctl_sched_migration_cost == 0)
5358 		return 0;
5359 
5360 	delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5361 
5362 	return delta < (s64)sysctl_sched_migration_cost;
5363 }
5364 
5365 #ifdef CONFIG_NUMA_BALANCING
5366 /* Returns true if the destination node has incurred more faults */
5367 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5368 {
5369 	struct numa_group *numa_group = rcu_dereference(p->numa_group);
5370 	int src_nid, dst_nid;
5371 
5372 	if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
5373 	    !(env->sd->flags & SD_NUMA)) {
5374 		return false;
5375 	}
5376 
5377 	src_nid = cpu_to_node(env->src_cpu);
5378 	dst_nid = cpu_to_node(env->dst_cpu);
5379 
5380 	if (src_nid == dst_nid)
5381 		return false;
5382 
5383 	if (numa_group) {
5384 		/* Task is already in the group's interleave set. */
5385 		if (node_isset(src_nid, numa_group->active_nodes))
5386 			return false;
5387 
5388 		/* Task is moving into the group's interleave set. */
5389 		if (node_isset(dst_nid, numa_group->active_nodes))
5390 			return true;
5391 
5392 		return group_faults(p, dst_nid) > group_faults(p, src_nid);
5393 	}
5394 
5395 	/* Encourage migration to the preferred node. */
5396 	if (dst_nid == p->numa_preferred_nid)
5397 		return true;
5398 
5399 	return task_faults(p, dst_nid) > task_faults(p, src_nid);
5400 }
5401 
5402 
5403 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5404 {
5405 	struct numa_group *numa_group = rcu_dereference(p->numa_group);
5406 	int src_nid, dst_nid;
5407 
5408 	if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5409 		return false;
5410 
5411 	if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5412 		return false;
5413 
5414 	src_nid = cpu_to_node(env->src_cpu);
5415 	dst_nid = cpu_to_node(env->dst_cpu);
5416 
5417 	if (src_nid == dst_nid)
5418 		return false;
5419 
5420 	if (numa_group) {
5421 		/* Task is moving within/into the group's interleave set. */
5422 		if (node_isset(dst_nid, numa_group->active_nodes))
5423 			return false;
5424 
5425 		/* Task is moving out of the group's interleave set. */
5426 		if (node_isset(src_nid, numa_group->active_nodes))
5427 			return true;
5428 
5429 		return group_faults(p, dst_nid) < group_faults(p, src_nid);
5430 	}
5431 
5432 	/* Migrating away from the preferred node is always bad. */
5433 	if (src_nid == p->numa_preferred_nid)
5434 		return true;
5435 
5436 	return task_faults(p, dst_nid) < task_faults(p, src_nid);
5437 }
5438 
5439 #else
5440 static inline bool migrate_improves_locality(struct task_struct *p,
5441 					     struct lb_env *env)
5442 {
5443 	return false;
5444 }
5445 
5446 static inline bool migrate_degrades_locality(struct task_struct *p,
5447 					     struct lb_env *env)
5448 {
5449 	return false;
5450 }
5451 #endif
5452 
5453 /*
5454  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5455  */
5456 static
5457 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5458 {
5459 	int tsk_cache_hot = 0;
5460 
5461 	lockdep_assert_held(&env->src_rq->lock);
5462 
5463 	/*
5464 	 * We do not migrate tasks that are:
5465 	 * 1) throttled_lb_pair, or
5466 	 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5467 	 * 3) running (obviously), or
5468 	 * 4) are cache-hot on their current CPU.
5469 	 */
5470 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5471 		return 0;
5472 
5473 	if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5474 		int cpu;
5475 
5476 		schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5477 
5478 		env->flags |= LBF_SOME_PINNED;
5479 
5480 		/*
5481 		 * Remember if this task can be migrated to any other cpu in
5482 		 * our sched_group. We may want to revisit it if we couldn't
5483 		 * meet load balance goals by pulling other tasks on src_cpu.
5484 		 *
5485 		 * Also avoid computing new_dst_cpu if we have already computed
5486 		 * one in current iteration.
5487 		 */
5488 		if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5489 			return 0;
5490 
5491 		/* Prevent to re-select dst_cpu via env's cpus */
5492 		for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5493 			if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5494 				env->flags |= LBF_DST_PINNED;
5495 				env->new_dst_cpu = cpu;
5496 				break;
5497 			}
5498 		}
5499 
5500 		return 0;
5501 	}
5502 
5503 	/* Record that we found atleast one task that could run on dst_cpu */
5504 	env->flags &= ~LBF_ALL_PINNED;
5505 
5506 	if (task_running(env->src_rq, p)) {
5507 		schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5508 		return 0;
5509 	}
5510 
5511 	/*
5512 	 * Aggressive migration if:
5513 	 * 1) destination numa is preferred
5514 	 * 2) task is cache cold, or
5515 	 * 3) too many balance attempts have failed.
5516 	 */
5517 	tsk_cache_hot = task_hot(p, env);
5518 	if (!tsk_cache_hot)
5519 		tsk_cache_hot = migrate_degrades_locality(p, env);
5520 
5521 	if (migrate_improves_locality(p, env) || !tsk_cache_hot ||
5522 	    env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5523 		if (tsk_cache_hot) {
5524 			schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5525 			schedstat_inc(p, se.statistics.nr_forced_migrations);
5526 		}
5527 		return 1;
5528 	}
5529 
5530 	schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5531 	return 0;
5532 }
5533 
5534 /*
5535  * detach_task() -- detach the task for the migration specified in env
5536  */
5537 static void detach_task(struct task_struct *p, struct lb_env *env)
5538 {
5539 	lockdep_assert_held(&env->src_rq->lock);
5540 
5541 	deactivate_task(env->src_rq, p, 0);
5542 	p->on_rq = TASK_ON_RQ_MIGRATING;
5543 	set_task_cpu(p, env->dst_cpu);
5544 }
5545 
5546 /*
5547  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5548  * part of active balancing operations within "domain".
5549  *
5550  * Returns a task if successful and NULL otherwise.
5551  */
5552 static struct task_struct *detach_one_task(struct lb_env *env)
5553 {
5554 	struct task_struct *p, *n;
5555 
5556 	lockdep_assert_held(&env->src_rq->lock);
5557 
5558 	list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5559 		if (!can_migrate_task(p, env))
5560 			continue;
5561 
5562 		detach_task(p, env);
5563 
5564 		/*
5565 		 * Right now, this is only the second place where
5566 		 * lb_gained[env->idle] is updated (other is detach_tasks)
5567 		 * so we can safely collect stats here rather than
5568 		 * inside detach_tasks().
5569 		 */
5570 		schedstat_inc(env->sd, lb_gained[env->idle]);
5571 		return p;
5572 	}
5573 	return NULL;
5574 }
5575 
5576 static const unsigned int sched_nr_migrate_break = 32;
5577 
5578 /*
5579  * detach_tasks() -- tries to detach up to imbalance weighted load from
5580  * busiest_rq, as part of a balancing operation within domain "sd".
5581  *
5582  * Returns number of detached tasks if successful and 0 otherwise.
5583  */
5584 static int detach_tasks(struct lb_env *env)
5585 {
5586 	struct list_head *tasks = &env->src_rq->cfs_tasks;
5587 	struct task_struct *p;
5588 	unsigned long load;
5589 	int detached = 0;
5590 
5591 	lockdep_assert_held(&env->src_rq->lock);
5592 
5593 	if (env->imbalance <= 0)
5594 		return 0;
5595 
5596 	while (!list_empty(tasks)) {
5597 		p = list_first_entry(tasks, struct task_struct, se.group_node);
5598 
5599 		env->loop++;
5600 		/* We've more or less seen every task there is, call it quits */
5601 		if (env->loop > env->loop_max)
5602 			break;
5603 
5604 		/* take a breather every nr_migrate tasks */
5605 		if (env->loop > env->loop_break) {
5606 			env->loop_break += sched_nr_migrate_break;
5607 			env->flags |= LBF_NEED_BREAK;
5608 			break;
5609 		}
5610 
5611 		if (!can_migrate_task(p, env))
5612 			goto next;
5613 
5614 		load = task_h_load(p);
5615 
5616 		if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5617 			goto next;
5618 
5619 		if ((load / 2) > env->imbalance)
5620 			goto next;
5621 
5622 		detach_task(p, env);
5623 		list_add(&p->se.group_node, &env->tasks);
5624 
5625 		detached++;
5626 		env->imbalance -= load;
5627 
5628 #ifdef CONFIG_PREEMPT
5629 		/*
5630 		 * NEWIDLE balancing is a source of latency, so preemptible
5631 		 * kernels will stop after the first task is detached to minimize
5632 		 * the critical section.
5633 		 */
5634 		if (env->idle == CPU_NEWLY_IDLE)
5635 			break;
5636 #endif
5637 
5638 		/*
5639 		 * We only want to steal up to the prescribed amount of
5640 		 * weighted load.
5641 		 */
5642 		if (env->imbalance <= 0)
5643 			break;
5644 
5645 		continue;
5646 next:
5647 		list_move_tail(&p->se.group_node, tasks);
5648 	}
5649 
5650 	/*
5651 	 * Right now, this is one of only two places we collect this stat
5652 	 * so we can safely collect detach_one_task() stats here rather
5653 	 * than inside detach_one_task().
5654 	 */
5655 	schedstat_add(env->sd, lb_gained[env->idle], detached);
5656 
5657 	return detached;
5658 }
5659 
5660 /*
5661  * attach_task() -- attach the task detached by detach_task() to its new rq.
5662  */
5663 static void attach_task(struct rq *rq, struct task_struct *p)
5664 {
5665 	lockdep_assert_held(&rq->lock);
5666 
5667 	BUG_ON(task_rq(p) != rq);
5668 	p->on_rq = TASK_ON_RQ_QUEUED;
5669 	activate_task(rq, p, 0);
5670 	check_preempt_curr(rq, p, 0);
5671 }
5672 
5673 /*
5674  * attach_one_task() -- attaches the task returned from detach_one_task() to
5675  * its new rq.
5676  */
5677 static void attach_one_task(struct rq *rq, struct task_struct *p)
5678 {
5679 	raw_spin_lock(&rq->lock);
5680 	attach_task(rq, p);
5681 	raw_spin_unlock(&rq->lock);
5682 }
5683 
5684 /*
5685  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5686  * new rq.
5687  */
5688 static void attach_tasks(struct lb_env *env)
5689 {
5690 	struct list_head *tasks = &env->tasks;
5691 	struct task_struct *p;
5692 
5693 	raw_spin_lock(&env->dst_rq->lock);
5694 
5695 	while (!list_empty(tasks)) {
5696 		p = list_first_entry(tasks, struct task_struct, se.group_node);
5697 		list_del_init(&p->se.group_node);
5698 
5699 		attach_task(env->dst_rq, p);
5700 	}
5701 
5702 	raw_spin_unlock(&env->dst_rq->lock);
5703 }
5704 
5705 #ifdef CONFIG_FAIR_GROUP_SCHED
5706 /*
5707  * update tg->load_weight by folding this cpu's load_avg
5708  */
5709 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5710 {
5711 	struct sched_entity *se = tg->se[cpu];
5712 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5713 
5714 	/* throttled entities do not contribute to load */
5715 	if (throttled_hierarchy(cfs_rq))
5716 		return;
5717 
5718 	update_cfs_rq_blocked_load(cfs_rq, 1);
5719 
5720 	if (se) {
5721 		update_entity_load_avg(se, 1);
5722 		/*
5723 		 * We pivot on our runnable average having decayed to zero for
5724 		 * list removal.  This generally implies that all our children
5725 		 * have also been removed (modulo rounding error or bandwidth
5726 		 * control); however, such cases are rare and we can fix these
5727 		 * at enqueue.
5728 		 *
5729 		 * TODO: fix up out-of-order children on enqueue.
5730 		 */
5731 		if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5732 			list_del_leaf_cfs_rq(cfs_rq);
5733 	} else {
5734 		struct rq *rq = rq_of(cfs_rq);
5735 		update_rq_runnable_avg(rq, rq->nr_running);
5736 	}
5737 }
5738 
5739 static void update_blocked_averages(int cpu)
5740 {
5741 	struct rq *rq = cpu_rq(cpu);
5742 	struct cfs_rq *cfs_rq;
5743 	unsigned long flags;
5744 
5745 	raw_spin_lock_irqsave(&rq->lock, flags);
5746 	update_rq_clock(rq);
5747 	/*
5748 	 * Iterates the task_group tree in a bottom up fashion, see
5749 	 * list_add_leaf_cfs_rq() for details.
5750 	 */
5751 	for_each_leaf_cfs_rq(rq, cfs_rq) {
5752 		/*
5753 		 * Note: We may want to consider periodically releasing
5754 		 * rq->lock about these updates so that creating many task
5755 		 * groups does not result in continually extending hold time.
5756 		 */
5757 		__update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5758 	}
5759 
5760 	raw_spin_unlock_irqrestore(&rq->lock, flags);
5761 }
5762 
5763 /*
5764  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5765  * This needs to be done in a top-down fashion because the load of a child
5766  * group is a fraction of its parents load.
5767  */
5768 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5769 {
5770 	struct rq *rq = rq_of(cfs_rq);
5771 	struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5772 	unsigned long now = jiffies;
5773 	unsigned long load;
5774 
5775 	if (cfs_rq->last_h_load_update == now)
5776 		return;
5777 
5778 	cfs_rq->h_load_next = NULL;
5779 	for_each_sched_entity(se) {
5780 		cfs_rq = cfs_rq_of(se);
5781 		cfs_rq->h_load_next = se;
5782 		if (cfs_rq->last_h_load_update == now)
5783 			break;
5784 	}
5785 
5786 	if (!se) {
5787 		cfs_rq->h_load = cfs_rq->runnable_load_avg;
5788 		cfs_rq->last_h_load_update = now;
5789 	}
5790 
5791 	while ((se = cfs_rq->h_load_next) != NULL) {
5792 		load = cfs_rq->h_load;
5793 		load = div64_ul(load * se->avg.load_avg_contrib,
5794 				cfs_rq->runnable_load_avg + 1);
5795 		cfs_rq = group_cfs_rq(se);
5796 		cfs_rq->h_load = load;
5797 		cfs_rq->last_h_load_update = now;
5798 	}
5799 }
5800 
5801 static unsigned long task_h_load(struct task_struct *p)
5802 {
5803 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
5804 
5805 	update_cfs_rq_h_load(cfs_rq);
5806 	return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5807 			cfs_rq->runnable_load_avg + 1);
5808 }
5809 #else
5810 static inline void update_blocked_averages(int cpu)
5811 {
5812 }
5813 
5814 static unsigned long task_h_load(struct task_struct *p)
5815 {
5816 	return p->se.avg.load_avg_contrib;
5817 }
5818 #endif
5819 
5820 /********** Helpers for find_busiest_group ************************/
5821 
5822 enum group_type {
5823 	group_other = 0,
5824 	group_imbalanced,
5825 	group_overloaded,
5826 };
5827 
5828 /*
5829  * sg_lb_stats - stats of a sched_group required for load_balancing
5830  */
5831 struct sg_lb_stats {
5832 	unsigned long avg_load; /*Avg load across the CPUs of the group */
5833 	unsigned long group_load; /* Total load over the CPUs of the group */
5834 	unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5835 	unsigned long load_per_task;
5836 	unsigned long group_capacity;
5837 	unsigned int sum_nr_running; /* Nr tasks running in the group */
5838 	unsigned int group_capacity_factor;
5839 	unsigned int idle_cpus;
5840 	unsigned int group_weight;
5841 	enum group_type group_type;
5842 	int group_has_free_capacity;
5843 #ifdef CONFIG_NUMA_BALANCING
5844 	unsigned int nr_numa_running;
5845 	unsigned int nr_preferred_running;
5846 #endif
5847 };
5848 
5849 /*
5850  * sd_lb_stats - Structure to store the statistics of a sched_domain
5851  *		 during load balancing.
5852  */
5853 struct sd_lb_stats {
5854 	struct sched_group *busiest;	/* Busiest group in this sd */
5855 	struct sched_group *local;	/* Local group in this sd */
5856 	unsigned long total_load;	/* Total load of all groups in sd */
5857 	unsigned long total_capacity;	/* Total capacity of all groups in sd */
5858 	unsigned long avg_load;	/* Average load across all groups in sd */
5859 
5860 	struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5861 	struct sg_lb_stats local_stat;	/* Statistics of the local group */
5862 };
5863 
5864 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5865 {
5866 	/*
5867 	 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5868 	 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5869 	 * We must however clear busiest_stat::avg_load because
5870 	 * update_sd_pick_busiest() reads this before assignment.
5871 	 */
5872 	*sds = (struct sd_lb_stats){
5873 		.busiest = NULL,
5874 		.local = NULL,
5875 		.total_load = 0UL,
5876 		.total_capacity = 0UL,
5877 		.busiest_stat = {
5878 			.avg_load = 0UL,
5879 			.sum_nr_running = 0,
5880 			.group_type = group_other,
5881 		},
5882 	};
5883 }
5884 
5885 /**
5886  * get_sd_load_idx - Obtain the load index for a given sched domain.
5887  * @sd: The sched_domain whose load_idx is to be obtained.
5888  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5889  *
5890  * Return: The load index.
5891  */
5892 static inline int get_sd_load_idx(struct sched_domain *sd,
5893 					enum cpu_idle_type idle)
5894 {
5895 	int load_idx;
5896 
5897 	switch (idle) {
5898 	case CPU_NOT_IDLE:
5899 		load_idx = sd->busy_idx;
5900 		break;
5901 
5902 	case CPU_NEWLY_IDLE:
5903 		load_idx = sd->newidle_idx;
5904 		break;
5905 	default:
5906 		load_idx = sd->idle_idx;
5907 		break;
5908 	}
5909 
5910 	return load_idx;
5911 }
5912 
5913 static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu)
5914 {
5915 	return SCHED_CAPACITY_SCALE;
5916 }
5917 
5918 unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
5919 {
5920 	return default_scale_capacity(sd, cpu);
5921 }
5922 
5923 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu)
5924 {
5925 	if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
5926 		return sd->smt_gain / sd->span_weight;
5927 
5928 	return SCHED_CAPACITY_SCALE;
5929 }
5930 
5931 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
5932 {
5933 	return default_scale_cpu_capacity(sd, cpu);
5934 }
5935 
5936 static unsigned long scale_rt_capacity(int cpu)
5937 {
5938 	struct rq *rq = cpu_rq(cpu);
5939 	u64 total, available, age_stamp, avg;
5940 	s64 delta;
5941 
5942 	/*
5943 	 * Since we're reading these variables without serialization make sure
5944 	 * we read them once before doing sanity checks on them.
5945 	 */
5946 	age_stamp = ACCESS_ONCE(rq->age_stamp);
5947 	avg = ACCESS_ONCE(rq->rt_avg);
5948 
5949 	delta = rq_clock(rq) - age_stamp;
5950 	if (unlikely(delta < 0))
5951 		delta = 0;
5952 
5953 	total = sched_avg_period() + delta;
5954 
5955 	if (unlikely(total < avg)) {
5956 		/* Ensures that capacity won't end up being negative */
5957 		available = 0;
5958 	} else {
5959 		available = total - avg;
5960 	}
5961 
5962 	if (unlikely((s64)total < SCHED_CAPACITY_SCALE))
5963 		total = SCHED_CAPACITY_SCALE;
5964 
5965 	total >>= SCHED_CAPACITY_SHIFT;
5966 
5967 	return div_u64(available, total);
5968 }
5969 
5970 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
5971 {
5972 	unsigned long capacity = SCHED_CAPACITY_SCALE;
5973 	struct sched_group *sdg = sd->groups;
5974 
5975 	if (sched_feat(ARCH_CAPACITY))
5976 		capacity *= arch_scale_cpu_capacity(sd, cpu);
5977 	else
5978 		capacity *= default_scale_cpu_capacity(sd, cpu);
5979 
5980 	capacity >>= SCHED_CAPACITY_SHIFT;
5981 
5982 	sdg->sgc->capacity_orig = capacity;
5983 
5984 	if (sched_feat(ARCH_CAPACITY))
5985 		capacity *= arch_scale_freq_capacity(sd, cpu);
5986 	else
5987 		capacity *= default_scale_capacity(sd, cpu);
5988 
5989 	capacity >>= SCHED_CAPACITY_SHIFT;
5990 
5991 	capacity *= scale_rt_capacity(cpu);
5992 	capacity >>= SCHED_CAPACITY_SHIFT;
5993 
5994 	if (!capacity)
5995 		capacity = 1;
5996 
5997 	cpu_rq(cpu)->cpu_capacity = capacity;
5998 	sdg->sgc->capacity = capacity;
5999 }
6000 
6001 void update_group_capacity(struct sched_domain *sd, int cpu)
6002 {
6003 	struct sched_domain *child = sd->child;
6004 	struct sched_group *group, *sdg = sd->groups;
6005 	unsigned long capacity, capacity_orig;
6006 	unsigned long interval;
6007 
6008 	interval = msecs_to_jiffies(sd->balance_interval);
6009 	interval = clamp(interval, 1UL, max_load_balance_interval);
6010 	sdg->sgc->next_update = jiffies + interval;
6011 
6012 	if (!child) {
6013 		update_cpu_capacity(sd, cpu);
6014 		return;
6015 	}
6016 
6017 	capacity_orig = capacity = 0;
6018 
6019 	if (child->flags & SD_OVERLAP) {
6020 		/*
6021 		 * SD_OVERLAP domains cannot assume that child groups
6022 		 * span the current group.
6023 		 */
6024 
6025 		for_each_cpu(cpu, sched_group_cpus(sdg)) {
6026 			struct sched_group_capacity *sgc;
6027 			struct rq *rq = cpu_rq(cpu);
6028 
6029 			/*
6030 			 * build_sched_domains() -> init_sched_groups_capacity()
6031 			 * gets here before we've attached the domains to the
6032 			 * runqueues.
6033 			 *
6034 			 * Use capacity_of(), which is set irrespective of domains
6035 			 * in update_cpu_capacity().
6036 			 *
6037 			 * This avoids capacity/capacity_orig from being 0 and
6038 			 * causing divide-by-zero issues on boot.
6039 			 *
6040 			 * Runtime updates will correct capacity_orig.
6041 			 */
6042 			if (unlikely(!rq->sd)) {
6043 				capacity_orig += capacity_of(cpu);
6044 				capacity += capacity_of(cpu);
6045 				continue;
6046 			}
6047 
6048 			sgc = rq->sd->groups->sgc;
6049 			capacity_orig += sgc->capacity_orig;
6050 			capacity += sgc->capacity;
6051 		}
6052 	} else  {
6053 		/*
6054 		 * !SD_OVERLAP domains can assume that child groups
6055 		 * span the current group.
6056 		 */
6057 
6058 		group = child->groups;
6059 		do {
6060 			capacity_orig += group->sgc->capacity_orig;
6061 			capacity += group->sgc->capacity;
6062 			group = group->next;
6063 		} while (group != child->groups);
6064 	}
6065 
6066 	sdg->sgc->capacity_orig = capacity_orig;
6067 	sdg->sgc->capacity = capacity;
6068 }
6069 
6070 /*
6071  * Try and fix up capacity for tiny siblings, this is needed when
6072  * things like SD_ASYM_PACKING need f_b_g to select another sibling
6073  * which on its own isn't powerful enough.
6074  *
6075  * See update_sd_pick_busiest() and check_asym_packing().
6076  */
6077 static inline int
6078 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
6079 {
6080 	/*
6081 	 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
6082 	 */
6083 	if (!(sd->flags & SD_SHARE_CPUCAPACITY))
6084 		return 0;
6085 
6086 	/*
6087 	 * If ~90% of the cpu_capacity is still there, we're good.
6088 	 */
6089 	if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29)
6090 		return 1;
6091 
6092 	return 0;
6093 }
6094 
6095 /*
6096  * Group imbalance indicates (and tries to solve) the problem where balancing
6097  * groups is inadequate due to tsk_cpus_allowed() constraints.
6098  *
6099  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6100  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6101  * Something like:
6102  *
6103  * 	{ 0 1 2 3 } { 4 5 6 7 }
6104  * 	        *     * * *
6105  *
6106  * If we were to balance group-wise we'd place two tasks in the first group and
6107  * two tasks in the second group. Clearly this is undesired as it will overload
6108  * cpu 3 and leave one of the cpus in the second group unused.
6109  *
6110  * The current solution to this issue is detecting the skew in the first group
6111  * by noticing the lower domain failed to reach balance and had difficulty
6112  * moving tasks due to affinity constraints.
6113  *
6114  * When this is so detected; this group becomes a candidate for busiest; see
6115  * update_sd_pick_busiest(). And calculate_imbalance() and
6116  * find_busiest_group() avoid some of the usual balance conditions to allow it
6117  * to create an effective group imbalance.
6118  *
6119  * This is a somewhat tricky proposition since the next run might not find the
6120  * group imbalance and decide the groups need to be balanced again. A most
6121  * subtle and fragile situation.
6122  */
6123 
6124 static inline int sg_imbalanced(struct sched_group *group)
6125 {
6126 	return group->sgc->imbalance;
6127 }
6128 
6129 /*
6130  * Compute the group capacity factor.
6131  *
6132  * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
6133  * first dividing out the smt factor and computing the actual number of cores
6134  * and limit unit capacity with that.
6135  */
6136 static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group)
6137 {
6138 	unsigned int capacity_factor, smt, cpus;
6139 	unsigned int capacity, capacity_orig;
6140 
6141 	capacity = group->sgc->capacity;
6142 	capacity_orig = group->sgc->capacity_orig;
6143 	cpus = group->group_weight;
6144 
6145 	/* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
6146 	smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig);
6147 	capacity_factor = cpus / smt; /* cores */
6148 
6149 	capacity_factor = min_t(unsigned,
6150 		capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE));
6151 	if (!capacity_factor)
6152 		capacity_factor = fix_small_capacity(env->sd, group);
6153 
6154 	return capacity_factor;
6155 }
6156 
6157 static enum group_type
6158 group_classify(struct sched_group *group, struct sg_lb_stats *sgs)
6159 {
6160 	if (sgs->sum_nr_running > sgs->group_capacity_factor)
6161 		return group_overloaded;
6162 
6163 	if (sg_imbalanced(group))
6164 		return group_imbalanced;
6165 
6166 	return group_other;
6167 }
6168 
6169 /**
6170  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6171  * @env: The load balancing environment.
6172  * @group: sched_group whose statistics are to be updated.
6173  * @load_idx: Load index of sched_domain of this_cpu for load calc.
6174  * @local_group: Does group contain this_cpu.
6175  * @sgs: variable to hold the statistics for this group.
6176  * @overload: Indicate more than one runnable task for any CPU.
6177  */
6178 static inline void update_sg_lb_stats(struct lb_env *env,
6179 			struct sched_group *group, int load_idx,
6180 			int local_group, struct sg_lb_stats *sgs,
6181 			bool *overload)
6182 {
6183 	unsigned long load;
6184 	int i;
6185 
6186 	memset(sgs, 0, sizeof(*sgs));
6187 
6188 	for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6189 		struct rq *rq = cpu_rq(i);
6190 
6191 		/* Bias balancing toward cpus of our domain */
6192 		if (local_group)
6193 			load = target_load(i, load_idx);
6194 		else
6195 			load = source_load(i, load_idx);
6196 
6197 		sgs->group_load += load;
6198 		sgs->sum_nr_running += rq->cfs.h_nr_running;
6199 
6200 		if (rq->nr_running > 1)
6201 			*overload = true;
6202 
6203 #ifdef CONFIG_NUMA_BALANCING
6204 		sgs->nr_numa_running += rq->nr_numa_running;
6205 		sgs->nr_preferred_running += rq->nr_preferred_running;
6206 #endif
6207 		sgs->sum_weighted_load += weighted_cpuload(i);
6208 		if (idle_cpu(i))
6209 			sgs->idle_cpus++;
6210 	}
6211 
6212 	/* Adjust by relative CPU capacity of the group */
6213 	sgs->group_capacity = group->sgc->capacity;
6214 	sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6215 
6216 	if (sgs->sum_nr_running)
6217 		sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6218 
6219 	sgs->group_weight = group->group_weight;
6220 	sgs->group_capacity_factor = sg_capacity_factor(env, group);
6221 	sgs->group_type = group_classify(group, sgs);
6222 
6223 	if (sgs->group_capacity_factor > sgs->sum_nr_running)
6224 		sgs->group_has_free_capacity = 1;
6225 }
6226 
6227 /**
6228  * update_sd_pick_busiest - return 1 on busiest group
6229  * @env: The load balancing environment.
6230  * @sds: sched_domain statistics
6231  * @sg: sched_group candidate to be checked for being the busiest
6232  * @sgs: sched_group statistics
6233  *
6234  * Determine if @sg is a busier group than the previously selected
6235  * busiest group.
6236  *
6237  * Return: %true if @sg is a busier group than the previously selected
6238  * busiest group. %false otherwise.
6239  */
6240 static bool update_sd_pick_busiest(struct lb_env *env,
6241 				   struct sd_lb_stats *sds,
6242 				   struct sched_group *sg,
6243 				   struct sg_lb_stats *sgs)
6244 {
6245 	struct sg_lb_stats *busiest = &sds->busiest_stat;
6246 
6247 	if (sgs->group_type > busiest->group_type)
6248 		return true;
6249 
6250 	if (sgs->group_type < busiest->group_type)
6251 		return false;
6252 
6253 	if (sgs->avg_load <= busiest->avg_load)
6254 		return false;
6255 
6256 	/* This is the busiest node in its class. */
6257 	if (!(env->sd->flags & SD_ASYM_PACKING))
6258 		return true;
6259 
6260 	/*
6261 	 * ASYM_PACKING needs to move all the work to the lowest
6262 	 * numbered CPUs in the group, therefore mark all groups
6263 	 * higher than ourself as busy.
6264 	 */
6265 	if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6266 		if (!sds->busiest)
6267 			return true;
6268 
6269 		if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6270 			return true;
6271 	}
6272 
6273 	return false;
6274 }
6275 
6276 #ifdef CONFIG_NUMA_BALANCING
6277 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6278 {
6279 	if (sgs->sum_nr_running > sgs->nr_numa_running)
6280 		return regular;
6281 	if (sgs->sum_nr_running > sgs->nr_preferred_running)
6282 		return remote;
6283 	return all;
6284 }
6285 
6286 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6287 {
6288 	if (rq->nr_running > rq->nr_numa_running)
6289 		return regular;
6290 	if (rq->nr_running > rq->nr_preferred_running)
6291 		return remote;
6292 	return all;
6293 }
6294 #else
6295 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6296 {
6297 	return all;
6298 }
6299 
6300 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6301 {
6302 	return regular;
6303 }
6304 #endif /* CONFIG_NUMA_BALANCING */
6305 
6306 /**
6307  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6308  * @env: The load balancing environment.
6309  * @sds: variable to hold the statistics for this sched_domain.
6310  */
6311 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6312 {
6313 	struct sched_domain *child = env->sd->child;
6314 	struct sched_group *sg = env->sd->groups;
6315 	struct sg_lb_stats tmp_sgs;
6316 	int load_idx, prefer_sibling = 0;
6317 	bool overload = false;
6318 
6319 	if (child && child->flags & SD_PREFER_SIBLING)
6320 		prefer_sibling = 1;
6321 
6322 	load_idx = get_sd_load_idx(env->sd, env->idle);
6323 
6324 	do {
6325 		struct sg_lb_stats *sgs = &tmp_sgs;
6326 		int local_group;
6327 
6328 		local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6329 		if (local_group) {
6330 			sds->local = sg;
6331 			sgs = &sds->local_stat;
6332 
6333 			if (env->idle != CPU_NEWLY_IDLE ||
6334 			    time_after_eq(jiffies, sg->sgc->next_update))
6335 				update_group_capacity(env->sd, env->dst_cpu);
6336 		}
6337 
6338 		update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6339 						&overload);
6340 
6341 		if (local_group)
6342 			goto next_group;
6343 
6344 		/*
6345 		 * In case the child domain prefers tasks go to siblings
6346 		 * first, lower the sg capacity factor to one so that we'll try
6347 		 * and move all the excess tasks away. We lower the capacity
6348 		 * of a group only if the local group has the capacity to fit
6349 		 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6350 		 * extra check prevents the case where you always pull from the
6351 		 * heaviest group when it is already under-utilized (possible
6352 		 * with a large weight task outweighs the tasks on the system).
6353 		 */
6354 		if (prefer_sibling && sds->local &&
6355 		    sds->local_stat.group_has_free_capacity) {
6356 			sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U);
6357 			sgs->group_type = group_classify(sg, sgs);
6358 		}
6359 
6360 		if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6361 			sds->busiest = sg;
6362 			sds->busiest_stat = *sgs;
6363 		}
6364 
6365 next_group:
6366 		/* Now, start updating sd_lb_stats */
6367 		sds->total_load += sgs->group_load;
6368 		sds->total_capacity += sgs->group_capacity;
6369 
6370 		sg = sg->next;
6371 	} while (sg != env->sd->groups);
6372 
6373 	if (env->sd->flags & SD_NUMA)
6374 		env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6375 
6376 	if (!env->sd->parent) {
6377 		/* update overload indicator if we are at root domain */
6378 		if (env->dst_rq->rd->overload != overload)
6379 			env->dst_rq->rd->overload = overload;
6380 	}
6381 
6382 }
6383 
6384 /**
6385  * check_asym_packing - Check to see if the group is packed into the
6386  *			sched doman.
6387  *
6388  * This is primarily intended to used at the sibling level.  Some
6389  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
6390  * case of POWER7, it can move to lower SMT modes only when higher
6391  * threads are idle.  When in lower SMT modes, the threads will
6392  * perform better since they share less core resources.  Hence when we
6393  * have idle threads, we want them to be the higher ones.
6394  *
6395  * This packing function is run on idle threads.  It checks to see if
6396  * the busiest CPU in this domain (core in the P7 case) has a higher
6397  * CPU number than the packing function is being run on.  Here we are
6398  * assuming lower CPU number will be equivalent to lower a SMT thread
6399  * number.
6400  *
6401  * Return: 1 when packing is required and a task should be moved to
6402  * this CPU.  The amount of the imbalance is returned in *imbalance.
6403  *
6404  * @env: The load balancing environment.
6405  * @sds: Statistics of the sched_domain which is to be packed
6406  */
6407 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6408 {
6409 	int busiest_cpu;
6410 
6411 	if (!(env->sd->flags & SD_ASYM_PACKING))
6412 		return 0;
6413 
6414 	if (!sds->busiest)
6415 		return 0;
6416 
6417 	busiest_cpu = group_first_cpu(sds->busiest);
6418 	if (env->dst_cpu > busiest_cpu)
6419 		return 0;
6420 
6421 	env->imbalance = DIV_ROUND_CLOSEST(
6422 		sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6423 		SCHED_CAPACITY_SCALE);
6424 
6425 	return 1;
6426 }
6427 
6428 /**
6429  * fix_small_imbalance - Calculate the minor imbalance that exists
6430  *			amongst the groups of a sched_domain, during
6431  *			load balancing.
6432  * @env: The load balancing environment.
6433  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6434  */
6435 static inline
6436 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6437 {
6438 	unsigned long tmp, capa_now = 0, capa_move = 0;
6439 	unsigned int imbn = 2;
6440 	unsigned long scaled_busy_load_per_task;
6441 	struct sg_lb_stats *local, *busiest;
6442 
6443 	local = &sds->local_stat;
6444 	busiest = &sds->busiest_stat;
6445 
6446 	if (!local->sum_nr_running)
6447 		local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6448 	else if (busiest->load_per_task > local->load_per_task)
6449 		imbn = 1;
6450 
6451 	scaled_busy_load_per_task =
6452 		(busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6453 		busiest->group_capacity;
6454 
6455 	if (busiest->avg_load + scaled_busy_load_per_task >=
6456 	    local->avg_load + (scaled_busy_load_per_task * imbn)) {
6457 		env->imbalance = busiest->load_per_task;
6458 		return;
6459 	}
6460 
6461 	/*
6462 	 * OK, we don't have enough imbalance to justify moving tasks,
6463 	 * however we may be able to increase total CPU capacity used by
6464 	 * moving them.
6465 	 */
6466 
6467 	capa_now += busiest->group_capacity *
6468 			min(busiest->load_per_task, busiest->avg_load);
6469 	capa_now += local->group_capacity *
6470 			min(local->load_per_task, local->avg_load);
6471 	capa_now /= SCHED_CAPACITY_SCALE;
6472 
6473 	/* Amount of load we'd subtract */
6474 	if (busiest->avg_load > scaled_busy_load_per_task) {
6475 		capa_move += busiest->group_capacity *
6476 			    min(busiest->load_per_task,
6477 				busiest->avg_load - scaled_busy_load_per_task);
6478 	}
6479 
6480 	/* Amount of load we'd add */
6481 	if (busiest->avg_load * busiest->group_capacity <
6482 	    busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6483 		tmp = (busiest->avg_load * busiest->group_capacity) /
6484 		      local->group_capacity;
6485 	} else {
6486 		tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6487 		      local->group_capacity;
6488 	}
6489 	capa_move += local->group_capacity *
6490 		    min(local->load_per_task, local->avg_load + tmp);
6491 	capa_move /= SCHED_CAPACITY_SCALE;
6492 
6493 	/* Move if we gain throughput */
6494 	if (capa_move > capa_now)
6495 		env->imbalance = busiest->load_per_task;
6496 }
6497 
6498 /**
6499  * calculate_imbalance - Calculate the amount of imbalance present within the
6500  *			 groups of a given sched_domain during load balance.
6501  * @env: load balance environment
6502  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6503  */
6504 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6505 {
6506 	unsigned long max_pull, load_above_capacity = ~0UL;
6507 	struct sg_lb_stats *local, *busiest;
6508 
6509 	local = &sds->local_stat;
6510 	busiest = &sds->busiest_stat;
6511 
6512 	if (busiest->group_type == group_imbalanced) {
6513 		/*
6514 		 * In the group_imb case we cannot rely on group-wide averages
6515 		 * to ensure cpu-load equilibrium, look at wider averages. XXX
6516 		 */
6517 		busiest->load_per_task =
6518 			min(busiest->load_per_task, sds->avg_load);
6519 	}
6520 
6521 	/*
6522 	 * In the presence of smp nice balancing, certain scenarios can have
6523 	 * max load less than avg load(as we skip the groups at or below
6524 	 * its cpu_capacity, while calculating max_load..)
6525 	 */
6526 	if (busiest->avg_load <= sds->avg_load ||
6527 	    local->avg_load >= sds->avg_load) {
6528 		env->imbalance = 0;
6529 		return fix_small_imbalance(env, sds);
6530 	}
6531 
6532 	/*
6533 	 * If there aren't any idle cpus, avoid creating some.
6534 	 */
6535 	if (busiest->group_type == group_overloaded &&
6536 	    local->group_type   == group_overloaded) {
6537 		load_above_capacity =
6538 			(busiest->sum_nr_running - busiest->group_capacity_factor);
6539 
6540 		load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE);
6541 		load_above_capacity /= busiest->group_capacity;
6542 	}
6543 
6544 	/*
6545 	 * We're trying to get all the cpus to the average_load, so we don't
6546 	 * want to push ourselves above the average load, nor do we wish to
6547 	 * reduce the max loaded cpu below the average load. At the same time,
6548 	 * we also don't want to reduce the group load below the group capacity
6549 	 * (so that we can implement power-savings policies etc). Thus we look
6550 	 * for the minimum possible imbalance.
6551 	 */
6552 	max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6553 
6554 	/* How much load to actually move to equalise the imbalance */
6555 	env->imbalance = min(
6556 		max_pull * busiest->group_capacity,
6557 		(sds->avg_load - local->avg_load) * local->group_capacity
6558 	) / SCHED_CAPACITY_SCALE;
6559 
6560 	/*
6561 	 * if *imbalance is less than the average load per runnable task
6562 	 * there is no guarantee that any tasks will be moved so we'll have
6563 	 * a think about bumping its value to force at least one task to be
6564 	 * moved
6565 	 */
6566 	if (env->imbalance < busiest->load_per_task)
6567 		return fix_small_imbalance(env, sds);
6568 }
6569 
6570 /******* find_busiest_group() helpers end here *********************/
6571 
6572 /**
6573  * find_busiest_group - Returns the busiest group within the sched_domain
6574  * if there is an imbalance. If there isn't an imbalance, and
6575  * the user has opted for power-savings, it returns a group whose
6576  * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6577  * such a group exists.
6578  *
6579  * Also calculates the amount of weighted load which should be moved
6580  * to restore balance.
6581  *
6582  * @env: The load balancing environment.
6583  *
6584  * Return:	- The busiest group if imbalance exists.
6585  *		- If no imbalance and user has opted for power-savings balance,
6586  *		   return the least loaded group whose CPUs can be
6587  *		   put to idle by rebalancing its tasks onto our group.
6588  */
6589 static struct sched_group *find_busiest_group(struct lb_env *env)
6590 {
6591 	struct sg_lb_stats *local, *busiest;
6592 	struct sd_lb_stats sds;
6593 
6594 	init_sd_lb_stats(&sds);
6595 
6596 	/*
6597 	 * Compute the various statistics relavent for load balancing at
6598 	 * this level.
6599 	 */
6600 	update_sd_lb_stats(env, &sds);
6601 	local = &sds.local_stat;
6602 	busiest = &sds.busiest_stat;
6603 
6604 	if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6605 	    check_asym_packing(env, &sds))
6606 		return sds.busiest;
6607 
6608 	/* There is no busy sibling group to pull tasks from */
6609 	if (!sds.busiest || busiest->sum_nr_running == 0)
6610 		goto out_balanced;
6611 
6612 	sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6613 						/ sds.total_capacity;
6614 
6615 	/*
6616 	 * If the busiest group is imbalanced the below checks don't
6617 	 * work because they assume all things are equal, which typically
6618 	 * isn't true due to cpus_allowed constraints and the like.
6619 	 */
6620 	if (busiest->group_type == group_imbalanced)
6621 		goto force_balance;
6622 
6623 	/* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6624 	if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity &&
6625 	    !busiest->group_has_free_capacity)
6626 		goto force_balance;
6627 
6628 	/*
6629 	 * If the local group is busier than the selected busiest group
6630 	 * don't try and pull any tasks.
6631 	 */
6632 	if (local->avg_load >= busiest->avg_load)
6633 		goto out_balanced;
6634 
6635 	/*
6636 	 * Don't pull any tasks if this group is already above the domain
6637 	 * average load.
6638 	 */
6639 	if (local->avg_load >= sds.avg_load)
6640 		goto out_balanced;
6641 
6642 	if (env->idle == CPU_IDLE) {
6643 		/*
6644 		 * This cpu is idle. If the busiest group is not overloaded
6645 		 * and there is no imbalance between this and busiest group
6646 		 * wrt idle cpus, it is balanced. The imbalance becomes
6647 		 * significant if the diff is greater than 1 otherwise we
6648 		 * might end up to just move the imbalance on another group
6649 		 */
6650 		if ((busiest->group_type != group_overloaded) &&
6651 				(local->idle_cpus <= (busiest->idle_cpus + 1)))
6652 			goto out_balanced;
6653 	} else {
6654 		/*
6655 		 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6656 		 * imbalance_pct to be conservative.
6657 		 */
6658 		if (100 * busiest->avg_load <=
6659 				env->sd->imbalance_pct * local->avg_load)
6660 			goto out_balanced;
6661 	}
6662 
6663 force_balance:
6664 	/* Looks like there is an imbalance. Compute it */
6665 	calculate_imbalance(env, &sds);
6666 	return sds.busiest;
6667 
6668 out_balanced:
6669 	env->imbalance = 0;
6670 	return NULL;
6671 }
6672 
6673 /*
6674  * find_busiest_queue - find the busiest runqueue among the cpus in group.
6675  */
6676 static struct rq *find_busiest_queue(struct lb_env *env,
6677 				     struct sched_group *group)
6678 {
6679 	struct rq *busiest = NULL, *rq;
6680 	unsigned long busiest_load = 0, busiest_capacity = 1;
6681 	int i;
6682 
6683 	for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6684 		unsigned long capacity, capacity_factor, wl;
6685 		enum fbq_type rt;
6686 
6687 		rq = cpu_rq(i);
6688 		rt = fbq_classify_rq(rq);
6689 
6690 		/*
6691 		 * We classify groups/runqueues into three groups:
6692 		 *  - regular: there are !numa tasks
6693 		 *  - remote:  there are numa tasks that run on the 'wrong' node
6694 		 *  - all:     there is no distinction
6695 		 *
6696 		 * In order to avoid migrating ideally placed numa tasks,
6697 		 * ignore those when there's better options.
6698 		 *
6699 		 * If we ignore the actual busiest queue to migrate another
6700 		 * task, the next balance pass can still reduce the busiest
6701 		 * queue by moving tasks around inside the node.
6702 		 *
6703 		 * If we cannot move enough load due to this classification
6704 		 * the next pass will adjust the group classification and
6705 		 * allow migration of more tasks.
6706 		 *
6707 		 * Both cases only affect the total convergence complexity.
6708 		 */
6709 		if (rt > env->fbq_type)
6710 			continue;
6711 
6712 		capacity = capacity_of(i);
6713 		capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE);
6714 		if (!capacity_factor)
6715 			capacity_factor = fix_small_capacity(env->sd, group);
6716 
6717 		wl = weighted_cpuload(i);
6718 
6719 		/*
6720 		 * When comparing with imbalance, use weighted_cpuload()
6721 		 * which is not scaled with the cpu capacity.
6722 		 */
6723 		if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance)
6724 			continue;
6725 
6726 		/*
6727 		 * For the load comparisons with the other cpu's, consider
6728 		 * the weighted_cpuload() scaled with the cpu capacity, so
6729 		 * that the load can be moved away from the cpu that is
6730 		 * potentially running at a lower capacity.
6731 		 *
6732 		 * Thus we're looking for max(wl_i / capacity_i), crosswise
6733 		 * multiplication to rid ourselves of the division works out
6734 		 * to: wl_i * capacity_j > wl_j * capacity_i;  where j is
6735 		 * our previous maximum.
6736 		 */
6737 		if (wl * busiest_capacity > busiest_load * capacity) {
6738 			busiest_load = wl;
6739 			busiest_capacity = capacity;
6740 			busiest = rq;
6741 		}
6742 	}
6743 
6744 	return busiest;
6745 }
6746 
6747 /*
6748  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6749  * so long as it is large enough.
6750  */
6751 #define MAX_PINNED_INTERVAL	512
6752 
6753 /* Working cpumask for load_balance and load_balance_newidle. */
6754 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6755 
6756 static int need_active_balance(struct lb_env *env)
6757 {
6758 	struct sched_domain *sd = env->sd;
6759 
6760 	if (env->idle == CPU_NEWLY_IDLE) {
6761 
6762 		/*
6763 		 * ASYM_PACKING needs to force migrate tasks from busy but
6764 		 * higher numbered CPUs in order to pack all tasks in the
6765 		 * lowest numbered CPUs.
6766 		 */
6767 		if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6768 			return 1;
6769 	}
6770 
6771 	return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6772 }
6773 
6774 static int active_load_balance_cpu_stop(void *data);
6775 
6776 static int should_we_balance(struct lb_env *env)
6777 {
6778 	struct sched_group *sg = env->sd->groups;
6779 	struct cpumask *sg_cpus, *sg_mask;
6780 	int cpu, balance_cpu = -1;
6781 
6782 	/*
6783 	 * In the newly idle case, we will allow all the cpu's
6784 	 * to do the newly idle load balance.
6785 	 */
6786 	if (env->idle == CPU_NEWLY_IDLE)
6787 		return 1;
6788 
6789 	sg_cpus = sched_group_cpus(sg);
6790 	sg_mask = sched_group_mask(sg);
6791 	/* Try to find first idle cpu */
6792 	for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6793 		if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6794 			continue;
6795 
6796 		balance_cpu = cpu;
6797 		break;
6798 	}
6799 
6800 	if (balance_cpu == -1)
6801 		balance_cpu = group_balance_cpu(sg);
6802 
6803 	/*
6804 	 * First idle cpu or the first cpu(busiest) in this sched group
6805 	 * is eligible for doing load balancing at this and above domains.
6806 	 */
6807 	return balance_cpu == env->dst_cpu;
6808 }
6809 
6810 /*
6811  * Check this_cpu to ensure it is balanced within domain. Attempt to move
6812  * tasks if there is an imbalance.
6813  */
6814 static int load_balance(int this_cpu, struct rq *this_rq,
6815 			struct sched_domain *sd, enum cpu_idle_type idle,
6816 			int *continue_balancing)
6817 {
6818 	int ld_moved, cur_ld_moved, active_balance = 0;
6819 	struct sched_domain *sd_parent = sd->parent;
6820 	struct sched_group *group;
6821 	struct rq *busiest;
6822 	unsigned long flags;
6823 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6824 
6825 	struct lb_env env = {
6826 		.sd		= sd,
6827 		.dst_cpu	= this_cpu,
6828 		.dst_rq		= this_rq,
6829 		.dst_grpmask    = sched_group_cpus(sd->groups),
6830 		.idle		= idle,
6831 		.loop_break	= sched_nr_migrate_break,
6832 		.cpus		= cpus,
6833 		.fbq_type	= all,
6834 		.tasks		= LIST_HEAD_INIT(env.tasks),
6835 	};
6836 
6837 	/*
6838 	 * For NEWLY_IDLE load_balancing, we don't need to consider
6839 	 * other cpus in our group
6840 	 */
6841 	if (idle == CPU_NEWLY_IDLE)
6842 		env.dst_grpmask = NULL;
6843 
6844 	cpumask_copy(cpus, cpu_active_mask);
6845 
6846 	schedstat_inc(sd, lb_count[idle]);
6847 
6848 redo:
6849 	if (!should_we_balance(&env)) {
6850 		*continue_balancing = 0;
6851 		goto out_balanced;
6852 	}
6853 
6854 	group = find_busiest_group(&env);
6855 	if (!group) {
6856 		schedstat_inc(sd, lb_nobusyg[idle]);
6857 		goto out_balanced;
6858 	}
6859 
6860 	busiest = find_busiest_queue(&env, group);
6861 	if (!busiest) {
6862 		schedstat_inc(sd, lb_nobusyq[idle]);
6863 		goto out_balanced;
6864 	}
6865 
6866 	BUG_ON(busiest == env.dst_rq);
6867 
6868 	schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6869 
6870 	ld_moved = 0;
6871 	if (busiest->nr_running > 1) {
6872 		/*
6873 		 * Attempt to move tasks. If find_busiest_group has found
6874 		 * an imbalance but busiest->nr_running <= 1, the group is
6875 		 * still unbalanced. ld_moved simply stays zero, so it is
6876 		 * correctly treated as an imbalance.
6877 		 */
6878 		env.flags |= LBF_ALL_PINNED;
6879 		env.src_cpu   = busiest->cpu;
6880 		env.src_rq    = busiest;
6881 		env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
6882 
6883 more_balance:
6884 		raw_spin_lock_irqsave(&busiest->lock, flags);
6885 
6886 		/*
6887 		 * cur_ld_moved - load moved in current iteration
6888 		 * ld_moved     - cumulative load moved across iterations
6889 		 */
6890 		cur_ld_moved = detach_tasks(&env);
6891 
6892 		/*
6893 		 * We've detached some tasks from busiest_rq. Every
6894 		 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6895 		 * unlock busiest->lock, and we are able to be sure
6896 		 * that nobody can manipulate the tasks in parallel.
6897 		 * See task_rq_lock() family for the details.
6898 		 */
6899 
6900 		raw_spin_unlock(&busiest->lock);
6901 
6902 		if (cur_ld_moved) {
6903 			attach_tasks(&env);
6904 			ld_moved += cur_ld_moved;
6905 		}
6906 
6907 		local_irq_restore(flags);
6908 
6909 		if (env.flags & LBF_NEED_BREAK) {
6910 			env.flags &= ~LBF_NEED_BREAK;
6911 			goto more_balance;
6912 		}
6913 
6914 		/*
6915 		 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6916 		 * us and move them to an alternate dst_cpu in our sched_group
6917 		 * where they can run. The upper limit on how many times we
6918 		 * iterate on same src_cpu is dependent on number of cpus in our
6919 		 * sched_group.
6920 		 *
6921 		 * This changes load balance semantics a bit on who can move
6922 		 * load to a given_cpu. In addition to the given_cpu itself
6923 		 * (or a ilb_cpu acting on its behalf where given_cpu is
6924 		 * nohz-idle), we now have balance_cpu in a position to move
6925 		 * load to given_cpu. In rare situations, this may cause
6926 		 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6927 		 * _independently_ and at _same_ time to move some load to
6928 		 * given_cpu) causing exceess load to be moved to given_cpu.
6929 		 * This however should not happen so much in practice and
6930 		 * moreover subsequent load balance cycles should correct the
6931 		 * excess load moved.
6932 		 */
6933 		if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6934 
6935 			/* Prevent to re-select dst_cpu via env's cpus */
6936 			cpumask_clear_cpu(env.dst_cpu, env.cpus);
6937 
6938 			env.dst_rq	 = cpu_rq(env.new_dst_cpu);
6939 			env.dst_cpu	 = env.new_dst_cpu;
6940 			env.flags	&= ~LBF_DST_PINNED;
6941 			env.loop	 = 0;
6942 			env.loop_break	 = sched_nr_migrate_break;
6943 
6944 			/*
6945 			 * Go back to "more_balance" rather than "redo" since we
6946 			 * need to continue with same src_cpu.
6947 			 */
6948 			goto more_balance;
6949 		}
6950 
6951 		/*
6952 		 * We failed to reach balance because of affinity.
6953 		 */
6954 		if (sd_parent) {
6955 			int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6956 
6957 			if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
6958 				*group_imbalance = 1;
6959 		}
6960 
6961 		/* All tasks on this runqueue were pinned by CPU affinity */
6962 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
6963 			cpumask_clear_cpu(cpu_of(busiest), cpus);
6964 			if (!cpumask_empty(cpus)) {
6965 				env.loop = 0;
6966 				env.loop_break = sched_nr_migrate_break;
6967 				goto redo;
6968 			}
6969 			goto out_all_pinned;
6970 		}
6971 	}
6972 
6973 	if (!ld_moved) {
6974 		schedstat_inc(sd, lb_failed[idle]);
6975 		/*
6976 		 * Increment the failure counter only on periodic balance.
6977 		 * We do not want newidle balance, which can be very
6978 		 * frequent, pollute the failure counter causing
6979 		 * excessive cache_hot migrations and active balances.
6980 		 */
6981 		if (idle != CPU_NEWLY_IDLE)
6982 			sd->nr_balance_failed++;
6983 
6984 		if (need_active_balance(&env)) {
6985 			raw_spin_lock_irqsave(&busiest->lock, flags);
6986 
6987 			/* don't kick the active_load_balance_cpu_stop,
6988 			 * if the curr task on busiest cpu can't be
6989 			 * moved to this_cpu
6990 			 */
6991 			if (!cpumask_test_cpu(this_cpu,
6992 					tsk_cpus_allowed(busiest->curr))) {
6993 				raw_spin_unlock_irqrestore(&busiest->lock,
6994 							    flags);
6995 				env.flags |= LBF_ALL_PINNED;
6996 				goto out_one_pinned;
6997 			}
6998 
6999 			/*
7000 			 * ->active_balance synchronizes accesses to
7001 			 * ->active_balance_work.  Once set, it's cleared
7002 			 * only after active load balance is finished.
7003 			 */
7004 			if (!busiest->active_balance) {
7005 				busiest->active_balance = 1;
7006 				busiest->push_cpu = this_cpu;
7007 				active_balance = 1;
7008 			}
7009 			raw_spin_unlock_irqrestore(&busiest->lock, flags);
7010 
7011 			if (active_balance) {
7012 				stop_one_cpu_nowait(cpu_of(busiest),
7013 					active_load_balance_cpu_stop, busiest,
7014 					&busiest->active_balance_work);
7015 			}
7016 
7017 			/*
7018 			 * We've kicked active balancing, reset the failure
7019 			 * counter.
7020 			 */
7021 			sd->nr_balance_failed = sd->cache_nice_tries+1;
7022 		}
7023 	} else
7024 		sd->nr_balance_failed = 0;
7025 
7026 	if (likely(!active_balance)) {
7027 		/* We were unbalanced, so reset the balancing interval */
7028 		sd->balance_interval = sd->min_interval;
7029 	} else {
7030 		/*
7031 		 * If we've begun active balancing, start to back off. This
7032 		 * case may not be covered by the all_pinned logic if there
7033 		 * is only 1 task on the busy runqueue (because we don't call
7034 		 * detach_tasks).
7035 		 */
7036 		if (sd->balance_interval < sd->max_interval)
7037 			sd->balance_interval *= 2;
7038 	}
7039 
7040 	goto out;
7041 
7042 out_balanced:
7043 	/*
7044 	 * We reach balance although we may have faced some affinity
7045 	 * constraints. Clear the imbalance flag if it was set.
7046 	 */
7047 	if (sd_parent) {
7048 		int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7049 
7050 		if (*group_imbalance)
7051 			*group_imbalance = 0;
7052 	}
7053 
7054 out_all_pinned:
7055 	/*
7056 	 * We reach balance because all tasks are pinned at this level so
7057 	 * we can't migrate them. Let the imbalance flag set so parent level
7058 	 * can try to migrate them.
7059 	 */
7060 	schedstat_inc(sd, lb_balanced[idle]);
7061 
7062 	sd->nr_balance_failed = 0;
7063 
7064 out_one_pinned:
7065 	/* tune up the balancing interval */
7066 	if (((env.flags & LBF_ALL_PINNED) &&
7067 			sd->balance_interval < MAX_PINNED_INTERVAL) ||
7068 			(sd->balance_interval < sd->max_interval))
7069 		sd->balance_interval *= 2;
7070 
7071 	ld_moved = 0;
7072 out:
7073 	return ld_moved;
7074 }
7075 
7076 static inline unsigned long
7077 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7078 {
7079 	unsigned long interval = sd->balance_interval;
7080 
7081 	if (cpu_busy)
7082 		interval *= sd->busy_factor;
7083 
7084 	/* scale ms to jiffies */
7085 	interval = msecs_to_jiffies(interval);
7086 	interval = clamp(interval, 1UL, max_load_balance_interval);
7087 
7088 	return interval;
7089 }
7090 
7091 static inline void
7092 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7093 {
7094 	unsigned long interval, next;
7095 
7096 	interval = get_sd_balance_interval(sd, cpu_busy);
7097 	next = sd->last_balance + interval;
7098 
7099 	if (time_after(*next_balance, next))
7100 		*next_balance = next;
7101 }
7102 
7103 /*
7104  * idle_balance is called by schedule() if this_cpu is about to become
7105  * idle. Attempts to pull tasks from other CPUs.
7106  */
7107 static int idle_balance(struct rq *this_rq)
7108 {
7109 	unsigned long next_balance = jiffies + HZ;
7110 	int this_cpu = this_rq->cpu;
7111 	struct sched_domain *sd;
7112 	int pulled_task = 0;
7113 	u64 curr_cost = 0;
7114 
7115 	idle_enter_fair(this_rq);
7116 
7117 	/*
7118 	 * We must set idle_stamp _before_ calling idle_balance(), such that we
7119 	 * measure the duration of idle_balance() as idle time.
7120 	 */
7121 	this_rq->idle_stamp = rq_clock(this_rq);
7122 
7123 	if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7124 	    !this_rq->rd->overload) {
7125 		rcu_read_lock();
7126 		sd = rcu_dereference_check_sched_domain(this_rq->sd);
7127 		if (sd)
7128 			update_next_balance(sd, 0, &next_balance);
7129 		rcu_read_unlock();
7130 
7131 		goto out;
7132 	}
7133 
7134 	/*
7135 	 * Drop the rq->lock, but keep IRQ/preempt disabled.
7136 	 */
7137 	raw_spin_unlock(&this_rq->lock);
7138 
7139 	update_blocked_averages(this_cpu);
7140 	rcu_read_lock();
7141 	for_each_domain(this_cpu, sd) {
7142 		int continue_balancing = 1;
7143 		u64 t0, domain_cost;
7144 
7145 		if (!(sd->flags & SD_LOAD_BALANCE))
7146 			continue;
7147 
7148 		if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7149 			update_next_balance(sd, 0, &next_balance);
7150 			break;
7151 		}
7152 
7153 		if (sd->flags & SD_BALANCE_NEWIDLE) {
7154 			t0 = sched_clock_cpu(this_cpu);
7155 
7156 			pulled_task = load_balance(this_cpu, this_rq,
7157 						   sd, CPU_NEWLY_IDLE,
7158 						   &continue_balancing);
7159 
7160 			domain_cost = sched_clock_cpu(this_cpu) - t0;
7161 			if (domain_cost > sd->max_newidle_lb_cost)
7162 				sd->max_newidle_lb_cost = domain_cost;
7163 
7164 			curr_cost += domain_cost;
7165 		}
7166 
7167 		update_next_balance(sd, 0, &next_balance);
7168 
7169 		/*
7170 		 * Stop searching for tasks to pull if there are
7171 		 * now runnable tasks on this rq.
7172 		 */
7173 		if (pulled_task || this_rq->nr_running > 0)
7174 			break;
7175 	}
7176 	rcu_read_unlock();
7177 
7178 	raw_spin_lock(&this_rq->lock);
7179 
7180 	if (curr_cost > this_rq->max_idle_balance_cost)
7181 		this_rq->max_idle_balance_cost = curr_cost;
7182 
7183 	/*
7184 	 * While browsing the domains, we released the rq lock, a task could
7185 	 * have been enqueued in the meantime. Since we're not going idle,
7186 	 * pretend we pulled a task.
7187 	 */
7188 	if (this_rq->cfs.h_nr_running && !pulled_task)
7189 		pulled_task = 1;
7190 
7191 out:
7192 	/* Move the next balance forward */
7193 	if (time_after(this_rq->next_balance, next_balance))
7194 		this_rq->next_balance = next_balance;
7195 
7196 	/* Is there a task of a high priority class? */
7197 	if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7198 		pulled_task = -1;
7199 
7200 	if (pulled_task) {
7201 		idle_exit_fair(this_rq);
7202 		this_rq->idle_stamp = 0;
7203 	}
7204 
7205 	return pulled_task;
7206 }
7207 
7208 /*
7209  * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7210  * running tasks off the busiest CPU onto idle CPUs. It requires at
7211  * least 1 task to be running on each physical CPU where possible, and
7212  * avoids physical / logical imbalances.
7213  */
7214 static int active_load_balance_cpu_stop(void *data)
7215 {
7216 	struct rq *busiest_rq = data;
7217 	int busiest_cpu = cpu_of(busiest_rq);
7218 	int target_cpu = busiest_rq->push_cpu;
7219 	struct rq *target_rq = cpu_rq(target_cpu);
7220 	struct sched_domain *sd;
7221 	struct task_struct *p = NULL;
7222 
7223 	raw_spin_lock_irq(&busiest_rq->lock);
7224 
7225 	/* make sure the requested cpu hasn't gone down in the meantime */
7226 	if (unlikely(busiest_cpu != smp_processor_id() ||
7227 		     !busiest_rq->active_balance))
7228 		goto out_unlock;
7229 
7230 	/* Is there any task to move? */
7231 	if (busiest_rq->nr_running <= 1)
7232 		goto out_unlock;
7233 
7234 	/*
7235 	 * This condition is "impossible", if it occurs
7236 	 * we need to fix it. Originally reported by
7237 	 * Bjorn Helgaas on a 128-cpu setup.
7238 	 */
7239 	BUG_ON(busiest_rq == target_rq);
7240 
7241 	/* Search for an sd spanning us and the target CPU. */
7242 	rcu_read_lock();
7243 	for_each_domain(target_cpu, sd) {
7244 		if ((sd->flags & SD_LOAD_BALANCE) &&
7245 		    cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7246 				break;
7247 	}
7248 
7249 	if (likely(sd)) {
7250 		struct lb_env env = {
7251 			.sd		= sd,
7252 			.dst_cpu	= target_cpu,
7253 			.dst_rq		= target_rq,
7254 			.src_cpu	= busiest_rq->cpu,
7255 			.src_rq		= busiest_rq,
7256 			.idle		= CPU_IDLE,
7257 		};
7258 
7259 		schedstat_inc(sd, alb_count);
7260 
7261 		p = detach_one_task(&env);
7262 		if (p)
7263 			schedstat_inc(sd, alb_pushed);
7264 		else
7265 			schedstat_inc(sd, alb_failed);
7266 	}
7267 	rcu_read_unlock();
7268 out_unlock:
7269 	busiest_rq->active_balance = 0;
7270 	raw_spin_unlock(&busiest_rq->lock);
7271 
7272 	if (p)
7273 		attach_one_task(target_rq, p);
7274 
7275 	local_irq_enable();
7276 
7277 	return 0;
7278 }
7279 
7280 static inline int on_null_domain(struct rq *rq)
7281 {
7282 	return unlikely(!rcu_dereference_sched(rq->sd));
7283 }
7284 
7285 #ifdef CONFIG_NO_HZ_COMMON
7286 /*
7287  * idle load balancing details
7288  * - When one of the busy CPUs notice that there may be an idle rebalancing
7289  *   needed, they will kick the idle load balancer, which then does idle
7290  *   load balancing for all the idle CPUs.
7291  */
7292 static struct {
7293 	cpumask_var_t idle_cpus_mask;
7294 	atomic_t nr_cpus;
7295 	unsigned long next_balance;     /* in jiffy units */
7296 } nohz ____cacheline_aligned;
7297 
7298 static inline int find_new_ilb(void)
7299 {
7300 	int ilb = cpumask_first(nohz.idle_cpus_mask);
7301 
7302 	if (ilb < nr_cpu_ids && idle_cpu(ilb))
7303 		return ilb;
7304 
7305 	return nr_cpu_ids;
7306 }
7307 
7308 /*
7309  * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7310  * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7311  * CPU (if there is one).
7312  */
7313 static void nohz_balancer_kick(void)
7314 {
7315 	int ilb_cpu;
7316 
7317 	nohz.next_balance++;
7318 
7319 	ilb_cpu = find_new_ilb();
7320 
7321 	if (ilb_cpu >= nr_cpu_ids)
7322 		return;
7323 
7324 	if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7325 		return;
7326 	/*
7327 	 * Use smp_send_reschedule() instead of resched_cpu().
7328 	 * This way we generate a sched IPI on the target cpu which
7329 	 * is idle. And the softirq performing nohz idle load balance
7330 	 * will be run before returning from the IPI.
7331 	 */
7332 	smp_send_reschedule(ilb_cpu);
7333 	return;
7334 }
7335 
7336 static inline void nohz_balance_exit_idle(int cpu)
7337 {
7338 	if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7339 		/*
7340 		 * Completely isolated CPUs don't ever set, so we must test.
7341 		 */
7342 		if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7343 			cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7344 			atomic_dec(&nohz.nr_cpus);
7345 		}
7346 		clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7347 	}
7348 }
7349 
7350 static inline void set_cpu_sd_state_busy(void)
7351 {
7352 	struct sched_domain *sd;
7353 	int cpu = smp_processor_id();
7354 
7355 	rcu_read_lock();
7356 	sd = rcu_dereference(per_cpu(sd_busy, cpu));
7357 
7358 	if (!sd || !sd->nohz_idle)
7359 		goto unlock;
7360 	sd->nohz_idle = 0;
7361 
7362 	atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7363 unlock:
7364 	rcu_read_unlock();
7365 }
7366 
7367 void set_cpu_sd_state_idle(void)
7368 {
7369 	struct sched_domain *sd;
7370 	int cpu = smp_processor_id();
7371 
7372 	rcu_read_lock();
7373 	sd = rcu_dereference(per_cpu(sd_busy, cpu));
7374 
7375 	if (!sd || sd->nohz_idle)
7376 		goto unlock;
7377 	sd->nohz_idle = 1;
7378 
7379 	atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7380 unlock:
7381 	rcu_read_unlock();
7382 }
7383 
7384 /*
7385  * This routine will record that the cpu is going idle with tick stopped.
7386  * This info will be used in performing idle load balancing in the future.
7387  */
7388 void nohz_balance_enter_idle(int cpu)
7389 {
7390 	/*
7391 	 * If this cpu is going down, then nothing needs to be done.
7392 	 */
7393 	if (!cpu_active(cpu))
7394 		return;
7395 
7396 	if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7397 		return;
7398 
7399 	/*
7400 	 * If we're a completely isolated CPU, we don't play.
7401 	 */
7402 	if (on_null_domain(cpu_rq(cpu)))
7403 		return;
7404 
7405 	cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7406 	atomic_inc(&nohz.nr_cpus);
7407 	set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7408 }
7409 
7410 static int sched_ilb_notifier(struct notifier_block *nfb,
7411 					unsigned long action, void *hcpu)
7412 {
7413 	switch (action & ~CPU_TASKS_FROZEN) {
7414 	case CPU_DYING:
7415 		nohz_balance_exit_idle(smp_processor_id());
7416 		return NOTIFY_OK;
7417 	default:
7418 		return NOTIFY_DONE;
7419 	}
7420 }
7421 #endif
7422 
7423 static DEFINE_SPINLOCK(balancing);
7424 
7425 /*
7426  * Scale the max load_balance interval with the number of CPUs in the system.
7427  * This trades load-balance latency on larger machines for less cross talk.
7428  */
7429 void update_max_interval(void)
7430 {
7431 	max_load_balance_interval = HZ*num_online_cpus()/10;
7432 }
7433 
7434 /*
7435  * It checks each scheduling domain to see if it is due to be balanced,
7436  * and initiates a balancing operation if so.
7437  *
7438  * Balancing parameters are set up in init_sched_domains.
7439  */
7440 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7441 {
7442 	int continue_balancing = 1;
7443 	int cpu = rq->cpu;
7444 	unsigned long interval;
7445 	struct sched_domain *sd;
7446 	/* Earliest time when we have to do rebalance again */
7447 	unsigned long next_balance = jiffies + 60*HZ;
7448 	int update_next_balance = 0;
7449 	int need_serialize, need_decay = 0;
7450 	u64 max_cost = 0;
7451 
7452 	update_blocked_averages(cpu);
7453 
7454 	rcu_read_lock();
7455 	for_each_domain(cpu, sd) {
7456 		/*
7457 		 * Decay the newidle max times here because this is a regular
7458 		 * visit to all the domains. Decay ~1% per second.
7459 		 */
7460 		if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7461 			sd->max_newidle_lb_cost =
7462 				(sd->max_newidle_lb_cost * 253) / 256;
7463 			sd->next_decay_max_lb_cost = jiffies + HZ;
7464 			need_decay = 1;
7465 		}
7466 		max_cost += sd->max_newidle_lb_cost;
7467 
7468 		if (!(sd->flags & SD_LOAD_BALANCE))
7469 			continue;
7470 
7471 		/*
7472 		 * Stop the load balance at this level. There is another
7473 		 * CPU in our sched group which is doing load balancing more
7474 		 * actively.
7475 		 */
7476 		if (!continue_balancing) {
7477 			if (need_decay)
7478 				continue;
7479 			break;
7480 		}
7481 
7482 		interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7483 
7484 		need_serialize = sd->flags & SD_SERIALIZE;
7485 		if (need_serialize) {
7486 			if (!spin_trylock(&balancing))
7487 				goto out;
7488 		}
7489 
7490 		if (time_after_eq(jiffies, sd->last_balance + interval)) {
7491 			if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7492 				/*
7493 				 * The LBF_DST_PINNED logic could have changed
7494 				 * env->dst_cpu, so we can't know our idle
7495 				 * state even if we migrated tasks. Update it.
7496 				 */
7497 				idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7498 			}
7499 			sd->last_balance = jiffies;
7500 			interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7501 		}
7502 		if (need_serialize)
7503 			spin_unlock(&balancing);
7504 out:
7505 		if (time_after(next_balance, sd->last_balance + interval)) {
7506 			next_balance = sd->last_balance + interval;
7507 			update_next_balance = 1;
7508 		}
7509 	}
7510 	if (need_decay) {
7511 		/*
7512 		 * Ensure the rq-wide value also decays but keep it at a
7513 		 * reasonable floor to avoid funnies with rq->avg_idle.
7514 		 */
7515 		rq->max_idle_balance_cost =
7516 			max((u64)sysctl_sched_migration_cost, max_cost);
7517 	}
7518 	rcu_read_unlock();
7519 
7520 	/*
7521 	 * next_balance will be updated only when there is a need.
7522 	 * When the cpu is attached to null domain for ex, it will not be
7523 	 * updated.
7524 	 */
7525 	if (likely(update_next_balance))
7526 		rq->next_balance = next_balance;
7527 }
7528 
7529 #ifdef CONFIG_NO_HZ_COMMON
7530 /*
7531  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7532  * rebalancing for all the cpus for whom scheduler ticks are stopped.
7533  */
7534 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7535 {
7536 	int this_cpu = this_rq->cpu;
7537 	struct rq *rq;
7538 	int balance_cpu;
7539 
7540 	if (idle != CPU_IDLE ||
7541 	    !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7542 		goto end;
7543 
7544 	for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7545 		if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7546 			continue;
7547 
7548 		/*
7549 		 * If this cpu gets work to do, stop the load balancing
7550 		 * work being done for other cpus. Next load
7551 		 * balancing owner will pick it up.
7552 		 */
7553 		if (need_resched())
7554 			break;
7555 
7556 		rq = cpu_rq(balance_cpu);
7557 
7558 		/*
7559 		 * If time for next balance is due,
7560 		 * do the balance.
7561 		 */
7562 		if (time_after_eq(jiffies, rq->next_balance)) {
7563 			raw_spin_lock_irq(&rq->lock);
7564 			update_rq_clock(rq);
7565 			update_idle_cpu_load(rq);
7566 			raw_spin_unlock_irq(&rq->lock);
7567 			rebalance_domains(rq, CPU_IDLE);
7568 		}
7569 
7570 		if (time_after(this_rq->next_balance, rq->next_balance))
7571 			this_rq->next_balance = rq->next_balance;
7572 	}
7573 	nohz.next_balance = this_rq->next_balance;
7574 end:
7575 	clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7576 }
7577 
7578 /*
7579  * Current heuristic for kicking the idle load balancer in the presence
7580  * of an idle cpu is the system.
7581  *   - This rq has more than one task.
7582  *   - At any scheduler domain level, this cpu's scheduler group has multiple
7583  *     busy cpu's exceeding the group's capacity.
7584  *   - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7585  *     domain span are idle.
7586  */
7587 static inline int nohz_kick_needed(struct rq *rq)
7588 {
7589 	unsigned long now = jiffies;
7590 	struct sched_domain *sd;
7591 	struct sched_group_capacity *sgc;
7592 	int nr_busy, cpu = rq->cpu;
7593 
7594 	if (unlikely(rq->idle_balance))
7595 		return 0;
7596 
7597        /*
7598 	* We may be recently in ticked or tickless idle mode. At the first
7599 	* busy tick after returning from idle, we will update the busy stats.
7600 	*/
7601 	set_cpu_sd_state_busy();
7602 	nohz_balance_exit_idle(cpu);
7603 
7604 	/*
7605 	 * None are in tickless mode and hence no need for NOHZ idle load
7606 	 * balancing.
7607 	 */
7608 	if (likely(!atomic_read(&nohz.nr_cpus)))
7609 		return 0;
7610 
7611 	if (time_before(now, nohz.next_balance))
7612 		return 0;
7613 
7614 	if (rq->nr_running >= 2)
7615 		goto need_kick;
7616 
7617 	rcu_read_lock();
7618 	sd = rcu_dereference(per_cpu(sd_busy, cpu));
7619 
7620 	if (sd) {
7621 		sgc = sd->groups->sgc;
7622 		nr_busy = atomic_read(&sgc->nr_busy_cpus);
7623 
7624 		if (nr_busy > 1)
7625 			goto need_kick_unlock;
7626 	}
7627 
7628 	sd = rcu_dereference(per_cpu(sd_asym, cpu));
7629 
7630 	if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7631 				  sched_domain_span(sd)) < cpu))
7632 		goto need_kick_unlock;
7633 
7634 	rcu_read_unlock();
7635 	return 0;
7636 
7637 need_kick_unlock:
7638 	rcu_read_unlock();
7639 need_kick:
7640 	return 1;
7641 }
7642 #else
7643 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7644 #endif
7645 
7646 /*
7647  * run_rebalance_domains is triggered when needed from the scheduler tick.
7648  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7649  */
7650 static void run_rebalance_domains(struct softirq_action *h)
7651 {
7652 	struct rq *this_rq = this_rq();
7653 	enum cpu_idle_type idle = this_rq->idle_balance ?
7654 						CPU_IDLE : CPU_NOT_IDLE;
7655 
7656 	rebalance_domains(this_rq, idle);
7657 
7658 	/*
7659 	 * If this cpu has a pending nohz_balance_kick, then do the
7660 	 * balancing on behalf of the other idle cpus whose ticks are
7661 	 * stopped.
7662 	 */
7663 	nohz_idle_balance(this_rq, idle);
7664 }
7665 
7666 /*
7667  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7668  */
7669 void trigger_load_balance(struct rq *rq)
7670 {
7671 	/* Don't need to rebalance while attached to NULL domain */
7672 	if (unlikely(on_null_domain(rq)))
7673 		return;
7674 
7675 	if (time_after_eq(jiffies, rq->next_balance))
7676 		raise_softirq(SCHED_SOFTIRQ);
7677 #ifdef CONFIG_NO_HZ_COMMON
7678 	if (nohz_kick_needed(rq))
7679 		nohz_balancer_kick();
7680 #endif
7681 }
7682 
7683 static void rq_online_fair(struct rq *rq)
7684 {
7685 	update_sysctl();
7686 
7687 	update_runtime_enabled(rq);
7688 }
7689 
7690 static void rq_offline_fair(struct rq *rq)
7691 {
7692 	update_sysctl();
7693 
7694 	/* Ensure any throttled groups are reachable by pick_next_task */
7695 	unthrottle_offline_cfs_rqs(rq);
7696 }
7697 
7698 #endif /* CONFIG_SMP */
7699 
7700 /*
7701  * scheduler tick hitting a task of our scheduling class:
7702  */
7703 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7704 {
7705 	struct cfs_rq *cfs_rq;
7706 	struct sched_entity *se = &curr->se;
7707 
7708 	for_each_sched_entity(se) {
7709 		cfs_rq = cfs_rq_of(se);
7710 		entity_tick(cfs_rq, se, queued);
7711 	}
7712 
7713 	if (numabalancing_enabled)
7714 		task_tick_numa(rq, curr);
7715 
7716 	update_rq_runnable_avg(rq, 1);
7717 }
7718 
7719 /*
7720  * called on fork with the child task as argument from the parent's context
7721  *  - child not yet on the tasklist
7722  *  - preemption disabled
7723  */
7724 static void task_fork_fair(struct task_struct *p)
7725 {
7726 	struct cfs_rq *cfs_rq;
7727 	struct sched_entity *se = &p->se, *curr;
7728 	int this_cpu = smp_processor_id();
7729 	struct rq *rq = this_rq();
7730 	unsigned long flags;
7731 
7732 	raw_spin_lock_irqsave(&rq->lock, flags);
7733 
7734 	update_rq_clock(rq);
7735 
7736 	cfs_rq = task_cfs_rq(current);
7737 	curr = cfs_rq->curr;
7738 
7739 	/*
7740 	 * Not only the cpu but also the task_group of the parent might have
7741 	 * been changed after parent->se.parent,cfs_rq were copied to
7742 	 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7743 	 * of child point to valid ones.
7744 	 */
7745 	rcu_read_lock();
7746 	__set_task_cpu(p, this_cpu);
7747 	rcu_read_unlock();
7748 
7749 	update_curr(cfs_rq);
7750 
7751 	if (curr)
7752 		se->vruntime = curr->vruntime;
7753 	place_entity(cfs_rq, se, 1);
7754 
7755 	if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7756 		/*
7757 		 * Upon rescheduling, sched_class::put_prev_task() will place
7758 		 * 'current' within the tree based on its new key value.
7759 		 */
7760 		swap(curr->vruntime, se->vruntime);
7761 		resched_curr(rq);
7762 	}
7763 
7764 	se->vruntime -= cfs_rq->min_vruntime;
7765 
7766 	raw_spin_unlock_irqrestore(&rq->lock, flags);
7767 }
7768 
7769 /*
7770  * Priority of the task has changed. Check to see if we preempt
7771  * the current task.
7772  */
7773 static void
7774 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7775 {
7776 	if (!task_on_rq_queued(p))
7777 		return;
7778 
7779 	/*
7780 	 * Reschedule if we are currently running on this runqueue and
7781 	 * our priority decreased, or if we are not currently running on
7782 	 * this runqueue and our priority is higher than the current's
7783 	 */
7784 	if (rq->curr == p) {
7785 		if (p->prio > oldprio)
7786 			resched_curr(rq);
7787 	} else
7788 		check_preempt_curr(rq, p, 0);
7789 }
7790 
7791 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7792 {
7793 	struct sched_entity *se = &p->se;
7794 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
7795 
7796 	/*
7797 	 * Ensure the task's vruntime is normalized, so that when it's
7798 	 * switched back to the fair class the enqueue_entity(.flags=0) will
7799 	 * do the right thing.
7800 	 *
7801 	 * If it's queued, then the dequeue_entity(.flags=0) will already
7802 	 * have normalized the vruntime, if it's !queued, then only when
7803 	 * the task is sleeping will it still have non-normalized vruntime.
7804 	 */
7805 	if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
7806 		/*
7807 		 * Fix up our vruntime so that the current sleep doesn't
7808 		 * cause 'unlimited' sleep bonus.
7809 		 */
7810 		place_entity(cfs_rq, se, 0);
7811 		se->vruntime -= cfs_rq->min_vruntime;
7812 	}
7813 
7814 #ifdef CONFIG_SMP
7815 	/*
7816 	* Remove our load from contribution when we leave sched_fair
7817 	* and ensure we don't carry in an old decay_count if we
7818 	* switch back.
7819 	*/
7820 	if (se->avg.decay_count) {
7821 		__synchronize_entity_decay(se);
7822 		subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7823 	}
7824 #endif
7825 }
7826 
7827 /*
7828  * We switched to the sched_fair class.
7829  */
7830 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7831 {
7832 #ifdef CONFIG_FAIR_GROUP_SCHED
7833 	struct sched_entity *se = &p->se;
7834 	/*
7835 	 * Since the real-depth could have been changed (only FAIR
7836 	 * class maintain depth value), reset depth properly.
7837 	 */
7838 	se->depth = se->parent ? se->parent->depth + 1 : 0;
7839 #endif
7840 	if (!task_on_rq_queued(p))
7841 		return;
7842 
7843 	/*
7844 	 * We were most likely switched from sched_rt, so
7845 	 * kick off the schedule if running, otherwise just see
7846 	 * if we can still preempt the current task.
7847 	 */
7848 	if (rq->curr == p)
7849 		resched_curr(rq);
7850 	else
7851 		check_preempt_curr(rq, p, 0);
7852 }
7853 
7854 /* Account for a task changing its policy or group.
7855  *
7856  * This routine is mostly called to set cfs_rq->curr field when a task
7857  * migrates between groups/classes.
7858  */
7859 static void set_curr_task_fair(struct rq *rq)
7860 {
7861 	struct sched_entity *se = &rq->curr->se;
7862 
7863 	for_each_sched_entity(se) {
7864 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
7865 
7866 		set_next_entity(cfs_rq, se);
7867 		/* ensure bandwidth has been allocated on our new cfs_rq */
7868 		account_cfs_rq_runtime(cfs_rq, 0);
7869 	}
7870 }
7871 
7872 void init_cfs_rq(struct cfs_rq *cfs_rq)
7873 {
7874 	cfs_rq->tasks_timeline = RB_ROOT;
7875 	cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7876 #ifndef CONFIG_64BIT
7877 	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7878 #endif
7879 #ifdef CONFIG_SMP
7880 	atomic64_set(&cfs_rq->decay_counter, 1);
7881 	atomic_long_set(&cfs_rq->removed_load, 0);
7882 #endif
7883 }
7884 
7885 #ifdef CONFIG_FAIR_GROUP_SCHED
7886 static void task_move_group_fair(struct task_struct *p, int queued)
7887 {
7888 	struct sched_entity *se = &p->se;
7889 	struct cfs_rq *cfs_rq;
7890 
7891 	/*
7892 	 * If the task was not on the rq at the time of this cgroup movement
7893 	 * it must have been asleep, sleeping tasks keep their ->vruntime
7894 	 * absolute on their old rq until wakeup (needed for the fair sleeper
7895 	 * bonus in place_entity()).
7896 	 *
7897 	 * If it was on the rq, we've just 'preempted' it, which does convert
7898 	 * ->vruntime to a relative base.
7899 	 *
7900 	 * Make sure both cases convert their relative position when migrating
7901 	 * to another cgroup's rq. This does somewhat interfere with the
7902 	 * fair sleeper stuff for the first placement, but who cares.
7903 	 */
7904 	/*
7905 	 * When !queued, vruntime of the task has usually NOT been normalized.
7906 	 * But there are some cases where it has already been normalized:
7907 	 *
7908 	 * - Moving a forked child which is waiting for being woken up by
7909 	 *   wake_up_new_task().
7910 	 * - Moving a task which has been woken up by try_to_wake_up() and
7911 	 *   waiting for actually being woken up by sched_ttwu_pending().
7912 	 *
7913 	 * To prevent boost or penalty in the new cfs_rq caused by delta
7914 	 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7915 	 */
7916 	if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7917 		queued = 1;
7918 
7919 	if (!queued)
7920 		se->vruntime -= cfs_rq_of(se)->min_vruntime;
7921 	set_task_rq(p, task_cpu(p));
7922 	se->depth = se->parent ? se->parent->depth + 1 : 0;
7923 	if (!queued) {
7924 		cfs_rq = cfs_rq_of(se);
7925 		se->vruntime += cfs_rq->min_vruntime;
7926 #ifdef CONFIG_SMP
7927 		/*
7928 		 * migrate_task_rq_fair() will have removed our previous
7929 		 * contribution, but we must synchronize for ongoing future
7930 		 * decay.
7931 		 */
7932 		se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7933 		cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7934 #endif
7935 	}
7936 }
7937 
7938 void free_fair_sched_group(struct task_group *tg)
7939 {
7940 	int i;
7941 
7942 	destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7943 
7944 	for_each_possible_cpu(i) {
7945 		if (tg->cfs_rq)
7946 			kfree(tg->cfs_rq[i]);
7947 		if (tg->se)
7948 			kfree(tg->se[i]);
7949 	}
7950 
7951 	kfree(tg->cfs_rq);
7952 	kfree(tg->se);
7953 }
7954 
7955 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7956 {
7957 	struct cfs_rq *cfs_rq;
7958 	struct sched_entity *se;
7959 	int i;
7960 
7961 	tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7962 	if (!tg->cfs_rq)
7963 		goto err;
7964 	tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7965 	if (!tg->se)
7966 		goto err;
7967 
7968 	tg->shares = NICE_0_LOAD;
7969 
7970 	init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7971 
7972 	for_each_possible_cpu(i) {
7973 		cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7974 				      GFP_KERNEL, cpu_to_node(i));
7975 		if (!cfs_rq)
7976 			goto err;
7977 
7978 		se = kzalloc_node(sizeof(struct sched_entity),
7979 				  GFP_KERNEL, cpu_to_node(i));
7980 		if (!se)
7981 			goto err_free_rq;
7982 
7983 		init_cfs_rq(cfs_rq);
7984 		init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7985 	}
7986 
7987 	return 1;
7988 
7989 err_free_rq:
7990 	kfree(cfs_rq);
7991 err:
7992 	return 0;
7993 }
7994 
7995 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7996 {
7997 	struct rq *rq = cpu_rq(cpu);
7998 	unsigned long flags;
7999 
8000 	/*
8001 	* Only empty task groups can be destroyed; so we can speculatively
8002 	* check on_list without danger of it being re-added.
8003 	*/
8004 	if (!tg->cfs_rq[cpu]->on_list)
8005 		return;
8006 
8007 	raw_spin_lock_irqsave(&rq->lock, flags);
8008 	list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8009 	raw_spin_unlock_irqrestore(&rq->lock, flags);
8010 }
8011 
8012 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8013 			struct sched_entity *se, int cpu,
8014 			struct sched_entity *parent)
8015 {
8016 	struct rq *rq = cpu_rq(cpu);
8017 
8018 	cfs_rq->tg = tg;
8019 	cfs_rq->rq = rq;
8020 	init_cfs_rq_runtime(cfs_rq);
8021 
8022 	tg->cfs_rq[cpu] = cfs_rq;
8023 	tg->se[cpu] = se;
8024 
8025 	/* se could be NULL for root_task_group */
8026 	if (!se)
8027 		return;
8028 
8029 	if (!parent) {
8030 		se->cfs_rq = &rq->cfs;
8031 		se->depth = 0;
8032 	} else {
8033 		se->cfs_rq = parent->my_q;
8034 		se->depth = parent->depth + 1;
8035 	}
8036 
8037 	se->my_q = cfs_rq;
8038 	/* guarantee group entities always have weight */
8039 	update_load_set(&se->load, NICE_0_LOAD);
8040 	se->parent = parent;
8041 }
8042 
8043 static DEFINE_MUTEX(shares_mutex);
8044 
8045 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8046 {
8047 	int i;
8048 	unsigned long flags;
8049 
8050 	/*
8051 	 * We can't change the weight of the root cgroup.
8052 	 */
8053 	if (!tg->se[0])
8054 		return -EINVAL;
8055 
8056 	shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8057 
8058 	mutex_lock(&shares_mutex);
8059 	if (tg->shares == shares)
8060 		goto done;
8061 
8062 	tg->shares = shares;
8063 	for_each_possible_cpu(i) {
8064 		struct rq *rq = cpu_rq(i);
8065 		struct sched_entity *se;
8066 
8067 		se = tg->se[i];
8068 		/* Propagate contribution to hierarchy */
8069 		raw_spin_lock_irqsave(&rq->lock, flags);
8070 
8071 		/* Possible calls to update_curr() need rq clock */
8072 		update_rq_clock(rq);
8073 		for_each_sched_entity(se)
8074 			update_cfs_shares(group_cfs_rq(se));
8075 		raw_spin_unlock_irqrestore(&rq->lock, flags);
8076 	}
8077 
8078 done:
8079 	mutex_unlock(&shares_mutex);
8080 	return 0;
8081 }
8082 #else /* CONFIG_FAIR_GROUP_SCHED */
8083 
8084 void free_fair_sched_group(struct task_group *tg) { }
8085 
8086 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8087 {
8088 	return 1;
8089 }
8090 
8091 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8092 
8093 #endif /* CONFIG_FAIR_GROUP_SCHED */
8094 
8095 
8096 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8097 {
8098 	struct sched_entity *se = &task->se;
8099 	unsigned int rr_interval = 0;
8100 
8101 	/*
8102 	 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8103 	 * idle runqueue:
8104 	 */
8105 	if (rq->cfs.load.weight)
8106 		rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8107 
8108 	return rr_interval;
8109 }
8110 
8111 /*
8112  * All the scheduling class methods:
8113  */
8114 const struct sched_class fair_sched_class = {
8115 	.next			= &idle_sched_class,
8116 	.enqueue_task		= enqueue_task_fair,
8117 	.dequeue_task		= dequeue_task_fair,
8118 	.yield_task		= yield_task_fair,
8119 	.yield_to_task		= yield_to_task_fair,
8120 
8121 	.check_preempt_curr	= check_preempt_wakeup,
8122 
8123 	.pick_next_task		= pick_next_task_fair,
8124 	.put_prev_task		= put_prev_task_fair,
8125 
8126 #ifdef CONFIG_SMP
8127 	.select_task_rq		= select_task_rq_fair,
8128 	.migrate_task_rq	= migrate_task_rq_fair,
8129 
8130 	.rq_online		= rq_online_fair,
8131 	.rq_offline		= rq_offline_fair,
8132 
8133 	.task_waking		= task_waking_fair,
8134 #endif
8135 
8136 	.set_curr_task          = set_curr_task_fair,
8137 	.task_tick		= task_tick_fair,
8138 	.task_fork		= task_fork_fair,
8139 
8140 	.prio_changed		= prio_changed_fair,
8141 	.switched_from		= switched_from_fair,
8142 	.switched_to		= switched_to_fair,
8143 
8144 	.get_rr_interval	= get_rr_interval_fair,
8145 
8146 	.update_curr		= update_curr_fair,
8147 
8148 #ifdef CONFIG_FAIR_GROUP_SCHED
8149 	.task_move_group	= task_move_group_fair,
8150 #endif
8151 };
8152 
8153 #ifdef CONFIG_SCHED_DEBUG
8154 void print_cfs_stats(struct seq_file *m, int cpu)
8155 {
8156 	struct cfs_rq *cfs_rq;
8157 
8158 	rcu_read_lock();
8159 	for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8160 		print_cfs_rq(m, cpu, cfs_rq);
8161 	rcu_read_unlock();
8162 }
8163 #endif
8164 
8165 __init void init_sched_fair_class(void)
8166 {
8167 #ifdef CONFIG_SMP
8168 	open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8169 
8170 #ifdef CONFIG_NO_HZ_COMMON
8171 	nohz.next_balance = jiffies;
8172 	zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8173 	cpu_notifier(sched_ilb_notifier, 0);
8174 #endif
8175 #endif /* SMP */
8176 
8177 }
8178