xref: /openbmc/linux/kernel/sched/fair.c (revision b802fb99)
1 /*
2  * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
3  *
4  *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
5  *
6  *  Interactivity improvements by Mike Galbraith
7  *  (C) 2007 Mike Galbraith <efault@gmx.de>
8  *
9  *  Various enhancements by Dmitry Adamushko.
10  *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
11  *
12  *  Group scheduling enhancements by Srivatsa Vaddagiri
13  *  Copyright IBM Corporation, 2007
14  *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
15  *
16  *  Scaled math optimizations by Thomas Gleixner
17  *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
18  *
19  *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
21  */
22 
23 #include <linux/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 unsigned int get_update_sysctl_factor(void)
145 {
146 	unsigned int cpus = min_t(unsigned 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 inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
287 {
288 	if (!cfs_rq->on_list) {
289 		/*
290 		 * Ensure we either appear before our parent (if already
291 		 * enqueued) or force our parent to appear after us when it is
292 		 * enqueued.  The fact that we always enqueue bottom-up
293 		 * reduces this to two cases.
294 		 */
295 		if (cfs_rq->tg->parent &&
296 		    cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 			list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 				&rq_of(cfs_rq)->leaf_cfs_rq_list);
299 		} else {
300 			list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 				&rq_of(cfs_rq)->leaf_cfs_rq_list);
302 		}
303 
304 		cfs_rq->on_list = 1;
305 	}
306 }
307 
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
309 {
310 	if (cfs_rq->on_list) {
311 		list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
312 		cfs_rq->on_list = 0;
313 	}
314 }
315 
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 	list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
319 
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
323 {
324 	if (se->cfs_rq == pse->cfs_rq)
325 		return se->cfs_rq;
326 
327 	return NULL;
328 }
329 
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
331 {
332 	return se->parent;
333 }
334 
335 static void
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
337 {
338 	int se_depth, pse_depth;
339 
340 	/*
341 	 * preemption test can be made between sibling entities who are in the
342 	 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 	 * both tasks until we find their ancestors who are siblings of common
344 	 * parent.
345 	 */
346 
347 	/* First walk up until both entities are at same depth */
348 	se_depth = (*se)->depth;
349 	pse_depth = (*pse)->depth;
350 
351 	while (se_depth > pse_depth) {
352 		se_depth--;
353 		*se = parent_entity(*se);
354 	}
355 
356 	while (pse_depth > se_depth) {
357 		pse_depth--;
358 		*pse = parent_entity(*pse);
359 	}
360 
361 	while (!is_same_group(*se, *pse)) {
362 		*se = parent_entity(*se);
363 		*pse = parent_entity(*pse);
364 	}
365 }
366 
367 #else	/* !CONFIG_FAIR_GROUP_SCHED */
368 
369 static inline struct task_struct *task_of(struct sched_entity *se)
370 {
371 	return container_of(se, struct task_struct, se);
372 }
373 
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
375 {
376 	return container_of(cfs_rq, struct rq, cfs);
377 }
378 
379 #define entity_is_task(se)	1
380 
381 #define for_each_sched_entity(se) \
382 		for (; se; se = NULL)
383 
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
385 {
386 	return &task_rq(p)->cfs;
387 }
388 
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
390 {
391 	struct task_struct *p = task_of(se);
392 	struct rq *rq = task_rq(p);
393 
394 	return &rq->cfs;
395 }
396 
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
399 {
400 	return NULL;
401 }
402 
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
404 {
405 }
406 
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
408 {
409 }
410 
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 		for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
413 
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
415 {
416 	return NULL;
417 }
418 
419 static inline void
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
421 {
422 }
423 
424 #endif	/* CONFIG_FAIR_GROUP_SCHED */
425 
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
428 
429 /**************************************************************
430  * Scheduling class tree data structure manipulation methods:
431  */
432 
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
434 {
435 	s64 delta = (s64)(vruntime - max_vruntime);
436 	if (delta > 0)
437 		max_vruntime = vruntime;
438 
439 	return max_vruntime;
440 }
441 
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
443 {
444 	s64 delta = (s64)(vruntime - min_vruntime);
445 	if (delta < 0)
446 		min_vruntime = vruntime;
447 
448 	return min_vruntime;
449 }
450 
451 static inline int entity_before(struct sched_entity *a,
452 				struct sched_entity *b)
453 {
454 	return (s64)(a->vruntime - b->vruntime) < 0;
455 }
456 
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
458 {
459 	u64 vruntime = cfs_rq->min_vruntime;
460 
461 	if (cfs_rq->curr)
462 		vruntime = cfs_rq->curr->vruntime;
463 
464 	if (cfs_rq->rb_leftmost) {
465 		struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
466 						   struct sched_entity,
467 						   run_node);
468 
469 		if (!cfs_rq->curr)
470 			vruntime = se->vruntime;
471 		else
472 			vruntime = min_vruntime(vruntime, se->vruntime);
473 	}
474 
475 	/* ensure we never gain time by being placed backwards. */
476 	cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
477 #ifndef CONFIG_64BIT
478 	smp_wmb();
479 	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
480 #endif
481 }
482 
483 /*
484  * Enqueue an entity into the rb-tree:
485  */
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
487 {
488 	struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 	struct rb_node *parent = NULL;
490 	struct sched_entity *entry;
491 	int leftmost = 1;
492 
493 	/*
494 	 * Find the right place in the rbtree:
495 	 */
496 	while (*link) {
497 		parent = *link;
498 		entry = rb_entry(parent, struct sched_entity, run_node);
499 		/*
500 		 * We dont care about collisions. Nodes with
501 		 * the same key stay together.
502 		 */
503 		if (entity_before(se, entry)) {
504 			link = &parent->rb_left;
505 		} else {
506 			link = &parent->rb_right;
507 			leftmost = 0;
508 		}
509 	}
510 
511 	/*
512 	 * Maintain a cache of leftmost tree entries (it is frequently
513 	 * used):
514 	 */
515 	if (leftmost)
516 		cfs_rq->rb_leftmost = &se->run_node;
517 
518 	rb_link_node(&se->run_node, parent, link);
519 	rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
520 }
521 
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
523 {
524 	if (cfs_rq->rb_leftmost == &se->run_node) {
525 		struct rb_node *next_node;
526 
527 		next_node = rb_next(&se->run_node);
528 		cfs_rq->rb_leftmost = next_node;
529 	}
530 
531 	rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
532 }
533 
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
535 {
536 	struct rb_node *left = cfs_rq->rb_leftmost;
537 
538 	if (!left)
539 		return NULL;
540 
541 	return rb_entry(left, struct sched_entity, run_node);
542 }
543 
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
545 {
546 	struct rb_node *next = rb_next(&se->run_node);
547 
548 	if (!next)
549 		return NULL;
550 
551 	return rb_entry(next, struct sched_entity, run_node);
552 }
553 
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
556 {
557 	struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
558 
559 	if (!last)
560 		return NULL;
561 
562 	return rb_entry(last, struct sched_entity, run_node);
563 }
564 
565 /**************************************************************
566  * Scheduling class statistics methods:
567  */
568 
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 		void __user *buffer, size_t *lenp,
571 		loff_t *ppos)
572 {
573 	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 	unsigned int factor = get_update_sysctl_factor();
575 
576 	if (ret || !write)
577 		return ret;
578 
579 	sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 					sysctl_sched_min_granularity);
581 
582 #define WRT_SYSCTL(name) \
583 	(normalized_sysctl_##name = sysctl_##name / (factor))
584 	WRT_SYSCTL(sched_min_granularity);
585 	WRT_SYSCTL(sched_latency);
586 	WRT_SYSCTL(sched_wakeup_granularity);
587 #undef WRT_SYSCTL
588 
589 	return 0;
590 }
591 #endif
592 
593 /*
594  * delta /= w
595  */
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
597 {
598 	if (unlikely(se->load.weight != NICE_0_LOAD))
599 		delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
600 
601 	return delta;
602 }
603 
604 /*
605  * The idea is to set a period in which each task runs once.
606  *
607  * When there are too many tasks (sched_nr_latency) we have to stretch
608  * this period because otherwise the slices get too small.
609  *
610  * p = (nr <= nl) ? l : l*nr/nl
611  */
612 static u64 __sched_period(unsigned long nr_running)
613 {
614 	if (unlikely(nr_running > sched_nr_latency))
615 		return nr_running * sysctl_sched_min_granularity;
616 	else
617 		return sysctl_sched_latency;
618 }
619 
620 /*
621  * We calculate the wall-time slice from the period by taking a part
622  * proportional to the weight.
623  *
624  * s = p*P[w/rw]
625  */
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
627 {
628 	u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
629 
630 	for_each_sched_entity(se) {
631 		struct load_weight *load;
632 		struct load_weight lw;
633 
634 		cfs_rq = cfs_rq_of(se);
635 		load = &cfs_rq->load;
636 
637 		if (unlikely(!se->on_rq)) {
638 			lw = cfs_rq->load;
639 
640 			update_load_add(&lw, se->load.weight);
641 			load = &lw;
642 		}
643 		slice = __calc_delta(slice, se->load.weight, load);
644 	}
645 	return slice;
646 }
647 
648 /*
649  * We calculate the vruntime slice of a to-be-inserted task.
650  *
651  * vs = s/w
652  */
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
654 {
655 	return calc_delta_fair(sched_slice(cfs_rq, se), se);
656 }
657 
658 #ifdef CONFIG_SMP
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
661 
662 /*
663  * We choose a half-life close to 1 scheduling period.
664  * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665  * dependent on this value.
666  */
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
670 
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity *se)
673 {
674 	struct sched_avg *sa = &se->avg;
675 
676 	sa->last_update_time = 0;
677 	/*
678 	 * sched_avg's period_contrib should be strictly less then 1024, so
679 	 * we give it 1023 to make sure it is almost a period (1024us), and
680 	 * will definitely be update (after enqueue).
681 	 */
682 	sa->period_contrib = 1023;
683 	sa->load_avg = scale_load_down(se->load.weight);
684 	sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
685 	sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
686 	sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
687 	/* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
688 }
689 
690 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
691 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
692 #else
693 void init_entity_runnable_average(struct sched_entity *se)
694 {
695 }
696 #endif
697 
698 /*
699  * Update the current task's runtime statistics.
700  */
701 static void update_curr(struct cfs_rq *cfs_rq)
702 {
703 	struct sched_entity *curr = cfs_rq->curr;
704 	u64 now = rq_clock_task(rq_of(cfs_rq));
705 	u64 delta_exec;
706 
707 	if (unlikely(!curr))
708 		return;
709 
710 	delta_exec = now - curr->exec_start;
711 	if (unlikely((s64)delta_exec <= 0))
712 		return;
713 
714 	curr->exec_start = now;
715 
716 	schedstat_set(curr->statistics.exec_max,
717 		      max(delta_exec, curr->statistics.exec_max));
718 
719 	curr->sum_exec_runtime += delta_exec;
720 	schedstat_add(cfs_rq, exec_clock, delta_exec);
721 
722 	curr->vruntime += calc_delta_fair(delta_exec, curr);
723 	update_min_vruntime(cfs_rq);
724 
725 	if (entity_is_task(curr)) {
726 		struct task_struct *curtask = task_of(curr);
727 
728 		trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
729 		cpuacct_charge(curtask, delta_exec);
730 		account_group_exec_runtime(curtask, delta_exec);
731 	}
732 
733 	account_cfs_rq_runtime(cfs_rq, delta_exec);
734 }
735 
736 static void update_curr_fair(struct rq *rq)
737 {
738 	update_curr(cfs_rq_of(&rq->curr->se));
739 }
740 
741 #ifdef CONFIG_SCHEDSTATS
742 static inline void
743 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
744 {
745 	u64 wait_start = rq_clock(rq_of(cfs_rq));
746 
747 	if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
748 	    likely(wait_start > se->statistics.wait_start))
749 		wait_start -= se->statistics.wait_start;
750 
751 	se->statistics.wait_start = wait_start;
752 }
753 
754 static void
755 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
756 {
757 	struct task_struct *p;
758 	u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
759 
760 	if (entity_is_task(se)) {
761 		p = task_of(se);
762 		if (task_on_rq_migrating(p)) {
763 			/*
764 			 * Preserve migrating task's wait time so wait_start
765 			 * time stamp can be adjusted to accumulate wait time
766 			 * prior to migration.
767 			 */
768 			se->statistics.wait_start = delta;
769 			return;
770 		}
771 		trace_sched_stat_wait(p, delta);
772 	}
773 
774 	se->statistics.wait_max = max(se->statistics.wait_max, delta);
775 	se->statistics.wait_count++;
776 	se->statistics.wait_sum += delta;
777 	se->statistics.wait_start = 0;
778 }
779 #else
780 static inline void
781 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
782 {
783 }
784 
785 static inline void
786 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
787 {
788 }
789 #endif
790 
791 /*
792  * Task is being enqueued - update stats:
793  */
794 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
795 {
796 	/*
797 	 * Are we enqueueing a waiting task? (for current tasks
798 	 * a dequeue/enqueue event is a NOP)
799 	 */
800 	if (se != cfs_rq->curr)
801 		update_stats_wait_start(cfs_rq, se);
802 }
803 
804 static inline void
805 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
806 {
807 	/*
808 	 * Mark the end of the wait period if dequeueing a
809 	 * waiting task:
810 	 */
811 	if (se != cfs_rq->curr)
812 		update_stats_wait_end(cfs_rq, se);
813 }
814 
815 /*
816  * We are picking a new current task - update its stats:
817  */
818 static inline void
819 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
820 {
821 	/*
822 	 * We are starting a new run period:
823 	 */
824 	se->exec_start = rq_clock_task(rq_of(cfs_rq));
825 }
826 
827 /**************************************************
828  * Scheduling class queueing methods:
829  */
830 
831 #ifdef CONFIG_NUMA_BALANCING
832 /*
833  * Approximate time to scan a full NUMA task in ms. The task scan period is
834  * calculated based on the tasks virtual memory size and
835  * numa_balancing_scan_size.
836  */
837 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
838 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
839 
840 /* Portion of address space to scan in MB */
841 unsigned int sysctl_numa_balancing_scan_size = 256;
842 
843 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
844 unsigned int sysctl_numa_balancing_scan_delay = 1000;
845 
846 static unsigned int task_nr_scan_windows(struct task_struct *p)
847 {
848 	unsigned long rss = 0;
849 	unsigned long nr_scan_pages;
850 
851 	/*
852 	 * Calculations based on RSS as non-present and empty pages are skipped
853 	 * by the PTE scanner and NUMA hinting faults should be trapped based
854 	 * on resident pages
855 	 */
856 	nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
857 	rss = get_mm_rss(p->mm);
858 	if (!rss)
859 		rss = nr_scan_pages;
860 
861 	rss = round_up(rss, nr_scan_pages);
862 	return rss / nr_scan_pages;
863 }
864 
865 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
866 #define MAX_SCAN_WINDOW 2560
867 
868 static unsigned int task_scan_min(struct task_struct *p)
869 {
870 	unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
871 	unsigned int scan, floor;
872 	unsigned int windows = 1;
873 
874 	if (scan_size < MAX_SCAN_WINDOW)
875 		windows = MAX_SCAN_WINDOW / scan_size;
876 	floor = 1000 / windows;
877 
878 	scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
879 	return max_t(unsigned int, floor, scan);
880 }
881 
882 static unsigned int task_scan_max(struct task_struct *p)
883 {
884 	unsigned int smin = task_scan_min(p);
885 	unsigned int smax;
886 
887 	/* Watch for min being lower than max due to floor calculations */
888 	smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
889 	return max(smin, smax);
890 }
891 
892 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
893 {
894 	rq->nr_numa_running += (p->numa_preferred_nid != -1);
895 	rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
896 }
897 
898 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
899 {
900 	rq->nr_numa_running -= (p->numa_preferred_nid != -1);
901 	rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
902 }
903 
904 struct numa_group {
905 	atomic_t refcount;
906 
907 	spinlock_t lock; /* nr_tasks, tasks */
908 	int nr_tasks;
909 	pid_t gid;
910 
911 	struct rcu_head rcu;
912 	nodemask_t active_nodes;
913 	unsigned long total_faults;
914 	/*
915 	 * Faults_cpu is used to decide whether memory should move
916 	 * towards the CPU. As a consequence, these stats are weighted
917 	 * more by CPU use than by memory faults.
918 	 */
919 	unsigned long *faults_cpu;
920 	unsigned long faults[0];
921 };
922 
923 /* Shared or private faults. */
924 #define NR_NUMA_HINT_FAULT_TYPES 2
925 
926 /* Memory and CPU locality */
927 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
928 
929 /* Averaged statistics, and temporary buffers. */
930 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
931 
932 pid_t task_numa_group_id(struct task_struct *p)
933 {
934 	return p->numa_group ? p->numa_group->gid : 0;
935 }
936 
937 /*
938  * The averaged statistics, shared & private, memory & cpu,
939  * occupy the first half of the array. The second half of the
940  * array is for current counters, which are averaged into the
941  * first set by task_numa_placement.
942  */
943 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
944 {
945 	return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
946 }
947 
948 static inline unsigned long task_faults(struct task_struct *p, int nid)
949 {
950 	if (!p->numa_faults)
951 		return 0;
952 
953 	return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
954 		p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
955 }
956 
957 static inline unsigned long group_faults(struct task_struct *p, int nid)
958 {
959 	if (!p->numa_group)
960 		return 0;
961 
962 	return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
963 		p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
964 }
965 
966 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
967 {
968 	return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
969 		group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
970 }
971 
972 /* Handle placement on systems where not all nodes are directly connected. */
973 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
974 					int maxdist, bool task)
975 {
976 	unsigned long score = 0;
977 	int node;
978 
979 	/*
980 	 * All nodes are directly connected, and the same distance
981 	 * from each other. No need for fancy placement algorithms.
982 	 */
983 	if (sched_numa_topology_type == NUMA_DIRECT)
984 		return 0;
985 
986 	/*
987 	 * This code is called for each node, introducing N^2 complexity,
988 	 * which should be ok given the number of nodes rarely exceeds 8.
989 	 */
990 	for_each_online_node(node) {
991 		unsigned long faults;
992 		int dist = node_distance(nid, node);
993 
994 		/*
995 		 * The furthest away nodes in the system are not interesting
996 		 * for placement; nid was already counted.
997 		 */
998 		if (dist == sched_max_numa_distance || node == nid)
999 			continue;
1000 
1001 		/*
1002 		 * On systems with a backplane NUMA topology, compare groups
1003 		 * of nodes, and move tasks towards the group with the most
1004 		 * memory accesses. When comparing two nodes at distance
1005 		 * "hoplimit", only nodes closer by than "hoplimit" are part
1006 		 * of each group. Skip other nodes.
1007 		 */
1008 		if (sched_numa_topology_type == NUMA_BACKPLANE &&
1009 					dist > maxdist)
1010 			continue;
1011 
1012 		/* Add up the faults from nearby nodes. */
1013 		if (task)
1014 			faults = task_faults(p, node);
1015 		else
1016 			faults = group_faults(p, node);
1017 
1018 		/*
1019 		 * On systems with a glueless mesh NUMA topology, there are
1020 		 * no fixed "groups of nodes". Instead, nodes that are not
1021 		 * directly connected bounce traffic through intermediate
1022 		 * nodes; a numa_group can occupy any set of nodes.
1023 		 * The further away a node is, the less the faults count.
1024 		 * This seems to result in good task placement.
1025 		 */
1026 		if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1027 			faults *= (sched_max_numa_distance - dist);
1028 			faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1029 		}
1030 
1031 		score += faults;
1032 	}
1033 
1034 	return score;
1035 }
1036 
1037 /*
1038  * These return the fraction of accesses done by a particular task, or
1039  * task group, on a particular numa node.  The group weight is given a
1040  * larger multiplier, in order to group tasks together that are almost
1041  * evenly spread out between numa nodes.
1042  */
1043 static inline unsigned long task_weight(struct task_struct *p, int nid,
1044 					int dist)
1045 {
1046 	unsigned long faults, total_faults;
1047 
1048 	if (!p->numa_faults)
1049 		return 0;
1050 
1051 	total_faults = p->total_numa_faults;
1052 
1053 	if (!total_faults)
1054 		return 0;
1055 
1056 	faults = task_faults(p, nid);
1057 	faults += score_nearby_nodes(p, nid, dist, true);
1058 
1059 	return 1000 * faults / total_faults;
1060 }
1061 
1062 static inline unsigned long group_weight(struct task_struct *p, int nid,
1063 					 int dist)
1064 {
1065 	unsigned long faults, total_faults;
1066 
1067 	if (!p->numa_group)
1068 		return 0;
1069 
1070 	total_faults = p->numa_group->total_faults;
1071 
1072 	if (!total_faults)
1073 		return 0;
1074 
1075 	faults = group_faults(p, nid);
1076 	faults += score_nearby_nodes(p, nid, dist, false);
1077 
1078 	return 1000 * faults / total_faults;
1079 }
1080 
1081 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1082 				int src_nid, int dst_cpu)
1083 {
1084 	struct numa_group *ng = p->numa_group;
1085 	int dst_nid = cpu_to_node(dst_cpu);
1086 	int last_cpupid, this_cpupid;
1087 
1088 	this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1089 
1090 	/*
1091 	 * Multi-stage node selection is used in conjunction with a periodic
1092 	 * migration fault to build a temporal task<->page relation. By using
1093 	 * a two-stage filter we remove short/unlikely relations.
1094 	 *
1095 	 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1096 	 * a task's usage of a particular page (n_p) per total usage of this
1097 	 * page (n_t) (in a given time-span) to a probability.
1098 	 *
1099 	 * Our periodic faults will sample this probability and getting the
1100 	 * same result twice in a row, given these samples are fully
1101 	 * independent, is then given by P(n)^2, provided our sample period
1102 	 * is sufficiently short compared to the usage pattern.
1103 	 *
1104 	 * This quadric squishes small probabilities, making it less likely we
1105 	 * act on an unlikely task<->page relation.
1106 	 */
1107 	last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1108 	if (!cpupid_pid_unset(last_cpupid) &&
1109 				cpupid_to_nid(last_cpupid) != dst_nid)
1110 		return false;
1111 
1112 	/* Always allow migrate on private faults */
1113 	if (cpupid_match_pid(p, last_cpupid))
1114 		return true;
1115 
1116 	/* A shared fault, but p->numa_group has not been set up yet. */
1117 	if (!ng)
1118 		return true;
1119 
1120 	/*
1121 	 * Do not migrate if the destination is not a node that
1122 	 * is actively used by this numa group.
1123 	 */
1124 	if (!node_isset(dst_nid, ng->active_nodes))
1125 		return false;
1126 
1127 	/*
1128 	 * Source is a node that is not actively used by this
1129 	 * numa group, while the destination is. Migrate.
1130 	 */
1131 	if (!node_isset(src_nid, ng->active_nodes))
1132 		return true;
1133 
1134 	/*
1135 	 * Both source and destination are nodes in active
1136 	 * use by this numa group. Maximize memory bandwidth
1137 	 * by migrating from more heavily used groups, to less
1138 	 * heavily used ones, spreading the load around.
1139 	 * Use a 1/4 hysteresis to avoid spurious page movement.
1140 	 */
1141 	return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1142 }
1143 
1144 static unsigned long weighted_cpuload(const int cpu);
1145 static unsigned long source_load(int cpu, int type);
1146 static unsigned long target_load(int cpu, int type);
1147 static unsigned long capacity_of(int cpu);
1148 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1149 
1150 /* Cached statistics for all CPUs within a node */
1151 struct numa_stats {
1152 	unsigned long nr_running;
1153 	unsigned long load;
1154 
1155 	/* Total compute capacity of CPUs on a node */
1156 	unsigned long compute_capacity;
1157 
1158 	/* Approximate capacity in terms of runnable tasks on a node */
1159 	unsigned long task_capacity;
1160 	int has_free_capacity;
1161 };
1162 
1163 /*
1164  * XXX borrowed from update_sg_lb_stats
1165  */
1166 static void update_numa_stats(struct numa_stats *ns, int nid)
1167 {
1168 	int smt, cpu, cpus = 0;
1169 	unsigned long capacity;
1170 
1171 	memset(ns, 0, sizeof(*ns));
1172 	for_each_cpu(cpu, cpumask_of_node(nid)) {
1173 		struct rq *rq = cpu_rq(cpu);
1174 
1175 		ns->nr_running += rq->nr_running;
1176 		ns->load += weighted_cpuload(cpu);
1177 		ns->compute_capacity += capacity_of(cpu);
1178 
1179 		cpus++;
1180 	}
1181 
1182 	/*
1183 	 * If we raced with hotplug and there are no CPUs left in our mask
1184 	 * the @ns structure is NULL'ed and task_numa_compare() will
1185 	 * not find this node attractive.
1186 	 *
1187 	 * We'll either bail at !has_free_capacity, or we'll detect a huge
1188 	 * imbalance and bail there.
1189 	 */
1190 	if (!cpus)
1191 		return;
1192 
1193 	/* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1194 	smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1195 	capacity = cpus / smt; /* cores */
1196 
1197 	ns->task_capacity = min_t(unsigned, capacity,
1198 		DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1199 	ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1200 }
1201 
1202 struct task_numa_env {
1203 	struct task_struct *p;
1204 
1205 	int src_cpu, src_nid;
1206 	int dst_cpu, dst_nid;
1207 
1208 	struct numa_stats src_stats, dst_stats;
1209 
1210 	int imbalance_pct;
1211 	int dist;
1212 
1213 	struct task_struct *best_task;
1214 	long best_imp;
1215 	int best_cpu;
1216 };
1217 
1218 static void task_numa_assign(struct task_numa_env *env,
1219 			     struct task_struct *p, long imp)
1220 {
1221 	if (env->best_task)
1222 		put_task_struct(env->best_task);
1223 
1224 	env->best_task = p;
1225 	env->best_imp = imp;
1226 	env->best_cpu = env->dst_cpu;
1227 }
1228 
1229 static bool load_too_imbalanced(long src_load, long dst_load,
1230 				struct task_numa_env *env)
1231 {
1232 	long imb, old_imb;
1233 	long orig_src_load, orig_dst_load;
1234 	long src_capacity, dst_capacity;
1235 
1236 	/*
1237 	 * The load is corrected for the CPU capacity available on each node.
1238 	 *
1239 	 * src_load        dst_load
1240 	 * ------------ vs ---------
1241 	 * src_capacity    dst_capacity
1242 	 */
1243 	src_capacity = env->src_stats.compute_capacity;
1244 	dst_capacity = env->dst_stats.compute_capacity;
1245 
1246 	/* We care about the slope of the imbalance, not the direction. */
1247 	if (dst_load < src_load)
1248 		swap(dst_load, src_load);
1249 
1250 	/* Is the difference below the threshold? */
1251 	imb = dst_load * src_capacity * 100 -
1252 	      src_load * dst_capacity * env->imbalance_pct;
1253 	if (imb <= 0)
1254 		return false;
1255 
1256 	/*
1257 	 * The imbalance is above the allowed threshold.
1258 	 * Compare it with the old imbalance.
1259 	 */
1260 	orig_src_load = env->src_stats.load;
1261 	orig_dst_load = env->dst_stats.load;
1262 
1263 	if (orig_dst_load < orig_src_load)
1264 		swap(orig_dst_load, orig_src_load);
1265 
1266 	old_imb = orig_dst_load * src_capacity * 100 -
1267 		  orig_src_load * dst_capacity * env->imbalance_pct;
1268 
1269 	/* Would this change make things worse? */
1270 	return (imb > old_imb);
1271 }
1272 
1273 /*
1274  * This checks if the overall compute and NUMA accesses of the system would
1275  * be improved if the source tasks was migrated to the target dst_cpu taking
1276  * into account that it might be best if task running on the dst_cpu should
1277  * be exchanged with the source task
1278  */
1279 static void task_numa_compare(struct task_numa_env *env,
1280 			      long taskimp, long groupimp)
1281 {
1282 	struct rq *src_rq = cpu_rq(env->src_cpu);
1283 	struct rq *dst_rq = cpu_rq(env->dst_cpu);
1284 	struct task_struct *cur;
1285 	long src_load, dst_load;
1286 	long load;
1287 	long imp = env->p->numa_group ? groupimp : taskimp;
1288 	long moveimp = imp;
1289 	int dist = env->dist;
1290 	bool assigned = false;
1291 
1292 	rcu_read_lock();
1293 
1294 	raw_spin_lock_irq(&dst_rq->lock);
1295 	cur = dst_rq->curr;
1296 	/*
1297 	 * No need to move the exiting task or idle task.
1298 	 */
1299 	if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1300 		cur = NULL;
1301 	else {
1302 		/*
1303 		 * The task_struct must be protected here to protect the
1304 		 * p->numa_faults access in the task_weight since the
1305 		 * numa_faults could already be freed in the following path:
1306 		 * finish_task_switch()
1307 		 *     --> put_task_struct()
1308 		 *         --> __put_task_struct()
1309 		 *             --> task_numa_free()
1310 		 */
1311 		get_task_struct(cur);
1312 	}
1313 
1314 	raw_spin_unlock_irq(&dst_rq->lock);
1315 
1316 	/*
1317 	 * Because we have preemption enabled we can get migrated around and
1318 	 * end try selecting ourselves (current == env->p) as a swap candidate.
1319 	 */
1320 	if (cur == env->p)
1321 		goto unlock;
1322 
1323 	/*
1324 	 * "imp" is the fault differential for the source task between the
1325 	 * source and destination node. Calculate the total differential for
1326 	 * the source task and potential destination task. The more negative
1327 	 * the value is, the more rmeote accesses that would be expected to
1328 	 * be incurred if the tasks were swapped.
1329 	 */
1330 	if (cur) {
1331 		/* Skip this swap candidate if cannot move to the source cpu */
1332 		if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1333 			goto unlock;
1334 
1335 		/*
1336 		 * If dst and source tasks are in the same NUMA group, or not
1337 		 * in any group then look only at task weights.
1338 		 */
1339 		if (cur->numa_group == env->p->numa_group) {
1340 			imp = taskimp + task_weight(cur, env->src_nid, dist) -
1341 			      task_weight(cur, env->dst_nid, dist);
1342 			/*
1343 			 * Add some hysteresis to prevent swapping the
1344 			 * tasks within a group over tiny differences.
1345 			 */
1346 			if (cur->numa_group)
1347 				imp -= imp/16;
1348 		} else {
1349 			/*
1350 			 * Compare the group weights. If a task is all by
1351 			 * itself (not part of a group), use the task weight
1352 			 * instead.
1353 			 */
1354 			if (cur->numa_group)
1355 				imp += group_weight(cur, env->src_nid, dist) -
1356 				       group_weight(cur, env->dst_nid, dist);
1357 			else
1358 				imp += task_weight(cur, env->src_nid, dist) -
1359 				       task_weight(cur, env->dst_nid, dist);
1360 		}
1361 	}
1362 
1363 	if (imp <= env->best_imp && moveimp <= env->best_imp)
1364 		goto unlock;
1365 
1366 	if (!cur) {
1367 		/* Is there capacity at our destination? */
1368 		if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1369 		    !env->dst_stats.has_free_capacity)
1370 			goto unlock;
1371 
1372 		goto balance;
1373 	}
1374 
1375 	/* Balance doesn't matter much if we're running a task per cpu */
1376 	if (imp > env->best_imp && src_rq->nr_running == 1 &&
1377 			dst_rq->nr_running == 1)
1378 		goto assign;
1379 
1380 	/*
1381 	 * In the overloaded case, try and keep the load balanced.
1382 	 */
1383 balance:
1384 	load = task_h_load(env->p);
1385 	dst_load = env->dst_stats.load + load;
1386 	src_load = env->src_stats.load - load;
1387 
1388 	if (moveimp > imp && moveimp > env->best_imp) {
1389 		/*
1390 		 * If the improvement from just moving env->p direction is
1391 		 * better than swapping tasks around, check if a move is
1392 		 * possible. Store a slightly smaller score than moveimp,
1393 		 * so an actually idle CPU will win.
1394 		 */
1395 		if (!load_too_imbalanced(src_load, dst_load, env)) {
1396 			imp = moveimp - 1;
1397 			put_task_struct(cur);
1398 			cur = NULL;
1399 			goto assign;
1400 		}
1401 	}
1402 
1403 	if (imp <= env->best_imp)
1404 		goto unlock;
1405 
1406 	if (cur) {
1407 		load = task_h_load(cur);
1408 		dst_load -= load;
1409 		src_load += load;
1410 	}
1411 
1412 	if (load_too_imbalanced(src_load, dst_load, env))
1413 		goto unlock;
1414 
1415 	/*
1416 	 * One idle CPU per node is evaluated for a task numa move.
1417 	 * Call select_idle_sibling to maybe find a better one.
1418 	 */
1419 	if (!cur)
1420 		env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1421 
1422 assign:
1423 	assigned = true;
1424 	task_numa_assign(env, cur, imp);
1425 unlock:
1426 	rcu_read_unlock();
1427 	/*
1428 	 * The dst_rq->curr isn't assigned. The protection for task_struct is
1429 	 * finished.
1430 	 */
1431 	if (cur && !assigned)
1432 		put_task_struct(cur);
1433 }
1434 
1435 static void task_numa_find_cpu(struct task_numa_env *env,
1436 				long taskimp, long groupimp)
1437 {
1438 	int cpu;
1439 
1440 	for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1441 		/* Skip this CPU if the source task cannot migrate */
1442 		if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1443 			continue;
1444 
1445 		env->dst_cpu = cpu;
1446 		task_numa_compare(env, taskimp, groupimp);
1447 	}
1448 }
1449 
1450 /* Only move tasks to a NUMA node less busy than the current node. */
1451 static bool numa_has_capacity(struct task_numa_env *env)
1452 {
1453 	struct numa_stats *src = &env->src_stats;
1454 	struct numa_stats *dst = &env->dst_stats;
1455 
1456 	if (src->has_free_capacity && !dst->has_free_capacity)
1457 		return false;
1458 
1459 	/*
1460 	 * Only consider a task move if the source has a higher load
1461 	 * than the destination, corrected for CPU capacity on each node.
1462 	 *
1463 	 *      src->load                dst->load
1464 	 * --------------------- vs ---------------------
1465 	 * src->compute_capacity    dst->compute_capacity
1466 	 */
1467 	if (src->load * dst->compute_capacity * env->imbalance_pct >
1468 
1469 	    dst->load * src->compute_capacity * 100)
1470 		return true;
1471 
1472 	return false;
1473 }
1474 
1475 static int task_numa_migrate(struct task_struct *p)
1476 {
1477 	struct task_numa_env env = {
1478 		.p = p,
1479 
1480 		.src_cpu = task_cpu(p),
1481 		.src_nid = task_node(p),
1482 
1483 		.imbalance_pct = 112,
1484 
1485 		.best_task = NULL,
1486 		.best_imp = 0,
1487 		.best_cpu = -1
1488 	};
1489 	struct sched_domain *sd;
1490 	unsigned long taskweight, groupweight;
1491 	int nid, ret, dist;
1492 	long taskimp, groupimp;
1493 
1494 	/*
1495 	 * Pick the lowest SD_NUMA domain, as that would have the smallest
1496 	 * imbalance and would be the first to start moving tasks about.
1497 	 *
1498 	 * And we want to avoid any moving of tasks about, as that would create
1499 	 * random movement of tasks -- counter the numa conditions we're trying
1500 	 * to satisfy here.
1501 	 */
1502 	rcu_read_lock();
1503 	sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1504 	if (sd)
1505 		env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1506 	rcu_read_unlock();
1507 
1508 	/*
1509 	 * Cpusets can break the scheduler domain tree into smaller
1510 	 * balance domains, some of which do not cross NUMA boundaries.
1511 	 * Tasks that are "trapped" in such domains cannot be migrated
1512 	 * elsewhere, so there is no point in (re)trying.
1513 	 */
1514 	if (unlikely(!sd)) {
1515 		p->numa_preferred_nid = task_node(p);
1516 		return -EINVAL;
1517 	}
1518 
1519 	env.dst_nid = p->numa_preferred_nid;
1520 	dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1521 	taskweight = task_weight(p, env.src_nid, dist);
1522 	groupweight = group_weight(p, env.src_nid, dist);
1523 	update_numa_stats(&env.src_stats, env.src_nid);
1524 	taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1525 	groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1526 	update_numa_stats(&env.dst_stats, env.dst_nid);
1527 
1528 	/* Try to find a spot on the preferred nid. */
1529 	if (numa_has_capacity(&env))
1530 		task_numa_find_cpu(&env, taskimp, groupimp);
1531 
1532 	/*
1533 	 * Look at other nodes in these cases:
1534 	 * - there is no space available on the preferred_nid
1535 	 * - the task is part of a numa_group that is interleaved across
1536 	 *   multiple NUMA nodes; in order to better consolidate the group,
1537 	 *   we need to check other locations.
1538 	 */
1539 	if (env.best_cpu == -1 || (p->numa_group &&
1540 			nodes_weight(p->numa_group->active_nodes) > 1)) {
1541 		for_each_online_node(nid) {
1542 			if (nid == env.src_nid || nid == p->numa_preferred_nid)
1543 				continue;
1544 
1545 			dist = node_distance(env.src_nid, env.dst_nid);
1546 			if (sched_numa_topology_type == NUMA_BACKPLANE &&
1547 						dist != env.dist) {
1548 				taskweight = task_weight(p, env.src_nid, dist);
1549 				groupweight = group_weight(p, env.src_nid, dist);
1550 			}
1551 
1552 			/* Only consider nodes where both task and groups benefit */
1553 			taskimp = task_weight(p, nid, dist) - taskweight;
1554 			groupimp = group_weight(p, nid, dist) - groupweight;
1555 			if (taskimp < 0 && groupimp < 0)
1556 				continue;
1557 
1558 			env.dist = dist;
1559 			env.dst_nid = nid;
1560 			update_numa_stats(&env.dst_stats, env.dst_nid);
1561 			if (numa_has_capacity(&env))
1562 				task_numa_find_cpu(&env, taskimp, groupimp);
1563 		}
1564 	}
1565 
1566 	/*
1567 	 * If the task is part of a workload that spans multiple NUMA nodes,
1568 	 * and is migrating into one of the workload's active nodes, remember
1569 	 * this node as the task's preferred numa node, so the workload can
1570 	 * settle down.
1571 	 * A task that migrated to a second choice node will be better off
1572 	 * trying for a better one later. Do not set the preferred node here.
1573 	 */
1574 	if (p->numa_group) {
1575 		if (env.best_cpu == -1)
1576 			nid = env.src_nid;
1577 		else
1578 			nid = env.dst_nid;
1579 
1580 		if (node_isset(nid, p->numa_group->active_nodes))
1581 			sched_setnuma(p, env.dst_nid);
1582 	}
1583 
1584 	/* No better CPU than the current one was found. */
1585 	if (env.best_cpu == -1)
1586 		return -EAGAIN;
1587 
1588 	/*
1589 	 * Reset the scan period if the task is being rescheduled on an
1590 	 * alternative node to recheck if the tasks is now properly placed.
1591 	 */
1592 	p->numa_scan_period = task_scan_min(p);
1593 
1594 	if (env.best_task == NULL) {
1595 		ret = migrate_task_to(p, env.best_cpu);
1596 		if (ret != 0)
1597 			trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1598 		return ret;
1599 	}
1600 
1601 	ret = migrate_swap(p, env.best_task);
1602 	if (ret != 0)
1603 		trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1604 	put_task_struct(env.best_task);
1605 	return ret;
1606 }
1607 
1608 /* Attempt to migrate a task to a CPU on the preferred node. */
1609 static void numa_migrate_preferred(struct task_struct *p)
1610 {
1611 	unsigned long interval = HZ;
1612 
1613 	/* This task has no NUMA fault statistics yet */
1614 	if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1615 		return;
1616 
1617 	/* Periodically retry migrating the task to the preferred node */
1618 	interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1619 	p->numa_migrate_retry = jiffies + interval;
1620 
1621 	/* Success if task is already running on preferred CPU */
1622 	if (task_node(p) == p->numa_preferred_nid)
1623 		return;
1624 
1625 	/* Otherwise, try migrate to a CPU on the preferred node */
1626 	task_numa_migrate(p);
1627 }
1628 
1629 /*
1630  * Find the nodes on which the workload is actively running. We do this by
1631  * tracking the nodes from which NUMA hinting faults are triggered. This can
1632  * be different from the set of nodes where the workload's memory is currently
1633  * located.
1634  *
1635  * The bitmask is used to make smarter decisions on when to do NUMA page
1636  * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1637  * are added when they cause over 6/16 of the maximum number of faults, but
1638  * only removed when they drop below 3/16.
1639  */
1640 static void update_numa_active_node_mask(struct numa_group *numa_group)
1641 {
1642 	unsigned long faults, max_faults = 0;
1643 	int nid;
1644 
1645 	for_each_online_node(nid) {
1646 		faults = group_faults_cpu(numa_group, nid);
1647 		if (faults > max_faults)
1648 			max_faults = faults;
1649 	}
1650 
1651 	for_each_online_node(nid) {
1652 		faults = group_faults_cpu(numa_group, nid);
1653 		if (!node_isset(nid, numa_group->active_nodes)) {
1654 			if (faults > max_faults * 6 / 16)
1655 				node_set(nid, numa_group->active_nodes);
1656 		} else if (faults < max_faults * 3 / 16)
1657 			node_clear(nid, numa_group->active_nodes);
1658 	}
1659 }
1660 
1661 /*
1662  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1663  * increments. The more local the fault statistics are, the higher the scan
1664  * period will be for the next scan window. If local/(local+remote) ratio is
1665  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1666  * the scan period will decrease. Aim for 70% local accesses.
1667  */
1668 #define NUMA_PERIOD_SLOTS 10
1669 #define NUMA_PERIOD_THRESHOLD 7
1670 
1671 /*
1672  * Increase the scan period (slow down scanning) if the majority of
1673  * our memory is already on our local node, or if the majority of
1674  * the page accesses are shared with other processes.
1675  * Otherwise, decrease the scan period.
1676  */
1677 static void update_task_scan_period(struct task_struct *p,
1678 			unsigned long shared, unsigned long private)
1679 {
1680 	unsigned int period_slot;
1681 	int ratio;
1682 	int diff;
1683 
1684 	unsigned long remote = p->numa_faults_locality[0];
1685 	unsigned long local = p->numa_faults_locality[1];
1686 
1687 	/*
1688 	 * If there were no record hinting faults then either the task is
1689 	 * completely idle or all activity is areas that are not of interest
1690 	 * to automatic numa balancing. Related to that, if there were failed
1691 	 * migration then it implies we are migrating too quickly or the local
1692 	 * node is overloaded. In either case, scan slower
1693 	 */
1694 	if (local + shared == 0 || p->numa_faults_locality[2]) {
1695 		p->numa_scan_period = min(p->numa_scan_period_max,
1696 			p->numa_scan_period << 1);
1697 
1698 		p->mm->numa_next_scan = jiffies +
1699 			msecs_to_jiffies(p->numa_scan_period);
1700 
1701 		return;
1702 	}
1703 
1704 	/*
1705 	 * Prepare to scale scan period relative to the current period.
1706 	 *	 == NUMA_PERIOD_THRESHOLD scan period stays the same
1707 	 *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1708 	 *	 >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1709 	 */
1710 	period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1711 	ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1712 	if (ratio >= NUMA_PERIOD_THRESHOLD) {
1713 		int slot = ratio - NUMA_PERIOD_THRESHOLD;
1714 		if (!slot)
1715 			slot = 1;
1716 		diff = slot * period_slot;
1717 	} else {
1718 		diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1719 
1720 		/*
1721 		 * Scale scan rate increases based on sharing. There is an
1722 		 * inverse relationship between the degree of sharing and
1723 		 * the adjustment made to the scanning period. Broadly
1724 		 * speaking the intent is that there is little point
1725 		 * scanning faster if shared accesses dominate as it may
1726 		 * simply bounce migrations uselessly
1727 		 */
1728 		ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1729 		diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1730 	}
1731 
1732 	p->numa_scan_period = clamp(p->numa_scan_period + diff,
1733 			task_scan_min(p), task_scan_max(p));
1734 	memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1735 }
1736 
1737 /*
1738  * Get the fraction of time the task has been running since the last
1739  * NUMA placement cycle. The scheduler keeps similar statistics, but
1740  * decays those on a 32ms period, which is orders of magnitude off
1741  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1742  * stats only if the task is so new there are no NUMA statistics yet.
1743  */
1744 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1745 {
1746 	u64 runtime, delta, now;
1747 	/* Use the start of this time slice to avoid calculations. */
1748 	now = p->se.exec_start;
1749 	runtime = p->se.sum_exec_runtime;
1750 
1751 	if (p->last_task_numa_placement) {
1752 		delta = runtime - p->last_sum_exec_runtime;
1753 		*period = now - p->last_task_numa_placement;
1754 	} else {
1755 		delta = p->se.avg.load_sum / p->se.load.weight;
1756 		*period = LOAD_AVG_MAX;
1757 	}
1758 
1759 	p->last_sum_exec_runtime = runtime;
1760 	p->last_task_numa_placement = now;
1761 
1762 	return delta;
1763 }
1764 
1765 /*
1766  * Determine the preferred nid for a task in a numa_group. This needs to
1767  * be done in a way that produces consistent results with group_weight,
1768  * otherwise workloads might not converge.
1769  */
1770 static int preferred_group_nid(struct task_struct *p, int nid)
1771 {
1772 	nodemask_t nodes;
1773 	int dist;
1774 
1775 	/* Direct connections between all NUMA nodes. */
1776 	if (sched_numa_topology_type == NUMA_DIRECT)
1777 		return nid;
1778 
1779 	/*
1780 	 * On a system with glueless mesh NUMA topology, group_weight
1781 	 * scores nodes according to the number of NUMA hinting faults on
1782 	 * both the node itself, and on nearby nodes.
1783 	 */
1784 	if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1785 		unsigned long score, max_score = 0;
1786 		int node, max_node = nid;
1787 
1788 		dist = sched_max_numa_distance;
1789 
1790 		for_each_online_node(node) {
1791 			score = group_weight(p, node, dist);
1792 			if (score > max_score) {
1793 				max_score = score;
1794 				max_node = node;
1795 			}
1796 		}
1797 		return max_node;
1798 	}
1799 
1800 	/*
1801 	 * Finding the preferred nid in a system with NUMA backplane
1802 	 * interconnect topology is more involved. The goal is to locate
1803 	 * tasks from numa_groups near each other in the system, and
1804 	 * untangle workloads from different sides of the system. This requires
1805 	 * searching down the hierarchy of node groups, recursively searching
1806 	 * inside the highest scoring group of nodes. The nodemask tricks
1807 	 * keep the complexity of the search down.
1808 	 */
1809 	nodes = node_online_map;
1810 	for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1811 		unsigned long max_faults = 0;
1812 		nodemask_t max_group = NODE_MASK_NONE;
1813 		int a, b;
1814 
1815 		/* Are there nodes at this distance from each other? */
1816 		if (!find_numa_distance(dist))
1817 			continue;
1818 
1819 		for_each_node_mask(a, nodes) {
1820 			unsigned long faults = 0;
1821 			nodemask_t this_group;
1822 			nodes_clear(this_group);
1823 
1824 			/* Sum group's NUMA faults; includes a==b case. */
1825 			for_each_node_mask(b, nodes) {
1826 				if (node_distance(a, b) < dist) {
1827 					faults += group_faults(p, b);
1828 					node_set(b, this_group);
1829 					node_clear(b, nodes);
1830 				}
1831 			}
1832 
1833 			/* Remember the top group. */
1834 			if (faults > max_faults) {
1835 				max_faults = faults;
1836 				max_group = this_group;
1837 				/*
1838 				 * subtle: at the smallest distance there is
1839 				 * just one node left in each "group", the
1840 				 * winner is the preferred nid.
1841 				 */
1842 				nid = a;
1843 			}
1844 		}
1845 		/* Next round, evaluate the nodes within max_group. */
1846 		if (!max_faults)
1847 			break;
1848 		nodes = max_group;
1849 	}
1850 	return nid;
1851 }
1852 
1853 static void task_numa_placement(struct task_struct *p)
1854 {
1855 	int seq, nid, max_nid = -1, max_group_nid = -1;
1856 	unsigned long max_faults = 0, max_group_faults = 0;
1857 	unsigned long fault_types[2] = { 0, 0 };
1858 	unsigned long total_faults;
1859 	u64 runtime, period;
1860 	spinlock_t *group_lock = NULL;
1861 
1862 	/*
1863 	 * The p->mm->numa_scan_seq field gets updated without
1864 	 * exclusive access. Use READ_ONCE() here to ensure
1865 	 * that the field is read in a single access:
1866 	 */
1867 	seq = READ_ONCE(p->mm->numa_scan_seq);
1868 	if (p->numa_scan_seq == seq)
1869 		return;
1870 	p->numa_scan_seq = seq;
1871 	p->numa_scan_period_max = task_scan_max(p);
1872 
1873 	total_faults = p->numa_faults_locality[0] +
1874 		       p->numa_faults_locality[1];
1875 	runtime = numa_get_avg_runtime(p, &period);
1876 
1877 	/* If the task is part of a group prevent parallel updates to group stats */
1878 	if (p->numa_group) {
1879 		group_lock = &p->numa_group->lock;
1880 		spin_lock_irq(group_lock);
1881 	}
1882 
1883 	/* Find the node with the highest number of faults */
1884 	for_each_online_node(nid) {
1885 		/* Keep track of the offsets in numa_faults array */
1886 		int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1887 		unsigned long faults = 0, group_faults = 0;
1888 		int priv;
1889 
1890 		for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1891 			long diff, f_diff, f_weight;
1892 
1893 			mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1894 			membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1895 			cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1896 			cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1897 
1898 			/* Decay existing window, copy faults since last scan */
1899 			diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1900 			fault_types[priv] += p->numa_faults[membuf_idx];
1901 			p->numa_faults[membuf_idx] = 0;
1902 
1903 			/*
1904 			 * Normalize the faults_from, so all tasks in a group
1905 			 * count according to CPU use, instead of by the raw
1906 			 * number of faults. Tasks with little runtime have
1907 			 * little over-all impact on throughput, and thus their
1908 			 * faults are less important.
1909 			 */
1910 			f_weight = div64_u64(runtime << 16, period + 1);
1911 			f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1912 				   (total_faults + 1);
1913 			f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1914 			p->numa_faults[cpubuf_idx] = 0;
1915 
1916 			p->numa_faults[mem_idx] += diff;
1917 			p->numa_faults[cpu_idx] += f_diff;
1918 			faults += p->numa_faults[mem_idx];
1919 			p->total_numa_faults += diff;
1920 			if (p->numa_group) {
1921 				/*
1922 				 * safe because we can only change our own group
1923 				 *
1924 				 * mem_idx represents the offset for a given
1925 				 * nid and priv in a specific region because it
1926 				 * is at the beginning of the numa_faults array.
1927 				 */
1928 				p->numa_group->faults[mem_idx] += diff;
1929 				p->numa_group->faults_cpu[mem_idx] += f_diff;
1930 				p->numa_group->total_faults += diff;
1931 				group_faults += p->numa_group->faults[mem_idx];
1932 			}
1933 		}
1934 
1935 		if (faults > max_faults) {
1936 			max_faults = faults;
1937 			max_nid = nid;
1938 		}
1939 
1940 		if (group_faults > max_group_faults) {
1941 			max_group_faults = group_faults;
1942 			max_group_nid = nid;
1943 		}
1944 	}
1945 
1946 	update_task_scan_period(p, fault_types[0], fault_types[1]);
1947 
1948 	if (p->numa_group) {
1949 		update_numa_active_node_mask(p->numa_group);
1950 		spin_unlock_irq(group_lock);
1951 		max_nid = preferred_group_nid(p, max_group_nid);
1952 	}
1953 
1954 	if (max_faults) {
1955 		/* Set the new preferred node */
1956 		if (max_nid != p->numa_preferred_nid)
1957 			sched_setnuma(p, max_nid);
1958 
1959 		if (task_node(p) != p->numa_preferred_nid)
1960 			numa_migrate_preferred(p);
1961 	}
1962 }
1963 
1964 static inline int get_numa_group(struct numa_group *grp)
1965 {
1966 	return atomic_inc_not_zero(&grp->refcount);
1967 }
1968 
1969 static inline void put_numa_group(struct numa_group *grp)
1970 {
1971 	if (atomic_dec_and_test(&grp->refcount))
1972 		kfree_rcu(grp, rcu);
1973 }
1974 
1975 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1976 			int *priv)
1977 {
1978 	struct numa_group *grp, *my_grp;
1979 	struct task_struct *tsk;
1980 	bool join = false;
1981 	int cpu = cpupid_to_cpu(cpupid);
1982 	int i;
1983 
1984 	if (unlikely(!p->numa_group)) {
1985 		unsigned int size = sizeof(struct numa_group) +
1986 				    4*nr_node_ids*sizeof(unsigned long);
1987 
1988 		grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1989 		if (!grp)
1990 			return;
1991 
1992 		atomic_set(&grp->refcount, 1);
1993 		spin_lock_init(&grp->lock);
1994 		grp->gid = p->pid;
1995 		/* Second half of the array tracks nids where faults happen */
1996 		grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1997 						nr_node_ids;
1998 
1999 		node_set(task_node(current), grp->active_nodes);
2000 
2001 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2002 			grp->faults[i] = p->numa_faults[i];
2003 
2004 		grp->total_faults = p->total_numa_faults;
2005 
2006 		grp->nr_tasks++;
2007 		rcu_assign_pointer(p->numa_group, grp);
2008 	}
2009 
2010 	rcu_read_lock();
2011 	tsk = READ_ONCE(cpu_rq(cpu)->curr);
2012 
2013 	if (!cpupid_match_pid(tsk, cpupid))
2014 		goto no_join;
2015 
2016 	grp = rcu_dereference(tsk->numa_group);
2017 	if (!grp)
2018 		goto no_join;
2019 
2020 	my_grp = p->numa_group;
2021 	if (grp == my_grp)
2022 		goto no_join;
2023 
2024 	/*
2025 	 * Only join the other group if its bigger; if we're the bigger group,
2026 	 * the other task will join us.
2027 	 */
2028 	if (my_grp->nr_tasks > grp->nr_tasks)
2029 		goto no_join;
2030 
2031 	/*
2032 	 * Tie-break on the grp address.
2033 	 */
2034 	if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2035 		goto no_join;
2036 
2037 	/* Always join threads in the same process. */
2038 	if (tsk->mm == current->mm)
2039 		join = true;
2040 
2041 	/* Simple filter to avoid false positives due to PID collisions */
2042 	if (flags & TNF_SHARED)
2043 		join = true;
2044 
2045 	/* Update priv based on whether false sharing was detected */
2046 	*priv = !join;
2047 
2048 	if (join && !get_numa_group(grp))
2049 		goto no_join;
2050 
2051 	rcu_read_unlock();
2052 
2053 	if (!join)
2054 		return;
2055 
2056 	BUG_ON(irqs_disabled());
2057 	double_lock_irq(&my_grp->lock, &grp->lock);
2058 
2059 	for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2060 		my_grp->faults[i] -= p->numa_faults[i];
2061 		grp->faults[i] += p->numa_faults[i];
2062 	}
2063 	my_grp->total_faults -= p->total_numa_faults;
2064 	grp->total_faults += p->total_numa_faults;
2065 
2066 	my_grp->nr_tasks--;
2067 	grp->nr_tasks++;
2068 
2069 	spin_unlock(&my_grp->lock);
2070 	spin_unlock_irq(&grp->lock);
2071 
2072 	rcu_assign_pointer(p->numa_group, grp);
2073 
2074 	put_numa_group(my_grp);
2075 	return;
2076 
2077 no_join:
2078 	rcu_read_unlock();
2079 	return;
2080 }
2081 
2082 void task_numa_free(struct task_struct *p)
2083 {
2084 	struct numa_group *grp = p->numa_group;
2085 	void *numa_faults = p->numa_faults;
2086 	unsigned long flags;
2087 	int i;
2088 
2089 	if (grp) {
2090 		spin_lock_irqsave(&grp->lock, flags);
2091 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2092 			grp->faults[i] -= p->numa_faults[i];
2093 		grp->total_faults -= p->total_numa_faults;
2094 
2095 		grp->nr_tasks--;
2096 		spin_unlock_irqrestore(&grp->lock, flags);
2097 		RCU_INIT_POINTER(p->numa_group, NULL);
2098 		put_numa_group(grp);
2099 	}
2100 
2101 	p->numa_faults = NULL;
2102 	kfree(numa_faults);
2103 }
2104 
2105 /*
2106  * Got a PROT_NONE fault for a page on @node.
2107  */
2108 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2109 {
2110 	struct task_struct *p = current;
2111 	bool migrated = flags & TNF_MIGRATED;
2112 	int cpu_node = task_node(current);
2113 	int local = !!(flags & TNF_FAULT_LOCAL);
2114 	int priv;
2115 
2116 	if (!static_branch_likely(&sched_numa_balancing))
2117 		return;
2118 
2119 	/* for example, ksmd faulting in a user's mm */
2120 	if (!p->mm)
2121 		return;
2122 
2123 	/* Allocate buffer to track faults on a per-node basis */
2124 	if (unlikely(!p->numa_faults)) {
2125 		int size = sizeof(*p->numa_faults) *
2126 			   NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2127 
2128 		p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2129 		if (!p->numa_faults)
2130 			return;
2131 
2132 		p->total_numa_faults = 0;
2133 		memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2134 	}
2135 
2136 	/*
2137 	 * First accesses are treated as private, otherwise consider accesses
2138 	 * to be private if the accessing pid has not changed
2139 	 */
2140 	if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2141 		priv = 1;
2142 	} else {
2143 		priv = cpupid_match_pid(p, last_cpupid);
2144 		if (!priv && !(flags & TNF_NO_GROUP))
2145 			task_numa_group(p, last_cpupid, flags, &priv);
2146 	}
2147 
2148 	/*
2149 	 * If a workload spans multiple NUMA nodes, a shared fault that
2150 	 * occurs wholly within the set of nodes that the workload is
2151 	 * actively using should be counted as local. This allows the
2152 	 * scan rate to slow down when a workload has settled down.
2153 	 */
2154 	if (!priv && !local && p->numa_group &&
2155 			node_isset(cpu_node, p->numa_group->active_nodes) &&
2156 			node_isset(mem_node, p->numa_group->active_nodes))
2157 		local = 1;
2158 
2159 	task_numa_placement(p);
2160 
2161 	/*
2162 	 * Retry task to preferred node migration periodically, in case it
2163 	 * case it previously failed, or the scheduler moved us.
2164 	 */
2165 	if (time_after(jiffies, p->numa_migrate_retry))
2166 		numa_migrate_preferred(p);
2167 
2168 	if (migrated)
2169 		p->numa_pages_migrated += pages;
2170 	if (flags & TNF_MIGRATE_FAIL)
2171 		p->numa_faults_locality[2] += pages;
2172 
2173 	p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2174 	p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2175 	p->numa_faults_locality[local] += pages;
2176 }
2177 
2178 static void reset_ptenuma_scan(struct task_struct *p)
2179 {
2180 	/*
2181 	 * We only did a read acquisition of the mmap sem, so
2182 	 * p->mm->numa_scan_seq is written to without exclusive access
2183 	 * and the update is not guaranteed to be atomic. That's not
2184 	 * much of an issue though, since this is just used for
2185 	 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2186 	 * expensive, to avoid any form of compiler optimizations:
2187 	 */
2188 	WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2189 	p->mm->numa_scan_offset = 0;
2190 }
2191 
2192 /*
2193  * The expensive part of numa migration is done from task_work context.
2194  * Triggered from task_tick_numa().
2195  */
2196 void task_numa_work(struct callback_head *work)
2197 {
2198 	unsigned long migrate, next_scan, now = jiffies;
2199 	struct task_struct *p = current;
2200 	struct mm_struct *mm = p->mm;
2201 	u64 runtime = p->se.sum_exec_runtime;
2202 	struct vm_area_struct *vma;
2203 	unsigned long start, end;
2204 	unsigned long nr_pte_updates = 0;
2205 	long pages, virtpages;
2206 
2207 	WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2208 
2209 	work->next = work; /* protect against double add */
2210 	/*
2211 	 * Who cares about NUMA placement when they're dying.
2212 	 *
2213 	 * NOTE: make sure not to dereference p->mm before this check,
2214 	 * exit_task_work() happens _after_ exit_mm() so we could be called
2215 	 * without p->mm even though we still had it when we enqueued this
2216 	 * work.
2217 	 */
2218 	if (p->flags & PF_EXITING)
2219 		return;
2220 
2221 	if (!mm->numa_next_scan) {
2222 		mm->numa_next_scan = now +
2223 			msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2224 	}
2225 
2226 	/*
2227 	 * Enforce maximal scan/migration frequency..
2228 	 */
2229 	migrate = mm->numa_next_scan;
2230 	if (time_before(now, migrate))
2231 		return;
2232 
2233 	if (p->numa_scan_period == 0) {
2234 		p->numa_scan_period_max = task_scan_max(p);
2235 		p->numa_scan_period = task_scan_min(p);
2236 	}
2237 
2238 	next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2239 	if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2240 		return;
2241 
2242 	/*
2243 	 * Delay this task enough that another task of this mm will likely win
2244 	 * the next time around.
2245 	 */
2246 	p->node_stamp += 2 * TICK_NSEC;
2247 
2248 	start = mm->numa_scan_offset;
2249 	pages = sysctl_numa_balancing_scan_size;
2250 	pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2251 	virtpages = pages * 8;	   /* Scan up to this much virtual space */
2252 	if (!pages)
2253 		return;
2254 
2255 
2256 	down_read(&mm->mmap_sem);
2257 	vma = find_vma(mm, start);
2258 	if (!vma) {
2259 		reset_ptenuma_scan(p);
2260 		start = 0;
2261 		vma = mm->mmap;
2262 	}
2263 	for (; vma; vma = vma->vm_next) {
2264 		if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2265 			is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2266 			continue;
2267 		}
2268 
2269 		/*
2270 		 * Shared library pages mapped by multiple processes are not
2271 		 * migrated as it is expected they are cache replicated. Avoid
2272 		 * hinting faults in read-only file-backed mappings or the vdso
2273 		 * as migrating the pages will be of marginal benefit.
2274 		 */
2275 		if (!vma->vm_mm ||
2276 		    (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2277 			continue;
2278 
2279 		/*
2280 		 * Skip inaccessible VMAs to avoid any confusion between
2281 		 * PROT_NONE and NUMA hinting ptes
2282 		 */
2283 		if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2284 			continue;
2285 
2286 		do {
2287 			start = max(start, vma->vm_start);
2288 			end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2289 			end = min(end, vma->vm_end);
2290 			nr_pte_updates = change_prot_numa(vma, start, end);
2291 
2292 			/*
2293 			 * Try to scan sysctl_numa_balancing_size worth of
2294 			 * hpages that have at least one present PTE that
2295 			 * is not already pte-numa. If the VMA contains
2296 			 * areas that are unused or already full of prot_numa
2297 			 * PTEs, scan up to virtpages, to skip through those
2298 			 * areas faster.
2299 			 */
2300 			if (nr_pte_updates)
2301 				pages -= (end - start) >> PAGE_SHIFT;
2302 			virtpages -= (end - start) >> PAGE_SHIFT;
2303 
2304 			start = end;
2305 			if (pages <= 0 || virtpages <= 0)
2306 				goto out;
2307 
2308 			cond_resched();
2309 		} while (end != vma->vm_end);
2310 	}
2311 
2312 out:
2313 	/*
2314 	 * It is possible to reach the end of the VMA list but the last few
2315 	 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2316 	 * would find the !migratable VMA on the next scan but not reset the
2317 	 * scanner to the start so check it now.
2318 	 */
2319 	if (vma)
2320 		mm->numa_scan_offset = start;
2321 	else
2322 		reset_ptenuma_scan(p);
2323 	up_read(&mm->mmap_sem);
2324 
2325 	/*
2326 	 * Make sure tasks use at least 32x as much time to run other code
2327 	 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2328 	 * Usually update_task_scan_period slows down scanning enough; on an
2329 	 * overloaded system we need to limit overhead on a per task basis.
2330 	 */
2331 	if (unlikely(p->se.sum_exec_runtime != runtime)) {
2332 		u64 diff = p->se.sum_exec_runtime - runtime;
2333 		p->node_stamp += 32 * diff;
2334 	}
2335 }
2336 
2337 /*
2338  * Drive the periodic memory faults..
2339  */
2340 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2341 {
2342 	struct callback_head *work = &curr->numa_work;
2343 	u64 period, now;
2344 
2345 	/*
2346 	 * We don't care about NUMA placement if we don't have memory.
2347 	 */
2348 	if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2349 		return;
2350 
2351 	/*
2352 	 * Using runtime rather than walltime has the dual advantage that
2353 	 * we (mostly) drive the selection from busy threads and that the
2354 	 * task needs to have done some actual work before we bother with
2355 	 * NUMA placement.
2356 	 */
2357 	now = curr->se.sum_exec_runtime;
2358 	period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2359 
2360 	if (now > curr->node_stamp + period) {
2361 		if (!curr->node_stamp)
2362 			curr->numa_scan_period = task_scan_min(curr);
2363 		curr->node_stamp += period;
2364 
2365 		if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2366 			init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2367 			task_work_add(curr, work, true);
2368 		}
2369 	}
2370 }
2371 #else
2372 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2373 {
2374 }
2375 
2376 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2377 {
2378 }
2379 
2380 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2381 {
2382 }
2383 #endif /* CONFIG_NUMA_BALANCING */
2384 
2385 static void
2386 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2387 {
2388 	update_load_add(&cfs_rq->load, se->load.weight);
2389 	if (!parent_entity(se))
2390 		update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2391 #ifdef CONFIG_SMP
2392 	if (entity_is_task(se)) {
2393 		struct rq *rq = rq_of(cfs_rq);
2394 
2395 		account_numa_enqueue(rq, task_of(se));
2396 		list_add(&se->group_node, &rq->cfs_tasks);
2397 	}
2398 #endif
2399 	cfs_rq->nr_running++;
2400 }
2401 
2402 static void
2403 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2404 {
2405 	update_load_sub(&cfs_rq->load, se->load.weight);
2406 	if (!parent_entity(se))
2407 		update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2408 	if (entity_is_task(se)) {
2409 		account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2410 		list_del_init(&se->group_node);
2411 	}
2412 	cfs_rq->nr_running--;
2413 }
2414 
2415 #ifdef CONFIG_FAIR_GROUP_SCHED
2416 # ifdef CONFIG_SMP
2417 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2418 {
2419 	long tg_weight;
2420 
2421 	/*
2422 	 * Use this CPU's real-time load instead of the last load contribution
2423 	 * as the updating of the contribution is delayed, and we will use the
2424 	 * the real-time load to calc the share. See update_tg_load_avg().
2425 	 */
2426 	tg_weight = atomic_long_read(&tg->load_avg);
2427 	tg_weight -= cfs_rq->tg_load_avg_contrib;
2428 	tg_weight += cfs_rq->load.weight;
2429 
2430 	return tg_weight;
2431 }
2432 
2433 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2434 {
2435 	long tg_weight, load, shares;
2436 
2437 	tg_weight = calc_tg_weight(tg, cfs_rq);
2438 	load = cfs_rq->load.weight;
2439 
2440 	shares = (tg->shares * load);
2441 	if (tg_weight)
2442 		shares /= tg_weight;
2443 
2444 	if (shares < MIN_SHARES)
2445 		shares = MIN_SHARES;
2446 	if (shares > tg->shares)
2447 		shares = tg->shares;
2448 
2449 	return shares;
2450 }
2451 # else /* CONFIG_SMP */
2452 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2453 {
2454 	return tg->shares;
2455 }
2456 # endif /* CONFIG_SMP */
2457 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2458 			    unsigned long weight)
2459 {
2460 	if (se->on_rq) {
2461 		/* commit outstanding execution time */
2462 		if (cfs_rq->curr == se)
2463 			update_curr(cfs_rq);
2464 		account_entity_dequeue(cfs_rq, se);
2465 	}
2466 
2467 	update_load_set(&se->load, weight);
2468 
2469 	if (se->on_rq)
2470 		account_entity_enqueue(cfs_rq, se);
2471 }
2472 
2473 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2474 
2475 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2476 {
2477 	struct task_group *tg;
2478 	struct sched_entity *se;
2479 	long shares;
2480 
2481 	tg = cfs_rq->tg;
2482 	se = tg->se[cpu_of(rq_of(cfs_rq))];
2483 	if (!se || throttled_hierarchy(cfs_rq))
2484 		return;
2485 #ifndef CONFIG_SMP
2486 	if (likely(se->load.weight == tg->shares))
2487 		return;
2488 #endif
2489 	shares = calc_cfs_shares(cfs_rq, tg);
2490 
2491 	reweight_entity(cfs_rq_of(se), se, shares);
2492 }
2493 #else /* CONFIG_FAIR_GROUP_SCHED */
2494 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2495 {
2496 }
2497 #endif /* CONFIG_FAIR_GROUP_SCHED */
2498 
2499 #ifdef CONFIG_SMP
2500 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2501 static const u32 runnable_avg_yN_inv[] = {
2502 	0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2503 	0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2504 	0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2505 	0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2506 	0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2507 	0x85aac367, 0x82cd8698,
2508 };
2509 
2510 /*
2511  * Precomputed \Sum y^k { 1<=k<=n }.  These are floor(true_value) to prevent
2512  * over-estimates when re-combining.
2513  */
2514 static const u32 runnable_avg_yN_sum[] = {
2515 	    0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2516 	 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2517 	17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2518 };
2519 
2520 /*
2521  * Approximate:
2522  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2523  */
2524 static __always_inline u64 decay_load(u64 val, u64 n)
2525 {
2526 	unsigned int local_n;
2527 
2528 	if (!n)
2529 		return val;
2530 	else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2531 		return 0;
2532 
2533 	/* after bounds checking we can collapse to 32-bit */
2534 	local_n = n;
2535 
2536 	/*
2537 	 * As y^PERIOD = 1/2, we can combine
2538 	 *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2539 	 * With a look-up table which covers y^n (n<PERIOD)
2540 	 *
2541 	 * To achieve constant time decay_load.
2542 	 */
2543 	if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2544 		val >>= local_n / LOAD_AVG_PERIOD;
2545 		local_n %= LOAD_AVG_PERIOD;
2546 	}
2547 
2548 	val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2549 	return val;
2550 }
2551 
2552 /*
2553  * For updates fully spanning n periods, the contribution to runnable
2554  * average will be: \Sum 1024*y^n
2555  *
2556  * We can compute this reasonably efficiently by combining:
2557  *   y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for  n <PERIOD}
2558  */
2559 static u32 __compute_runnable_contrib(u64 n)
2560 {
2561 	u32 contrib = 0;
2562 
2563 	if (likely(n <= LOAD_AVG_PERIOD))
2564 		return runnable_avg_yN_sum[n];
2565 	else if (unlikely(n >= LOAD_AVG_MAX_N))
2566 		return LOAD_AVG_MAX;
2567 
2568 	/* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2569 	do {
2570 		contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2571 		contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2572 
2573 		n -= LOAD_AVG_PERIOD;
2574 	} while (n > LOAD_AVG_PERIOD);
2575 
2576 	contrib = decay_load(contrib, n);
2577 	return contrib + runnable_avg_yN_sum[n];
2578 }
2579 
2580 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2581 #error "load tracking assumes 2^10 as unit"
2582 #endif
2583 
2584 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2585 
2586 /*
2587  * We can represent the historical contribution to runnable average as the
2588  * coefficients of a geometric series.  To do this we sub-divide our runnable
2589  * history into segments of approximately 1ms (1024us); label the segment that
2590  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2591  *
2592  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2593  *      p0            p1           p2
2594  *     (now)       (~1ms ago)  (~2ms ago)
2595  *
2596  * Let u_i denote the fraction of p_i that the entity was runnable.
2597  *
2598  * We then designate the fractions u_i as our co-efficients, yielding the
2599  * following representation of historical load:
2600  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2601  *
2602  * We choose y based on the with of a reasonably scheduling period, fixing:
2603  *   y^32 = 0.5
2604  *
2605  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2606  * approximately half as much as the contribution to load within the last ms
2607  * (u_0).
2608  *
2609  * When a period "rolls over" and we have new u_0`, multiplying the previous
2610  * sum again by y is sufficient to update:
2611  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2612  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2613  */
2614 static __always_inline int
2615 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2616 		  unsigned long weight, int running, struct cfs_rq *cfs_rq)
2617 {
2618 	u64 delta, scaled_delta, periods;
2619 	u32 contrib;
2620 	unsigned int delta_w, scaled_delta_w, decayed = 0;
2621 	unsigned long scale_freq, scale_cpu;
2622 
2623 	delta = now - sa->last_update_time;
2624 	/*
2625 	 * This should only happen when time goes backwards, which it
2626 	 * unfortunately does during sched clock init when we swap over to TSC.
2627 	 */
2628 	if ((s64)delta < 0) {
2629 		sa->last_update_time = now;
2630 		return 0;
2631 	}
2632 
2633 	/*
2634 	 * Use 1024ns as the unit of measurement since it's a reasonable
2635 	 * approximation of 1us and fast to compute.
2636 	 */
2637 	delta >>= 10;
2638 	if (!delta)
2639 		return 0;
2640 	sa->last_update_time = now;
2641 
2642 	scale_freq = arch_scale_freq_capacity(NULL, cpu);
2643 	scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2644 
2645 	/* delta_w is the amount already accumulated against our next period */
2646 	delta_w = sa->period_contrib;
2647 	if (delta + delta_w >= 1024) {
2648 		decayed = 1;
2649 
2650 		/* how much left for next period will start over, we don't know yet */
2651 		sa->period_contrib = 0;
2652 
2653 		/*
2654 		 * Now that we know we're crossing a period boundary, figure
2655 		 * out how much from delta we need to complete the current
2656 		 * period and accrue it.
2657 		 */
2658 		delta_w = 1024 - delta_w;
2659 		scaled_delta_w = cap_scale(delta_w, scale_freq);
2660 		if (weight) {
2661 			sa->load_sum += weight * scaled_delta_w;
2662 			if (cfs_rq) {
2663 				cfs_rq->runnable_load_sum +=
2664 						weight * scaled_delta_w;
2665 			}
2666 		}
2667 		if (running)
2668 			sa->util_sum += scaled_delta_w * scale_cpu;
2669 
2670 		delta -= delta_w;
2671 
2672 		/* Figure out how many additional periods this update spans */
2673 		periods = delta / 1024;
2674 		delta %= 1024;
2675 
2676 		sa->load_sum = decay_load(sa->load_sum, periods + 1);
2677 		if (cfs_rq) {
2678 			cfs_rq->runnable_load_sum =
2679 				decay_load(cfs_rq->runnable_load_sum, periods + 1);
2680 		}
2681 		sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2682 
2683 		/* Efficiently calculate \sum (1..n_period) 1024*y^i */
2684 		contrib = __compute_runnable_contrib(periods);
2685 		contrib = cap_scale(contrib, scale_freq);
2686 		if (weight) {
2687 			sa->load_sum += weight * contrib;
2688 			if (cfs_rq)
2689 				cfs_rq->runnable_load_sum += weight * contrib;
2690 		}
2691 		if (running)
2692 			sa->util_sum += contrib * scale_cpu;
2693 	}
2694 
2695 	/* Remainder of delta accrued against u_0` */
2696 	scaled_delta = cap_scale(delta, scale_freq);
2697 	if (weight) {
2698 		sa->load_sum += weight * scaled_delta;
2699 		if (cfs_rq)
2700 			cfs_rq->runnable_load_sum += weight * scaled_delta;
2701 	}
2702 	if (running)
2703 		sa->util_sum += scaled_delta * scale_cpu;
2704 
2705 	sa->period_contrib += delta;
2706 
2707 	if (decayed) {
2708 		sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2709 		if (cfs_rq) {
2710 			cfs_rq->runnable_load_avg =
2711 				div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2712 		}
2713 		sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2714 	}
2715 
2716 	return decayed;
2717 }
2718 
2719 #ifdef CONFIG_FAIR_GROUP_SCHED
2720 /*
2721  * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2722  * and effective_load (which is not done because it is too costly).
2723  */
2724 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2725 {
2726 	long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2727 
2728 	/*
2729 	 * No need to update load_avg for root_task_group as it is not used.
2730 	 */
2731 	if (cfs_rq->tg == &root_task_group)
2732 		return;
2733 
2734 	if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2735 		atomic_long_add(delta, &cfs_rq->tg->load_avg);
2736 		cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2737 	}
2738 }
2739 
2740 /*
2741  * Called within set_task_rq() right before setting a task's cpu. The
2742  * caller only guarantees p->pi_lock is held; no other assumptions,
2743  * including the state of rq->lock, should be made.
2744  */
2745 void set_task_rq_fair(struct sched_entity *se,
2746 		      struct cfs_rq *prev, struct cfs_rq *next)
2747 {
2748 	if (!sched_feat(ATTACH_AGE_LOAD))
2749 		return;
2750 
2751 	/*
2752 	 * We are supposed to update the task to "current" time, then its up to
2753 	 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2754 	 * getting what current time is, so simply throw away the out-of-date
2755 	 * time. This will result in the wakee task is less decayed, but giving
2756 	 * the wakee more load sounds not bad.
2757 	 */
2758 	if (se->avg.last_update_time && prev) {
2759 		u64 p_last_update_time;
2760 		u64 n_last_update_time;
2761 
2762 #ifndef CONFIG_64BIT
2763 		u64 p_last_update_time_copy;
2764 		u64 n_last_update_time_copy;
2765 
2766 		do {
2767 			p_last_update_time_copy = prev->load_last_update_time_copy;
2768 			n_last_update_time_copy = next->load_last_update_time_copy;
2769 
2770 			smp_rmb();
2771 
2772 			p_last_update_time = prev->avg.last_update_time;
2773 			n_last_update_time = next->avg.last_update_time;
2774 
2775 		} while (p_last_update_time != p_last_update_time_copy ||
2776 			 n_last_update_time != n_last_update_time_copy);
2777 #else
2778 		p_last_update_time = prev->avg.last_update_time;
2779 		n_last_update_time = next->avg.last_update_time;
2780 #endif
2781 		__update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2782 				  &se->avg, 0, 0, NULL);
2783 		se->avg.last_update_time = n_last_update_time;
2784 	}
2785 }
2786 #else /* CONFIG_FAIR_GROUP_SCHED */
2787 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2788 #endif /* CONFIG_FAIR_GROUP_SCHED */
2789 
2790 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2791 
2792 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2793 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2794 {
2795 	struct sched_avg *sa = &cfs_rq->avg;
2796 	int decayed, removed = 0;
2797 
2798 	if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2799 		s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2800 		sa->load_avg = max_t(long, sa->load_avg - r, 0);
2801 		sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2802 		removed = 1;
2803 	}
2804 
2805 	if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2806 		long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2807 		sa->util_avg = max_t(long, sa->util_avg - r, 0);
2808 		sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
2809 	}
2810 
2811 	decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2812 		scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2813 
2814 #ifndef CONFIG_64BIT
2815 	smp_wmb();
2816 	cfs_rq->load_last_update_time_copy = sa->last_update_time;
2817 #endif
2818 
2819 	return decayed || removed;
2820 }
2821 
2822 /* Update task and its cfs_rq load average */
2823 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2824 {
2825 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
2826 	u64 now = cfs_rq_clock_task(cfs_rq);
2827 	int cpu = cpu_of(rq_of(cfs_rq));
2828 
2829 	/*
2830 	 * Track task load average for carrying it to new CPU after migrated, and
2831 	 * track group sched_entity load average for task_h_load calc in migration
2832 	 */
2833 	__update_load_avg(now, cpu, &se->avg,
2834 			  se->on_rq * scale_load_down(se->load.weight),
2835 			  cfs_rq->curr == se, NULL);
2836 
2837 	if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2838 		update_tg_load_avg(cfs_rq, 0);
2839 }
2840 
2841 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2842 {
2843 	if (!sched_feat(ATTACH_AGE_LOAD))
2844 		goto skip_aging;
2845 
2846 	/*
2847 	 * If we got migrated (either between CPUs or between cgroups) we'll
2848 	 * have aged the average right before clearing @last_update_time.
2849 	 */
2850 	if (se->avg.last_update_time) {
2851 		__update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2852 				  &se->avg, 0, 0, NULL);
2853 
2854 		/*
2855 		 * XXX: we could have just aged the entire load away if we've been
2856 		 * absent from the fair class for too long.
2857 		 */
2858 	}
2859 
2860 skip_aging:
2861 	se->avg.last_update_time = cfs_rq->avg.last_update_time;
2862 	cfs_rq->avg.load_avg += se->avg.load_avg;
2863 	cfs_rq->avg.load_sum += se->avg.load_sum;
2864 	cfs_rq->avg.util_avg += se->avg.util_avg;
2865 	cfs_rq->avg.util_sum += se->avg.util_sum;
2866 }
2867 
2868 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2869 {
2870 	__update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2871 			  &se->avg, se->on_rq * scale_load_down(se->load.weight),
2872 			  cfs_rq->curr == se, NULL);
2873 
2874 	cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2875 	cfs_rq->avg.load_sum = max_t(s64,  cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2876 	cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2877 	cfs_rq->avg.util_sum = max_t(s32,  cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2878 }
2879 
2880 /* Add the load generated by se into cfs_rq's load average */
2881 static inline void
2882 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2883 {
2884 	struct sched_avg *sa = &se->avg;
2885 	u64 now = cfs_rq_clock_task(cfs_rq);
2886 	int migrated, decayed;
2887 
2888 	migrated = !sa->last_update_time;
2889 	if (!migrated) {
2890 		__update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2891 			se->on_rq * scale_load_down(se->load.weight),
2892 			cfs_rq->curr == se, NULL);
2893 	}
2894 
2895 	decayed = update_cfs_rq_load_avg(now, cfs_rq);
2896 
2897 	cfs_rq->runnable_load_avg += sa->load_avg;
2898 	cfs_rq->runnable_load_sum += sa->load_sum;
2899 
2900 	if (migrated)
2901 		attach_entity_load_avg(cfs_rq, se);
2902 
2903 	if (decayed || migrated)
2904 		update_tg_load_avg(cfs_rq, 0);
2905 }
2906 
2907 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2908 static inline void
2909 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2910 {
2911 	update_load_avg(se, 1);
2912 
2913 	cfs_rq->runnable_load_avg =
2914 		max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2915 	cfs_rq->runnable_load_sum =
2916 		max_t(s64,  cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2917 }
2918 
2919 #ifndef CONFIG_64BIT
2920 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2921 {
2922 	u64 last_update_time_copy;
2923 	u64 last_update_time;
2924 
2925 	do {
2926 		last_update_time_copy = cfs_rq->load_last_update_time_copy;
2927 		smp_rmb();
2928 		last_update_time = cfs_rq->avg.last_update_time;
2929 	} while (last_update_time != last_update_time_copy);
2930 
2931 	return last_update_time;
2932 }
2933 #else
2934 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2935 {
2936 	return cfs_rq->avg.last_update_time;
2937 }
2938 #endif
2939 
2940 /*
2941  * Task first catches up with cfs_rq, and then subtract
2942  * itself from the cfs_rq (task must be off the queue now).
2943  */
2944 void remove_entity_load_avg(struct sched_entity *se)
2945 {
2946 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
2947 	u64 last_update_time;
2948 
2949 	/*
2950 	 * Newly created task or never used group entity should not be removed
2951 	 * from its (source) cfs_rq
2952 	 */
2953 	if (se->avg.last_update_time == 0)
2954 		return;
2955 
2956 	last_update_time = cfs_rq_last_update_time(cfs_rq);
2957 
2958 	__update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2959 	atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2960 	atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2961 }
2962 
2963 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2964 {
2965 	return cfs_rq->runnable_load_avg;
2966 }
2967 
2968 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2969 {
2970 	return cfs_rq->avg.load_avg;
2971 }
2972 
2973 static int idle_balance(struct rq *this_rq);
2974 
2975 #else /* CONFIG_SMP */
2976 
2977 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2978 static inline void
2979 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2980 static inline void
2981 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2982 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2983 
2984 static inline void
2985 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2986 static inline void
2987 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2988 
2989 static inline int idle_balance(struct rq *rq)
2990 {
2991 	return 0;
2992 }
2993 
2994 #endif /* CONFIG_SMP */
2995 
2996 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2997 {
2998 #ifdef CONFIG_SCHEDSTATS
2999 	struct task_struct *tsk = NULL;
3000 
3001 	if (entity_is_task(se))
3002 		tsk = task_of(se);
3003 
3004 	if (se->statistics.sleep_start) {
3005 		u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3006 
3007 		if ((s64)delta < 0)
3008 			delta = 0;
3009 
3010 		if (unlikely(delta > se->statistics.sleep_max))
3011 			se->statistics.sleep_max = delta;
3012 
3013 		se->statistics.sleep_start = 0;
3014 		se->statistics.sum_sleep_runtime += delta;
3015 
3016 		if (tsk) {
3017 			account_scheduler_latency(tsk, delta >> 10, 1);
3018 			trace_sched_stat_sleep(tsk, delta);
3019 		}
3020 	}
3021 	if (se->statistics.block_start) {
3022 		u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3023 
3024 		if ((s64)delta < 0)
3025 			delta = 0;
3026 
3027 		if (unlikely(delta > se->statistics.block_max))
3028 			se->statistics.block_max = delta;
3029 
3030 		se->statistics.block_start = 0;
3031 		se->statistics.sum_sleep_runtime += delta;
3032 
3033 		if (tsk) {
3034 			if (tsk->in_iowait) {
3035 				se->statistics.iowait_sum += delta;
3036 				se->statistics.iowait_count++;
3037 				trace_sched_stat_iowait(tsk, delta);
3038 			}
3039 
3040 			trace_sched_stat_blocked(tsk, delta);
3041 
3042 			/*
3043 			 * Blocking time is in units of nanosecs, so shift by
3044 			 * 20 to get a milliseconds-range estimation of the
3045 			 * amount of time that the task spent sleeping:
3046 			 */
3047 			if (unlikely(prof_on == SLEEP_PROFILING)) {
3048 				profile_hits(SLEEP_PROFILING,
3049 						(void *)get_wchan(tsk),
3050 						delta >> 20);
3051 			}
3052 			account_scheduler_latency(tsk, delta >> 10, 0);
3053 		}
3054 	}
3055 #endif
3056 }
3057 
3058 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3059 {
3060 #ifdef CONFIG_SCHED_DEBUG
3061 	s64 d = se->vruntime - cfs_rq->min_vruntime;
3062 
3063 	if (d < 0)
3064 		d = -d;
3065 
3066 	if (d > 3*sysctl_sched_latency)
3067 		schedstat_inc(cfs_rq, nr_spread_over);
3068 #endif
3069 }
3070 
3071 static void
3072 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3073 {
3074 	u64 vruntime = cfs_rq->min_vruntime;
3075 
3076 	/*
3077 	 * The 'current' period is already promised to the current tasks,
3078 	 * however the extra weight of the new task will slow them down a
3079 	 * little, place the new task so that it fits in the slot that
3080 	 * stays open at the end.
3081 	 */
3082 	if (initial && sched_feat(START_DEBIT))
3083 		vruntime += sched_vslice(cfs_rq, se);
3084 
3085 	/* sleeps up to a single latency don't count. */
3086 	if (!initial) {
3087 		unsigned long thresh = sysctl_sched_latency;
3088 
3089 		/*
3090 		 * Halve their sleep time's effect, to allow
3091 		 * for a gentler effect of sleepers:
3092 		 */
3093 		if (sched_feat(GENTLE_FAIR_SLEEPERS))
3094 			thresh >>= 1;
3095 
3096 		vruntime -= thresh;
3097 	}
3098 
3099 	/* ensure we never gain time by being placed backwards. */
3100 	se->vruntime = max_vruntime(se->vruntime, vruntime);
3101 }
3102 
3103 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3104 
3105 static void
3106 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3107 {
3108 	/*
3109 	 * Update the normalized vruntime before updating min_vruntime
3110 	 * through calling update_curr().
3111 	 */
3112 	if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3113 		se->vruntime += cfs_rq->min_vruntime;
3114 
3115 	/*
3116 	 * Update run-time statistics of the 'current'.
3117 	 */
3118 	update_curr(cfs_rq);
3119 	enqueue_entity_load_avg(cfs_rq, se);
3120 	account_entity_enqueue(cfs_rq, se);
3121 	update_cfs_shares(cfs_rq);
3122 
3123 	if (flags & ENQUEUE_WAKEUP) {
3124 		place_entity(cfs_rq, se, 0);
3125 		enqueue_sleeper(cfs_rq, se);
3126 	}
3127 
3128 	update_stats_enqueue(cfs_rq, se);
3129 	check_spread(cfs_rq, se);
3130 	if (se != cfs_rq->curr)
3131 		__enqueue_entity(cfs_rq, se);
3132 	se->on_rq = 1;
3133 
3134 	if (cfs_rq->nr_running == 1) {
3135 		list_add_leaf_cfs_rq(cfs_rq);
3136 		check_enqueue_throttle(cfs_rq);
3137 	}
3138 }
3139 
3140 static void __clear_buddies_last(struct sched_entity *se)
3141 {
3142 	for_each_sched_entity(se) {
3143 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
3144 		if (cfs_rq->last != se)
3145 			break;
3146 
3147 		cfs_rq->last = NULL;
3148 	}
3149 }
3150 
3151 static void __clear_buddies_next(struct sched_entity *se)
3152 {
3153 	for_each_sched_entity(se) {
3154 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
3155 		if (cfs_rq->next != se)
3156 			break;
3157 
3158 		cfs_rq->next = NULL;
3159 	}
3160 }
3161 
3162 static void __clear_buddies_skip(struct sched_entity *se)
3163 {
3164 	for_each_sched_entity(se) {
3165 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
3166 		if (cfs_rq->skip != se)
3167 			break;
3168 
3169 		cfs_rq->skip = NULL;
3170 	}
3171 }
3172 
3173 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3174 {
3175 	if (cfs_rq->last == se)
3176 		__clear_buddies_last(se);
3177 
3178 	if (cfs_rq->next == se)
3179 		__clear_buddies_next(se);
3180 
3181 	if (cfs_rq->skip == se)
3182 		__clear_buddies_skip(se);
3183 }
3184 
3185 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3186 
3187 static void
3188 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3189 {
3190 	/*
3191 	 * Update run-time statistics of the 'current'.
3192 	 */
3193 	update_curr(cfs_rq);
3194 	dequeue_entity_load_avg(cfs_rq, se);
3195 
3196 	update_stats_dequeue(cfs_rq, se);
3197 	if (flags & DEQUEUE_SLEEP) {
3198 #ifdef CONFIG_SCHEDSTATS
3199 		if (entity_is_task(se)) {
3200 			struct task_struct *tsk = task_of(se);
3201 
3202 			if (tsk->state & TASK_INTERRUPTIBLE)
3203 				se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3204 			if (tsk->state & TASK_UNINTERRUPTIBLE)
3205 				se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3206 		}
3207 #endif
3208 	}
3209 
3210 	clear_buddies(cfs_rq, se);
3211 
3212 	if (se != cfs_rq->curr)
3213 		__dequeue_entity(cfs_rq, se);
3214 	se->on_rq = 0;
3215 	account_entity_dequeue(cfs_rq, se);
3216 
3217 	/*
3218 	 * Normalize the entity after updating the min_vruntime because the
3219 	 * update can refer to the ->curr item and we need to reflect this
3220 	 * movement in our normalized position.
3221 	 */
3222 	if (!(flags & DEQUEUE_SLEEP))
3223 		se->vruntime -= cfs_rq->min_vruntime;
3224 
3225 	/* return excess runtime on last dequeue */
3226 	return_cfs_rq_runtime(cfs_rq);
3227 
3228 	update_min_vruntime(cfs_rq);
3229 	update_cfs_shares(cfs_rq);
3230 }
3231 
3232 /*
3233  * Preempt the current task with a newly woken task if needed:
3234  */
3235 static void
3236 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3237 {
3238 	unsigned long ideal_runtime, delta_exec;
3239 	struct sched_entity *se;
3240 	s64 delta;
3241 
3242 	ideal_runtime = sched_slice(cfs_rq, curr);
3243 	delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3244 	if (delta_exec > ideal_runtime) {
3245 		resched_curr(rq_of(cfs_rq));
3246 		/*
3247 		 * The current task ran long enough, ensure it doesn't get
3248 		 * re-elected due to buddy favours.
3249 		 */
3250 		clear_buddies(cfs_rq, curr);
3251 		return;
3252 	}
3253 
3254 	/*
3255 	 * Ensure that a task that missed wakeup preemption by a
3256 	 * narrow margin doesn't have to wait for a full slice.
3257 	 * This also mitigates buddy induced latencies under load.
3258 	 */
3259 	if (delta_exec < sysctl_sched_min_granularity)
3260 		return;
3261 
3262 	se = __pick_first_entity(cfs_rq);
3263 	delta = curr->vruntime - se->vruntime;
3264 
3265 	if (delta < 0)
3266 		return;
3267 
3268 	if (delta > ideal_runtime)
3269 		resched_curr(rq_of(cfs_rq));
3270 }
3271 
3272 static void
3273 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3274 {
3275 	/* 'current' is not kept within the tree. */
3276 	if (se->on_rq) {
3277 		/*
3278 		 * Any task has to be enqueued before it get to execute on
3279 		 * a CPU. So account for the time it spent waiting on the
3280 		 * runqueue.
3281 		 */
3282 		update_stats_wait_end(cfs_rq, se);
3283 		__dequeue_entity(cfs_rq, se);
3284 		update_load_avg(se, 1);
3285 	}
3286 
3287 	update_stats_curr_start(cfs_rq, se);
3288 	cfs_rq->curr = se;
3289 #ifdef CONFIG_SCHEDSTATS
3290 	/*
3291 	 * Track our maximum slice length, if the CPU's load is at
3292 	 * least twice that of our own weight (i.e. dont track it
3293 	 * when there are only lesser-weight tasks around):
3294 	 */
3295 	if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3296 		se->statistics.slice_max = max(se->statistics.slice_max,
3297 			se->sum_exec_runtime - se->prev_sum_exec_runtime);
3298 	}
3299 #endif
3300 	se->prev_sum_exec_runtime = se->sum_exec_runtime;
3301 }
3302 
3303 static int
3304 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3305 
3306 /*
3307  * Pick the next process, keeping these things in mind, in this order:
3308  * 1) keep things fair between processes/task groups
3309  * 2) pick the "next" process, since someone really wants that to run
3310  * 3) pick the "last" process, for cache locality
3311  * 4) do not run the "skip" process, if something else is available
3312  */
3313 static struct sched_entity *
3314 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3315 {
3316 	struct sched_entity *left = __pick_first_entity(cfs_rq);
3317 	struct sched_entity *se;
3318 
3319 	/*
3320 	 * If curr is set we have to see if its left of the leftmost entity
3321 	 * still in the tree, provided there was anything in the tree at all.
3322 	 */
3323 	if (!left || (curr && entity_before(curr, left)))
3324 		left = curr;
3325 
3326 	se = left; /* ideally we run the leftmost entity */
3327 
3328 	/*
3329 	 * Avoid running the skip buddy, if running something else can
3330 	 * be done without getting too unfair.
3331 	 */
3332 	if (cfs_rq->skip == se) {
3333 		struct sched_entity *second;
3334 
3335 		if (se == curr) {
3336 			second = __pick_first_entity(cfs_rq);
3337 		} else {
3338 			second = __pick_next_entity(se);
3339 			if (!second || (curr && entity_before(curr, second)))
3340 				second = curr;
3341 		}
3342 
3343 		if (second && wakeup_preempt_entity(second, left) < 1)
3344 			se = second;
3345 	}
3346 
3347 	/*
3348 	 * Prefer last buddy, try to return the CPU to a preempted task.
3349 	 */
3350 	if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3351 		se = cfs_rq->last;
3352 
3353 	/*
3354 	 * Someone really wants this to run. If it's not unfair, run it.
3355 	 */
3356 	if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3357 		se = cfs_rq->next;
3358 
3359 	clear_buddies(cfs_rq, se);
3360 
3361 	return se;
3362 }
3363 
3364 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3365 
3366 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3367 {
3368 	/*
3369 	 * If still on the runqueue then deactivate_task()
3370 	 * was not called and update_curr() has to be done:
3371 	 */
3372 	if (prev->on_rq)
3373 		update_curr(cfs_rq);
3374 
3375 	/* throttle cfs_rqs exceeding runtime */
3376 	check_cfs_rq_runtime(cfs_rq);
3377 
3378 	check_spread(cfs_rq, prev);
3379 	if (prev->on_rq) {
3380 		update_stats_wait_start(cfs_rq, prev);
3381 		/* Put 'current' back into the tree. */
3382 		__enqueue_entity(cfs_rq, prev);
3383 		/* in !on_rq case, update occurred at dequeue */
3384 		update_load_avg(prev, 0);
3385 	}
3386 	cfs_rq->curr = NULL;
3387 }
3388 
3389 static void
3390 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3391 {
3392 	/*
3393 	 * Update run-time statistics of the 'current'.
3394 	 */
3395 	update_curr(cfs_rq);
3396 
3397 	/*
3398 	 * Ensure that runnable average is periodically updated.
3399 	 */
3400 	update_load_avg(curr, 1);
3401 	update_cfs_shares(cfs_rq);
3402 
3403 #ifdef CONFIG_SCHED_HRTICK
3404 	/*
3405 	 * queued ticks are scheduled to match the slice, so don't bother
3406 	 * validating it and just reschedule.
3407 	 */
3408 	if (queued) {
3409 		resched_curr(rq_of(cfs_rq));
3410 		return;
3411 	}
3412 	/*
3413 	 * don't let the period tick interfere with the hrtick preemption
3414 	 */
3415 	if (!sched_feat(DOUBLE_TICK) &&
3416 			hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3417 		return;
3418 #endif
3419 
3420 	if (cfs_rq->nr_running > 1)
3421 		check_preempt_tick(cfs_rq, curr);
3422 }
3423 
3424 
3425 /**************************************************
3426  * CFS bandwidth control machinery
3427  */
3428 
3429 #ifdef CONFIG_CFS_BANDWIDTH
3430 
3431 #ifdef HAVE_JUMP_LABEL
3432 static struct static_key __cfs_bandwidth_used;
3433 
3434 static inline bool cfs_bandwidth_used(void)
3435 {
3436 	return static_key_false(&__cfs_bandwidth_used);
3437 }
3438 
3439 void cfs_bandwidth_usage_inc(void)
3440 {
3441 	static_key_slow_inc(&__cfs_bandwidth_used);
3442 }
3443 
3444 void cfs_bandwidth_usage_dec(void)
3445 {
3446 	static_key_slow_dec(&__cfs_bandwidth_used);
3447 }
3448 #else /* HAVE_JUMP_LABEL */
3449 static bool cfs_bandwidth_used(void)
3450 {
3451 	return true;
3452 }
3453 
3454 void cfs_bandwidth_usage_inc(void) {}
3455 void cfs_bandwidth_usage_dec(void) {}
3456 #endif /* HAVE_JUMP_LABEL */
3457 
3458 /*
3459  * default period for cfs group bandwidth.
3460  * default: 0.1s, units: nanoseconds
3461  */
3462 static inline u64 default_cfs_period(void)
3463 {
3464 	return 100000000ULL;
3465 }
3466 
3467 static inline u64 sched_cfs_bandwidth_slice(void)
3468 {
3469 	return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3470 }
3471 
3472 /*
3473  * Replenish runtime according to assigned quota and update expiration time.
3474  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3475  * additional synchronization around rq->lock.
3476  *
3477  * requires cfs_b->lock
3478  */
3479 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3480 {
3481 	u64 now;
3482 
3483 	if (cfs_b->quota == RUNTIME_INF)
3484 		return;
3485 
3486 	now = sched_clock_cpu(smp_processor_id());
3487 	cfs_b->runtime = cfs_b->quota;
3488 	cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3489 }
3490 
3491 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3492 {
3493 	return &tg->cfs_bandwidth;
3494 }
3495 
3496 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3497 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3498 {
3499 	if (unlikely(cfs_rq->throttle_count))
3500 		return cfs_rq->throttled_clock_task;
3501 
3502 	return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3503 }
3504 
3505 /* returns 0 on failure to allocate runtime */
3506 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3507 {
3508 	struct task_group *tg = cfs_rq->tg;
3509 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3510 	u64 amount = 0, min_amount, expires;
3511 
3512 	/* note: this is a positive sum as runtime_remaining <= 0 */
3513 	min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3514 
3515 	raw_spin_lock(&cfs_b->lock);
3516 	if (cfs_b->quota == RUNTIME_INF)
3517 		amount = min_amount;
3518 	else {
3519 		start_cfs_bandwidth(cfs_b);
3520 
3521 		if (cfs_b->runtime > 0) {
3522 			amount = min(cfs_b->runtime, min_amount);
3523 			cfs_b->runtime -= amount;
3524 			cfs_b->idle = 0;
3525 		}
3526 	}
3527 	expires = cfs_b->runtime_expires;
3528 	raw_spin_unlock(&cfs_b->lock);
3529 
3530 	cfs_rq->runtime_remaining += amount;
3531 	/*
3532 	 * we may have advanced our local expiration to account for allowed
3533 	 * spread between our sched_clock and the one on which runtime was
3534 	 * issued.
3535 	 */
3536 	if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3537 		cfs_rq->runtime_expires = expires;
3538 
3539 	return cfs_rq->runtime_remaining > 0;
3540 }
3541 
3542 /*
3543  * Note: This depends on the synchronization provided by sched_clock and the
3544  * fact that rq->clock snapshots this value.
3545  */
3546 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3547 {
3548 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3549 
3550 	/* if the deadline is ahead of our clock, nothing to do */
3551 	if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3552 		return;
3553 
3554 	if (cfs_rq->runtime_remaining < 0)
3555 		return;
3556 
3557 	/*
3558 	 * If the local deadline has passed we have to consider the
3559 	 * possibility that our sched_clock is 'fast' and the global deadline
3560 	 * has not truly expired.
3561 	 *
3562 	 * Fortunately we can check determine whether this the case by checking
3563 	 * whether the global deadline has advanced. It is valid to compare
3564 	 * cfs_b->runtime_expires without any locks since we only care about
3565 	 * exact equality, so a partial write will still work.
3566 	 */
3567 
3568 	if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3569 		/* extend local deadline, drift is bounded above by 2 ticks */
3570 		cfs_rq->runtime_expires += TICK_NSEC;
3571 	} else {
3572 		/* global deadline is ahead, expiration has passed */
3573 		cfs_rq->runtime_remaining = 0;
3574 	}
3575 }
3576 
3577 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3578 {
3579 	/* dock delta_exec before expiring quota (as it could span periods) */
3580 	cfs_rq->runtime_remaining -= delta_exec;
3581 	expire_cfs_rq_runtime(cfs_rq);
3582 
3583 	if (likely(cfs_rq->runtime_remaining > 0))
3584 		return;
3585 
3586 	/*
3587 	 * if we're unable to extend our runtime we resched so that the active
3588 	 * hierarchy can be throttled
3589 	 */
3590 	if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3591 		resched_curr(rq_of(cfs_rq));
3592 }
3593 
3594 static __always_inline
3595 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3596 {
3597 	if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3598 		return;
3599 
3600 	__account_cfs_rq_runtime(cfs_rq, delta_exec);
3601 }
3602 
3603 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3604 {
3605 	return cfs_bandwidth_used() && cfs_rq->throttled;
3606 }
3607 
3608 /* check whether cfs_rq, or any parent, is throttled */
3609 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3610 {
3611 	return cfs_bandwidth_used() && cfs_rq->throttle_count;
3612 }
3613 
3614 /*
3615  * Ensure that neither of the group entities corresponding to src_cpu or
3616  * dest_cpu are members of a throttled hierarchy when performing group
3617  * load-balance operations.
3618  */
3619 static inline int throttled_lb_pair(struct task_group *tg,
3620 				    int src_cpu, int dest_cpu)
3621 {
3622 	struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3623 
3624 	src_cfs_rq = tg->cfs_rq[src_cpu];
3625 	dest_cfs_rq = tg->cfs_rq[dest_cpu];
3626 
3627 	return throttled_hierarchy(src_cfs_rq) ||
3628 	       throttled_hierarchy(dest_cfs_rq);
3629 }
3630 
3631 /* updated child weight may affect parent so we have to do this bottom up */
3632 static int tg_unthrottle_up(struct task_group *tg, void *data)
3633 {
3634 	struct rq *rq = data;
3635 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3636 
3637 	cfs_rq->throttle_count--;
3638 #ifdef CONFIG_SMP
3639 	if (!cfs_rq->throttle_count) {
3640 		/* adjust cfs_rq_clock_task() */
3641 		cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3642 					     cfs_rq->throttled_clock_task;
3643 	}
3644 #endif
3645 
3646 	return 0;
3647 }
3648 
3649 static int tg_throttle_down(struct task_group *tg, void *data)
3650 {
3651 	struct rq *rq = data;
3652 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3653 
3654 	/* group is entering throttled state, stop time */
3655 	if (!cfs_rq->throttle_count)
3656 		cfs_rq->throttled_clock_task = rq_clock_task(rq);
3657 	cfs_rq->throttle_count++;
3658 
3659 	return 0;
3660 }
3661 
3662 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3663 {
3664 	struct rq *rq = rq_of(cfs_rq);
3665 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3666 	struct sched_entity *se;
3667 	long task_delta, dequeue = 1;
3668 	bool empty;
3669 
3670 	se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3671 
3672 	/* freeze hierarchy runnable averages while throttled */
3673 	rcu_read_lock();
3674 	walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3675 	rcu_read_unlock();
3676 
3677 	task_delta = cfs_rq->h_nr_running;
3678 	for_each_sched_entity(se) {
3679 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3680 		/* throttled entity or throttle-on-deactivate */
3681 		if (!se->on_rq)
3682 			break;
3683 
3684 		if (dequeue)
3685 			dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3686 		qcfs_rq->h_nr_running -= task_delta;
3687 
3688 		if (qcfs_rq->load.weight)
3689 			dequeue = 0;
3690 	}
3691 
3692 	if (!se)
3693 		sub_nr_running(rq, task_delta);
3694 
3695 	cfs_rq->throttled = 1;
3696 	cfs_rq->throttled_clock = rq_clock(rq);
3697 	raw_spin_lock(&cfs_b->lock);
3698 	empty = list_empty(&cfs_b->throttled_cfs_rq);
3699 
3700 	/*
3701 	 * Add to the _head_ of the list, so that an already-started
3702 	 * distribute_cfs_runtime will not see us
3703 	 */
3704 	list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3705 
3706 	/*
3707 	 * If we're the first throttled task, make sure the bandwidth
3708 	 * timer is running.
3709 	 */
3710 	if (empty)
3711 		start_cfs_bandwidth(cfs_b);
3712 
3713 	raw_spin_unlock(&cfs_b->lock);
3714 }
3715 
3716 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3717 {
3718 	struct rq *rq = rq_of(cfs_rq);
3719 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3720 	struct sched_entity *se;
3721 	int enqueue = 1;
3722 	long task_delta;
3723 
3724 	se = cfs_rq->tg->se[cpu_of(rq)];
3725 
3726 	cfs_rq->throttled = 0;
3727 
3728 	update_rq_clock(rq);
3729 
3730 	raw_spin_lock(&cfs_b->lock);
3731 	cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3732 	list_del_rcu(&cfs_rq->throttled_list);
3733 	raw_spin_unlock(&cfs_b->lock);
3734 
3735 	/* update hierarchical throttle state */
3736 	walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3737 
3738 	if (!cfs_rq->load.weight)
3739 		return;
3740 
3741 	task_delta = cfs_rq->h_nr_running;
3742 	for_each_sched_entity(se) {
3743 		if (se->on_rq)
3744 			enqueue = 0;
3745 
3746 		cfs_rq = cfs_rq_of(se);
3747 		if (enqueue)
3748 			enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3749 		cfs_rq->h_nr_running += task_delta;
3750 
3751 		if (cfs_rq_throttled(cfs_rq))
3752 			break;
3753 	}
3754 
3755 	if (!se)
3756 		add_nr_running(rq, task_delta);
3757 
3758 	/* determine whether we need to wake up potentially idle cpu */
3759 	if (rq->curr == rq->idle && rq->cfs.nr_running)
3760 		resched_curr(rq);
3761 }
3762 
3763 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3764 		u64 remaining, u64 expires)
3765 {
3766 	struct cfs_rq *cfs_rq;
3767 	u64 runtime;
3768 	u64 starting_runtime = remaining;
3769 
3770 	rcu_read_lock();
3771 	list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3772 				throttled_list) {
3773 		struct rq *rq = rq_of(cfs_rq);
3774 
3775 		raw_spin_lock(&rq->lock);
3776 		if (!cfs_rq_throttled(cfs_rq))
3777 			goto next;
3778 
3779 		runtime = -cfs_rq->runtime_remaining + 1;
3780 		if (runtime > remaining)
3781 			runtime = remaining;
3782 		remaining -= runtime;
3783 
3784 		cfs_rq->runtime_remaining += runtime;
3785 		cfs_rq->runtime_expires = expires;
3786 
3787 		/* we check whether we're throttled above */
3788 		if (cfs_rq->runtime_remaining > 0)
3789 			unthrottle_cfs_rq(cfs_rq);
3790 
3791 next:
3792 		raw_spin_unlock(&rq->lock);
3793 
3794 		if (!remaining)
3795 			break;
3796 	}
3797 	rcu_read_unlock();
3798 
3799 	return starting_runtime - remaining;
3800 }
3801 
3802 /*
3803  * Responsible for refilling a task_group's bandwidth and unthrottling its
3804  * cfs_rqs as appropriate. If there has been no activity within the last
3805  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3806  * used to track this state.
3807  */
3808 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3809 {
3810 	u64 runtime, runtime_expires;
3811 	int throttled;
3812 
3813 	/* no need to continue the timer with no bandwidth constraint */
3814 	if (cfs_b->quota == RUNTIME_INF)
3815 		goto out_deactivate;
3816 
3817 	throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3818 	cfs_b->nr_periods += overrun;
3819 
3820 	/*
3821 	 * idle depends on !throttled (for the case of a large deficit), and if
3822 	 * we're going inactive then everything else can be deferred
3823 	 */
3824 	if (cfs_b->idle && !throttled)
3825 		goto out_deactivate;
3826 
3827 	__refill_cfs_bandwidth_runtime(cfs_b);
3828 
3829 	if (!throttled) {
3830 		/* mark as potentially idle for the upcoming period */
3831 		cfs_b->idle = 1;
3832 		return 0;
3833 	}
3834 
3835 	/* account preceding periods in which throttling occurred */
3836 	cfs_b->nr_throttled += overrun;
3837 
3838 	runtime_expires = cfs_b->runtime_expires;
3839 
3840 	/*
3841 	 * This check is repeated as we are holding onto the new bandwidth while
3842 	 * we unthrottle. This can potentially race with an unthrottled group
3843 	 * trying to acquire new bandwidth from the global pool. This can result
3844 	 * in us over-using our runtime if it is all used during this loop, but
3845 	 * only by limited amounts in that extreme case.
3846 	 */
3847 	while (throttled && cfs_b->runtime > 0) {
3848 		runtime = cfs_b->runtime;
3849 		raw_spin_unlock(&cfs_b->lock);
3850 		/* we can't nest cfs_b->lock while distributing bandwidth */
3851 		runtime = distribute_cfs_runtime(cfs_b, runtime,
3852 						 runtime_expires);
3853 		raw_spin_lock(&cfs_b->lock);
3854 
3855 		throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3856 
3857 		cfs_b->runtime -= min(runtime, cfs_b->runtime);
3858 	}
3859 
3860 	/*
3861 	 * While we are ensured activity in the period following an
3862 	 * unthrottle, this also covers the case in which the new bandwidth is
3863 	 * insufficient to cover the existing bandwidth deficit.  (Forcing the
3864 	 * timer to remain active while there are any throttled entities.)
3865 	 */
3866 	cfs_b->idle = 0;
3867 
3868 	return 0;
3869 
3870 out_deactivate:
3871 	return 1;
3872 }
3873 
3874 /* a cfs_rq won't donate quota below this amount */
3875 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3876 /* minimum remaining period time to redistribute slack quota */
3877 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3878 /* how long we wait to gather additional slack before distributing */
3879 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3880 
3881 /*
3882  * Are we near the end of the current quota period?
3883  *
3884  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3885  * hrtimer base being cleared by hrtimer_start. In the case of
3886  * migrate_hrtimers, base is never cleared, so we are fine.
3887  */
3888 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3889 {
3890 	struct hrtimer *refresh_timer = &cfs_b->period_timer;
3891 	u64 remaining;
3892 
3893 	/* if the call-back is running a quota refresh is already occurring */
3894 	if (hrtimer_callback_running(refresh_timer))
3895 		return 1;
3896 
3897 	/* is a quota refresh about to occur? */
3898 	remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3899 	if (remaining < min_expire)
3900 		return 1;
3901 
3902 	return 0;
3903 }
3904 
3905 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3906 {
3907 	u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3908 
3909 	/* if there's a quota refresh soon don't bother with slack */
3910 	if (runtime_refresh_within(cfs_b, min_left))
3911 		return;
3912 
3913 	hrtimer_start(&cfs_b->slack_timer,
3914 			ns_to_ktime(cfs_bandwidth_slack_period),
3915 			HRTIMER_MODE_REL);
3916 }
3917 
3918 /* we know any runtime found here is valid as update_curr() precedes return */
3919 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3920 {
3921 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3922 	s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3923 
3924 	if (slack_runtime <= 0)
3925 		return;
3926 
3927 	raw_spin_lock(&cfs_b->lock);
3928 	if (cfs_b->quota != RUNTIME_INF &&
3929 	    cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3930 		cfs_b->runtime += slack_runtime;
3931 
3932 		/* we are under rq->lock, defer unthrottling using a timer */
3933 		if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3934 		    !list_empty(&cfs_b->throttled_cfs_rq))
3935 			start_cfs_slack_bandwidth(cfs_b);
3936 	}
3937 	raw_spin_unlock(&cfs_b->lock);
3938 
3939 	/* even if it's not valid for return we don't want to try again */
3940 	cfs_rq->runtime_remaining -= slack_runtime;
3941 }
3942 
3943 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3944 {
3945 	if (!cfs_bandwidth_used())
3946 		return;
3947 
3948 	if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3949 		return;
3950 
3951 	__return_cfs_rq_runtime(cfs_rq);
3952 }
3953 
3954 /*
3955  * This is done with a timer (instead of inline with bandwidth return) since
3956  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3957  */
3958 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3959 {
3960 	u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3961 	u64 expires;
3962 
3963 	/* confirm we're still not at a refresh boundary */
3964 	raw_spin_lock(&cfs_b->lock);
3965 	if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3966 		raw_spin_unlock(&cfs_b->lock);
3967 		return;
3968 	}
3969 
3970 	if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3971 		runtime = cfs_b->runtime;
3972 
3973 	expires = cfs_b->runtime_expires;
3974 	raw_spin_unlock(&cfs_b->lock);
3975 
3976 	if (!runtime)
3977 		return;
3978 
3979 	runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3980 
3981 	raw_spin_lock(&cfs_b->lock);
3982 	if (expires == cfs_b->runtime_expires)
3983 		cfs_b->runtime -= min(runtime, cfs_b->runtime);
3984 	raw_spin_unlock(&cfs_b->lock);
3985 }
3986 
3987 /*
3988  * When a group wakes up we want to make sure that its quota is not already
3989  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3990  * runtime as update_curr() throttling can not not trigger until it's on-rq.
3991  */
3992 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3993 {
3994 	if (!cfs_bandwidth_used())
3995 		return;
3996 
3997 	/* an active group must be handled by the update_curr()->put() path */
3998 	if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3999 		return;
4000 
4001 	/* ensure the group is not already throttled */
4002 	if (cfs_rq_throttled(cfs_rq))
4003 		return;
4004 
4005 	/* update runtime allocation */
4006 	account_cfs_rq_runtime(cfs_rq, 0);
4007 	if (cfs_rq->runtime_remaining <= 0)
4008 		throttle_cfs_rq(cfs_rq);
4009 }
4010 
4011 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4012 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4013 {
4014 	if (!cfs_bandwidth_used())
4015 		return false;
4016 
4017 	if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4018 		return false;
4019 
4020 	/*
4021 	 * it's possible for a throttled entity to be forced into a running
4022 	 * state (e.g. set_curr_task), in this case we're finished.
4023 	 */
4024 	if (cfs_rq_throttled(cfs_rq))
4025 		return true;
4026 
4027 	throttle_cfs_rq(cfs_rq);
4028 	return true;
4029 }
4030 
4031 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4032 {
4033 	struct cfs_bandwidth *cfs_b =
4034 		container_of(timer, struct cfs_bandwidth, slack_timer);
4035 
4036 	do_sched_cfs_slack_timer(cfs_b);
4037 
4038 	return HRTIMER_NORESTART;
4039 }
4040 
4041 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4042 {
4043 	struct cfs_bandwidth *cfs_b =
4044 		container_of(timer, struct cfs_bandwidth, period_timer);
4045 	int overrun;
4046 	int idle = 0;
4047 
4048 	raw_spin_lock(&cfs_b->lock);
4049 	for (;;) {
4050 		overrun = hrtimer_forward_now(timer, cfs_b->period);
4051 		if (!overrun)
4052 			break;
4053 
4054 		idle = do_sched_cfs_period_timer(cfs_b, overrun);
4055 	}
4056 	if (idle)
4057 		cfs_b->period_active = 0;
4058 	raw_spin_unlock(&cfs_b->lock);
4059 
4060 	return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4061 }
4062 
4063 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4064 {
4065 	raw_spin_lock_init(&cfs_b->lock);
4066 	cfs_b->runtime = 0;
4067 	cfs_b->quota = RUNTIME_INF;
4068 	cfs_b->period = ns_to_ktime(default_cfs_period());
4069 
4070 	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4071 	hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4072 	cfs_b->period_timer.function = sched_cfs_period_timer;
4073 	hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4074 	cfs_b->slack_timer.function = sched_cfs_slack_timer;
4075 }
4076 
4077 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4078 {
4079 	cfs_rq->runtime_enabled = 0;
4080 	INIT_LIST_HEAD(&cfs_rq->throttled_list);
4081 }
4082 
4083 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4084 {
4085 	lockdep_assert_held(&cfs_b->lock);
4086 
4087 	if (!cfs_b->period_active) {
4088 		cfs_b->period_active = 1;
4089 		hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4090 		hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4091 	}
4092 }
4093 
4094 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4095 {
4096 	/* init_cfs_bandwidth() was not called */
4097 	if (!cfs_b->throttled_cfs_rq.next)
4098 		return;
4099 
4100 	hrtimer_cancel(&cfs_b->period_timer);
4101 	hrtimer_cancel(&cfs_b->slack_timer);
4102 }
4103 
4104 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4105 {
4106 	struct cfs_rq *cfs_rq;
4107 
4108 	for_each_leaf_cfs_rq(rq, cfs_rq) {
4109 		struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4110 
4111 		raw_spin_lock(&cfs_b->lock);
4112 		cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4113 		raw_spin_unlock(&cfs_b->lock);
4114 	}
4115 }
4116 
4117 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4118 {
4119 	struct cfs_rq *cfs_rq;
4120 
4121 	for_each_leaf_cfs_rq(rq, cfs_rq) {
4122 		if (!cfs_rq->runtime_enabled)
4123 			continue;
4124 
4125 		/*
4126 		 * clock_task is not advancing so we just need to make sure
4127 		 * there's some valid quota amount
4128 		 */
4129 		cfs_rq->runtime_remaining = 1;
4130 		/*
4131 		 * Offline rq is schedulable till cpu is completely disabled
4132 		 * in take_cpu_down(), so we prevent new cfs throttling here.
4133 		 */
4134 		cfs_rq->runtime_enabled = 0;
4135 
4136 		if (cfs_rq_throttled(cfs_rq))
4137 			unthrottle_cfs_rq(cfs_rq);
4138 	}
4139 }
4140 
4141 #else /* CONFIG_CFS_BANDWIDTH */
4142 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4143 {
4144 	return rq_clock_task(rq_of(cfs_rq));
4145 }
4146 
4147 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4148 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4149 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4150 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4151 
4152 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4153 {
4154 	return 0;
4155 }
4156 
4157 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4158 {
4159 	return 0;
4160 }
4161 
4162 static inline int throttled_lb_pair(struct task_group *tg,
4163 				    int src_cpu, int dest_cpu)
4164 {
4165 	return 0;
4166 }
4167 
4168 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4169 
4170 #ifdef CONFIG_FAIR_GROUP_SCHED
4171 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4172 #endif
4173 
4174 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4175 {
4176 	return NULL;
4177 }
4178 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4179 static inline void update_runtime_enabled(struct rq *rq) {}
4180 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4181 
4182 #endif /* CONFIG_CFS_BANDWIDTH */
4183 
4184 /**************************************************
4185  * CFS operations on tasks:
4186  */
4187 
4188 #ifdef CONFIG_SCHED_HRTICK
4189 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4190 {
4191 	struct sched_entity *se = &p->se;
4192 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4193 
4194 	WARN_ON(task_rq(p) != rq);
4195 
4196 	if (cfs_rq->nr_running > 1) {
4197 		u64 slice = sched_slice(cfs_rq, se);
4198 		u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4199 		s64 delta = slice - ran;
4200 
4201 		if (delta < 0) {
4202 			if (rq->curr == p)
4203 				resched_curr(rq);
4204 			return;
4205 		}
4206 		hrtick_start(rq, delta);
4207 	}
4208 }
4209 
4210 /*
4211  * called from enqueue/dequeue and updates the hrtick when the
4212  * current task is from our class and nr_running is low enough
4213  * to matter.
4214  */
4215 static void hrtick_update(struct rq *rq)
4216 {
4217 	struct task_struct *curr = rq->curr;
4218 
4219 	if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4220 		return;
4221 
4222 	if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4223 		hrtick_start_fair(rq, curr);
4224 }
4225 #else /* !CONFIG_SCHED_HRTICK */
4226 static inline void
4227 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4228 {
4229 }
4230 
4231 static inline void hrtick_update(struct rq *rq)
4232 {
4233 }
4234 #endif
4235 
4236 /*
4237  * The enqueue_task method is called before nr_running is
4238  * increased. Here we update the fair scheduling stats and
4239  * then put the task into the rbtree:
4240  */
4241 static void
4242 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4243 {
4244 	struct cfs_rq *cfs_rq;
4245 	struct sched_entity *se = &p->se;
4246 
4247 	for_each_sched_entity(se) {
4248 		if (se->on_rq)
4249 			break;
4250 		cfs_rq = cfs_rq_of(se);
4251 		enqueue_entity(cfs_rq, se, flags);
4252 
4253 		/*
4254 		 * end evaluation on encountering a throttled cfs_rq
4255 		 *
4256 		 * note: in the case of encountering a throttled cfs_rq we will
4257 		 * post the final h_nr_running increment below.
4258 		*/
4259 		if (cfs_rq_throttled(cfs_rq))
4260 			break;
4261 		cfs_rq->h_nr_running++;
4262 
4263 		flags = ENQUEUE_WAKEUP;
4264 	}
4265 
4266 	for_each_sched_entity(se) {
4267 		cfs_rq = cfs_rq_of(se);
4268 		cfs_rq->h_nr_running++;
4269 
4270 		if (cfs_rq_throttled(cfs_rq))
4271 			break;
4272 
4273 		update_load_avg(se, 1);
4274 		update_cfs_shares(cfs_rq);
4275 	}
4276 
4277 	if (!se)
4278 		add_nr_running(rq, 1);
4279 
4280 	hrtick_update(rq);
4281 }
4282 
4283 static void set_next_buddy(struct sched_entity *se);
4284 
4285 /*
4286  * The dequeue_task method is called before nr_running is
4287  * decreased. We remove the task from the rbtree and
4288  * update the fair scheduling stats:
4289  */
4290 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4291 {
4292 	struct cfs_rq *cfs_rq;
4293 	struct sched_entity *se = &p->se;
4294 	int task_sleep = flags & DEQUEUE_SLEEP;
4295 
4296 	for_each_sched_entity(se) {
4297 		cfs_rq = cfs_rq_of(se);
4298 		dequeue_entity(cfs_rq, se, flags);
4299 
4300 		/*
4301 		 * end evaluation on encountering a throttled cfs_rq
4302 		 *
4303 		 * note: in the case of encountering a throttled cfs_rq we will
4304 		 * post the final h_nr_running decrement below.
4305 		*/
4306 		if (cfs_rq_throttled(cfs_rq))
4307 			break;
4308 		cfs_rq->h_nr_running--;
4309 
4310 		/* Don't dequeue parent if it has other entities besides us */
4311 		if (cfs_rq->load.weight) {
4312 			/*
4313 			 * Bias pick_next to pick a task from this cfs_rq, as
4314 			 * p is sleeping when it is within its sched_slice.
4315 			 */
4316 			if (task_sleep && parent_entity(se))
4317 				set_next_buddy(parent_entity(se));
4318 
4319 			/* avoid re-evaluating load for this entity */
4320 			se = parent_entity(se);
4321 			break;
4322 		}
4323 		flags |= DEQUEUE_SLEEP;
4324 	}
4325 
4326 	for_each_sched_entity(se) {
4327 		cfs_rq = cfs_rq_of(se);
4328 		cfs_rq->h_nr_running--;
4329 
4330 		if (cfs_rq_throttled(cfs_rq))
4331 			break;
4332 
4333 		update_load_avg(se, 1);
4334 		update_cfs_shares(cfs_rq);
4335 	}
4336 
4337 	if (!se)
4338 		sub_nr_running(rq, 1);
4339 
4340 	hrtick_update(rq);
4341 }
4342 
4343 #ifdef CONFIG_SMP
4344 
4345 /*
4346  * per rq 'load' arrray crap; XXX kill this.
4347  */
4348 
4349 /*
4350  * The exact cpuload calculated at every tick would be:
4351  *
4352  *   load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4353  *
4354  * If a cpu misses updates for n ticks (as it was idle) and update gets
4355  * called on the n+1-th tick when cpu may be busy, then we have:
4356  *
4357  *   load_n   = (1 - 1/2^i)^n * load_0
4358  *   load_n+1 = (1 - 1/2^i)   * load_n + (1/2^i) * cur_load
4359  *
4360  * decay_load_missed() below does efficient calculation of
4361  *
4362  *   load' = (1 - 1/2^i)^n * load
4363  *
4364  * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4365  * This allows us to precompute the above in said factors, thereby allowing the
4366  * reduction of an arbitrary n in O(log_2 n) steps. (See also
4367  * fixed_power_int())
4368  *
4369  * The calculation is approximated on a 128 point scale.
4370  */
4371 #define DEGRADE_SHIFT		7
4372 
4373 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4374 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4375 	{   0,   0,  0,  0,  0,  0, 0, 0 },
4376 	{  64,  32,  8,  0,  0,  0, 0, 0 },
4377 	{  96,  72, 40, 12,  1,  0, 0, 0 },
4378 	{ 112,  98, 75, 43, 15,  1, 0, 0 },
4379 	{ 120, 112, 98, 76, 45, 16, 2, 0 }
4380 };
4381 
4382 /*
4383  * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4384  * would be when CPU is idle and so we just decay the old load without
4385  * adding any new load.
4386  */
4387 static unsigned long
4388 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4389 {
4390 	int j = 0;
4391 
4392 	if (!missed_updates)
4393 		return load;
4394 
4395 	if (missed_updates >= degrade_zero_ticks[idx])
4396 		return 0;
4397 
4398 	if (idx == 1)
4399 		return load >> missed_updates;
4400 
4401 	while (missed_updates) {
4402 		if (missed_updates % 2)
4403 			load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4404 
4405 		missed_updates >>= 1;
4406 		j++;
4407 	}
4408 	return load;
4409 }
4410 
4411 /**
4412  * __update_cpu_load - update the rq->cpu_load[] statistics
4413  * @this_rq: The rq to update statistics for
4414  * @this_load: The current load
4415  * @pending_updates: The number of missed updates
4416  * @active: !0 for NOHZ_FULL
4417  *
4418  * Update rq->cpu_load[] statistics. This function is usually called every
4419  * scheduler tick (TICK_NSEC).
4420  *
4421  * This function computes a decaying average:
4422  *
4423  *   load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4424  *
4425  * Because of NOHZ it might not get called on every tick which gives need for
4426  * the @pending_updates argument.
4427  *
4428  *   load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4429  *             = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4430  *             = A * (A * load[i]_n-2 + B) + B
4431  *             = A * (A * (A * load[i]_n-3 + B) + B) + B
4432  *             = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4433  *             = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4434  *             = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4435  *             = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4436  *
4437  * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4438  * any change in load would have resulted in the tick being turned back on.
4439  *
4440  * For regular NOHZ, this reduces to:
4441  *
4442  *   load[i]_n = (1 - 1/2^i)^n * load[i]_0
4443  *
4444  * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4445  * term. See the @active paramter.
4446  */
4447 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4448 			      unsigned long pending_updates, int active)
4449 {
4450 	unsigned long tickless_load = active ? this_rq->cpu_load[0] : 0;
4451 	int i, scale;
4452 
4453 	this_rq->nr_load_updates++;
4454 
4455 	/* Update our load: */
4456 	this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4457 	for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4458 		unsigned long old_load, new_load;
4459 
4460 		/* scale is effectively 1 << i now, and >> i divides by scale */
4461 
4462 		old_load = this_rq->cpu_load[i] - tickless_load;
4463 		old_load = decay_load_missed(old_load, pending_updates - 1, i);
4464 		old_load += tickless_load;
4465 		new_load = this_load;
4466 		/*
4467 		 * Round up the averaging division if load is increasing. This
4468 		 * prevents us from getting stuck on 9 if the load is 10, for
4469 		 * example.
4470 		 */
4471 		if (new_load > old_load)
4472 			new_load += scale - 1;
4473 
4474 		this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4475 	}
4476 
4477 	sched_avg_update(this_rq);
4478 }
4479 
4480 /* Used instead of source_load when we know the type == 0 */
4481 static unsigned long weighted_cpuload(const int cpu)
4482 {
4483 	return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4484 }
4485 
4486 #ifdef CONFIG_NO_HZ_COMMON
4487 /*
4488  * There is no sane way to deal with nohz on smp when using jiffies because the
4489  * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4490  * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4491  *
4492  * Therefore we cannot use the delta approach from the regular tick since that
4493  * would seriously skew the load calculation. However we'll make do for those
4494  * updates happening while idle (nohz_idle_balance) or coming out of idle
4495  * (tick_nohz_idle_exit).
4496  *
4497  * This means we might still be one tick off for nohz periods.
4498  */
4499 
4500 /*
4501  * Called from nohz_idle_balance() to update the load ratings before doing the
4502  * idle balance.
4503  */
4504 static void update_idle_cpu_load(struct rq *this_rq)
4505 {
4506 	unsigned long curr_jiffies = READ_ONCE(jiffies);
4507 	unsigned long load = weighted_cpuload(cpu_of(this_rq));
4508 	unsigned long pending_updates;
4509 
4510 	/*
4511 	 * bail if there's load or we're actually up-to-date.
4512 	 */
4513 	if (load || curr_jiffies == this_rq->last_load_update_tick)
4514 		return;
4515 
4516 	pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4517 	this_rq->last_load_update_tick = curr_jiffies;
4518 
4519 	__update_cpu_load(this_rq, load, pending_updates, 0);
4520 }
4521 
4522 /*
4523  * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4524  */
4525 void update_cpu_load_nohz(int active)
4526 {
4527 	struct rq *this_rq = this_rq();
4528 	unsigned long curr_jiffies = READ_ONCE(jiffies);
4529 	unsigned long load = active ? weighted_cpuload(cpu_of(this_rq)) : 0;
4530 	unsigned long pending_updates;
4531 
4532 	if (curr_jiffies == this_rq->last_load_update_tick)
4533 		return;
4534 
4535 	raw_spin_lock(&this_rq->lock);
4536 	pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4537 	if (pending_updates) {
4538 		this_rq->last_load_update_tick = curr_jiffies;
4539 		/*
4540 		 * In the regular NOHZ case, we were idle, this means load 0.
4541 		 * In the NOHZ_FULL case, we were non-idle, we should consider
4542 		 * its weighted load.
4543 		 */
4544 		__update_cpu_load(this_rq, load, pending_updates, active);
4545 	}
4546 	raw_spin_unlock(&this_rq->lock);
4547 }
4548 #endif /* CONFIG_NO_HZ */
4549 
4550 /*
4551  * Called from scheduler_tick()
4552  */
4553 void update_cpu_load_active(struct rq *this_rq)
4554 {
4555 	unsigned long load = weighted_cpuload(cpu_of(this_rq));
4556 	/*
4557 	 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4558 	 */
4559 	this_rq->last_load_update_tick = jiffies;
4560 	__update_cpu_load(this_rq, load, 1, 1);
4561 }
4562 
4563 /*
4564  * Return a low guess at the load of a migration-source cpu weighted
4565  * according to the scheduling class and "nice" value.
4566  *
4567  * We want to under-estimate the load of migration sources, to
4568  * balance conservatively.
4569  */
4570 static unsigned long source_load(int cpu, int type)
4571 {
4572 	struct rq *rq = cpu_rq(cpu);
4573 	unsigned long total = weighted_cpuload(cpu);
4574 
4575 	if (type == 0 || !sched_feat(LB_BIAS))
4576 		return total;
4577 
4578 	return min(rq->cpu_load[type-1], total);
4579 }
4580 
4581 /*
4582  * Return a high guess at the load of a migration-target cpu weighted
4583  * according to the scheduling class and "nice" value.
4584  */
4585 static unsigned long target_load(int cpu, int type)
4586 {
4587 	struct rq *rq = cpu_rq(cpu);
4588 	unsigned long total = weighted_cpuload(cpu);
4589 
4590 	if (type == 0 || !sched_feat(LB_BIAS))
4591 		return total;
4592 
4593 	return max(rq->cpu_load[type-1], total);
4594 }
4595 
4596 static unsigned long capacity_of(int cpu)
4597 {
4598 	return cpu_rq(cpu)->cpu_capacity;
4599 }
4600 
4601 static unsigned long capacity_orig_of(int cpu)
4602 {
4603 	return cpu_rq(cpu)->cpu_capacity_orig;
4604 }
4605 
4606 static unsigned long cpu_avg_load_per_task(int cpu)
4607 {
4608 	struct rq *rq = cpu_rq(cpu);
4609 	unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4610 	unsigned long load_avg = weighted_cpuload(cpu);
4611 
4612 	if (nr_running)
4613 		return load_avg / nr_running;
4614 
4615 	return 0;
4616 }
4617 
4618 static void record_wakee(struct task_struct *p)
4619 {
4620 	/*
4621 	 * Rough decay (wiping) for cost saving, don't worry
4622 	 * about the boundary, really active task won't care
4623 	 * about the loss.
4624 	 */
4625 	if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4626 		current->wakee_flips >>= 1;
4627 		current->wakee_flip_decay_ts = jiffies;
4628 	}
4629 
4630 	if (current->last_wakee != p) {
4631 		current->last_wakee = p;
4632 		current->wakee_flips++;
4633 	}
4634 }
4635 
4636 static void task_waking_fair(struct task_struct *p)
4637 {
4638 	struct sched_entity *se = &p->se;
4639 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4640 	u64 min_vruntime;
4641 
4642 #ifndef CONFIG_64BIT
4643 	u64 min_vruntime_copy;
4644 
4645 	do {
4646 		min_vruntime_copy = cfs_rq->min_vruntime_copy;
4647 		smp_rmb();
4648 		min_vruntime = cfs_rq->min_vruntime;
4649 	} while (min_vruntime != min_vruntime_copy);
4650 #else
4651 	min_vruntime = cfs_rq->min_vruntime;
4652 #endif
4653 
4654 	se->vruntime -= min_vruntime;
4655 	record_wakee(p);
4656 }
4657 
4658 #ifdef CONFIG_FAIR_GROUP_SCHED
4659 /*
4660  * effective_load() calculates the load change as seen from the root_task_group
4661  *
4662  * Adding load to a group doesn't make a group heavier, but can cause movement
4663  * of group shares between cpus. Assuming the shares were perfectly aligned one
4664  * can calculate the shift in shares.
4665  *
4666  * Calculate the effective load difference if @wl is added (subtracted) to @tg
4667  * on this @cpu and results in a total addition (subtraction) of @wg to the
4668  * total group weight.
4669  *
4670  * Given a runqueue weight distribution (rw_i) we can compute a shares
4671  * distribution (s_i) using:
4672  *
4673  *   s_i = rw_i / \Sum rw_j						(1)
4674  *
4675  * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4676  * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4677  * shares distribution (s_i):
4678  *
4679  *   rw_i = {   2,   4,   1,   0 }
4680  *   s_i  = { 2/7, 4/7, 1/7,   0 }
4681  *
4682  * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4683  * task used to run on and the CPU the waker is running on), we need to
4684  * compute the effect of waking a task on either CPU and, in case of a sync
4685  * wakeup, compute the effect of the current task going to sleep.
4686  *
4687  * So for a change of @wl to the local @cpu with an overall group weight change
4688  * of @wl we can compute the new shares distribution (s'_i) using:
4689  *
4690  *   s'_i = (rw_i + @wl) / (@wg + \Sum rw_j)				(2)
4691  *
4692  * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4693  * differences in waking a task to CPU 0. The additional task changes the
4694  * weight and shares distributions like:
4695  *
4696  *   rw'_i = {   3,   4,   1,   0 }
4697  *   s'_i  = { 3/8, 4/8, 1/8,   0 }
4698  *
4699  * We can then compute the difference in effective weight by using:
4700  *
4701  *   dw_i = S * (s'_i - s_i)						(3)
4702  *
4703  * Where 'S' is the group weight as seen by its parent.
4704  *
4705  * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4706  * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4707  * 4/7) times the weight of the group.
4708  */
4709 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4710 {
4711 	struct sched_entity *se = tg->se[cpu];
4712 
4713 	if (!tg->parent)	/* the trivial, non-cgroup case */
4714 		return wl;
4715 
4716 	for_each_sched_entity(se) {
4717 		long w, W;
4718 
4719 		tg = se->my_q->tg;
4720 
4721 		/*
4722 		 * W = @wg + \Sum rw_j
4723 		 */
4724 		W = wg + calc_tg_weight(tg, se->my_q);
4725 
4726 		/*
4727 		 * w = rw_i + @wl
4728 		 */
4729 		w = cfs_rq_load_avg(se->my_q) + wl;
4730 
4731 		/*
4732 		 * wl = S * s'_i; see (2)
4733 		 */
4734 		if (W > 0 && w < W)
4735 			wl = (w * (long)tg->shares) / W;
4736 		else
4737 			wl = tg->shares;
4738 
4739 		/*
4740 		 * Per the above, wl is the new se->load.weight value; since
4741 		 * those are clipped to [MIN_SHARES, ...) do so now. See
4742 		 * calc_cfs_shares().
4743 		 */
4744 		if (wl < MIN_SHARES)
4745 			wl = MIN_SHARES;
4746 
4747 		/*
4748 		 * wl = dw_i = S * (s'_i - s_i); see (3)
4749 		 */
4750 		wl -= se->avg.load_avg;
4751 
4752 		/*
4753 		 * Recursively apply this logic to all parent groups to compute
4754 		 * the final effective load change on the root group. Since
4755 		 * only the @tg group gets extra weight, all parent groups can
4756 		 * only redistribute existing shares. @wl is the shift in shares
4757 		 * resulting from this level per the above.
4758 		 */
4759 		wg = 0;
4760 	}
4761 
4762 	return wl;
4763 }
4764 #else
4765 
4766 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4767 {
4768 	return wl;
4769 }
4770 
4771 #endif
4772 
4773 /*
4774  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4775  * A waker of many should wake a different task than the one last awakened
4776  * at a frequency roughly N times higher than one of its wakees.  In order
4777  * to determine whether we should let the load spread vs consolodating to
4778  * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4779  * partner, and a factor of lls_size higher frequency in the other.  With
4780  * both conditions met, we can be relatively sure that the relationship is
4781  * non-monogamous, with partner count exceeding socket size.  Waker/wakee
4782  * being client/server, worker/dispatcher, interrupt source or whatever is
4783  * irrelevant, spread criteria is apparent partner count exceeds socket size.
4784  */
4785 static int wake_wide(struct task_struct *p)
4786 {
4787 	unsigned int master = current->wakee_flips;
4788 	unsigned int slave = p->wakee_flips;
4789 	int factor = this_cpu_read(sd_llc_size);
4790 
4791 	if (master < slave)
4792 		swap(master, slave);
4793 	if (slave < factor || master < slave * factor)
4794 		return 0;
4795 	return 1;
4796 }
4797 
4798 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4799 {
4800 	s64 this_load, load;
4801 	s64 this_eff_load, prev_eff_load;
4802 	int idx, this_cpu, prev_cpu;
4803 	struct task_group *tg;
4804 	unsigned long weight;
4805 	int balanced;
4806 
4807 	idx	  = sd->wake_idx;
4808 	this_cpu  = smp_processor_id();
4809 	prev_cpu  = task_cpu(p);
4810 	load	  = source_load(prev_cpu, idx);
4811 	this_load = target_load(this_cpu, idx);
4812 
4813 	/*
4814 	 * If sync wakeup then subtract the (maximum possible)
4815 	 * effect of the currently running task from the load
4816 	 * of the current CPU:
4817 	 */
4818 	if (sync) {
4819 		tg = task_group(current);
4820 		weight = current->se.avg.load_avg;
4821 
4822 		this_load += effective_load(tg, this_cpu, -weight, -weight);
4823 		load += effective_load(tg, prev_cpu, 0, -weight);
4824 	}
4825 
4826 	tg = task_group(p);
4827 	weight = p->se.avg.load_avg;
4828 
4829 	/*
4830 	 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4831 	 * due to the sync cause above having dropped this_load to 0, we'll
4832 	 * always have an imbalance, but there's really nothing you can do
4833 	 * about that, so that's good too.
4834 	 *
4835 	 * Otherwise check if either cpus are near enough in load to allow this
4836 	 * task to be woken on this_cpu.
4837 	 */
4838 	this_eff_load = 100;
4839 	this_eff_load *= capacity_of(prev_cpu);
4840 
4841 	prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4842 	prev_eff_load *= capacity_of(this_cpu);
4843 
4844 	if (this_load > 0) {
4845 		this_eff_load *= this_load +
4846 			effective_load(tg, this_cpu, weight, weight);
4847 
4848 		prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4849 	}
4850 
4851 	balanced = this_eff_load <= prev_eff_load;
4852 
4853 	schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4854 
4855 	if (!balanced)
4856 		return 0;
4857 
4858 	schedstat_inc(sd, ttwu_move_affine);
4859 	schedstat_inc(p, se.statistics.nr_wakeups_affine);
4860 
4861 	return 1;
4862 }
4863 
4864 /*
4865  * find_idlest_group finds and returns the least busy CPU group within the
4866  * domain.
4867  */
4868 static struct sched_group *
4869 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4870 		  int this_cpu, int sd_flag)
4871 {
4872 	struct sched_group *idlest = NULL, *group = sd->groups;
4873 	unsigned long min_load = ULONG_MAX, this_load = 0;
4874 	int load_idx = sd->forkexec_idx;
4875 	int imbalance = 100 + (sd->imbalance_pct-100)/2;
4876 
4877 	if (sd_flag & SD_BALANCE_WAKE)
4878 		load_idx = sd->wake_idx;
4879 
4880 	do {
4881 		unsigned long load, avg_load;
4882 		int local_group;
4883 		int i;
4884 
4885 		/* Skip over this group if it has no CPUs allowed */
4886 		if (!cpumask_intersects(sched_group_cpus(group),
4887 					tsk_cpus_allowed(p)))
4888 			continue;
4889 
4890 		local_group = cpumask_test_cpu(this_cpu,
4891 					       sched_group_cpus(group));
4892 
4893 		/* Tally up the load of all CPUs in the group */
4894 		avg_load = 0;
4895 
4896 		for_each_cpu(i, sched_group_cpus(group)) {
4897 			/* Bias balancing toward cpus of our domain */
4898 			if (local_group)
4899 				load = source_load(i, load_idx);
4900 			else
4901 				load = target_load(i, load_idx);
4902 
4903 			avg_load += load;
4904 		}
4905 
4906 		/* Adjust by relative CPU capacity of the group */
4907 		avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4908 
4909 		if (local_group) {
4910 			this_load = avg_load;
4911 		} else if (avg_load < min_load) {
4912 			min_load = avg_load;
4913 			idlest = group;
4914 		}
4915 	} while (group = group->next, group != sd->groups);
4916 
4917 	if (!idlest || 100*this_load < imbalance*min_load)
4918 		return NULL;
4919 	return idlest;
4920 }
4921 
4922 /*
4923  * find_idlest_cpu - find the idlest cpu among the cpus in group.
4924  */
4925 static int
4926 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4927 {
4928 	unsigned long load, min_load = ULONG_MAX;
4929 	unsigned int min_exit_latency = UINT_MAX;
4930 	u64 latest_idle_timestamp = 0;
4931 	int least_loaded_cpu = this_cpu;
4932 	int shallowest_idle_cpu = -1;
4933 	int i;
4934 
4935 	/* Traverse only the allowed CPUs */
4936 	for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4937 		if (idle_cpu(i)) {
4938 			struct rq *rq = cpu_rq(i);
4939 			struct cpuidle_state *idle = idle_get_state(rq);
4940 			if (idle && idle->exit_latency < min_exit_latency) {
4941 				/*
4942 				 * We give priority to a CPU whose idle state
4943 				 * has the smallest exit latency irrespective
4944 				 * of any idle timestamp.
4945 				 */
4946 				min_exit_latency = idle->exit_latency;
4947 				latest_idle_timestamp = rq->idle_stamp;
4948 				shallowest_idle_cpu = i;
4949 			} else if ((!idle || idle->exit_latency == min_exit_latency) &&
4950 				   rq->idle_stamp > latest_idle_timestamp) {
4951 				/*
4952 				 * If equal or no active idle state, then
4953 				 * the most recently idled CPU might have
4954 				 * a warmer cache.
4955 				 */
4956 				latest_idle_timestamp = rq->idle_stamp;
4957 				shallowest_idle_cpu = i;
4958 			}
4959 		} else if (shallowest_idle_cpu == -1) {
4960 			load = weighted_cpuload(i);
4961 			if (load < min_load || (load == min_load && i == this_cpu)) {
4962 				min_load = load;
4963 				least_loaded_cpu = i;
4964 			}
4965 		}
4966 	}
4967 
4968 	return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4969 }
4970 
4971 /*
4972  * Try and locate an idle CPU in the sched_domain.
4973  */
4974 static int select_idle_sibling(struct task_struct *p, int target)
4975 {
4976 	struct sched_domain *sd;
4977 	struct sched_group *sg;
4978 	int i = task_cpu(p);
4979 
4980 	if (idle_cpu(target))
4981 		return target;
4982 
4983 	/*
4984 	 * If the prevous cpu is cache affine and idle, don't be stupid.
4985 	 */
4986 	if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4987 		return i;
4988 
4989 	/*
4990 	 * Otherwise, iterate the domains and find an elegible idle cpu.
4991 	 */
4992 	sd = rcu_dereference(per_cpu(sd_llc, target));
4993 	for_each_lower_domain(sd) {
4994 		sg = sd->groups;
4995 		do {
4996 			if (!cpumask_intersects(sched_group_cpus(sg),
4997 						tsk_cpus_allowed(p)))
4998 				goto next;
4999 
5000 			for_each_cpu(i, sched_group_cpus(sg)) {
5001 				if (i == target || !idle_cpu(i))
5002 					goto next;
5003 			}
5004 
5005 			target = cpumask_first_and(sched_group_cpus(sg),
5006 					tsk_cpus_allowed(p));
5007 			goto done;
5008 next:
5009 			sg = sg->next;
5010 		} while (sg != sd->groups);
5011 	}
5012 done:
5013 	return target;
5014 }
5015 
5016 /*
5017  * cpu_util returns the amount of capacity of a CPU that is used by CFS
5018  * tasks. The unit of the return value must be the one of capacity so we can
5019  * compare the utilization with the capacity of the CPU that is available for
5020  * CFS task (ie cpu_capacity).
5021  *
5022  * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5023  * recent utilization of currently non-runnable tasks on a CPU. It represents
5024  * the amount of utilization of a CPU in the range [0..capacity_orig] where
5025  * capacity_orig is the cpu_capacity available at the highest frequency
5026  * (arch_scale_freq_capacity()).
5027  * The utilization of a CPU converges towards a sum equal to or less than the
5028  * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5029  * the running time on this CPU scaled by capacity_curr.
5030  *
5031  * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5032  * higher than capacity_orig because of unfortunate rounding in
5033  * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5034  * the average stabilizes with the new running time. We need to check that the
5035  * utilization stays within the range of [0..capacity_orig] and cap it if
5036  * necessary. Without utilization capping, a group could be seen as overloaded
5037  * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5038  * available capacity. We allow utilization to overshoot capacity_curr (but not
5039  * capacity_orig) as it useful for predicting the capacity required after task
5040  * migrations (scheduler-driven DVFS).
5041  */
5042 static int cpu_util(int cpu)
5043 {
5044 	unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5045 	unsigned long capacity = capacity_orig_of(cpu);
5046 
5047 	return (util >= capacity) ? capacity : util;
5048 }
5049 
5050 /*
5051  * select_task_rq_fair: Select target runqueue for the waking task in domains
5052  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5053  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5054  *
5055  * Balances load by selecting the idlest cpu in the idlest group, or under
5056  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5057  *
5058  * Returns the target cpu number.
5059  *
5060  * preempt must be disabled.
5061  */
5062 static int
5063 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5064 {
5065 	struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5066 	int cpu = smp_processor_id();
5067 	int new_cpu = prev_cpu;
5068 	int want_affine = 0;
5069 	int sync = wake_flags & WF_SYNC;
5070 
5071 	if (sd_flag & SD_BALANCE_WAKE)
5072 		want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5073 
5074 	rcu_read_lock();
5075 	for_each_domain(cpu, tmp) {
5076 		if (!(tmp->flags & SD_LOAD_BALANCE))
5077 			break;
5078 
5079 		/*
5080 		 * If both cpu and prev_cpu are part of this domain,
5081 		 * cpu is a valid SD_WAKE_AFFINE target.
5082 		 */
5083 		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5084 		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5085 			affine_sd = tmp;
5086 			break;
5087 		}
5088 
5089 		if (tmp->flags & sd_flag)
5090 			sd = tmp;
5091 		else if (!want_affine)
5092 			break;
5093 	}
5094 
5095 	if (affine_sd) {
5096 		sd = NULL; /* Prefer wake_affine over balance flags */
5097 		if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5098 			new_cpu = cpu;
5099 	}
5100 
5101 	if (!sd) {
5102 		if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5103 			new_cpu = select_idle_sibling(p, new_cpu);
5104 
5105 	} else while (sd) {
5106 		struct sched_group *group;
5107 		int weight;
5108 
5109 		if (!(sd->flags & sd_flag)) {
5110 			sd = sd->child;
5111 			continue;
5112 		}
5113 
5114 		group = find_idlest_group(sd, p, cpu, sd_flag);
5115 		if (!group) {
5116 			sd = sd->child;
5117 			continue;
5118 		}
5119 
5120 		new_cpu = find_idlest_cpu(group, p, cpu);
5121 		if (new_cpu == -1 || new_cpu == cpu) {
5122 			/* Now try balancing at a lower domain level of cpu */
5123 			sd = sd->child;
5124 			continue;
5125 		}
5126 
5127 		/* Now try balancing at a lower domain level of new_cpu */
5128 		cpu = new_cpu;
5129 		weight = sd->span_weight;
5130 		sd = NULL;
5131 		for_each_domain(cpu, tmp) {
5132 			if (weight <= tmp->span_weight)
5133 				break;
5134 			if (tmp->flags & sd_flag)
5135 				sd = tmp;
5136 		}
5137 		/* while loop will break here if sd == NULL */
5138 	}
5139 	rcu_read_unlock();
5140 
5141 	return new_cpu;
5142 }
5143 
5144 /*
5145  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5146  * cfs_rq_of(p) references at time of call are still valid and identify the
5147  * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5148  */
5149 static void migrate_task_rq_fair(struct task_struct *p)
5150 {
5151 	/*
5152 	 * We are supposed to update the task to "current" time, then its up to date
5153 	 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5154 	 * what current time is, so simply throw away the out-of-date time. This
5155 	 * will result in the wakee task is less decayed, but giving the wakee more
5156 	 * load sounds not bad.
5157 	 */
5158 	remove_entity_load_avg(&p->se);
5159 
5160 	/* Tell new CPU we are migrated */
5161 	p->se.avg.last_update_time = 0;
5162 
5163 	/* We have migrated, no longer consider this task hot */
5164 	p->se.exec_start = 0;
5165 }
5166 
5167 static void task_dead_fair(struct task_struct *p)
5168 {
5169 	remove_entity_load_avg(&p->se);
5170 }
5171 #endif /* CONFIG_SMP */
5172 
5173 static unsigned long
5174 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5175 {
5176 	unsigned long gran = sysctl_sched_wakeup_granularity;
5177 
5178 	/*
5179 	 * Since its curr running now, convert the gran from real-time
5180 	 * to virtual-time in his units.
5181 	 *
5182 	 * By using 'se' instead of 'curr' we penalize light tasks, so
5183 	 * they get preempted easier. That is, if 'se' < 'curr' then
5184 	 * the resulting gran will be larger, therefore penalizing the
5185 	 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5186 	 * be smaller, again penalizing the lighter task.
5187 	 *
5188 	 * This is especially important for buddies when the leftmost
5189 	 * task is higher priority than the buddy.
5190 	 */
5191 	return calc_delta_fair(gran, se);
5192 }
5193 
5194 /*
5195  * Should 'se' preempt 'curr'.
5196  *
5197  *             |s1
5198  *        |s2
5199  *   |s3
5200  *         g
5201  *      |<--->|c
5202  *
5203  *  w(c, s1) = -1
5204  *  w(c, s2) =  0
5205  *  w(c, s3) =  1
5206  *
5207  */
5208 static int
5209 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5210 {
5211 	s64 gran, vdiff = curr->vruntime - se->vruntime;
5212 
5213 	if (vdiff <= 0)
5214 		return -1;
5215 
5216 	gran = wakeup_gran(curr, se);
5217 	if (vdiff > gran)
5218 		return 1;
5219 
5220 	return 0;
5221 }
5222 
5223 static void set_last_buddy(struct sched_entity *se)
5224 {
5225 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5226 		return;
5227 
5228 	for_each_sched_entity(se)
5229 		cfs_rq_of(se)->last = se;
5230 }
5231 
5232 static void set_next_buddy(struct sched_entity *se)
5233 {
5234 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5235 		return;
5236 
5237 	for_each_sched_entity(se)
5238 		cfs_rq_of(se)->next = se;
5239 }
5240 
5241 static void set_skip_buddy(struct sched_entity *se)
5242 {
5243 	for_each_sched_entity(se)
5244 		cfs_rq_of(se)->skip = se;
5245 }
5246 
5247 /*
5248  * Preempt the current task with a newly woken task if needed:
5249  */
5250 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5251 {
5252 	struct task_struct *curr = rq->curr;
5253 	struct sched_entity *se = &curr->se, *pse = &p->se;
5254 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5255 	int scale = cfs_rq->nr_running >= sched_nr_latency;
5256 	int next_buddy_marked = 0;
5257 
5258 	if (unlikely(se == pse))
5259 		return;
5260 
5261 	/*
5262 	 * This is possible from callers such as attach_tasks(), in which we
5263 	 * unconditionally check_prempt_curr() after an enqueue (which may have
5264 	 * lead to a throttle).  This both saves work and prevents false
5265 	 * next-buddy nomination below.
5266 	 */
5267 	if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5268 		return;
5269 
5270 	if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5271 		set_next_buddy(pse);
5272 		next_buddy_marked = 1;
5273 	}
5274 
5275 	/*
5276 	 * We can come here with TIF_NEED_RESCHED already set from new task
5277 	 * wake up path.
5278 	 *
5279 	 * Note: this also catches the edge-case of curr being in a throttled
5280 	 * group (e.g. via set_curr_task), since update_curr() (in the
5281 	 * enqueue of curr) will have resulted in resched being set.  This
5282 	 * prevents us from potentially nominating it as a false LAST_BUDDY
5283 	 * below.
5284 	 */
5285 	if (test_tsk_need_resched(curr))
5286 		return;
5287 
5288 	/* Idle tasks are by definition preempted by non-idle tasks. */
5289 	if (unlikely(curr->policy == SCHED_IDLE) &&
5290 	    likely(p->policy != SCHED_IDLE))
5291 		goto preempt;
5292 
5293 	/*
5294 	 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5295 	 * is driven by the tick):
5296 	 */
5297 	if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5298 		return;
5299 
5300 	find_matching_se(&se, &pse);
5301 	update_curr(cfs_rq_of(se));
5302 	BUG_ON(!pse);
5303 	if (wakeup_preempt_entity(se, pse) == 1) {
5304 		/*
5305 		 * Bias pick_next to pick the sched entity that is
5306 		 * triggering this preemption.
5307 		 */
5308 		if (!next_buddy_marked)
5309 			set_next_buddy(pse);
5310 		goto preempt;
5311 	}
5312 
5313 	return;
5314 
5315 preempt:
5316 	resched_curr(rq);
5317 	/*
5318 	 * Only set the backward buddy when the current task is still
5319 	 * on the rq. This can happen when a wakeup gets interleaved
5320 	 * with schedule on the ->pre_schedule() or idle_balance()
5321 	 * point, either of which can * drop the rq lock.
5322 	 *
5323 	 * Also, during early boot the idle thread is in the fair class,
5324 	 * for obvious reasons its a bad idea to schedule back to it.
5325 	 */
5326 	if (unlikely(!se->on_rq || curr == rq->idle))
5327 		return;
5328 
5329 	if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5330 		set_last_buddy(se);
5331 }
5332 
5333 static struct task_struct *
5334 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5335 {
5336 	struct cfs_rq *cfs_rq = &rq->cfs;
5337 	struct sched_entity *se;
5338 	struct task_struct *p;
5339 	int new_tasks;
5340 
5341 again:
5342 #ifdef CONFIG_FAIR_GROUP_SCHED
5343 	if (!cfs_rq->nr_running)
5344 		goto idle;
5345 
5346 	if (prev->sched_class != &fair_sched_class)
5347 		goto simple;
5348 
5349 	/*
5350 	 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5351 	 * likely that a next task is from the same cgroup as the current.
5352 	 *
5353 	 * Therefore attempt to avoid putting and setting the entire cgroup
5354 	 * hierarchy, only change the part that actually changes.
5355 	 */
5356 
5357 	do {
5358 		struct sched_entity *curr = cfs_rq->curr;
5359 
5360 		/*
5361 		 * Since we got here without doing put_prev_entity() we also
5362 		 * have to consider cfs_rq->curr. If it is still a runnable
5363 		 * entity, update_curr() will update its vruntime, otherwise
5364 		 * forget we've ever seen it.
5365 		 */
5366 		if (curr) {
5367 			if (curr->on_rq)
5368 				update_curr(cfs_rq);
5369 			else
5370 				curr = NULL;
5371 
5372 			/*
5373 			 * This call to check_cfs_rq_runtime() will do the
5374 			 * throttle and dequeue its entity in the parent(s).
5375 			 * Therefore the 'simple' nr_running test will indeed
5376 			 * be correct.
5377 			 */
5378 			if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5379 				goto simple;
5380 		}
5381 
5382 		se = pick_next_entity(cfs_rq, curr);
5383 		cfs_rq = group_cfs_rq(se);
5384 	} while (cfs_rq);
5385 
5386 	p = task_of(se);
5387 
5388 	/*
5389 	 * Since we haven't yet done put_prev_entity and if the selected task
5390 	 * is a different task than we started out with, try and touch the
5391 	 * least amount of cfs_rqs.
5392 	 */
5393 	if (prev != p) {
5394 		struct sched_entity *pse = &prev->se;
5395 
5396 		while (!(cfs_rq = is_same_group(se, pse))) {
5397 			int se_depth = se->depth;
5398 			int pse_depth = pse->depth;
5399 
5400 			if (se_depth <= pse_depth) {
5401 				put_prev_entity(cfs_rq_of(pse), pse);
5402 				pse = parent_entity(pse);
5403 			}
5404 			if (se_depth >= pse_depth) {
5405 				set_next_entity(cfs_rq_of(se), se);
5406 				se = parent_entity(se);
5407 			}
5408 		}
5409 
5410 		put_prev_entity(cfs_rq, pse);
5411 		set_next_entity(cfs_rq, se);
5412 	}
5413 
5414 	if (hrtick_enabled(rq))
5415 		hrtick_start_fair(rq, p);
5416 
5417 	return p;
5418 simple:
5419 	cfs_rq = &rq->cfs;
5420 #endif
5421 
5422 	if (!cfs_rq->nr_running)
5423 		goto idle;
5424 
5425 	put_prev_task(rq, prev);
5426 
5427 	do {
5428 		se = pick_next_entity(cfs_rq, NULL);
5429 		set_next_entity(cfs_rq, se);
5430 		cfs_rq = group_cfs_rq(se);
5431 	} while (cfs_rq);
5432 
5433 	p = task_of(se);
5434 
5435 	if (hrtick_enabled(rq))
5436 		hrtick_start_fair(rq, p);
5437 
5438 	return p;
5439 
5440 idle:
5441 	/*
5442 	 * This is OK, because current is on_cpu, which avoids it being picked
5443 	 * for load-balance and preemption/IRQs are still disabled avoiding
5444 	 * further scheduler activity on it and we're being very careful to
5445 	 * re-start the picking loop.
5446 	 */
5447 	lockdep_unpin_lock(&rq->lock);
5448 	new_tasks = idle_balance(rq);
5449 	lockdep_pin_lock(&rq->lock);
5450 	/*
5451 	 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5452 	 * possible for any higher priority task to appear. In that case we
5453 	 * must re-start the pick_next_entity() loop.
5454 	 */
5455 	if (new_tasks < 0)
5456 		return RETRY_TASK;
5457 
5458 	if (new_tasks > 0)
5459 		goto again;
5460 
5461 	return NULL;
5462 }
5463 
5464 /*
5465  * Account for a descheduled task:
5466  */
5467 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5468 {
5469 	struct sched_entity *se = &prev->se;
5470 	struct cfs_rq *cfs_rq;
5471 
5472 	for_each_sched_entity(se) {
5473 		cfs_rq = cfs_rq_of(se);
5474 		put_prev_entity(cfs_rq, se);
5475 	}
5476 }
5477 
5478 /*
5479  * sched_yield() is very simple
5480  *
5481  * The magic of dealing with the ->skip buddy is in pick_next_entity.
5482  */
5483 static void yield_task_fair(struct rq *rq)
5484 {
5485 	struct task_struct *curr = rq->curr;
5486 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5487 	struct sched_entity *se = &curr->se;
5488 
5489 	/*
5490 	 * Are we the only task in the tree?
5491 	 */
5492 	if (unlikely(rq->nr_running == 1))
5493 		return;
5494 
5495 	clear_buddies(cfs_rq, se);
5496 
5497 	if (curr->policy != SCHED_BATCH) {
5498 		update_rq_clock(rq);
5499 		/*
5500 		 * Update run-time statistics of the 'current'.
5501 		 */
5502 		update_curr(cfs_rq);
5503 		/*
5504 		 * Tell update_rq_clock() that we've just updated,
5505 		 * so we don't do microscopic update in schedule()
5506 		 * and double the fastpath cost.
5507 		 */
5508 		rq_clock_skip_update(rq, true);
5509 	}
5510 
5511 	set_skip_buddy(se);
5512 }
5513 
5514 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5515 {
5516 	struct sched_entity *se = &p->se;
5517 
5518 	/* throttled hierarchies are not runnable */
5519 	if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5520 		return false;
5521 
5522 	/* Tell the scheduler that we'd really like pse to run next. */
5523 	set_next_buddy(se);
5524 
5525 	yield_task_fair(rq);
5526 
5527 	return true;
5528 }
5529 
5530 #ifdef CONFIG_SMP
5531 /**************************************************
5532  * Fair scheduling class load-balancing methods.
5533  *
5534  * BASICS
5535  *
5536  * The purpose of load-balancing is to achieve the same basic fairness the
5537  * per-cpu scheduler provides, namely provide a proportional amount of compute
5538  * time to each task. This is expressed in the following equation:
5539  *
5540  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
5541  *
5542  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5543  * W_i,0 is defined as:
5544  *
5545  *   W_i,0 = \Sum_j w_i,j                                             (2)
5546  *
5547  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5548  * is derived from the nice value as per prio_to_weight[].
5549  *
5550  * The weight average is an exponential decay average of the instantaneous
5551  * weight:
5552  *
5553  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
5554  *
5555  * C_i is the compute capacity of cpu i, typically it is the
5556  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5557  * can also include other factors [XXX].
5558  *
5559  * To achieve this balance we define a measure of imbalance which follows
5560  * directly from (1):
5561  *
5562  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
5563  *
5564  * We them move tasks around to minimize the imbalance. In the continuous
5565  * function space it is obvious this converges, in the discrete case we get
5566  * a few fun cases generally called infeasible weight scenarios.
5567  *
5568  * [XXX expand on:
5569  *     - infeasible weights;
5570  *     - local vs global optima in the discrete case. ]
5571  *
5572  *
5573  * SCHED DOMAINS
5574  *
5575  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5576  * for all i,j solution, we create a tree of cpus that follows the hardware
5577  * topology where each level pairs two lower groups (or better). This results
5578  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5579  * tree to only the first of the previous level and we decrease the frequency
5580  * of load-balance at each level inv. proportional to the number of cpus in
5581  * the groups.
5582  *
5583  * This yields:
5584  *
5585  *     log_2 n     1     n
5586  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
5587  *     i = 0      2^i   2^i
5588  *                               `- size of each group
5589  *         |         |     `- number of cpus doing load-balance
5590  *         |         `- freq
5591  *         `- sum over all levels
5592  *
5593  * Coupled with a limit on how many tasks we can migrate every balance pass,
5594  * this makes (5) the runtime complexity of the balancer.
5595  *
5596  * An important property here is that each CPU is still (indirectly) connected
5597  * to every other cpu in at most O(log n) steps:
5598  *
5599  * The adjacency matrix of the resulting graph is given by:
5600  *
5601  *             log_2 n
5602  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
5603  *             k = 0
5604  *
5605  * And you'll find that:
5606  *
5607  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
5608  *
5609  * Showing there's indeed a path between every cpu in at most O(log n) steps.
5610  * The task movement gives a factor of O(m), giving a convergence complexity
5611  * of:
5612  *
5613  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
5614  *
5615  *
5616  * WORK CONSERVING
5617  *
5618  * In order to avoid CPUs going idle while there's still work to do, new idle
5619  * balancing is more aggressive and has the newly idle cpu iterate up the domain
5620  * tree itself instead of relying on other CPUs to bring it work.
5621  *
5622  * This adds some complexity to both (5) and (8) but it reduces the total idle
5623  * time.
5624  *
5625  * [XXX more?]
5626  *
5627  *
5628  * CGROUPS
5629  *
5630  * Cgroups make a horror show out of (2), instead of a simple sum we get:
5631  *
5632  *                                s_k,i
5633  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
5634  *                                 S_k
5635  *
5636  * Where
5637  *
5638  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
5639  *
5640  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5641  *
5642  * The big problem is S_k, its a global sum needed to compute a local (W_i)
5643  * property.
5644  *
5645  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5646  *      rewrite all of this once again.]
5647  */
5648 
5649 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5650 
5651 enum fbq_type { regular, remote, all };
5652 
5653 #define LBF_ALL_PINNED	0x01
5654 #define LBF_NEED_BREAK	0x02
5655 #define LBF_DST_PINNED  0x04
5656 #define LBF_SOME_PINNED	0x08
5657 
5658 struct lb_env {
5659 	struct sched_domain	*sd;
5660 
5661 	struct rq		*src_rq;
5662 	int			src_cpu;
5663 
5664 	int			dst_cpu;
5665 	struct rq		*dst_rq;
5666 
5667 	struct cpumask		*dst_grpmask;
5668 	int			new_dst_cpu;
5669 	enum cpu_idle_type	idle;
5670 	long			imbalance;
5671 	/* The set of CPUs under consideration for load-balancing */
5672 	struct cpumask		*cpus;
5673 
5674 	unsigned int		flags;
5675 
5676 	unsigned int		loop;
5677 	unsigned int		loop_break;
5678 	unsigned int		loop_max;
5679 
5680 	enum fbq_type		fbq_type;
5681 	struct list_head	tasks;
5682 };
5683 
5684 /*
5685  * Is this task likely cache-hot:
5686  */
5687 static int task_hot(struct task_struct *p, struct lb_env *env)
5688 {
5689 	s64 delta;
5690 
5691 	lockdep_assert_held(&env->src_rq->lock);
5692 
5693 	if (p->sched_class != &fair_sched_class)
5694 		return 0;
5695 
5696 	if (unlikely(p->policy == SCHED_IDLE))
5697 		return 0;
5698 
5699 	/*
5700 	 * Buddy candidates are cache hot:
5701 	 */
5702 	if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5703 			(&p->se == cfs_rq_of(&p->se)->next ||
5704 			 &p->se == cfs_rq_of(&p->se)->last))
5705 		return 1;
5706 
5707 	if (sysctl_sched_migration_cost == -1)
5708 		return 1;
5709 	if (sysctl_sched_migration_cost == 0)
5710 		return 0;
5711 
5712 	delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5713 
5714 	return delta < (s64)sysctl_sched_migration_cost;
5715 }
5716 
5717 #ifdef CONFIG_NUMA_BALANCING
5718 /*
5719  * Returns 1, if task migration degrades locality
5720  * Returns 0, if task migration improves locality i.e migration preferred.
5721  * Returns -1, if task migration is not affected by locality.
5722  */
5723 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5724 {
5725 	struct numa_group *numa_group = rcu_dereference(p->numa_group);
5726 	unsigned long src_faults, dst_faults;
5727 	int src_nid, dst_nid;
5728 
5729 	if (!static_branch_likely(&sched_numa_balancing))
5730 		return -1;
5731 
5732 	if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5733 		return -1;
5734 
5735 	src_nid = cpu_to_node(env->src_cpu);
5736 	dst_nid = cpu_to_node(env->dst_cpu);
5737 
5738 	if (src_nid == dst_nid)
5739 		return -1;
5740 
5741 	/* Migrating away from the preferred node is always bad. */
5742 	if (src_nid == p->numa_preferred_nid) {
5743 		if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5744 			return 1;
5745 		else
5746 			return -1;
5747 	}
5748 
5749 	/* Encourage migration to the preferred node. */
5750 	if (dst_nid == p->numa_preferred_nid)
5751 		return 0;
5752 
5753 	if (numa_group) {
5754 		src_faults = group_faults(p, src_nid);
5755 		dst_faults = group_faults(p, dst_nid);
5756 	} else {
5757 		src_faults = task_faults(p, src_nid);
5758 		dst_faults = task_faults(p, dst_nid);
5759 	}
5760 
5761 	return dst_faults < src_faults;
5762 }
5763 
5764 #else
5765 static inline int migrate_degrades_locality(struct task_struct *p,
5766 					     struct lb_env *env)
5767 {
5768 	return -1;
5769 }
5770 #endif
5771 
5772 /*
5773  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5774  */
5775 static
5776 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5777 {
5778 	int tsk_cache_hot;
5779 
5780 	lockdep_assert_held(&env->src_rq->lock);
5781 
5782 	/*
5783 	 * We do not migrate tasks that are:
5784 	 * 1) throttled_lb_pair, or
5785 	 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5786 	 * 3) running (obviously), or
5787 	 * 4) are cache-hot on their current CPU.
5788 	 */
5789 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5790 		return 0;
5791 
5792 	if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5793 		int cpu;
5794 
5795 		schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5796 
5797 		env->flags |= LBF_SOME_PINNED;
5798 
5799 		/*
5800 		 * Remember if this task can be migrated to any other cpu in
5801 		 * our sched_group. We may want to revisit it if we couldn't
5802 		 * meet load balance goals by pulling other tasks on src_cpu.
5803 		 *
5804 		 * Also avoid computing new_dst_cpu if we have already computed
5805 		 * one in current iteration.
5806 		 */
5807 		if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5808 			return 0;
5809 
5810 		/* Prevent to re-select dst_cpu via env's cpus */
5811 		for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5812 			if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5813 				env->flags |= LBF_DST_PINNED;
5814 				env->new_dst_cpu = cpu;
5815 				break;
5816 			}
5817 		}
5818 
5819 		return 0;
5820 	}
5821 
5822 	/* Record that we found atleast one task that could run on dst_cpu */
5823 	env->flags &= ~LBF_ALL_PINNED;
5824 
5825 	if (task_running(env->src_rq, p)) {
5826 		schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5827 		return 0;
5828 	}
5829 
5830 	/*
5831 	 * Aggressive migration if:
5832 	 * 1) destination numa is preferred
5833 	 * 2) task is cache cold, or
5834 	 * 3) too many balance attempts have failed.
5835 	 */
5836 	tsk_cache_hot = migrate_degrades_locality(p, env);
5837 	if (tsk_cache_hot == -1)
5838 		tsk_cache_hot = task_hot(p, env);
5839 
5840 	if (tsk_cache_hot <= 0 ||
5841 	    env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5842 		if (tsk_cache_hot == 1) {
5843 			schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5844 			schedstat_inc(p, se.statistics.nr_forced_migrations);
5845 		}
5846 		return 1;
5847 	}
5848 
5849 	schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5850 	return 0;
5851 }
5852 
5853 /*
5854  * detach_task() -- detach the task for the migration specified in env
5855  */
5856 static void detach_task(struct task_struct *p, struct lb_env *env)
5857 {
5858 	lockdep_assert_held(&env->src_rq->lock);
5859 
5860 	p->on_rq = TASK_ON_RQ_MIGRATING;
5861 	deactivate_task(env->src_rq, p, 0);
5862 	set_task_cpu(p, env->dst_cpu);
5863 }
5864 
5865 /*
5866  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5867  * part of active balancing operations within "domain".
5868  *
5869  * Returns a task if successful and NULL otherwise.
5870  */
5871 static struct task_struct *detach_one_task(struct lb_env *env)
5872 {
5873 	struct task_struct *p, *n;
5874 
5875 	lockdep_assert_held(&env->src_rq->lock);
5876 
5877 	list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5878 		if (!can_migrate_task(p, env))
5879 			continue;
5880 
5881 		detach_task(p, env);
5882 
5883 		/*
5884 		 * Right now, this is only the second place where
5885 		 * lb_gained[env->idle] is updated (other is detach_tasks)
5886 		 * so we can safely collect stats here rather than
5887 		 * inside detach_tasks().
5888 		 */
5889 		schedstat_inc(env->sd, lb_gained[env->idle]);
5890 		return p;
5891 	}
5892 	return NULL;
5893 }
5894 
5895 static const unsigned int sched_nr_migrate_break = 32;
5896 
5897 /*
5898  * detach_tasks() -- tries to detach up to imbalance weighted load from
5899  * busiest_rq, as part of a balancing operation within domain "sd".
5900  *
5901  * Returns number of detached tasks if successful and 0 otherwise.
5902  */
5903 static int detach_tasks(struct lb_env *env)
5904 {
5905 	struct list_head *tasks = &env->src_rq->cfs_tasks;
5906 	struct task_struct *p;
5907 	unsigned long load;
5908 	int detached = 0;
5909 
5910 	lockdep_assert_held(&env->src_rq->lock);
5911 
5912 	if (env->imbalance <= 0)
5913 		return 0;
5914 
5915 	while (!list_empty(tasks)) {
5916 		/*
5917 		 * We don't want to steal all, otherwise we may be treated likewise,
5918 		 * which could at worst lead to a livelock crash.
5919 		 */
5920 		if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5921 			break;
5922 
5923 		p = list_first_entry(tasks, struct task_struct, se.group_node);
5924 
5925 		env->loop++;
5926 		/* We've more or less seen every task there is, call it quits */
5927 		if (env->loop > env->loop_max)
5928 			break;
5929 
5930 		/* take a breather every nr_migrate tasks */
5931 		if (env->loop > env->loop_break) {
5932 			env->loop_break += sched_nr_migrate_break;
5933 			env->flags |= LBF_NEED_BREAK;
5934 			break;
5935 		}
5936 
5937 		if (!can_migrate_task(p, env))
5938 			goto next;
5939 
5940 		load = task_h_load(p);
5941 
5942 		if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5943 			goto next;
5944 
5945 		if ((load / 2) > env->imbalance)
5946 			goto next;
5947 
5948 		detach_task(p, env);
5949 		list_add(&p->se.group_node, &env->tasks);
5950 
5951 		detached++;
5952 		env->imbalance -= load;
5953 
5954 #ifdef CONFIG_PREEMPT
5955 		/*
5956 		 * NEWIDLE balancing is a source of latency, so preemptible
5957 		 * kernels will stop after the first task is detached to minimize
5958 		 * the critical section.
5959 		 */
5960 		if (env->idle == CPU_NEWLY_IDLE)
5961 			break;
5962 #endif
5963 
5964 		/*
5965 		 * We only want to steal up to the prescribed amount of
5966 		 * weighted load.
5967 		 */
5968 		if (env->imbalance <= 0)
5969 			break;
5970 
5971 		continue;
5972 next:
5973 		list_move_tail(&p->se.group_node, tasks);
5974 	}
5975 
5976 	/*
5977 	 * Right now, this is one of only two places we collect this stat
5978 	 * so we can safely collect detach_one_task() stats here rather
5979 	 * than inside detach_one_task().
5980 	 */
5981 	schedstat_add(env->sd, lb_gained[env->idle], detached);
5982 
5983 	return detached;
5984 }
5985 
5986 /*
5987  * attach_task() -- attach the task detached by detach_task() to its new rq.
5988  */
5989 static void attach_task(struct rq *rq, struct task_struct *p)
5990 {
5991 	lockdep_assert_held(&rq->lock);
5992 
5993 	BUG_ON(task_rq(p) != rq);
5994 	activate_task(rq, p, 0);
5995 	p->on_rq = TASK_ON_RQ_QUEUED;
5996 	check_preempt_curr(rq, p, 0);
5997 }
5998 
5999 /*
6000  * attach_one_task() -- attaches the task returned from detach_one_task() to
6001  * its new rq.
6002  */
6003 static void attach_one_task(struct rq *rq, struct task_struct *p)
6004 {
6005 	raw_spin_lock(&rq->lock);
6006 	attach_task(rq, p);
6007 	raw_spin_unlock(&rq->lock);
6008 }
6009 
6010 /*
6011  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6012  * new rq.
6013  */
6014 static void attach_tasks(struct lb_env *env)
6015 {
6016 	struct list_head *tasks = &env->tasks;
6017 	struct task_struct *p;
6018 
6019 	raw_spin_lock(&env->dst_rq->lock);
6020 
6021 	while (!list_empty(tasks)) {
6022 		p = list_first_entry(tasks, struct task_struct, se.group_node);
6023 		list_del_init(&p->se.group_node);
6024 
6025 		attach_task(env->dst_rq, p);
6026 	}
6027 
6028 	raw_spin_unlock(&env->dst_rq->lock);
6029 }
6030 
6031 #ifdef CONFIG_FAIR_GROUP_SCHED
6032 static void update_blocked_averages(int cpu)
6033 {
6034 	struct rq *rq = cpu_rq(cpu);
6035 	struct cfs_rq *cfs_rq;
6036 	unsigned long flags;
6037 
6038 	raw_spin_lock_irqsave(&rq->lock, flags);
6039 	update_rq_clock(rq);
6040 
6041 	/*
6042 	 * Iterates the task_group tree in a bottom up fashion, see
6043 	 * list_add_leaf_cfs_rq() for details.
6044 	 */
6045 	for_each_leaf_cfs_rq(rq, cfs_rq) {
6046 		/* throttled entities do not contribute to load */
6047 		if (throttled_hierarchy(cfs_rq))
6048 			continue;
6049 
6050 		if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6051 			update_tg_load_avg(cfs_rq, 0);
6052 	}
6053 	raw_spin_unlock_irqrestore(&rq->lock, flags);
6054 }
6055 
6056 /*
6057  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6058  * This needs to be done in a top-down fashion because the load of a child
6059  * group is a fraction of its parents load.
6060  */
6061 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6062 {
6063 	struct rq *rq = rq_of(cfs_rq);
6064 	struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6065 	unsigned long now = jiffies;
6066 	unsigned long load;
6067 
6068 	if (cfs_rq->last_h_load_update == now)
6069 		return;
6070 
6071 	cfs_rq->h_load_next = NULL;
6072 	for_each_sched_entity(se) {
6073 		cfs_rq = cfs_rq_of(se);
6074 		cfs_rq->h_load_next = se;
6075 		if (cfs_rq->last_h_load_update == now)
6076 			break;
6077 	}
6078 
6079 	if (!se) {
6080 		cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6081 		cfs_rq->last_h_load_update = now;
6082 	}
6083 
6084 	while ((se = cfs_rq->h_load_next) != NULL) {
6085 		load = cfs_rq->h_load;
6086 		load = div64_ul(load * se->avg.load_avg,
6087 			cfs_rq_load_avg(cfs_rq) + 1);
6088 		cfs_rq = group_cfs_rq(se);
6089 		cfs_rq->h_load = load;
6090 		cfs_rq->last_h_load_update = now;
6091 	}
6092 }
6093 
6094 static unsigned long task_h_load(struct task_struct *p)
6095 {
6096 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
6097 
6098 	update_cfs_rq_h_load(cfs_rq);
6099 	return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6100 			cfs_rq_load_avg(cfs_rq) + 1);
6101 }
6102 #else
6103 static inline void update_blocked_averages(int cpu)
6104 {
6105 	struct rq *rq = cpu_rq(cpu);
6106 	struct cfs_rq *cfs_rq = &rq->cfs;
6107 	unsigned long flags;
6108 
6109 	raw_spin_lock_irqsave(&rq->lock, flags);
6110 	update_rq_clock(rq);
6111 	update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6112 	raw_spin_unlock_irqrestore(&rq->lock, flags);
6113 }
6114 
6115 static unsigned long task_h_load(struct task_struct *p)
6116 {
6117 	return p->se.avg.load_avg;
6118 }
6119 #endif
6120 
6121 /********** Helpers for find_busiest_group ************************/
6122 
6123 enum group_type {
6124 	group_other = 0,
6125 	group_imbalanced,
6126 	group_overloaded,
6127 };
6128 
6129 /*
6130  * sg_lb_stats - stats of a sched_group required for load_balancing
6131  */
6132 struct sg_lb_stats {
6133 	unsigned long avg_load; /*Avg load across the CPUs of the group */
6134 	unsigned long group_load; /* Total load over the CPUs of the group */
6135 	unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6136 	unsigned long load_per_task;
6137 	unsigned long group_capacity;
6138 	unsigned long group_util; /* Total utilization of the group */
6139 	unsigned int sum_nr_running; /* Nr tasks running in the group */
6140 	unsigned int idle_cpus;
6141 	unsigned int group_weight;
6142 	enum group_type group_type;
6143 	int group_no_capacity;
6144 #ifdef CONFIG_NUMA_BALANCING
6145 	unsigned int nr_numa_running;
6146 	unsigned int nr_preferred_running;
6147 #endif
6148 };
6149 
6150 /*
6151  * sd_lb_stats - Structure to store the statistics of a sched_domain
6152  *		 during load balancing.
6153  */
6154 struct sd_lb_stats {
6155 	struct sched_group *busiest;	/* Busiest group in this sd */
6156 	struct sched_group *local;	/* Local group in this sd */
6157 	unsigned long total_load;	/* Total load of all groups in sd */
6158 	unsigned long total_capacity;	/* Total capacity of all groups in sd */
6159 	unsigned long avg_load;	/* Average load across all groups in sd */
6160 
6161 	struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6162 	struct sg_lb_stats local_stat;	/* Statistics of the local group */
6163 };
6164 
6165 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6166 {
6167 	/*
6168 	 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6169 	 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6170 	 * We must however clear busiest_stat::avg_load because
6171 	 * update_sd_pick_busiest() reads this before assignment.
6172 	 */
6173 	*sds = (struct sd_lb_stats){
6174 		.busiest = NULL,
6175 		.local = NULL,
6176 		.total_load = 0UL,
6177 		.total_capacity = 0UL,
6178 		.busiest_stat = {
6179 			.avg_load = 0UL,
6180 			.sum_nr_running = 0,
6181 			.group_type = group_other,
6182 		},
6183 	};
6184 }
6185 
6186 /**
6187  * get_sd_load_idx - Obtain the load index for a given sched domain.
6188  * @sd: The sched_domain whose load_idx is to be obtained.
6189  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6190  *
6191  * Return: The load index.
6192  */
6193 static inline int get_sd_load_idx(struct sched_domain *sd,
6194 					enum cpu_idle_type idle)
6195 {
6196 	int load_idx;
6197 
6198 	switch (idle) {
6199 	case CPU_NOT_IDLE:
6200 		load_idx = sd->busy_idx;
6201 		break;
6202 
6203 	case CPU_NEWLY_IDLE:
6204 		load_idx = sd->newidle_idx;
6205 		break;
6206 	default:
6207 		load_idx = sd->idle_idx;
6208 		break;
6209 	}
6210 
6211 	return load_idx;
6212 }
6213 
6214 static unsigned long scale_rt_capacity(int cpu)
6215 {
6216 	struct rq *rq = cpu_rq(cpu);
6217 	u64 total, used, age_stamp, avg;
6218 	s64 delta;
6219 
6220 	/*
6221 	 * Since we're reading these variables without serialization make sure
6222 	 * we read them once before doing sanity checks on them.
6223 	 */
6224 	age_stamp = READ_ONCE(rq->age_stamp);
6225 	avg = READ_ONCE(rq->rt_avg);
6226 	delta = __rq_clock_broken(rq) - age_stamp;
6227 
6228 	if (unlikely(delta < 0))
6229 		delta = 0;
6230 
6231 	total = sched_avg_period() + delta;
6232 
6233 	used = div_u64(avg, total);
6234 
6235 	if (likely(used < SCHED_CAPACITY_SCALE))
6236 		return SCHED_CAPACITY_SCALE - used;
6237 
6238 	return 1;
6239 }
6240 
6241 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6242 {
6243 	unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6244 	struct sched_group *sdg = sd->groups;
6245 
6246 	cpu_rq(cpu)->cpu_capacity_orig = capacity;
6247 
6248 	capacity *= scale_rt_capacity(cpu);
6249 	capacity >>= SCHED_CAPACITY_SHIFT;
6250 
6251 	if (!capacity)
6252 		capacity = 1;
6253 
6254 	cpu_rq(cpu)->cpu_capacity = capacity;
6255 	sdg->sgc->capacity = capacity;
6256 }
6257 
6258 void update_group_capacity(struct sched_domain *sd, int cpu)
6259 {
6260 	struct sched_domain *child = sd->child;
6261 	struct sched_group *group, *sdg = sd->groups;
6262 	unsigned long capacity;
6263 	unsigned long interval;
6264 
6265 	interval = msecs_to_jiffies(sd->balance_interval);
6266 	interval = clamp(interval, 1UL, max_load_balance_interval);
6267 	sdg->sgc->next_update = jiffies + interval;
6268 
6269 	if (!child) {
6270 		update_cpu_capacity(sd, cpu);
6271 		return;
6272 	}
6273 
6274 	capacity = 0;
6275 
6276 	if (child->flags & SD_OVERLAP) {
6277 		/*
6278 		 * SD_OVERLAP domains cannot assume that child groups
6279 		 * span the current group.
6280 		 */
6281 
6282 		for_each_cpu(cpu, sched_group_cpus(sdg)) {
6283 			struct sched_group_capacity *sgc;
6284 			struct rq *rq = cpu_rq(cpu);
6285 
6286 			/*
6287 			 * build_sched_domains() -> init_sched_groups_capacity()
6288 			 * gets here before we've attached the domains to the
6289 			 * runqueues.
6290 			 *
6291 			 * Use capacity_of(), which is set irrespective of domains
6292 			 * in update_cpu_capacity().
6293 			 *
6294 			 * This avoids capacity from being 0 and
6295 			 * causing divide-by-zero issues on boot.
6296 			 */
6297 			if (unlikely(!rq->sd)) {
6298 				capacity += capacity_of(cpu);
6299 				continue;
6300 			}
6301 
6302 			sgc = rq->sd->groups->sgc;
6303 			capacity += sgc->capacity;
6304 		}
6305 	} else  {
6306 		/*
6307 		 * !SD_OVERLAP domains can assume that child groups
6308 		 * span the current group.
6309 		 */
6310 
6311 		group = child->groups;
6312 		do {
6313 			capacity += group->sgc->capacity;
6314 			group = group->next;
6315 		} while (group != child->groups);
6316 	}
6317 
6318 	sdg->sgc->capacity = capacity;
6319 }
6320 
6321 /*
6322  * Check whether the capacity of the rq has been noticeably reduced by side
6323  * activity. The imbalance_pct is used for the threshold.
6324  * Return true is the capacity is reduced
6325  */
6326 static inline int
6327 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6328 {
6329 	return ((rq->cpu_capacity * sd->imbalance_pct) <
6330 				(rq->cpu_capacity_orig * 100));
6331 }
6332 
6333 /*
6334  * Group imbalance indicates (and tries to solve) the problem where balancing
6335  * groups is inadequate due to tsk_cpus_allowed() constraints.
6336  *
6337  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6338  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6339  * Something like:
6340  *
6341  * 	{ 0 1 2 3 } { 4 5 6 7 }
6342  * 	        *     * * *
6343  *
6344  * If we were to balance group-wise we'd place two tasks in the first group and
6345  * two tasks in the second group. Clearly this is undesired as it will overload
6346  * cpu 3 and leave one of the cpus in the second group unused.
6347  *
6348  * The current solution to this issue is detecting the skew in the first group
6349  * by noticing the lower domain failed to reach balance and had difficulty
6350  * moving tasks due to affinity constraints.
6351  *
6352  * When this is so detected; this group becomes a candidate for busiest; see
6353  * update_sd_pick_busiest(). And calculate_imbalance() and
6354  * find_busiest_group() avoid some of the usual balance conditions to allow it
6355  * to create an effective group imbalance.
6356  *
6357  * This is a somewhat tricky proposition since the next run might not find the
6358  * group imbalance and decide the groups need to be balanced again. A most
6359  * subtle and fragile situation.
6360  */
6361 
6362 static inline int sg_imbalanced(struct sched_group *group)
6363 {
6364 	return group->sgc->imbalance;
6365 }
6366 
6367 /*
6368  * group_has_capacity returns true if the group has spare capacity that could
6369  * be used by some tasks.
6370  * We consider that a group has spare capacity if the  * number of task is
6371  * smaller than the number of CPUs or if the utilization is lower than the
6372  * available capacity for CFS tasks.
6373  * For the latter, we use a threshold to stabilize the state, to take into
6374  * account the variance of the tasks' load and to return true if the available
6375  * capacity in meaningful for the load balancer.
6376  * As an example, an available capacity of 1% can appear but it doesn't make
6377  * any benefit for the load balance.
6378  */
6379 static inline bool
6380 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6381 {
6382 	if (sgs->sum_nr_running < sgs->group_weight)
6383 		return true;
6384 
6385 	if ((sgs->group_capacity * 100) >
6386 			(sgs->group_util * env->sd->imbalance_pct))
6387 		return true;
6388 
6389 	return false;
6390 }
6391 
6392 /*
6393  *  group_is_overloaded returns true if the group has more tasks than it can
6394  *  handle.
6395  *  group_is_overloaded is not equals to !group_has_capacity because a group
6396  *  with the exact right number of tasks, has no more spare capacity but is not
6397  *  overloaded so both group_has_capacity and group_is_overloaded return
6398  *  false.
6399  */
6400 static inline bool
6401 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6402 {
6403 	if (sgs->sum_nr_running <= sgs->group_weight)
6404 		return false;
6405 
6406 	if ((sgs->group_capacity * 100) <
6407 			(sgs->group_util * env->sd->imbalance_pct))
6408 		return true;
6409 
6410 	return false;
6411 }
6412 
6413 static inline enum
6414 group_type group_classify(struct sched_group *group,
6415 			  struct sg_lb_stats *sgs)
6416 {
6417 	if (sgs->group_no_capacity)
6418 		return group_overloaded;
6419 
6420 	if (sg_imbalanced(group))
6421 		return group_imbalanced;
6422 
6423 	return group_other;
6424 }
6425 
6426 /**
6427  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6428  * @env: The load balancing environment.
6429  * @group: sched_group whose statistics are to be updated.
6430  * @load_idx: Load index of sched_domain of this_cpu for load calc.
6431  * @local_group: Does group contain this_cpu.
6432  * @sgs: variable to hold the statistics for this group.
6433  * @overload: Indicate more than one runnable task for any CPU.
6434  */
6435 static inline void update_sg_lb_stats(struct lb_env *env,
6436 			struct sched_group *group, int load_idx,
6437 			int local_group, struct sg_lb_stats *sgs,
6438 			bool *overload)
6439 {
6440 	unsigned long load;
6441 	int i, nr_running;
6442 
6443 	memset(sgs, 0, sizeof(*sgs));
6444 
6445 	for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6446 		struct rq *rq = cpu_rq(i);
6447 
6448 		/* Bias balancing toward cpus of our domain */
6449 		if (local_group)
6450 			load = target_load(i, load_idx);
6451 		else
6452 			load = source_load(i, load_idx);
6453 
6454 		sgs->group_load += load;
6455 		sgs->group_util += cpu_util(i);
6456 		sgs->sum_nr_running += rq->cfs.h_nr_running;
6457 
6458 		nr_running = rq->nr_running;
6459 		if (nr_running > 1)
6460 			*overload = true;
6461 
6462 #ifdef CONFIG_NUMA_BALANCING
6463 		sgs->nr_numa_running += rq->nr_numa_running;
6464 		sgs->nr_preferred_running += rq->nr_preferred_running;
6465 #endif
6466 		sgs->sum_weighted_load += weighted_cpuload(i);
6467 		/*
6468 		 * No need to call idle_cpu() if nr_running is not 0
6469 		 */
6470 		if (!nr_running && idle_cpu(i))
6471 			sgs->idle_cpus++;
6472 	}
6473 
6474 	/* Adjust by relative CPU capacity of the group */
6475 	sgs->group_capacity = group->sgc->capacity;
6476 	sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6477 
6478 	if (sgs->sum_nr_running)
6479 		sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6480 
6481 	sgs->group_weight = group->group_weight;
6482 
6483 	sgs->group_no_capacity = group_is_overloaded(env, sgs);
6484 	sgs->group_type = group_classify(group, sgs);
6485 }
6486 
6487 /**
6488  * update_sd_pick_busiest - return 1 on busiest group
6489  * @env: The load balancing environment.
6490  * @sds: sched_domain statistics
6491  * @sg: sched_group candidate to be checked for being the busiest
6492  * @sgs: sched_group statistics
6493  *
6494  * Determine if @sg is a busier group than the previously selected
6495  * busiest group.
6496  *
6497  * Return: %true if @sg is a busier group than the previously selected
6498  * busiest group. %false otherwise.
6499  */
6500 static bool update_sd_pick_busiest(struct lb_env *env,
6501 				   struct sd_lb_stats *sds,
6502 				   struct sched_group *sg,
6503 				   struct sg_lb_stats *sgs)
6504 {
6505 	struct sg_lb_stats *busiest = &sds->busiest_stat;
6506 
6507 	if (sgs->group_type > busiest->group_type)
6508 		return true;
6509 
6510 	if (sgs->group_type < busiest->group_type)
6511 		return false;
6512 
6513 	if (sgs->avg_load <= busiest->avg_load)
6514 		return false;
6515 
6516 	/* This is the busiest node in its class. */
6517 	if (!(env->sd->flags & SD_ASYM_PACKING))
6518 		return true;
6519 
6520 	/*
6521 	 * ASYM_PACKING needs to move all the work to the lowest
6522 	 * numbered CPUs in the group, therefore mark all groups
6523 	 * higher than ourself as busy.
6524 	 */
6525 	if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6526 		if (!sds->busiest)
6527 			return true;
6528 
6529 		if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6530 			return true;
6531 	}
6532 
6533 	return false;
6534 }
6535 
6536 #ifdef CONFIG_NUMA_BALANCING
6537 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6538 {
6539 	if (sgs->sum_nr_running > sgs->nr_numa_running)
6540 		return regular;
6541 	if (sgs->sum_nr_running > sgs->nr_preferred_running)
6542 		return remote;
6543 	return all;
6544 }
6545 
6546 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6547 {
6548 	if (rq->nr_running > rq->nr_numa_running)
6549 		return regular;
6550 	if (rq->nr_running > rq->nr_preferred_running)
6551 		return remote;
6552 	return all;
6553 }
6554 #else
6555 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6556 {
6557 	return all;
6558 }
6559 
6560 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6561 {
6562 	return regular;
6563 }
6564 #endif /* CONFIG_NUMA_BALANCING */
6565 
6566 /**
6567  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6568  * @env: The load balancing environment.
6569  * @sds: variable to hold the statistics for this sched_domain.
6570  */
6571 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6572 {
6573 	struct sched_domain *child = env->sd->child;
6574 	struct sched_group *sg = env->sd->groups;
6575 	struct sg_lb_stats tmp_sgs;
6576 	int load_idx, prefer_sibling = 0;
6577 	bool overload = false;
6578 
6579 	if (child && child->flags & SD_PREFER_SIBLING)
6580 		prefer_sibling = 1;
6581 
6582 	load_idx = get_sd_load_idx(env->sd, env->idle);
6583 
6584 	do {
6585 		struct sg_lb_stats *sgs = &tmp_sgs;
6586 		int local_group;
6587 
6588 		local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6589 		if (local_group) {
6590 			sds->local = sg;
6591 			sgs = &sds->local_stat;
6592 
6593 			if (env->idle != CPU_NEWLY_IDLE ||
6594 			    time_after_eq(jiffies, sg->sgc->next_update))
6595 				update_group_capacity(env->sd, env->dst_cpu);
6596 		}
6597 
6598 		update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6599 						&overload);
6600 
6601 		if (local_group)
6602 			goto next_group;
6603 
6604 		/*
6605 		 * In case the child domain prefers tasks go to siblings
6606 		 * first, lower the sg capacity so that we'll try
6607 		 * and move all the excess tasks away. We lower the capacity
6608 		 * of a group only if the local group has the capacity to fit
6609 		 * these excess tasks. The extra check prevents the case where
6610 		 * you always pull from the heaviest group when it is already
6611 		 * under-utilized (possible with a large weight task outweighs
6612 		 * the tasks on the system).
6613 		 */
6614 		if (prefer_sibling && sds->local &&
6615 		    group_has_capacity(env, &sds->local_stat) &&
6616 		    (sgs->sum_nr_running > 1)) {
6617 			sgs->group_no_capacity = 1;
6618 			sgs->group_type = group_classify(sg, sgs);
6619 		}
6620 
6621 		if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6622 			sds->busiest = sg;
6623 			sds->busiest_stat = *sgs;
6624 		}
6625 
6626 next_group:
6627 		/* Now, start updating sd_lb_stats */
6628 		sds->total_load += sgs->group_load;
6629 		sds->total_capacity += sgs->group_capacity;
6630 
6631 		sg = sg->next;
6632 	} while (sg != env->sd->groups);
6633 
6634 	if (env->sd->flags & SD_NUMA)
6635 		env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6636 
6637 	if (!env->sd->parent) {
6638 		/* update overload indicator if we are at root domain */
6639 		if (env->dst_rq->rd->overload != overload)
6640 			env->dst_rq->rd->overload = overload;
6641 	}
6642 
6643 }
6644 
6645 /**
6646  * check_asym_packing - Check to see if the group is packed into the
6647  *			sched doman.
6648  *
6649  * This is primarily intended to used at the sibling level.  Some
6650  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
6651  * case of POWER7, it can move to lower SMT modes only when higher
6652  * threads are idle.  When in lower SMT modes, the threads will
6653  * perform better since they share less core resources.  Hence when we
6654  * have idle threads, we want them to be the higher ones.
6655  *
6656  * This packing function is run on idle threads.  It checks to see if
6657  * the busiest CPU in this domain (core in the P7 case) has a higher
6658  * CPU number than the packing function is being run on.  Here we are
6659  * assuming lower CPU number will be equivalent to lower a SMT thread
6660  * number.
6661  *
6662  * Return: 1 when packing is required and a task should be moved to
6663  * this CPU.  The amount of the imbalance is returned in *imbalance.
6664  *
6665  * @env: The load balancing environment.
6666  * @sds: Statistics of the sched_domain which is to be packed
6667  */
6668 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6669 {
6670 	int busiest_cpu;
6671 
6672 	if (!(env->sd->flags & SD_ASYM_PACKING))
6673 		return 0;
6674 
6675 	if (!sds->busiest)
6676 		return 0;
6677 
6678 	busiest_cpu = group_first_cpu(sds->busiest);
6679 	if (env->dst_cpu > busiest_cpu)
6680 		return 0;
6681 
6682 	env->imbalance = DIV_ROUND_CLOSEST(
6683 		sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6684 		SCHED_CAPACITY_SCALE);
6685 
6686 	return 1;
6687 }
6688 
6689 /**
6690  * fix_small_imbalance - Calculate the minor imbalance that exists
6691  *			amongst the groups of a sched_domain, during
6692  *			load balancing.
6693  * @env: The load balancing environment.
6694  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6695  */
6696 static inline
6697 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6698 {
6699 	unsigned long tmp, capa_now = 0, capa_move = 0;
6700 	unsigned int imbn = 2;
6701 	unsigned long scaled_busy_load_per_task;
6702 	struct sg_lb_stats *local, *busiest;
6703 
6704 	local = &sds->local_stat;
6705 	busiest = &sds->busiest_stat;
6706 
6707 	if (!local->sum_nr_running)
6708 		local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6709 	else if (busiest->load_per_task > local->load_per_task)
6710 		imbn = 1;
6711 
6712 	scaled_busy_load_per_task =
6713 		(busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6714 		busiest->group_capacity;
6715 
6716 	if (busiest->avg_load + scaled_busy_load_per_task >=
6717 	    local->avg_load + (scaled_busy_load_per_task * imbn)) {
6718 		env->imbalance = busiest->load_per_task;
6719 		return;
6720 	}
6721 
6722 	/*
6723 	 * OK, we don't have enough imbalance to justify moving tasks,
6724 	 * however we may be able to increase total CPU capacity used by
6725 	 * moving them.
6726 	 */
6727 
6728 	capa_now += busiest->group_capacity *
6729 			min(busiest->load_per_task, busiest->avg_load);
6730 	capa_now += local->group_capacity *
6731 			min(local->load_per_task, local->avg_load);
6732 	capa_now /= SCHED_CAPACITY_SCALE;
6733 
6734 	/* Amount of load we'd subtract */
6735 	if (busiest->avg_load > scaled_busy_load_per_task) {
6736 		capa_move += busiest->group_capacity *
6737 			    min(busiest->load_per_task,
6738 				busiest->avg_load - scaled_busy_load_per_task);
6739 	}
6740 
6741 	/* Amount of load we'd add */
6742 	if (busiest->avg_load * busiest->group_capacity <
6743 	    busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6744 		tmp = (busiest->avg_load * busiest->group_capacity) /
6745 		      local->group_capacity;
6746 	} else {
6747 		tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6748 		      local->group_capacity;
6749 	}
6750 	capa_move += local->group_capacity *
6751 		    min(local->load_per_task, local->avg_load + tmp);
6752 	capa_move /= SCHED_CAPACITY_SCALE;
6753 
6754 	/* Move if we gain throughput */
6755 	if (capa_move > capa_now)
6756 		env->imbalance = busiest->load_per_task;
6757 }
6758 
6759 /**
6760  * calculate_imbalance - Calculate the amount of imbalance present within the
6761  *			 groups of a given sched_domain during load balance.
6762  * @env: load balance environment
6763  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6764  */
6765 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6766 {
6767 	unsigned long max_pull, load_above_capacity = ~0UL;
6768 	struct sg_lb_stats *local, *busiest;
6769 
6770 	local = &sds->local_stat;
6771 	busiest = &sds->busiest_stat;
6772 
6773 	if (busiest->group_type == group_imbalanced) {
6774 		/*
6775 		 * In the group_imb case we cannot rely on group-wide averages
6776 		 * to ensure cpu-load equilibrium, look at wider averages. XXX
6777 		 */
6778 		busiest->load_per_task =
6779 			min(busiest->load_per_task, sds->avg_load);
6780 	}
6781 
6782 	/*
6783 	 * In the presence of smp nice balancing, certain scenarios can have
6784 	 * max load less than avg load(as we skip the groups at or below
6785 	 * its cpu_capacity, while calculating max_load..)
6786 	 */
6787 	if (busiest->avg_load <= sds->avg_load ||
6788 	    local->avg_load >= sds->avg_load) {
6789 		env->imbalance = 0;
6790 		return fix_small_imbalance(env, sds);
6791 	}
6792 
6793 	/*
6794 	 * If there aren't any idle cpus, avoid creating some.
6795 	 */
6796 	if (busiest->group_type == group_overloaded &&
6797 	    local->group_type   == group_overloaded) {
6798 		load_above_capacity = busiest->sum_nr_running *
6799 					SCHED_LOAD_SCALE;
6800 		if (load_above_capacity > busiest->group_capacity)
6801 			load_above_capacity -= busiest->group_capacity;
6802 		else
6803 			load_above_capacity = ~0UL;
6804 	}
6805 
6806 	/*
6807 	 * We're trying to get all the cpus to the average_load, so we don't
6808 	 * want to push ourselves above the average load, nor do we wish to
6809 	 * reduce the max loaded cpu below the average load. At the same time,
6810 	 * we also don't want to reduce the group load below the group capacity
6811 	 * (so that we can implement power-savings policies etc). Thus we look
6812 	 * for the minimum possible imbalance.
6813 	 */
6814 	max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6815 
6816 	/* How much load to actually move to equalise the imbalance */
6817 	env->imbalance = min(
6818 		max_pull * busiest->group_capacity,
6819 		(sds->avg_load - local->avg_load) * local->group_capacity
6820 	) / SCHED_CAPACITY_SCALE;
6821 
6822 	/*
6823 	 * if *imbalance is less than the average load per runnable task
6824 	 * there is no guarantee that any tasks will be moved so we'll have
6825 	 * a think about bumping its value to force at least one task to be
6826 	 * moved
6827 	 */
6828 	if (env->imbalance < busiest->load_per_task)
6829 		return fix_small_imbalance(env, sds);
6830 }
6831 
6832 /******* find_busiest_group() helpers end here *********************/
6833 
6834 /**
6835  * find_busiest_group - Returns the busiest group within the sched_domain
6836  * if there is an imbalance. If there isn't an imbalance, and
6837  * the user has opted for power-savings, it returns a group whose
6838  * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6839  * such a group exists.
6840  *
6841  * Also calculates the amount of weighted load which should be moved
6842  * to restore balance.
6843  *
6844  * @env: The load balancing environment.
6845  *
6846  * Return:	- The busiest group if imbalance exists.
6847  *		- If no imbalance and user has opted for power-savings balance,
6848  *		   return the least loaded group whose CPUs can be
6849  *		   put to idle by rebalancing its tasks onto our group.
6850  */
6851 static struct sched_group *find_busiest_group(struct lb_env *env)
6852 {
6853 	struct sg_lb_stats *local, *busiest;
6854 	struct sd_lb_stats sds;
6855 
6856 	init_sd_lb_stats(&sds);
6857 
6858 	/*
6859 	 * Compute the various statistics relavent for load balancing at
6860 	 * this level.
6861 	 */
6862 	update_sd_lb_stats(env, &sds);
6863 	local = &sds.local_stat;
6864 	busiest = &sds.busiest_stat;
6865 
6866 	/* ASYM feature bypasses nice load balance check */
6867 	if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6868 	    check_asym_packing(env, &sds))
6869 		return sds.busiest;
6870 
6871 	/* There is no busy sibling group to pull tasks from */
6872 	if (!sds.busiest || busiest->sum_nr_running == 0)
6873 		goto out_balanced;
6874 
6875 	sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6876 						/ sds.total_capacity;
6877 
6878 	/*
6879 	 * If the busiest group is imbalanced the below checks don't
6880 	 * work because they assume all things are equal, which typically
6881 	 * isn't true due to cpus_allowed constraints and the like.
6882 	 */
6883 	if (busiest->group_type == group_imbalanced)
6884 		goto force_balance;
6885 
6886 	/* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6887 	if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6888 	    busiest->group_no_capacity)
6889 		goto force_balance;
6890 
6891 	/*
6892 	 * If the local group is busier than the selected busiest group
6893 	 * don't try and pull any tasks.
6894 	 */
6895 	if (local->avg_load >= busiest->avg_load)
6896 		goto out_balanced;
6897 
6898 	/*
6899 	 * Don't pull any tasks if this group is already above the domain
6900 	 * average load.
6901 	 */
6902 	if (local->avg_load >= sds.avg_load)
6903 		goto out_balanced;
6904 
6905 	if (env->idle == CPU_IDLE) {
6906 		/*
6907 		 * This cpu is idle. If the busiest group is not overloaded
6908 		 * and there is no imbalance between this and busiest group
6909 		 * wrt idle cpus, it is balanced. The imbalance becomes
6910 		 * significant if the diff is greater than 1 otherwise we
6911 		 * might end up to just move the imbalance on another group
6912 		 */
6913 		if ((busiest->group_type != group_overloaded) &&
6914 				(local->idle_cpus <= (busiest->idle_cpus + 1)))
6915 			goto out_balanced;
6916 	} else {
6917 		/*
6918 		 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6919 		 * imbalance_pct to be conservative.
6920 		 */
6921 		if (100 * busiest->avg_load <=
6922 				env->sd->imbalance_pct * local->avg_load)
6923 			goto out_balanced;
6924 	}
6925 
6926 force_balance:
6927 	/* Looks like there is an imbalance. Compute it */
6928 	calculate_imbalance(env, &sds);
6929 	return sds.busiest;
6930 
6931 out_balanced:
6932 	env->imbalance = 0;
6933 	return NULL;
6934 }
6935 
6936 /*
6937  * find_busiest_queue - find the busiest runqueue among the cpus in group.
6938  */
6939 static struct rq *find_busiest_queue(struct lb_env *env,
6940 				     struct sched_group *group)
6941 {
6942 	struct rq *busiest = NULL, *rq;
6943 	unsigned long busiest_load = 0, busiest_capacity = 1;
6944 	int i;
6945 
6946 	for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6947 		unsigned long capacity, wl;
6948 		enum fbq_type rt;
6949 
6950 		rq = cpu_rq(i);
6951 		rt = fbq_classify_rq(rq);
6952 
6953 		/*
6954 		 * We classify groups/runqueues into three groups:
6955 		 *  - regular: there are !numa tasks
6956 		 *  - remote:  there are numa tasks that run on the 'wrong' node
6957 		 *  - all:     there is no distinction
6958 		 *
6959 		 * In order to avoid migrating ideally placed numa tasks,
6960 		 * ignore those when there's better options.
6961 		 *
6962 		 * If we ignore the actual busiest queue to migrate another
6963 		 * task, the next balance pass can still reduce the busiest
6964 		 * queue by moving tasks around inside the node.
6965 		 *
6966 		 * If we cannot move enough load due to this classification
6967 		 * the next pass will adjust the group classification and
6968 		 * allow migration of more tasks.
6969 		 *
6970 		 * Both cases only affect the total convergence complexity.
6971 		 */
6972 		if (rt > env->fbq_type)
6973 			continue;
6974 
6975 		capacity = capacity_of(i);
6976 
6977 		wl = weighted_cpuload(i);
6978 
6979 		/*
6980 		 * When comparing with imbalance, use weighted_cpuload()
6981 		 * which is not scaled with the cpu capacity.
6982 		 */
6983 
6984 		if (rq->nr_running == 1 && wl > env->imbalance &&
6985 		    !check_cpu_capacity(rq, env->sd))
6986 			continue;
6987 
6988 		/*
6989 		 * For the load comparisons with the other cpu's, consider
6990 		 * the weighted_cpuload() scaled with the cpu capacity, so
6991 		 * that the load can be moved away from the cpu that is
6992 		 * potentially running at a lower capacity.
6993 		 *
6994 		 * Thus we're looking for max(wl_i / capacity_i), crosswise
6995 		 * multiplication to rid ourselves of the division works out
6996 		 * to: wl_i * capacity_j > wl_j * capacity_i;  where j is
6997 		 * our previous maximum.
6998 		 */
6999 		if (wl * busiest_capacity > busiest_load * capacity) {
7000 			busiest_load = wl;
7001 			busiest_capacity = capacity;
7002 			busiest = rq;
7003 		}
7004 	}
7005 
7006 	return busiest;
7007 }
7008 
7009 /*
7010  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7011  * so long as it is large enough.
7012  */
7013 #define MAX_PINNED_INTERVAL	512
7014 
7015 /* Working cpumask for load_balance and load_balance_newidle. */
7016 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7017 
7018 static int need_active_balance(struct lb_env *env)
7019 {
7020 	struct sched_domain *sd = env->sd;
7021 
7022 	if (env->idle == CPU_NEWLY_IDLE) {
7023 
7024 		/*
7025 		 * ASYM_PACKING needs to force migrate tasks from busy but
7026 		 * higher numbered CPUs in order to pack all tasks in the
7027 		 * lowest numbered CPUs.
7028 		 */
7029 		if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7030 			return 1;
7031 	}
7032 
7033 	/*
7034 	 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7035 	 * It's worth migrating the task if the src_cpu's capacity is reduced
7036 	 * because of other sched_class or IRQs if more capacity stays
7037 	 * available on dst_cpu.
7038 	 */
7039 	if ((env->idle != CPU_NOT_IDLE) &&
7040 	    (env->src_rq->cfs.h_nr_running == 1)) {
7041 		if ((check_cpu_capacity(env->src_rq, sd)) &&
7042 		    (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7043 			return 1;
7044 	}
7045 
7046 	return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7047 }
7048 
7049 static int active_load_balance_cpu_stop(void *data);
7050 
7051 static int should_we_balance(struct lb_env *env)
7052 {
7053 	struct sched_group *sg = env->sd->groups;
7054 	struct cpumask *sg_cpus, *sg_mask;
7055 	int cpu, balance_cpu = -1;
7056 
7057 	/*
7058 	 * In the newly idle case, we will allow all the cpu's
7059 	 * to do the newly idle load balance.
7060 	 */
7061 	if (env->idle == CPU_NEWLY_IDLE)
7062 		return 1;
7063 
7064 	sg_cpus = sched_group_cpus(sg);
7065 	sg_mask = sched_group_mask(sg);
7066 	/* Try to find first idle cpu */
7067 	for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7068 		if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7069 			continue;
7070 
7071 		balance_cpu = cpu;
7072 		break;
7073 	}
7074 
7075 	if (balance_cpu == -1)
7076 		balance_cpu = group_balance_cpu(sg);
7077 
7078 	/*
7079 	 * First idle cpu or the first cpu(busiest) in this sched group
7080 	 * is eligible for doing load balancing at this and above domains.
7081 	 */
7082 	return balance_cpu == env->dst_cpu;
7083 }
7084 
7085 /*
7086  * Check this_cpu to ensure it is balanced within domain. Attempt to move
7087  * tasks if there is an imbalance.
7088  */
7089 static int load_balance(int this_cpu, struct rq *this_rq,
7090 			struct sched_domain *sd, enum cpu_idle_type idle,
7091 			int *continue_balancing)
7092 {
7093 	int ld_moved, cur_ld_moved, active_balance = 0;
7094 	struct sched_domain *sd_parent = sd->parent;
7095 	struct sched_group *group;
7096 	struct rq *busiest;
7097 	unsigned long flags;
7098 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7099 
7100 	struct lb_env env = {
7101 		.sd		= sd,
7102 		.dst_cpu	= this_cpu,
7103 		.dst_rq		= this_rq,
7104 		.dst_grpmask    = sched_group_cpus(sd->groups),
7105 		.idle		= idle,
7106 		.loop_break	= sched_nr_migrate_break,
7107 		.cpus		= cpus,
7108 		.fbq_type	= all,
7109 		.tasks		= LIST_HEAD_INIT(env.tasks),
7110 	};
7111 
7112 	/*
7113 	 * For NEWLY_IDLE load_balancing, we don't need to consider
7114 	 * other cpus in our group
7115 	 */
7116 	if (idle == CPU_NEWLY_IDLE)
7117 		env.dst_grpmask = NULL;
7118 
7119 	cpumask_copy(cpus, cpu_active_mask);
7120 
7121 	schedstat_inc(sd, lb_count[idle]);
7122 
7123 redo:
7124 	if (!should_we_balance(&env)) {
7125 		*continue_balancing = 0;
7126 		goto out_balanced;
7127 	}
7128 
7129 	group = find_busiest_group(&env);
7130 	if (!group) {
7131 		schedstat_inc(sd, lb_nobusyg[idle]);
7132 		goto out_balanced;
7133 	}
7134 
7135 	busiest = find_busiest_queue(&env, group);
7136 	if (!busiest) {
7137 		schedstat_inc(sd, lb_nobusyq[idle]);
7138 		goto out_balanced;
7139 	}
7140 
7141 	BUG_ON(busiest == env.dst_rq);
7142 
7143 	schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7144 
7145 	env.src_cpu = busiest->cpu;
7146 	env.src_rq = busiest;
7147 
7148 	ld_moved = 0;
7149 	if (busiest->nr_running > 1) {
7150 		/*
7151 		 * Attempt to move tasks. If find_busiest_group has found
7152 		 * an imbalance but busiest->nr_running <= 1, the group is
7153 		 * still unbalanced. ld_moved simply stays zero, so it is
7154 		 * correctly treated as an imbalance.
7155 		 */
7156 		env.flags |= LBF_ALL_PINNED;
7157 		env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
7158 
7159 more_balance:
7160 		raw_spin_lock_irqsave(&busiest->lock, flags);
7161 
7162 		/*
7163 		 * cur_ld_moved - load moved in current iteration
7164 		 * ld_moved     - cumulative load moved across iterations
7165 		 */
7166 		cur_ld_moved = detach_tasks(&env);
7167 
7168 		/*
7169 		 * We've detached some tasks from busiest_rq. Every
7170 		 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7171 		 * unlock busiest->lock, and we are able to be sure
7172 		 * that nobody can manipulate the tasks in parallel.
7173 		 * See task_rq_lock() family for the details.
7174 		 */
7175 
7176 		raw_spin_unlock(&busiest->lock);
7177 
7178 		if (cur_ld_moved) {
7179 			attach_tasks(&env);
7180 			ld_moved += cur_ld_moved;
7181 		}
7182 
7183 		local_irq_restore(flags);
7184 
7185 		if (env.flags & LBF_NEED_BREAK) {
7186 			env.flags &= ~LBF_NEED_BREAK;
7187 			goto more_balance;
7188 		}
7189 
7190 		/*
7191 		 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7192 		 * us and move them to an alternate dst_cpu in our sched_group
7193 		 * where they can run. The upper limit on how many times we
7194 		 * iterate on same src_cpu is dependent on number of cpus in our
7195 		 * sched_group.
7196 		 *
7197 		 * This changes load balance semantics a bit on who can move
7198 		 * load to a given_cpu. In addition to the given_cpu itself
7199 		 * (or a ilb_cpu acting on its behalf where given_cpu is
7200 		 * nohz-idle), we now have balance_cpu in a position to move
7201 		 * load to given_cpu. In rare situations, this may cause
7202 		 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7203 		 * _independently_ and at _same_ time to move some load to
7204 		 * given_cpu) causing exceess load to be moved to given_cpu.
7205 		 * This however should not happen so much in practice and
7206 		 * moreover subsequent load balance cycles should correct the
7207 		 * excess load moved.
7208 		 */
7209 		if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7210 
7211 			/* Prevent to re-select dst_cpu via env's cpus */
7212 			cpumask_clear_cpu(env.dst_cpu, env.cpus);
7213 
7214 			env.dst_rq	 = cpu_rq(env.new_dst_cpu);
7215 			env.dst_cpu	 = env.new_dst_cpu;
7216 			env.flags	&= ~LBF_DST_PINNED;
7217 			env.loop	 = 0;
7218 			env.loop_break	 = sched_nr_migrate_break;
7219 
7220 			/*
7221 			 * Go back to "more_balance" rather than "redo" since we
7222 			 * need to continue with same src_cpu.
7223 			 */
7224 			goto more_balance;
7225 		}
7226 
7227 		/*
7228 		 * We failed to reach balance because of affinity.
7229 		 */
7230 		if (sd_parent) {
7231 			int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7232 
7233 			if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7234 				*group_imbalance = 1;
7235 		}
7236 
7237 		/* All tasks on this runqueue were pinned by CPU affinity */
7238 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
7239 			cpumask_clear_cpu(cpu_of(busiest), cpus);
7240 			if (!cpumask_empty(cpus)) {
7241 				env.loop = 0;
7242 				env.loop_break = sched_nr_migrate_break;
7243 				goto redo;
7244 			}
7245 			goto out_all_pinned;
7246 		}
7247 	}
7248 
7249 	if (!ld_moved) {
7250 		schedstat_inc(sd, lb_failed[idle]);
7251 		/*
7252 		 * Increment the failure counter only on periodic balance.
7253 		 * We do not want newidle balance, which can be very
7254 		 * frequent, pollute the failure counter causing
7255 		 * excessive cache_hot migrations and active balances.
7256 		 */
7257 		if (idle != CPU_NEWLY_IDLE)
7258 			sd->nr_balance_failed++;
7259 
7260 		if (need_active_balance(&env)) {
7261 			raw_spin_lock_irqsave(&busiest->lock, flags);
7262 
7263 			/* don't kick the active_load_balance_cpu_stop,
7264 			 * if the curr task on busiest cpu can't be
7265 			 * moved to this_cpu
7266 			 */
7267 			if (!cpumask_test_cpu(this_cpu,
7268 					tsk_cpus_allowed(busiest->curr))) {
7269 				raw_spin_unlock_irqrestore(&busiest->lock,
7270 							    flags);
7271 				env.flags |= LBF_ALL_PINNED;
7272 				goto out_one_pinned;
7273 			}
7274 
7275 			/*
7276 			 * ->active_balance synchronizes accesses to
7277 			 * ->active_balance_work.  Once set, it's cleared
7278 			 * only after active load balance is finished.
7279 			 */
7280 			if (!busiest->active_balance) {
7281 				busiest->active_balance = 1;
7282 				busiest->push_cpu = this_cpu;
7283 				active_balance = 1;
7284 			}
7285 			raw_spin_unlock_irqrestore(&busiest->lock, flags);
7286 
7287 			if (active_balance) {
7288 				stop_one_cpu_nowait(cpu_of(busiest),
7289 					active_load_balance_cpu_stop, busiest,
7290 					&busiest->active_balance_work);
7291 			}
7292 
7293 			/*
7294 			 * We've kicked active balancing, reset the failure
7295 			 * counter.
7296 			 */
7297 			sd->nr_balance_failed = sd->cache_nice_tries+1;
7298 		}
7299 	} else
7300 		sd->nr_balance_failed = 0;
7301 
7302 	if (likely(!active_balance)) {
7303 		/* We were unbalanced, so reset the balancing interval */
7304 		sd->balance_interval = sd->min_interval;
7305 	} else {
7306 		/*
7307 		 * If we've begun active balancing, start to back off. This
7308 		 * case may not be covered by the all_pinned logic if there
7309 		 * is only 1 task on the busy runqueue (because we don't call
7310 		 * detach_tasks).
7311 		 */
7312 		if (sd->balance_interval < sd->max_interval)
7313 			sd->balance_interval *= 2;
7314 	}
7315 
7316 	goto out;
7317 
7318 out_balanced:
7319 	/*
7320 	 * We reach balance although we may have faced some affinity
7321 	 * constraints. Clear the imbalance flag if it was set.
7322 	 */
7323 	if (sd_parent) {
7324 		int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7325 
7326 		if (*group_imbalance)
7327 			*group_imbalance = 0;
7328 	}
7329 
7330 out_all_pinned:
7331 	/*
7332 	 * We reach balance because all tasks are pinned at this level so
7333 	 * we can't migrate them. Let the imbalance flag set so parent level
7334 	 * can try to migrate them.
7335 	 */
7336 	schedstat_inc(sd, lb_balanced[idle]);
7337 
7338 	sd->nr_balance_failed = 0;
7339 
7340 out_one_pinned:
7341 	/* tune up the balancing interval */
7342 	if (((env.flags & LBF_ALL_PINNED) &&
7343 			sd->balance_interval < MAX_PINNED_INTERVAL) ||
7344 			(sd->balance_interval < sd->max_interval))
7345 		sd->balance_interval *= 2;
7346 
7347 	ld_moved = 0;
7348 out:
7349 	return ld_moved;
7350 }
7351 
7352 static inline unsigned long
7353 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7354 {
7355 	unsigned long interval = sd->balance_interval;
7356 
7357 	if (cpu_busy)
7358 		interval *= sd->busy_factor;
7359 
7360 	/* scale ms to jiffies */
7361 	interval = msecs_to_jiffies(interval);
7362 	interval = clamp(interval, 1UL, max_load_balance_interval);
7363 
7364 	return interval;
7365 }
7366 
7367 static inline void
7368 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7369 {
7370 	unsigned long interval, next;
7371 
7372 	interval = get_sd_balance_interval(sd, cpu_busy);
7373 	next = sd->last_balance + interval;
7374 
7375 	if (time_after(*next_balance, next))
7376 		*next_balance = next;
7377 }
7378 
7379 /*
7380  * idle_balance is called by schedule() if this_cpu is about to become
7381  * idle. Attempts to pull tasks from other CPUs.
7382  */
7383 static int idle_balance(struct rq *this_rq)
7384 {
7385 	unsigned long next_balance = jiffies + HZ;
7386 	int this_cpu = this_rq->cpu;
7387 	struct sched_domain *sd;
7388 	int pulled_task = 0;
7389 	u64 curr_cost = 0;
7390 
7391 	/*
7392 	 * We must set idle_stamp _before_ calling idle_balance(), such that we
7393 	 * measure the duration of idle_balance() as idle time.
7394 	 */
7395 	this_rq->idle_stamp = rq_clock(this_rq);
7396 
7397 	if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7398 	    !this_rq->rd->overload) {
7399 		rcu_read_lock();
7400 		sd = rcu_dereference_check_sched_domain(this_rq->sd);
7401 		if (sd)
7402 			update_next_balance(sd, 0, &next_balance);
7403 		rcu_read_unlock();
7404 
7405 		goto out;
7406 	}
7407 
7408 	raw_spin_unlock(&this_rq->lock);
7409 
7410 	update_blocked_averages(this_cpu);
7411 	rcu_read_lock();
7412 	for_each_domain(this_cpu, sd) {
7413 		int continue_balancing = 1;
7414 		u64 t0, domain_cost;
7415 
7416 		if (!(sd->flags & SD_LOAD_BALANCE))
7417 			continue;
7418 
7419 		if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7420 			update_next_balance(sd, 0, &next_balance);
7421 			break;
7422 		}
7423 
7424 		if (sd->flags & SD_BALANCE_NEWIDLE) {
7425 			t0 = sched_clock_cpu(this_cpu);
7426 
7427 			pulled_task = load_balance(this_cpu, this_rq,
7428 						   sd, CPU_NEWLY_IDLE,
7429 						   &continue_balancing);
7430 
7431 			domain_cost = sched_clock_cpu(this_cpu) - t0;
7432 			if (domain_cost > sd->max_newidle_lb_cost)
7433 				sd->max_newidle_lb_cost = domain_cost;
7434 
7435 			curr_cost += domain_cost;
7436 		}
7437 
7438 		update_next_balance(sd, 0, &next_balance);
7439 
7440 		/*
7441 		 * Stop searching for tasks to pull if there are
7442 		 * now runnable tasks on this rq.
7443 		 */
7444 		if (pulled_task || this_rq->nr_running > 0)
7445 			break;
7446 	}
7447 	rcu_read_unlock();
7448 
7449 	raw_spin_lock(&this_rq->lock);
7450 
7451 	if (curr_cost > this_rq->max_idle_balance_cost)
7452 		this_rq->max_idle_balance_cost = curr_cost;
7453 
7454 	/*
7455 	 * While browsing the domains, we released the rq lock, a task could
7456 	 * have been enqueued in the meantime. Since we're not going idle,
7457 	 * pretend we pulled a task.
7458 	 */
7459 	if (this_rq->cfs.h_nr_running && !pulled_task)
7460 		pulled_task = 1;
7461 
7462 out:
7463 	/* Move the next balance forward */
7464 	if (time_after(this_rq->next_balance, next_balance))
7465 		this_rq->next_balance = next_balance;
7466 
7467 	/* Is there a task of a high priority class? */
7468 	if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7469 		pulled_task = -1;
7470 
7471 	if (pulled_task)
7472 		this_rq->idle_stamp = 0;
7473 
7474 	return pulled_task;
7475 }
7476 
7477 /*
7478  * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7479  * running tasks off the busiest CPU onto idle CPUs. It requires at
7480  * least 1 task to be running on each physical CPU where possible, and
7481  * avoids physical / logical imbalances.
7482  */
7483 static int active_load_balance_cpu_stop(void *data)
7484 {
7485 	struct rq *busiest_rq = data;
7486 	int busiest_cpu = cpu_of(busiest_rq);
7487 	int target_cpu = busiest_rq->push_cpu;
7488 	struct rq *target_rq = cpu_rq(target_cpu);
7489 	struct sched_domain *sd;
7490 	struct task_struct *p = NULL;
7491 
7492 	raw_spin_lock_irq(&busiest_rq->lock);
7493 
7494 	/* make sure the requested cpu hasn't gone down in the meantime */
7495 	if (unlikely(busiest_cpu != smp_processor_id() ||
7496 		     !busiest_rq->active_balance))
7497 		goto out_unlock;
7498 
7499 	/* Is there any task to move? */
7500 	if (busiest_rq->nr_running <= 1)
7501 		goto out_unlock;
7502 
7503 	/*
7504 	 * This condition is "impossible", if it occurs
7505 	 * we need to fix it. Originally reported by
7506 	 * Bjorn Helgaas on a 128-cpu setup.
7507 	 */
7508 	BUG_ON(busiest_rq == target_rq);
7509 
7510 	/* Search for an sd spanning us and the target CPU. */
7511 	rcu_read_lock();
7512 	for_each_domain(target_cpu, sd) {
7513 		if ((sd->flags & SD_LOAD_BALANCE) &&
7514 		    cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7515 				break;
7516 	}
7517 
7518 	if (likely(sd)) {
7519 		struct lb_env env = {
7520 			.sd		= sd,
7521 			.dst_cpu	= target_cpu,
7522 			.dst_rq		= target_rq,
7523 			.src_cpu	= busiest_rq->cpu,
7524 			.src_rq		= busiest_rq,
7525 			.idle		= CPU_IDLE,
7526 		};
7527 
7528 		schedstat_inc(sd, alb_count);
7529 
7530 		p = detach_one_task(&env);
7531 		if (p)
7532 			schedstat_inc(sd, alb_pushed);
7533 		else
7534 			schedstat_inc(sd, alb_failed);
7535 	}
7536 	rcu_read_unlock();
7537 out_unlock:
7538 	busiest_rq->active_balance = 0;
7539 	raw_spin_unlock(&busiest_rq->lock);
7540 
7541 	if (p)
7542 		attach_one_task(target_rq, p);
7543 
7544 	local_irq_enable();
7545 
7546 	return 0;
7547 }
7548 
7549 static inline int on_null_domain(struct rq *rq)
7550 {
7551 	return unlikely(!rcu_dereference_sched(rq->sd));
7552 }
7553 
7554 #ifdef CONFIG_NO_HZ_COMMON
7555 /*
7556  * idle load balancing details
7557  * - When one of the busy CPUs notice that there may be an idle rebalancing
7558  *   needed, they will kick the idle load balancer, which then does idle
7559  *   load balancing for all the idle CPUs.
7560  */
7561 static struct {
7562 	cpumask_var_t idle_cpus_mask;
7563 	atomic_t nr_cpus;
7564 	unsigned long next_balance;     /* in jiffy units */
7565 } nohz ____cacheline_aligned;
7566 
7567 static inline int find_new_ilb(void)
7568 {
7569 	int ilb = cpumask_first(nohz.idle_cpus_mask);
7570 
7571 	if (ilb < nr_cpu_ids && idle_cpu(ilb))
7572 		return ilb;
7573 
7574 	return nr_cpu_ids;
7575 }
7576 
7577 /*
7578  * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7579  * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7580  * CPU (if there is one).
7581  */
7582 static void nohz_balancer_kick(void)
7583 {
7584 	int ilb_cpu;
7585 
7586 	nohz.next_balance++;
7587 
7588 	ilb_cpu = find_new_ilb();
7589 
7590 	if (ilb_cpu >= nr_cpu_ids)
7591 		return;
7592 
7593 	if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7594 		return;
7595 	/*
7596 	 * Use smp_send_reschedule() instead of resched_cpu().
7597 	 * This way we generate a sched IPI on the target cpu which
7598 	 * is idle. And the softirq performing nohz idle load balance
7599 	 * will be run before returning from the IPI.
7600 	 */
7601 	smp_send_reschedule(ilb_cpu);
7602 	return;
7603 }
7604 
7605 static inline void nohz_balance_exit_idle(int cpu)
7606 {
7607 	if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7608 		/*
7609 		 * Completely isolated CPUs don't ever set, so we must test.
7610 		 */
7611 		if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7612 			cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7613 			atomic_dec(&nohz.nr_cpus);
7614 		}
7615 		clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7616 	}
7617 }
7618 
7619 static inline void set_cpu_sd_state_busy(void)
7620 {
7621 	struct sched_domain *sd;
7622 	int cpu = smp_processor_id();
7623 
7624 	rcu_read_lock();
7625 	sd = rcu_dereference(per_cpu(sd_busy, cpu));
7626 
7627 	if (!sd || !sd->nohz_idle)
7628 		goto unlock;
7629 	sd->nohz_idle = 0;
7630 
7631 	atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7632 unlock:
7633 	rcu_read_unlock();
7634 }
7635 
7636 void set_cpu_sd_state_idle(void)
7637 {
7638 	struct sched_domain *sd;
7639 	int cpu = smp_processor_id();
7640 
7641 	rcu_read_lock();
7642 	sd = rcu_dereference(per_cpu(sd_busy, cpu));
7643 
7644 	if (!sd || sd->nohz_idle)
7645 		goto unlock;
7646 	sd->nohz_idle = 1;
7647 
7648 	atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7649 unlock:
7650 	rcu_read_unlock();
7651 }
7652 
7653 /*
7654  * This routine will record that the cpu is going idle with tick stopped.
7655  * This info will be used in performing idle load balancing in the future.
7656  */
7657 void nohz_balance_enter_idle(int cpu)
7658 {
7659 	/*
7660 	 * If this cpu is going down, then nothing needs to be done.
7661 	 */
7662 	if (!cpu_active(cpu))
7663 		return;
7664 
7665 	if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7666 		return;
7667 
7668 	/*
7669 	 * If we're a completely isolated CPU, we don't play.
7670 	 */
7671 	if (on_null_domain(cpu_rq(cpu)))
7672 		return;
7673 
7674 	cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7675 	atomic_inc(&nohz.nr_cpus);
7676 	set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7677 }
7678 
7679 static int sched_ilb_notifier(struct notifier_block *nfb,
7680 					unsigned long action, void *hcpu)
7681 {
7682 	switch (action & ~CPU_TASKS_FROZEN) {
7683 	case CPU_DYING:
7684 		nohz_balance_exit_idle(smp_processor_id());
7685 		return NOTIFY_OK;
7686 	default:
7687 		return NOTIFY_DONE;
7688 	}
7689 }
7690 #endif
7691 
7692 static DEFINE_SPINLOCK(balancing);
7693 
7694 /*
7695  * Scale the max load_balance interval with the number of CPUs in the system.
7696  * This trades load-balance latency on larger machines for less cross talk.
7697  */
7698 void update_max_interval(void)
7699 {
7700 	max_load_balance_interval = HZ*num_online_cpus()/10;
7701 }
7702 
7703 /*
7704  * It checks each scheduling domain to see if it is due to be balanced,
7705  * and initiates a balancing operation if so.
7706  *
7707  * Balancing parameters are set up in init_sched_domains.
7708  */
7709 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7710 {
7711 	int continue_balancing = 1;
7712 	int cpu = rq->cpu;
7713 	unsigned long interval;
7714 	struct sched_domain *sd;
7715 	/* Earliest time when we have to do rebalance again */
7716 	unsigned long next_balance = jiffies + 60*HZ;
7717 	int update_next_balance = 0;
7718 	int need_serialize, need_decay = 0;
7719 	u64 max_cost = 0;
7720 
7721 	update_blocked_averages(cpu);
7722 
7723 	rcu_read_lock();
7724 	for_each_domain(cpu, sd) {
7725 		/*
7726 		 * Decay the newidle max times here because this is a regular
7727 		 * visit to all the domains. Decay ~1% per second.
7728 		 */
7729 		if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7730 			sd->max_newidle_lb_cost =
7731 				(sd->max_newidle_lb_cost * 253) / 256;
7732 			sd->next_decay_max_lb_cost = jiffies + HZ;
7733 			need_decay = 1;
7734 		}
7735 		max_cost += sd->max_newidle_lb_cost;
7736 
7737 		if (!(sd->flags & SD_LOAD_BALANCE))
7738 			continue;
7739 
7740 		/*
7741 		 * Stop the load balance at this level. There is another
7742 		 * CPU in our sched group which is doing load balancing more
7743 		 * actively.
7744 		 */
7745 		if (!continue_balancing) {
7746 			if (need_decay)
7747 				continue;
7748 			break;
7749 		}
7750 
7751 		interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7752 
7753 		need_serialize = sd->flags & SD_SERIALIZE;
7754 		if (need_serialize) {
7755 			if (!spin_trylock(&balancing))
7756 				goto out;
7757 		}
7758 
7759 		if (time_after_eq(jiffies, sd->last_balance + interval)) {
7760 			if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7761 				/*
7762 				 * The LBF_DST_PINNED logic could have changed
7763 				 * env->dst_cpu, so we can't know our idle
7764 				 * state even if we migrated tasks. Update it.
7765 				 */
7766 				idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7767 			}
7768 			sd->last_balance = jiffies;
7769 			interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7770 		}
7771 		if (need_serialize)
7772 			spin_unlock(&balancing);
7773 out:
7774 		if (time_after(next_balance, sd->last_balance + interval)) {
7775 			next_balance = sd->last_balance + interval;
7776 			update_next_balance = 1;
7777 		}
7778 	}
7779 	if (need_decay) {
7780 		/*
7781 		 * Ensure the rq-wide value also decays but keep it at a
7782 		 * reasonable floor to avoid funnies with rq->avg_idle.
7783 		 */
7784 		rq->max_idle_balance_cost =
7785 			max((u64)sysctl_sched_migration_cost, max_cost);
7786 	}
7787 	rcu_read_unlock();
7788 
7789 	/*
7790 	 * next_balance will be updated only when there is a need.
7791 	 * When the cpu is attached to null domain for ex, it will not be
7792 	 * updated.
7793 	 */
7794 	if (likely(update_next_balance)) {
7795 		rq->next_balance = next_balance;
7796 
7797 #ifdef CONFIG_NO_HZ_COMMON
7798 		/*
7799 		 * If this CPU has been elected to perform the nohz idle
7800 		 * balance. Other idle CPUs have already rebalanced with
7801 		 * nohz_idle_balance() and nohz.next_balance has been
7802 		 * updated accordingly. This CPU is now running the idle load
7803 		 * balance for itself and we need to update the
7804 		 * nohz.next_balance accordingly.
7805 		 */
7806 		if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7807 			nohz.next_balance = rq->next_balance;
7808 #endif
7809 	}
7810 }
7811 
7812 #ifdef CONFIG_NO_HZ_COMMON
7813 /*
7814  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7815  * rebalancing for all the cpus for whom scheduler ticks are stopped.
7816  */
7817 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7818 {
7819 	int this_cpu = this_rq->cpu;
7820 	struct rq *rq;
7821 	int balance_cpu;
7822 	/* Earliest time when we have to do rebalance again */
7823 	unsigned long next_balance = jiffies + 60*HZ;
7824 	int update_next_balance = 0;
7825 
7826 	if (idle != CPU_IDLE ||
7827 	    !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7828 		goto end;
7829 
7830 	for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7831 		if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7832 			continue;
7833 
7834 		/*
7835 		 * If this cpu gets work to do, stop the load balancing
7836 		 * work being done for other cpus. Next load
7837 		 * balancing owner will pick it up.
7838 		 */
7839 		if (need_resched())
7840 			break;
7841 
7842 		rq = cpu_rq(balance_cpu);
7843 
7844 		/*
7845 		 * If time for next balance is due,
7846 		 * do the balance.
7847 		 */
7848 		if (time_after_eq(jiffies, rq->next_balance)) {
7849 			raw_spin_lock_irq(&rq->lock);
7850 			update_rq_clock(rq);
7851 			update_idle_cpu_load(rq);
7852 			raw_spin_unlock_irq(&rq->lock);
7853 			rebalance_domains(rq, CPU_IDLE);
7854 		}
7855 
7856 		if (time_after(next_balance, rq->next_balance)) {
7857 			next_balance = rq->next_balance;
7858 			update_next_balance = 1;
7859 		}
7860 	}
7861 
7862 	/*
7863 	 * next_balance will be updated only when there is a need.
7864 	 * When the CPU is attached to null domain for ex, it will not be
7865 	 * updated.
7866 	 */
7867 	if (likely(update_next_balance))
7868 		nohz.next_balance = next_balance;
7869 end:
7870 	clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7871 }
7872 
7873 /*
7874  * Current heuristic for kicking the idle load balancer in the presence
7875  * of an idle cpu in the system.
7876  *   - This rq has more than one task.
7877  *   - This rq has at least one CFS task and the capacity of the CPU is
7878  *     significantly reduced because of RT tasks or IRQs.
7879  *   - At parent of LLC scheduler domain level, this cpu's scheduler group has
7880  *     multiple busy cpu.
7881  *   - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7882  *     domain span are idle.
7883  */
7884 static inline bool nohz_kick_needed(struct rq *rq)
7885 {
7886 	unsigned long now = jiffies;
7887 	struct sched_domain *sd;
7888 	struct sched_group_capacity *sgc;
7889 	int nr_busy, cpu = rq->cpu;
7890 	bool kick = false;
7891 
7892 	if (unlikely(rq->idle_balance))
7893 		return false;
7894 
7895        /*
7896 	* We may be recently in ticked or tickless idle mode. At the first
7897 	* busy tick after returning from idle, we will update the busy stats.
7898 	*/
7899 	set_cpu_sd_state_busy();
7900 	nohz_balance_exit_idle(cpu);
7901 
7902 	/*
7903 	 * None are in tickless mode and hence no need for NOHZ idle load
7904 	 * balancing.
7905 	 */
7906 	if (likely(!atomic_read(&nohz.nr_cpus)))
7907 		return false;
7908 
7909 	if (time_before(now, nohz.next_balance))
7910 		return false;
7911 
7912 	if (rq->nr_running >= 2)
7913 		return true;
7914 
7915 	rcu_read_lock();
7916 	sd = rcu_dereference(per_cpu(sd_busy, cpu));
7917 	if (sd) {
7918 		sgc = sd->groups->sgc;
7919 		nr_busy = atomic_read(&sgc->nr_busy_cpus);
7920 
7921 		if (nr_busy > 1) {
7922 			kick = true;
7923 			goto unlock;
7924 		}
7925 
7926 	}
7927 
7928 	sd = rcu_dereference(rq->sd);
7929 	if (sd) {
7930 		if ((rq->cfs.h_nr_running >= 1) &&
7931 				check_cpu_capacity(rq, sd)) {
7932 			kick = true;
7933 			goto unlock;
7934 		}
7935 	}
7936 
7937 	sd = rcu_dereference(per_cpu(sd_asym, cpu));
7938 	if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7939 				  sched_domain_span(sd)) < cpu)) {
7940 		kick = true;
7941 		goto unlock;
7942 	}
7943 
7944 unlock:
7945 	rcu_read_unlock();
7946 	return kick;
7947 }
7948 #else
7949 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7950 #endif
7951 
7952 /*
7953  * run_rebalance_domains is triggered when needed from the scheduler tick.
7954  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7955  */
7956 static void run_rebalance_domains(struct softirq_action *h)
7957 {
7958 	struct rq *this_rq = this_rq();
7959 	enum cpu_idle_type idle = this_rq->idle_balance ?
7960 						CPU_IDLE : CPU_NOT_IDLE;
7961 
7962 	/*
7963 	 * If this cpu has a pending nohz_balance_kick, then do the
7964 	 * balancing on behalf of the other idle cpus whose ticks are
7965 	 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7966 	 * give the idle cpus a chance to load balance. Else we may
7967 	 * load balance only within the local sched_domain hierarchy
7968 	 * and abort nohz_idle_balance altogether if we pull some load.
7969 	 */
7970 	nohz_idle_balance(this_rq, idle);
7971 	rebalance_domains(this_rq, idle);
7972 }
7973 
7974 /*
7975  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7976  */
7977 void trigger_load_balance(struct rq *rq)
7978 {
7979 	/* Don't need to rebalance while attached to NULL domain */
7980 	if (unlikely(on_null_domain(rq)))
7981 		return;
7982 
7983 	if (time_after_eq(jiffies, rq->next_balance))
7984 		raise_softirq(SCHED_SOFTIRQ);
7985 #ifdef CONFIG_NO_HZ_COMMON
7986 	if (nohz_kick_needed(rq))
7987 		nohz_balancer_kick();
7988 #endif
7989 }
7990 
7991 static void rq_online_fair(struct rq *rq)
7992 {
7993 	update_sysctl();
7994 
7995 	update_runtime_enabled(rq);
7996 }
7997 
7998 static void rq_offline_fair(struct rq *rq)
7999 {
8000 	update_sysctl();
8001 
8002 	/* Ensure any throttled groups are reachable by pick_next_task */
8003 	unthrottle_offline_cfs_rqs(rq);
8004 }
8005 
8006 #endif /* CONFIG_SMP */
8007 
8008 /*
8009  * scheduler tick hitting a task of our scheduling class:
8010  */
8011 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8012 {
8013 	struct cfs_rq *cfs_rq;
8014 	struct sched_entity *se = &curr->se;
8015 
8016 	for_each_sched_entity(se) {
8017 		cfs_rq = cfs_rq_of(se);
8018 		entity_tick(cfs_rq, se, queued);
8019 	}
8020 
8021 	if (static_branch_unlikely(&sched_numa_balancing))
8022 		task_tick_numa(rq, curr);
8023 }
8024 
8025 /*
8026  * called on fork with the child task as argument from the parent's context
8027  *  - child not yet on the tasklist
8028  *  - preemption disabled
8029  */
8030 static void task_fork_fair(struct task_struct *p)
8031 {
8032 	struct cfs_rq *cfs_rq;
8033 	struct sched_entity *se = &p->se, *curr;
8034 	int this_cpu = smp_processor_id();
8035 	struct rq *rq = this_rq();
8036 	unsigned long flags;
8037 
8038 	raw_spin_lock_irqsave(&rq->lock, flags);
8039 
8040 	update_rq_clock(rq);
8041 
8042 	cfs_rq = task_cfs_rq(current);
8043 	curr = cfs_rq->curr;
8044 
8045 	/*
8046 	 * Not only the cpu but also the task_group of the parent might have
8047 	 * been changed after parent->se.parent,cfs_rq were copied to
8048 	 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8049 	 * of child point to valid ones.
8050 	 */
8051 	rcu_read_lock();
8052 	__set_task_cpu(p, this_cpu);
8053 	rcu_read_unlock();
8054 
8055 	update_curr(cfs_rq);
8056 
8057 	if (curr)
8058 		se->vruntime = curr->vruntime;
8059 	place_entity(cfs_rq, se, 1);
8060 
8061 	if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8062 		/*
8063 		 * Upon rescheduling, sched_class::put_prev_task() will place
8064 		 * 'current' within the tree based on its new key value.
8065 		 */
8066 		swap(curr->vruntime, se->vruntime);
8067 		resched_curr(rq);
8068 	}
8069 
8070 	se->vruntime -= cfs_rq->min_vruntime;
8071 
8072 	raw_spin_unlock_irqrestore(&rq->lock, flags);
8073 }
8074 
8075 /*
8076  * Priority of the task has changed. Check to see if we preempt
8077  * the current task.
8078  */
8079 static void
8080 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8081 {
8082 	if (!task_on_rq_queued(p))
8083 		return;
8084 
8085 	/*
8086 	 * Reschedule if we are currently running on this runqueue and
8087 	 * our priority decreased, or if we are not currently running on
8088 	 * this runqueue and our priority is higher than the current's
8089 	 */
8090 	if (rq->curr == p) {
8091 		if (p->prio > oldprio)
8092 			resched_curr(rq);
8093 	} else
8094 		check_preempt_curr(rq, p, 0);
8095 }
8096 
8097 static inline bool vruntime_normalized(struct task_struct *p)
8098 {
8099 	struct sched_entity *se = &p->se;
8100 
8101 	/*
8102 	 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8103 	 * the dequeue_entity(.flags=0) will already have normalized the
8104 	 * vruntime.
8105 	 */
8106 	if (p->on_rq)
8107 		return true;
8108 
8109 	/*
8110 	 * When !on_rq, vruntime of the task has usually NOT been normalized.
8111 	 * But there are some cases where it has already been normalized:
8112 	 *
8113 	 * - A forked child which is waiting for being woken up by
8114 	 *   wake_up_new_task().
8115 	 * - A task which has been woken up by try_to_wake_up() and
8116 	 *   waiting for actually being woken up by sched_ttwu_pending().
8117 	 */
8118 	if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8119 		return true;
8120 
8121 	return false;
8122 }
8123 
8124 static void detach_task_cfs_rq(struct task_struct *p)
8125 {
8126 	struct sched_entity *se = &p->se;
8127 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
8128 
8129 	if (!vruntime_normalized(p)) {
8130 		/*
8131 		 * Fix up our vruntime so that the current sleep doesn't
8132 		 * cause 'unlimited' sleep bonus.
8133 		 */
8134 		place_entity(cfs_rq, se, 0);
8135 		se->vruntime -= cfs_rq->min_vruntime;
8136 	}
8137 
8138 	/* Catch up with the cfs_rq and remove our load when we leave */
8139 	detach_entity_load_avg(cfs_rq, se);
8140 }
8141 
8142 static void attach_task_cfs_rq(struct task_struct *p)
8143 {
8144 	struct sched_entity *se = &p->se;
8145 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
8146 
8147 #ifdef CONFIG_FAIR_GROUP_SCHED
8148 	/*
8149 	 * Since the real-depth could have been changed (only FAIR
8150 	 * class maintain depth value), reset depth properly.
8151 	 */
8152 	se->depth = se->parent ? se->parent->depth + 1 : 0;
8153 #endif
8154 
8155 	/* Synchronize task with its cfs_rq */
8156 	attach_entity_load_avg(cfs_rq, se);
8157 
8158 	if (!vruntime_normalized(p))
8159 		se->vruntime += cfs_rq->min_vruntime;
8160 }
8161 
8162 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8163 {
8164 	detach_task_cfs_rq(p);
8165 }
8166 
8167 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8168 {
8169 	attach_task_cfs_rq(p);
8170 
8171 	if (task_on_rq_queued(p)) {
8172 		/*
8173 		 * We were most likely switched from sched_rt, so
8174 		 * kick off the schedule if running, otherwise just see
8175 		 * if we can still preempt the current task.
8176 		 */
8177 		if (rq->curr == p)
8178 			resched_curr(rq);
8179 		else
8180 			check_preempt_curr(rq, p, 0);
8181 	}
8182 }
8183 
8184 /* Account for a task changing its policy or group.
8185  *
8186  * This routine is mostly called to set cfs_rq->curr field when a task
8187  * migrates between groups/classes.
8188  */
8189 static void set_curr_task_fair(struct rq *rq)
8190 {
8191 	struct sched_entity *se = &rq->curr->se;
8192 
8193 	for_each_sched_entity(se) {
8194 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
8195 
8196 		set_next_entity(cfs_rq, se);
8197 		/* ensure bandwidth has been allocated on our new cfs_rq */
8198 		account_cfs_rq_runtime(cfs_rq, 0);
8199 	}
8200 }
8201 
8202 void init_cfs_rq(struct cfs_rq *cfs_rq)
8203 {
8204 	cfs_rq->tasks_timeline = RB_ROOT;
8205 	cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8206 #ifndef CONFIG_64BIT
8207 	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8208 #endif
8209 #ifdef CONFIG_SMP
8210 	atomic_long_set(&cfs_rq->removed_load_avg, 0);
8211 	atomic_long_set(&cfs_rq->removed_util_avg, 0);
8212 #endif
8213 }
8214 
8215 #ifdef CONFIG_FAIR_GROUP_SCHED
8216 static void task_move_group_fair(struct task_struct *p)
8217 {
8218 	detach_task_cfs_rq(p);
8219 	set_task_rq(p, task_cpu(p));
8220 
8221 #ifdef CONFIG_SMP
8222 	/* Tell se's cfs_rq has been changed -- migrated */
8223 	p->se.avg.last_update_time = 0;
8224 #endif
8225 	attach_task_cfs_rq(p);
8226 }
8227 
8228 void free_fair_sched_group(struct task_group *tg)
8229 {
8230 	int i;
8231 
8232 	destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8233 
8234 	for_each_possible_cpu(i) {
8235 		if (tg->cfs_rq)
8236 			kfree(tg->cfs_rq[i]);
8237 		if (tg->se) {
8238 			if (tg->se[i])
8239 				remove_entity_load_avg(tg->se[i]);
8240 			kfree(tg->se[i]);
8241 		}
8242 	}
8243 
8244 	kfree(tg->cfs_rq);
8245 	kfree(tg->se);
8246 }
8247 
8248 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8249 {
8250 	struct cfs_rq *cfs_rq;
8251 	struct sched_entity *se;
8252 	int i;
8253 
8254 	tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8255 	if (!tg->cfs_rq)
8256 		goto err;
8257 	tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8258 	if (!tg->se)
8259 		goto err;
8260 
8261 	tg->shares = NICE_0_LOAD;
8262 
8263 	init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8264 
8265 	for_each_possible_cpu(i) {
8266 		cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8267 				      GFP_KERNEL, cpu_to_node(i));
8268 		if (!cfs_rq)
8269 			goto err;
8270 
8271 		se = kzalloc_node(sizeof(struct sched_entity),
8272 				  GFP_KERNEL, cpu_to_node(i));
8273 		if (!se)
8274 			goto err_free_rq;
8275 
8276 		init_cfs_rq(cfs_rq);
8277 		init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8278 		init_entity_runnable_average(se);
8279 	}
8280 
8281 	return 1;
8282 
8283 err_free_rq:
8284 	kfree(cfs_rq);
8285 err:
8286 	return 0;
8287 }
8288 
8289 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8290 {
8291 	struct rq *rq = cpu_rq(cpu);
8292 	unsigned long flags;
8293 
8294 	/*
8295 	* Only empty task groups can be destroyed; so we can speculatively
8296 	* check on_list without danger of it being re-added.
8297 	*/
8298 	if (!tg->cfs_rq[cpu]->on_list)
8299 		return;
8300 
8301 	raw_spin_lock_irqsave(&rq->lock, flags);
8302 	list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8303 	raw_spin_unlock_irqrestore(&rq->lock, flags);
8304 }
8305 
8306 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8307 			struct sched_entity *se, int cpu,
8308 			struct sched_entity *parent)
8309 {
8310 	struct rq *rq = cpu_rq(cpu);
8311 
8312 	cfs_rq->tg = tg;
8313 	cfs_rq->rq = rq;
8314 	init_cfs_rq_runtime(cfs_rq);
8315 
8316 	tg->cfs_rq[cpu] = cfs_rq;
8317 	tg->se[cpu] = se;
8318 
8319 	/* se could be NULL for root_task_group */
8320 	if (!se)
8321 		return;
8322 
8323 	if (!parent) {
8324 		se->cfs_rq = &rq->cfs;
8325 		se->depth = 0;
8326 	} else {
8327 		se->cfs_rq = parent->my_q;
8328 		se->depth = parent->depth + 1;
8329 	}
8330 
8331 	se->my_q = cfs_rq;
8332 	/* guarantee group entities always have weight */
8333 	update_load_set(&se->load, NICE_0_LOAD);
8334 	se->parent = parent;
8335 }
8336 
8337 static DEFINE_MUTEX(shares_mutex);
8338 
8339 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8340 {
8341 	int i;
8342 	unsigned long flags;
8343 
8344 	/*
8345 	 * We can't change the weight of the root cgroup.
8346 	 */
8347 	if (!tg->se[0])
8348 		return -EINVAL;
8349 
8350 	shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8351 
8352 	mutex_lock(&shares_mutex);
8353 	if (tg->shares == shares)
8354 		goto done;
8355 
8356 	tg->shares = shares;
8357 	for_each_possible_cpu(i) {
8358 		struct rq *rq = cpu_rq(i);
8359 		struct sched_entity *se;
8360 
8361 		se = tg->se[i];
8362 		/* Propagate contribution to hierarchy */
8363 		raw_spin_lock_irqsave(&rq->lock, flags);
8364 
8365 		/* Possible calls to update_curr() need rq clock */
8366 		update_rq_clock(rq);
8367 		for_each_sched_entity(se)
8368 			update_cfs_shares(group_cfs_rq(se));
8369 		raw_spin_unlock_irqrestore(&rq->lock, flags);
8370 	}
8371 
8372 done:
8373 	mutex_unlock(&shares_mutex);
8374 	return 0;
8375 }
8376 #else /* CONFIG_FAIR_GROUP_SCHED */
8377 
8378 void free_fair_sched_group(struct task_group *tg) { }
8379 
8380 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8381 {
8382 	return 1;
8383 }
8384 
8385 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8386 
8387 #endif /* CONFIG_FAIR_GROUP_SCHED */
8388 
8389 
8390 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8391 {
8392 	struct sched_entity *se = &task->se;
8393 	unsigned int rr_interval = 0;
8394 
8395 	/*
8396 	 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8397 	 * idle runqueue:
8398 	 */
8399 	if (rq->cfs.load.weight)
8400 		rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8401 
8402 	return rr_interval;
8403 }
8404 
8405 /*
8406  * All the scheduling class methods:
8407  */
8408 const struct sched_class fair_sched_class = {
8409 	.next			= &idle_sched_class,
8410 	.enqueue_task		= enqueue_task_fair,
8411 	.dequeue_task		= dequeue_task_fair,
8412 	.yield_task		= yield_task_fair,
8413 	.yield_to_task		= yield_to_task_fair,
8414 
8415 	.check_preempt_curr	= check_preempt_wakeup,
8416 
8417 	.pick_next_task		= pick_next_task_fair,
8418 	.put_prev_task		= put_prev_task_fair,
8419 
8420 #ifdef CONFIG_SMP
8421 	.select_task_rq		= select_task_rq_fair,
8422 	.migrate_task_rq	= migrate_task_rq_fair,
8423 
8424 	.rq_online		= rq_online_fair,
8425 	.rq_offline		= rq_offline_fair,
8426 
8427 	.task_waking		= task_waking_fair,
8428 	.task_dead		= task_dead_fair,
8429 	.set_cpus_allowed	= set_cpus_allowed_common,
8430 #endif
8431 
8432 	.set_curr_task          = set_curr_task_fair,
8433 	.task_tick		= task_tick_fair,
8434 	.task_fork		= task_fork_fair,
8435 
8436 	.prio_changed		= prio_changed_fair,
8437 	.switched_from		= switched_from_fair,
8438 	.switched_to		= switched_to_fair,
8439 
8440 	.get_rr_interval	= get_rr_interval_fair,
8441 
8442 	.update_curr		= update_curr_fair,
8443 
8444 #ifdef CONFIG_FAIR_GROUP_SCHED
8445 	.task_move_group	= task_move_group_fair,
8446 #endif
8447 };
8448 
8449 #ifdef CONFIG_SCHED_DEBUG
8450 void print_cfs_stats(struct seq_file *m, int cpu)
8451 {
8452 	struct cfs_rq *cfs_rq;
8453 
8454 	rcu_read_lock();
8455 	for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8456 		print_cfs_rq(m, cpu, cfs_rq);
8457 	rcu_read_unlock();
8458 }
8459 
8460 #ifdef CONFIG_NUMA_BALANCING
8461 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8462 {
8463 	int node;
8464 	unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8465 
8466 	for_each_online_node(node) {
8467 		if (p->numa_faults) {
8468 			tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8469 			tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8470 		}
8471 		if (p->numa_group) {
8472 			gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8473 			gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8474 		}
8475 		print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8476 	}
8477 }
8478 #endif /* CONFIG_NUMA_BALANCING */
8479 #endif /* CONFIG_SCHED_DEBUG */
8480 
8481 __init void init_sched_fair_class(void)
8482 {
8483 #ifdef CONFIG_SMP
8484 	open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8485 
8486 #ifdef CONFIG_NO_HZ_COMMON
8487 	nohz.next_balance = jiffies;
8488 	zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8489 	cpu_notifier(sched_ilb_notifier, 0);
8490 #endif
8491 #endif /* SMP */
8492 
8493 }
8494