xref: /openbmc/linux/kernel/sched/fair.c (revision f7018c21)
1 /*
2  * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
3  *
4  *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
5  *
6  *  Interactivity improvements by Mike Galbraith
7  *  (C) 2007 Mike Galbraith <efault@gmx.de>
8  *
9  *  Various enhancements by Dmitry Adamushko.
10  *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
11  *
12  *  Group scheduling enhancements by Srivatsa Vaddagiri
13  *  Copyright IBM Corporation, 2007
14  *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
15  *
16  *  Scaled math optimizations by Thomas Gleixner
17  *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
18  *
19  *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
21  */
22 
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
32 
33 #include <trace/events/sched.h>
34 
35 #include "sched.h"
36 
37 /*
38  * Targeted preemption latency for CPU-bound tasks:
39  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40  *
41  * NOTE: this latency value is not the same as the concept of
42  * 'timeslice length' - timeslices in CFS are of variable length
43  * and have no persistent notion like in traditional, time-slice
44  * based scheduling concepts.
45  *
46  * (to see the precise effective timeslice length of your workload,
47  *  run vmstat and monitor the context-switches (cs) field)
48  */
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
51 
52 /*
53  * The initial- and re-scaling of tunables is configurable
54  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
55  *
56  * Options are:
57  * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58  * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59  * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
60  */
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 	= SCHED_TUNABLESCALING_LOG;
63 
64 /*
65  * Minimal preemption granularity for CPU-bound tasks:
66  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
67  */
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
70 
71 /*
72  * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
73  */
74 static unsigned int sched_nr_latency = 8;
75 
76 /*
77  * After fork, child runs first. If set to 0 (default) then
78  * parent will (try to) run first.
79  */
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
81 
82 /*
83  * SCHED_OTHER wake-up granularity.
84  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
85  *
86  * This option delays the preemption effects of decoupled workloads
87  * and reduces their over-scheduling. Synchronous workloads will still
88  * have immediate wakeup/sleep latencies.
89  */
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
92 
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
94 
95 /*
96  * The exponential sliding  window over which load is averaged for shares
97  * distribution.
98  * (default: 10msec)
99  */
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
101 
102 #ifdef CONFIG_CFS_BANDWIDTH
103 /*
104  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105  * each time a cfs_rq requests quota.
106  *
107  * Note: in the case that the slice exceeds the runtime remaining (either due
108  * to consumption or the quota being specified to be smaller than the slice)
109  * we will always only issue the remaining available time.
110  *
111  * default: 5 msec, units: microseconds
112   */
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
114 #endif
115 
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
117 {
118 	lw->weight += inc;
119 	lw->inv_weight = 0;
120 }
121 
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
123 {
124 	lw->weight -= dec;
125 	lw->inv_weight = 0;
126 }
127 
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
129 {
130 	lw->weight = w;
131 	lw->inv_weight = 0;
132 }
133 
134 /*
135  * Increase the granularity value when there are more CPUs,
136  * because with more CPUs the 'effective latency' as visible
137  * to users decreases. But the relationship is not linear,
138  * so pick a second-best guess by going with the log2 of the
139  * number of CPUs.
140  *
141  * This idea comes from the SD scheduler of Con Kolivas:
142  */
143 static int get_update_sysctl_factor(void)
144 {
145 	unsigned int cpus = min_t(int, num_online_cpus(), 8);
146 	unsigned int factor;
147 
148 	switch (sysctl_sched_tunable_scaling) {
149 	case SCHED_TUNABLESCALING_NONE:
150 		factor = 1;
151 		break;
152 	case SCHED_TUNABLESCALING_LINEAR:
153 		factor = cpus;
154 		break;
155 	case SCHED_TUNABLESCALING_LOG:
156 	default:
157 		factor = 1 + ilog2(cpus);
158 		break;
159 	}
160 
161 	return factor;
162 }
163 
164 static void update_sysctl(void)
165 {
166 	unsigned int factor = get_update_sysctl_factor();
167 
168 #define SET_SYSCTL(name) \
169 	(sysctl_##name = (factor) * normalized_sysctl_##name)
170 	SET_SYSCTL(sched_min_granularity);
171 	SET_SYSCTL(sched_latency);
172 	SET_SYSCTL(sched_wakeup_granularity);
173 #undef SET_SYSCTL
174 }
175 
176 void sched_init_granularity(void)
177 {
178 	update_sysctl();
179 }
180 
181 #define WMULT_CONST	(~0U)
182 #define WMULT_SHIFT	32
183 
184 static void __update_inv_weight(struct load_weight *lw)
185 {
186 	unsigned long w;
187 
188 	if (likely(lw->inv_weight))
189 		return;
190 
191 	w = scale_load_down(lw->weight);
192 
193 	if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
194 		lw->inv_weight = 1;
195 	else if (unlikely(!w))
196 		lw->inv_weight = WMULT_CONST;
197 	else
198 		lw->inv_weight = WMULT_CONST / w;
199 }
200 
201 /*
202  * delta_exec * weight / lw.weight
203  *   OR
204  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
205  *
206  * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207  * we're guaranteed shift stays positive because inv_weight is guaranteed to
208  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
209  *
210  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211  * weight/lw.weight <= 1, and therefore our shift will also be positive.
212  */
213 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
214 {
215 	u64 fact = scale_load_down(weight);
216 	int shift = WMULT_SHIFT;
217 
218 	__update_inv_weight(lw);
219 
220 	if (unlikely(fact >> 32)) {
221 		while (fact >> 32) {
222 			fact >>= 1;
223 			shift--;
224 		}
225 	}
226 
227 	/* hint to use a 32x32->64 mul */
228 	fact = (u64)(u32)fact * lw->inv_weight;
229 
230 	while (fact >> 32) {
231 		fact >>= 1;
232 		shift--;
233 	}
234 
235 	return mul_u64_u32_shr(delta_exec, fact, shift);
236 }
237 
238 
239 const struct sched_class fair_sched_class;
240 
241 /**************************************************************
242  * CFS operations on generic schedulable entities:
243  */
244 
245 #ifdef CONFIG_FAIR_GROUP_SCHED
246 
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
249 {
250 	return cfs_rq->rq;
251 }
252 
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se)	(!se->my_q)
255 
256 static inline struct task_struct *task_of(struct sched_entity *se)
257 {
258 #ifdef CONFIG_SCHED_DEBUG
259 	WARN_ON_ONCE(!entity_is_task(se));
260 #endif
261 	return container_of(se, struct task_struct, se);
262 }
263 
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 		for (; se; se = se->parent)
267 
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
269 {
270 	return p->se.cfs_rq;
271 }
272 
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
275 {
276 	return se->cfs_rq;
277 }
278 
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
281 {
282 	return grp->my_q;
283 }
284 
285 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 				       int force_update);
287 
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
289 {
290 	if (!cfs_rq->on_list) {
291 		/*
292 		 * Ensure we either appear before our parent (if already
293 		 * enqueued) or force our parent to appear after us when it is
294 		 * enqueued.  The fact that we always enqueue bottom-up
295 		 * reduces this to two cases.
296 		 */
297 		if (cfs_rq->tg->parent &&
298 		    cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
299 			list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
300 				&rq_of(cfs_rq)->leaf_cfs_rq_list);
301 		} else {
302 			list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
303 				&rq_of(cfs_rq)->leaf_cfs_rq_list);
304 		}
305 
306 		cfs_rq->on_list = 1;
307 		/* We should have no load, but we need to update last_decay. */
308 		update_cfs_rq_blocked_load(cfs_rq, 0);
309 	}
310 }
311 
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
313 {
314 	if (cfs_rq->on_list) {
315 		list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 		cfs_rq->on_list = 0;
317 	}
318 }
319 
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 	list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
323 
324 /* Do the two (enqueued) entities belong to the same group ? */
325 static inline struct cfs_rq *
326 is_same_group(struct sched_entity *se, struct sched_entity *pse)
327 {
328 	if (se->cfs_rq == pse->cfs_rq)
329 		return se->cfs_rq;
330 
331 	return NULL;
332 }
333 
334 static inline struct sched_entity *parent_entity(struct sched_entity *se)
335 {
336 	return se->parent;
337 }
338 
339 static void
340 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
341 {
342 	int se_depth, pse_depth;
343 
344 	/*
345 	 * preemption test can be made between sibling entities who are in the
346 	 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 	 * both tasks until we find their ancestors who are siblings of common
348 	 * parent.
349 	 */
350 
351 	/* First walk up until both entities are at same depth */
352 	se_depth = (*se)->depth;
353 	pse_depth = (*pse)->depth;
354 
355 	while (se_depth > pse_depth) {
356 		se_depth--;
357 		*se = parent_entity(*se);
358 	}
359 
360 	while (pse_depth > se_depth) {
361 		pse_depth--;
362 		*pse = parent_entity(*pse);
363 	}
364 
365 	while (!is_same_group(*se, *pse)) {
366 		*se = parent_entity(*se);
367 		*pse = parent_entity(*pse);
368 	}
369 }
370 
371 #else	/* !CONFIG_FAIR_GROUP_SCHED */
372 
373 static inline struct task_struct *task_of(struct sched_entity *se)
374 {
375 	return container_of(se, struct task_struct, se);
376 }
377 
378 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
379 {
380 	return container_of(cfs_rq, struct rq, cfs);
381 }
382 
383 #define entity_is_task(se)	1
384 
385 #define for_each_sched_entity(se) \
386 		for (; se; se = NULL)
387 
388 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
389 {
390 	return &task_rq(p)->cfs;
391 }
392 
393 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
394 {
395 	struct task_struct *p = task_of(se);
396 	struct rq *rq = task_rq(p);
397 
398 	return &rq->cfs;
399 }
400 
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 {
404 	return NULL;
405 }
406 
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
408 {
409 }
410 
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
412 {
413 }
414 
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 		for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
417 
418 static inline struct sched_entity *parent_entity(struct sched_entity *se)
419 {
420 	return NULL;
421 }
422 
423 static inline void
424 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
425 {
426 }
427 
428 #endif	/* CONFIG_FAIR_GROUP_SCHED */
429 
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
432 
433 /**************************************************************
434  * Scheduling class tree data structure manipulation methods:
435  */
436 
437 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
438 {
439 	s64 delta = (s64)(vruntime - max_vruntime);
440 	if (delta > 0)
441 		max_vruntime = vruntime;
442 
443 	return max_vruntime;
444 }
445 
446 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
447 {
448 	s64 delta = (s64)(vruntime - min_vruntime);
449 	if (delta < 0)
450 		min_vruntime = vruntime;
451 
452 	return min_vruntime;
453 }
454 
455 static inline int entity_before(struct sched_entity *a,
456 				struct sched_entity *b)
457 {
458 	return (s64)(a->vruntime - b->vruntime) < 0;
459 }
460 
461 static void update_min_vruntime(struct cfs_rq *cfs_rq)
462 {
463 	u64 vruntime = cfs_rq->min_vruntime;
464 
465 	if (cfs_rq->curr)
466 		vruntime = cfs_rq->curr->vruntime;
467 
468 	if (cfs_rq->rb_leftmost) {
469 		struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 						   struct sched_entity,
471 						   run_node);
472 
473 		if (!cfs_rq->curr)
474 			vruntime = se->vruntime;
475 		else
476 			vruntime = min_vruntime(vruntime, se->vruntime);
477 	}
478 
479 	/* ensure we never gain time by being placed backwards. */
480 	cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
481 #ifndef CONFIG_64BIT
482 	smp_wmb();
483 	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 #endif
485 }
486 
487 /*
488  * Enqueue an entity into the rb-tree:
489  */
490 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
491 {
492 	struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
493 	struct rb_node *parent = NULL;
494 	struct sched_entity *entry;
495 	int leftmost = 1;
496 
497 	/*
498 	 * Find the right place in the rbtree:
499 	 */
500 	while (*link) {
501 		parent = *link;
502 		entry = rb_entry(parent, struct sched_entity, run_node);
503 		/*
504 		 * We dont care about collisions. Nodes with
505 		 * the same key stay together.
506 		 */
507 		if (entity_before(se, entry)) {
508 			link = &parent->rb_left;
509 		} else {
510 			link = &parent->rb_right;
511 			leftmost = 0;
512 		}
513 	}
514 
515 	/*
516 	 * Maintain a cache of leftmost tree entries (it is frequently
517 	 * used):
518 	 */
519 	if (leftmost)
520 		cfs_rq->rb_leftmost = &se->run_node;
521 
522 	rb_link_node(&se->run_node, parent, link);
523 	rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
524 }
525 
526 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
527 {
528 	if (cfs_rq->rb_leftmost == &se->run_node) {
529 		struct rb_node *next_node;
530 
531 		next_node = rb_next(&se->run_node);
532 		cfs_rq->rb_leftmost = next_node;
533 	}
534 
535 	rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
536 }
537 
538 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
539 {
540 	struct rb_node *left = cfs_rq->rb_leftmost;
541 
542 	if (!left)
543 		return NULL;
544 
545 	return rb_entry(left, struct sched_entity, run_node);
546 }
547 
548 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
549 {
550 	struct rb_node *next = rb_next(&se->run_node);
551 
552 	if (!next)
553 		return NULL;
554 
555 	return rb_entry(next, struct sched_entity, run_node);
556 }
557 
558 #ifdef CONFIG_SCHED_DEBUG
559 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
560 {
561 	struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 
563 	if (!last)
564 		return NULL;
565 
566 	return rb_entry(last, struct sched_entity, run_node);
567 }
568 
569 /**************************************************************
570  * Scheduling class statistics methods:
571  */
572 
573 int sched_proc_update_handler(struct ctl_table *table, int write,
574 		void __user *buffer, size_t *lenp,
575 		loff_t *ppos)
576 {
577 	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
578 	int factor = get_update_sysctl_factor();
579 
580 	if (ret || !write)
581 		return ret;
582 
583 	sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
584 					sysctl_sched_min_granularity);
585 
586 #define WRT_SYSCTL(name) \
587 	(normalized_sysctl_##name = sysctl_##name / (factor))
588 	WRT_SYSCTL(sched_min_granularity);
589 	WRT_SYSCTL(sched_latency);
590 	WRT_SYSCTL(sched_wakeup_granularity);
591 #undef WRT_SYSCTL
592 
593 	return 0;
594 }
595 #endif
596 
597 /*
598  * delta /= w
599  */
600 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
601 {
602 	if (unlikely(se->load.weight != NICE_0_LOAD))
603 		delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
604 
605 	return delta;
606 }
607 
608 /*
609  * The idea is to set a period in which each task runs once.
610  *
611  * When there are too many tasks (sched_nr_latency) we have to stretch
612  * this period because otherwise the slices get too small.
613  *
614  * p = (nr <= nl) ? l : l*nr/nl
615  */
616 static u64 __sched_period(unsigned long nr_running)
617 {
618 	u64 period = sysctl_sched_latency;
619 	unsigned long nr_latency = sched_nr_latency;
620 
621 	if (unlikely(nr_running > nr_latency)) {
622 		period = sysctl_sched_min_granularity;
623 		period *= nr_running;
624 	}
625 
626 	return period;
627 }
628 
629 /*
630  * We calculate the wall-time slice from the period by taking a part
631  * proportional to the weight.
632  *
633  * s = p*P[w/rw]
634  */
635 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
636 {
637 	u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
638 
639 	for_each_sched_entity(se) {
640 		struct load_weight *load;
641 		struct load_weight lw;
642 
643 		cfs_rq = cfs_rq_of(se);
644 		load = &cfs_rq->load;
645 
646 		if (unlikely(!se->on_rq)) {
647 			lw = cfs_rq->load;
648 
649 			update_load_add(&lw, se->load.weight);
650 			load = &lw;
651 		}
652 		slice = __calc_delta(slice, se->load.weight, load);
653 	}
654 	return slice;
655 }
656 
657 /*
658  * We calculate the vruntime slice of a to-be-inserted task.
659  *
660  * vs = s/w
661  */
662 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
663 {
664 	return calc_delta_fair(sched_slice(cfs_rq, se), se);
665 }
666 
667 #ifdef CONFIG_SMP
668 static unsigned long task_h_load(struct task_struct *p);
669 
670 static inline void __update_task_entity_contrib(struct sched_entity *se);
671 
672 /* Give new task start runnable values to heavy its load in infant time */
673 void init_task_runnable_average(struct task_struct *p)
674 {
675 	u32 slice;
676 
677 	p->se.avg.decay_count = 0;
678 	slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
679 	p->se.avg.runnable_avg_sum = slice;
680 	p->se.avg.runnable_avg_period = slice;
681 	__update_task_entity_contrib(&p->se);
682 }
683 #else
684 void init_task_runnable_average(struct task_struct *p)
685 {
686 }
687 #endif
688 
689 /*
690  * Update the current task's runtime statistics.
691  */
692 static void update_curr(struct cfs_rq *cfs_rq)
693 {
694 	struct sched_entity *curr = cfs_rq->curr;
695 	u64 now = rq_clock_task(rq_of(cfs_rq));
696 	u64 delta_exec;
697 
698 	if (unlikely(!curr))
699 		return;
700 
701 	delta_exec = now - curr->exec_start;
702 	if (unlikely((s64)delta_exec <= 0))
703 		return;
704 
705 	curr->exec_start = now;
706 
707 	schedstat_set(curr->statistics.exec_max,
708 		      max(delta_exec, curr->statistics.exec_max));
709 
710 	curr->sum_exec_runtime += delta_exec;
711 	schedstat_add(cfs_rq, exec_clock, delta_exec);
712 
713 	curr->vruntime += calc_delta_fair(delta_exec, curr);
714 	update_min_vruntime(cfs_rq);
715 
716 	if (entity_is_task(curr)) {
717 		struct task_struct *curtask = task_of(curr);
718 
719 		trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
720 		cpuacct_charge(curtask, delta_exec);
721 		account_group_exec_runtime(curtask, delta_exec);
722 	}
723 
724 	account_cfs_rq_runtime(cfs_rq, delta_exec);
725 }
726 
727 static inline void
728 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
729 {
730 	schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
731 }
732 
733 /*
734  * Task is being enqueued - update stats:
735  */
736 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
737 {
738 	/*
739 	 * Are we enqueueing a waiting task? (for current tasks
740 	 * a dequeue/enqueue event is a NOP)
741 	 */
742 	if (se != cfs_rq->curr)
743 		update_stats_wait_start(cfs_rq, se);
744 }
745 
746 static void
747 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
748 {
749 	schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
750 			rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
751 	schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
752 	schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
753 			rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
754 #ifdef CONFIG_SCHEDSTATS
755 	if (entity_is_task(se)) {
756 		trace_sched_stat_wait(task_of(se),
757 			rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
758 	}
759 #endif
760 	schedstat_set(se->statistics.wait_start, 0);
761 }
762 
763 static inline void
764 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
765 {
766 	/*
767 	 * Mark the end of the wait period if dequeueing a
768 	 * waiting task:
769 	 */
770 	if (se != cfs_rq->curr)
771 		update_stats_wait_end(cfs_rq, se);
772 }
773 
774 /*
775  * We are picking a new current task - update its stats:
776  */
777 static inline void
778 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
779 {
780 	/*
781 	 * We are starting a new run period:
782 	 */
783 	se->exec_start = rq_clock_task(rq_of(cfs_rq));
784 }
785 
786 /**************************************************
787  * Scheduling class queueing methods:
788  */
789 
790 #ifdef CONFIG_NUMA_BALANCING
791 /*
792  * Approximate time to scan a full NUMA task in ms. The task scan period is
793  * calculated based on the tasks virtual memory size and
794  * numa_balancing_scan_size.
795  */
796 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
797 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
798 
799 /* Portion of address space to scan in MB */
800 unsigned int sysctl_numa_balancing_scan_size = 256;
801 
802 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
803 unsigned int sysctl_numa_balancing_scan_delay = 1000;
804 
805 static unsigned int task_nr_scan_windows(struct task_struct *p)
806 {
807 	unsigned long rss = 0;
808 	unsigned long nr_scan_pages;
809 
810 	/*
811 	 * Calculations based on RSS as non-present and empty pages are skipped
812 	 * by the PTE scanner and NUMA hinting faults should be trapped based
813 	 * on resident pages
814 	 */
815 	nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
816 	rss = get_mm_rss(p->mm);
817 	if (!rss)
818 		rss = nr_scan_pages;
819 
820 	rss = round_up(rss, nr_scan_pages);
821 	return rss / nr_scan_pages;
822 }
823 
824 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
825 #define MAX_SCAN_WINDOW 2560
826 
827 static unsigned int task_scan_min(struct task_struct *p)
828 {
829 	unsigned int scan, floor;
830 	unsigned int windows = 1;
831 
832 	if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
833 		windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
834 	floor = 1000 / windows;
835 
836 	scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
837 	return max_t(unsigned int, floor, scan);
838 }
839 
840 static unsigned int task_scan_max(struct task_struct *p)
841 {
842 	unsigned int smin = task_scan_min(p);
843 	unsigned int smax;
844 
845 	/* Watch for min being lower than max due to floor calculations */
846 	smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
847 	return max(smin, smax);
848 }
849 
850 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
851 {
852 	rq->nr_numa_running += (p->numa_preferred_nid != -1);
853 	rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
854 }
855 
856 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
857 {
858 	rq->nr_numa_running -= (p->numa_preferred_nid != -1);
859 	rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
860 }
861 
862 struct numa_group {
863 	atomic_t refcount;
864 
865 	spinlock_t lock; /* nr_tasks, tasks */
866 	int nr_tasks;
867 	pid_t gid;
868 	struct list_head task_list;
869 
870 	struct rcu_head rcu;
871 	nodemask_t active_nodes;
872 	unsigned long total_faults;
873 	/*
874 	 * Faults_cpu is used to decide whether memory should move
875 	 * towards the CPU. As a consequence, these stats are weighted
876 	 * more by CPU use than by memory faults.
877 	 */
878 	unsigned long *faults_cpu;
879 	unsigned long faults[0];
880 };
881 
882 /* Shared or private faults. */
883 #define NR_NUMA_HINT_FAULT_TYPES 2
884 
885 /* Memory and CPU locality */
886 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
887 
888 /* Averaged statistics, and temporary buffers. */
889 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
890 
891 pid_t task_numa_group_id(struct task_struct *p)
892 {
893 	return p->numa_group ? p->numa_group->gid : 0;
894 }
895 
896 static inline int task_faults_idx(int nid, int priv)
897 {
898 	return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
899 }
900 
901 static inline unsigned long task_faults(struct task_struct *p, int nid)
902 {
903 	if (!p->numa_faults_memory)
904 		return 0;
905 
906 	return p->numa_faults_memory[task_faults_idx(nid, 0)] +
907 		p->numa_faults_memory[task_faults_idx(nid, 1)];
908 }
909 
910 static inline unsigned long group_faults(struct task_struct *p, int nid)
911 {
912 	if (!p->numa_group)
913 		return 0;
914 
915 	return p->numa_group->faults[task_faults_idx(nid, 0)] +
916 		p->numa_group->faults[task_faults_idx(nid, 1)];
917 }
918 
919 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
920 {
921 	return group->faults_cpu[task_faults_idx(nid, 0)] +
922 		group->faults_cpu[task_faults_idx(nid, 1)];
923 }
924 
925 /*
926  * These return the fraction of accesses done by a particular task, or
927  * task group, on a particular numa node.  The group weight is given a
928  * larger multiplier, in order to group tasks together that are almost
929  * evenly spread out between numa nodes.
930  */
931 static inline unsigned long task_weight(struct task_struct *p, int nid)
932 {
933 	unsigned long total_faults;
934 
935 	if (!p->numa_faults_memory)
936 		return 0;
937 
938 	total_faults = p->total_numa_faults;
939 
940 	if (!total_faults)
941 		return 0;
942 
943 	return 1000 * task_faults(p, nid) / total_faults;
944 }
945 
946 static inline unsigned long group_weight(struct task_struct *p, int nid)
947 {
948 	if (!p->numa_group || !p->numa_group->total_faults)
949 		return 0;
950 
951 	return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
952 }
953 
954 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
955 				int src_nid, int dst_cpu)
956 {
957 	struct numa_group *ng = p->numa_group;
958 	int dst_nid = cpu_to_node(dst_cpu);
959 	int last_cpupid, this_cpupid;
960 
961 	this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
962 
963 	/*
964 	 * Multi-stage node selection is used in conjunction with a periodic
965 	 * migration fault to build a temporal task<->page relation. By using
966 	 * a two-stage filter we remove short/unlikely relations.
967 	 *
968 	 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
969 	 * a task's usage of a particular page (n_p) per total usage of this
970 	 * page (n_t) (in a given time-span) to a probability.
971 	 *
972 	 * Our periodic faults will sample this probability and getting the
973 	 * same result twice in a row, given these samples are fully
974 	 * independent, is then given by P(n)^2, provided our sample period
975 	 * is sufficiently short compared to the usage pattern.
976 	 *
977 	 * This quadric squishes small probabilities, making it less likely we
978 	 * act on an unlikely task<->page relation.
979 	 */
980 	last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
981 	if (!cpupid_pid_unset(last_cpupid) &&
982 				cpupid_to_nid(last_cpupid) != dst_nid)
983 		return false;
984 
985 	/* Always allow migrate on private faults */
986 	if (cpupid_match_pid(p, last_cpupid))
987 		return true;
988 
989 	/* A shared fault, but p->numa_group has not been set up yet. */
990 	if (!ng)
991 		return true;
992 
993 	/*
994 	 * Do not migrate if the destination is not a node that
995 	 * is actively used by this numa group.
996 	 */
997 	if (!node_isset(dst_nid, ng->active_nodes))
998 		return false;
999 
1000 	/*
1001 	 * Source is a node that is not actively used by this
1002 	 * numa group, while the destination is. Migrate.
1003 	 */
1004 	if (!node_isset(src_nid, ng->active_nodes))
1005 		return true;
1006 
1007 	/*
1008 	 * Both source and destination are nodes in active
1009 	 * use by this numa group. Maximize memory bandwidth
1010 	 * by migrating from more heavily used groups, to less
1011 	 * heavily used ones, spreading the load around.
1012 	 * Use a 1/4 hysteresis to avoid spurious page movement.
1013 	 */
1014 	return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1015 }
1016 
1017 static unsigned long weighted_cpuload(const int cpu);
1018 static unsigned long source_load(int cpu, int type);
1019 static unsigned long target_load(int cpu, int type);
1020 static unsigned long power_of(int cpu);
1021 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1022 
1023 /* Cached statistics for all CPUs within a node */
1024 struct numa_stats {
1025 	unsigned long nr_running;
1026 	unsigned long load;
1027 
1028 	/* Total compute capacity of CPUs on a node */
1029 	unsigned long power;
1030 
1031 	/* Approximate capacity in terms of runnable tasks on a node */
1032 	unsigned long capacity;
1033 	int has_capacity;
1034 };
1035 
1036 /*
1037  * XXX borrowed from update_sg_lb_stats
1038  */
1039 static void update_numa_stats(struct numa_stats *ns, int nid)
1040 {
1041 	int cpu, cpus = 0;
1042 
1043 	memset(ns, 0, sizeof(*ns));
1044 	for_each_cpu(cpu, cpumask_of_node(nid)) {
1045 		struct rq *rq = cpu_rq(cpu);
1046 
1047 		ns->nr_running += rq->nr_running;
1048 		ns->load += weighted_cpuload(cpu);
1049 		ns->power += power_of(cpu);
1050 
1051 		cpus++;
1052 	}
1053 
1054 	/*
1055 	 * If we raced with hotplug and there are no CPUs left in our mask
1056 	 * the @ns structure is NULL'ed and task_numa_compare() will
1057 	 * not find this node attractive.
1058 	 *
1059 	 * We'll either bail at !has_capacity, or we'll detect a huge imbalance
1060 	 * and bail there.
1061 	 */
1062 	if (!cpus)
1063 		return;
1064 
1065 	ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1066 	ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1067 	ns->has_capacity = (ns->nr_running < ns->capacity);
1068 }
1069 
1070 struct task_numa_env {
1071 	struct task_struct *p;
1072 
1073 	int src_cpu, src_nid;
1074 	int dst_cpu, dst_nid;
1075 
1076 	struct numa_stats src_stats, dst_stats;
1077 
1078 	int imbalance_pct;
1079 
1080 	struct task_struct *best_task;
1081 	long best_imp;
1082 	int best_cpu;
1083 };
1084 
1085 static void task_numa_assign(struct task_numa_env *env,
1086 			     struct task_struct *p, long imp)
1087 {
1088 	if (env->best_task)
1089 		put_task_struct(env->best_task);
1090 	if (p)
1091 		get_task_struct(p);
1092 
1093 	env->best_task = p;
1094 	env->best_imp = imp;
1095 	env->best_cpu = env->dst_cpu;
1096 }
1097 
1098 /*
1099  * This checks if the overall compute and NUMA accesses of the system would
1100  * be improved if the source tasks was migrated to the target dst_cpu taking
1101  * into account that it might be best if task running on the dst_cpu should
1102  * be exchanged with the source task
1103  */
1104 static void task_numa_compare(struct task_numa_env *env,
1105 			      long taskimp, long groupimp)
1106 {
1107 	struct rq *src_rq = cpu_rq(env->src_cpu);
1108 	struct rq *dst_rq = cpu_rq(env->dst_cpu);
1109 	struct task_struct *cur;
1110 	long dst_load, src_load;
1111 	long load;
1112 	long imp = (groupimp > 0) ? groupimp : taskimp;
1113 
1114 	rcu_read_lock();
1115 	cur = ACCESS_ONCE(dst_rq->curr);
1116 	if (cur->pid == 0) /* idle */
1117 		cur = NULL;
1118 
1119 	/*
1120 	 * "imp" is the fault differential for the source task between the
1121 	 * source and destination node. Calculate the total differential for
1122 	 * the source task and potential destination task. The more negative
1123 	 * the value is, the more rmeote accesses that would be expected to
1124 	 * be incurred if the tasks were swapped.
1125 	 */
1126 	if (cur) {
1127 		/* Skip this swap candidate if cannot move to the source cpu */
1128 		if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1129 			goto unlock;
1130 
1131 		/*
1132 		 * If dst and source tasks are in the same NUMA group, or not
1133 		 * in any group then look only at task weights.
1134 		 */
1135 		if (cur->numa_group == env->p->numa_group) {
1136 			imp = taskimp + task_weight(cur, env->src_nid) -
1137 			      task_weight(cur, env->dst_nid);
1138 			/*
1139 			 * Add some hysteresis to prevent swapping the
1140 			 * tasks within a group over tiny differences.
1141 			 */
1142 			if (cur->numa_group)
1143 				imp -= imp/16;
1144 		} else {
1145 			/*
1146 			 * Compare the group weights. If a task is all by
1147 			 * itself (not part of a group), use the task weight
1148 			 * instead.
1149 			 */
1150 			if (env->p->numa_group)
1151 				imp = groupimp;
1152 			else
1153 				imp = taskimp;
1154 
1155 			if (cur->numa_group)
1156 				imp += group_weight(cur, env->src_nid) -
1157 				       group_weight(cur, env->dst_nid);
1158 			else
1159 				imp += task_weight(cur, env->src_nid) -
1160 				       task_weight(cur, env->dst_nid);
1161 		}
1162 	}
1163 
1164 	if (imp < env->best_imp)
1165 		goto unlock;
1166 
1167 	if (!cur) {
1168 		/* Is there capacity at our destination? */
1169 		if (env->src_stats.has_capacity &&
1170 		    !env->dst_stats.has_capacity)
1171 			goto unlock;
1172 
1173 		goto balance;
1174 	}
1175 
1176 	/* Balance doesn't matter much if we're running a task per cpu */
1177 	if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1178 		goto assign;
1179 
1180 	/*
1181 	 * In the overloaded case, try and keep the load balanced.
1182 	 */
1183 balance:
1184 	dst_load = env->dst_stats.load;
1185 	src_load = env->src_stats.load;
1186 
1187 	/* XXX missing power terms */
1188 	load = task_h_load(env->p);
1189 	dst_load += load;
1190 	src_load -= load;
1191 
1192 	if (cur) {
1193 		load = task_h_load(cur);
1194 		dst_load -= load;
1195 		src_load += load;
1196 	}
1197 
1198 	/* make src_load the smaller */
1199 	if (dst_load < src_load)
1200 		swap(dst_load, src_load);
1201 
1202 	if (src_load * env->imbalance_pct < dst_load * 100)
1203 		goto unlock;
1204 
1205 assign:
1206 	task_numa_assign(env, cur, imp);
1207 unlock:
1208 	rcu_read_unlock();
1209 }
1210 
1211 static void task_numa_find_cpu(struct task_numa_env *env,
1212 				long taskimp, long groupimp)
1213 {
1214 	int cpu;
1215 
1216 	for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1217 		/* Skip this CPU if the source task cannot migrate */
1218 		if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1219 			continue;
1220 
1221 		env->dst_cpu = cpu;
1222 		task_numa_compare(env, taskimp, groupimp);
1223 	}
1224 }
1225 
1226 static int task_numa_migrate(struct task_struct *p)
1227 {
1228 	struct task_numa_env env = {
1229 		.p = p,
1230 
1231 		.src_cpu = task_cpu(p),
1232 		.src_nid = task_node(p),
1233 
1234 		.imbalance_pct = 112,
1235 
1236 		.best_task = NULL,
1237 		.best_imp = 0,
1238 		.best_cpu = -1
1239 	};
1240 	struct sched_domain *sd;
1241 	unsigned long taskweight, groupweight;
1242 	int nid, ret;
1243 	long taskimp, groupimp;
1244 
1245 	/*
1246 	 * Pick the lowest SD_NUMA domain, as that would have the smallest
1247 	 * imbalance and would be the first to start moving tasks about.
1248 	 *
1249 	 * And we want to avoid any moving of tasks about, as that would create
1250 	 * random movement of tasks -- counter the numa conditions we're trying
1251 	 * to satisfy here.
1252 	 */
1253 	rcu_read_lock();
1254 	sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1255 	if (sd)
1256 		env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1257 	rcu_read_unlock();
1258 
1259 	/*
1260 	 * Cpusets can break the scheduler domain tree into smaller
1261 	 * balance domains, some of which do not cross NUMA boundaries.
1262 	 * Tasks that are "trapped" in such domains cannot be migrated
1263 	 * elsewhere, so there is no point in (re)trying.
1264 	 */
1265 	if (unlikely(!sd)) {
1266 		p->numa_preferred_nid = task_node(p);
1267 		return -EINVAL;
1268 	}
1269 
1270 	taskweight = task_weight(p, env.src_nid);
1271 	groupweight = group_weight(p, env.src_nid);
1272 	update_numa_stats(&env.src_stats, env.src_nid);
1273 	env.dst_nid = p->numa_preferred_nid;
1274 	taskimp = task_weight(p, env.dst_nid) - taskweight;
1275 	groupimp = group_weight(p, env.dst_nid) - groupweight;
1276 	update_numa_stats(&env.dst_stats, env.dst_nid);
1277 
1278 	/* If the preferred nid has capacity, try to use it. */
1279 	if (env.dst_stats.has_capacity)
1280 		task_numa_find_cpu(&env, taskimp, groupimp);
1281 
1282 	/* No space available on the preferred nid. Look elsewhere. */
1283 	if (env.best_cpu == -1) {
1284 		for_each_online_node(nid) {
1285 			if (nid == env.src_nid || nid == p->numa_preferred_nid)
1286 				continue;
1287 
1288 			/* Only consider nodes where both task and groups benefit */
1289 			taskimp = task_weight(p, nid) - taskweight;
1290 			groupimp = group_weight(p, nid) - groupweight;
1291 			if (taskimp < 0 && groupimp < 0)
1292 				continue;
1293 
1294 			env.dst_nid = nid;
1295 			update_numa_stats(&env.dst_stats, env.dst_nid);
1296 			task_numa_find_cpu(&env, taskimp, groupimp);
1297 		}
1298 	}
1299 
1300 	/* No better CPU than the current one was found. */
1301 	if (env.best_cpu == -1)
1302 		return -EAGAIN;
1303 
1304 	sched_setnuma(p, env.dst_nid);
1305 
1306 	/*
1307 	 * Reset the scan period if the task is being rescheduled on an
1308 	 * alternative node to recheck if the tasks is now properly placed.
1309 	 */
1310 	p->numa_scan_period = task_scan_min(p);
1311 
1312 	if (env.best_task == NULL) {
1313 		ret = migrate_task_to(p, env.best_cpu);
1314 		if (ret != 0)
1315 			trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1316 		return ret;
1317 	}
1318 
1319 	ret = migrate_swap(p, env.best_task);
1320 	if (ret != 0)
1321 		trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1322 	put_task_struct(env.best_task);
1323 	return ret;
1324 }
1325 
1326 /* Attempt to migrate a task to a CPU on the preferred node. */
1327 static void numa_migrate_preferred(struct task_struct *p)
1328 {
1329 	/* This task has no NUMA fault statistics yet */
1330 	if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1331 		return;
1332 
1333 	/* Periodically retry migrating the task to the preferred node */
1334 	p->numa_migrate_retry = jiffies + HZ;
1335 
1336 	/* Success if task is already running on preferred CPU */
1337 	if (task_node(p) == p->numa_preferred_nid)
1338 		return;
1339 
1340 	/* Otherwise, try migrate to a CPU on the preferred node */
1341 	task_numa_migrate(p);
1342 }
1343 
1344 /*
1345  * Find the nodes on which the workload is actively running. We do this by
1346  * tracking the nodes from which NUMA hinting faults are triggered. This can
1347  * be different from the set of nodes where the workload's memory is currently
1348  * located.
1349  *
1350  * The bitmask is used to make smarter decisions on when to do NUMA page
1351  * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1352  * are added when they cause over 6/16 of the maximum number of faults, but
1353  * only removed when they drop below 3/16.
1354  */
1355 static void update_numa_active_node_mask(struct numa_group *numa_group)
1356 {
1357 	unsigned long faults, max_faults = 0;
1358 	int nid;
1359 
1360 	for_each_online_node(nid) {
1361 		faults = group_faults_cpu(numa_group, nid);
1362 		if (faults > max_faults)
1363 			max_faults = faults;
1364 	}
1365 
1366 	for_each_online_node(nid) {
1367 		faults = group_faults_cpu(numa_group, nid);
1368 		if (!node_isset(nid, numa_group->active_nodes)) {
1369 			if (faults > max_faults * 6 / 16)
1370 				node_set(nid, numa_group->active_nodes);
1371 		} else if (faults < max_faults * 3 / 16)
1372 			node_clear(nid, numa_group->active_nodes);
1373 	}
1374 }
1375 
1376 /*
1377  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1378  * increments. The more local the fault statistics are, the higher the scan
1379  * period will be for the next scan window. If local/remote ratio is below
1380  * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1381  * scan period will decrease
1382  */
1383 #define NUMA_PERIOD_SLOTS 10
1384 #define NUMA_PERIOD_THRESHOLD 3
1385 
1386 /*
1387  * Increase the scan period (slow down scanning) if the majority of
1388  * our memory is already on our local node, or if the majority of
1389  * the page accesses are shared with other processes.
1390  * Otherwise, decrease the scan period.
1391  */
1392 static void update_task_scan_period(struct task_struct *p,
1393 			unsigned long shared, unsigned long private)
1394 {
1395 	unsigned int period_slot;
1396 	int ratio;
1397 	int diff;
1398 
1399 	unsigned long remote = p->numa_faults_locality[0];
1400 	unsigned long local = p->numa_faults_locality[1];
1401 
1402 	/*
1403 	 * If there were no record hinting faults then either the task is
1404 	 * completely idle or all activity is areas that are not of interest
1405 	 * to automatic numa balancing. Scan slower
1406 	 */
1407 	if (local + shared == 0) {
1408 		p->numa_scan_period = min(p->numa_scan_period_max,
1409 			p->numa_scan_period << 1);
1410 
1411 		p->mm->numa_next_scan = jiffies +
1412 			msecs_to_jiffies(p->numa_scan_period);
1413 
1414 		return;
1415 	}
1416 
1417 	/*
1418 	 * Prepare to scale scan period relative to the current period.
1419 	 *	 == NUMA_PERIOD_THRESHOLD scan period stays the same
1420 	 *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1421 	 *	 >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1422 	 */
1423 	period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1424 	ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1425 	if (ratio >= NUMA_PERIOD_THRESHOLD) {
1426 		int slot = ratio - NUMA_PERIOD_THRESHOLD;
1427 		if (!slot)
1428 			slot = 1;
1429 		diff = slot * period_slot;
1430 	} else {
1431 		diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1432 
1433 		/*
1434 		 * Scale scan rate increases based on sharing. There is an
1435 		 * inverse relationship between the degree of sharing and
1436 		 * the adjustment made to the scanning period. Broadly
1437 		 * speaking the intent is that there is little point
1438 		 * scanning faster if shared accesses dominate as it may
1439 		 * simply bounce migrations uselessly
1440 		 */
1441 		ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1442 		diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1443 	}
1444 
1445 	p->numa_scan_period = clamp(p->numa_scan_period + diff,
1446 			task_scan_min(p), task_scan_max(p));
1447 	memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1448 }
1449 
1450 /*
1451  * Get the fraction of time the task has been running since the last
1452  * NUMA placement cycle. The scheduler keeps similar statistics, but
1453  * decays those on a 32ms period, which is orders of magnitude off
1454  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1455  * stats only if the task is so new there are no NUMA statistics yet.
1456  */
1457 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1458 {
1459 	u64 runtime, delta, now;
1460 	/* Use the start of this time slice to avoid calculations. */
1461 	now = p->se.exec_start;
1462 	runtime = p->se.sum_exec_runtime;
1463 
1464 	if (p->last_task_numa_placement) {
1465 		delta = runtime - p->last_sum_exec_runtime;
1466 		*period = now - p->last_task_numa_placement;
1467 	} else {
1468 		delta = p->se.avg.runnable_avg_sum;
1469 		*period = p->se.avg.runnable_avg_period;
1470 	}
1471 
1472 	p->last_sum_exec_runtime = runtime;
1473 	p->last_task_numa_placement = now;
1474 
1475 	return delta;
1476 }
1477 
1478 static void task_numa_placement(struct task_struct *p)
1479 {
1480 	int seq, nid, max_nid = -1, max_group_nid = -1;
1481 	unsigned long max_faults = 0, max_group_faults = 0;
1482 	unsigned long fault_types[2] = { 0, 0 };
1483 	unsigned long total_faults;
1484 	u64 runtime, period;
1485 	spinlock_t *group_lock = NULL;
1486 
1487 	seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1488 	if (p->numa_scan_seq == seq)
1489 		return;
1490 	p->numa_scan_seq = seq;
1491 	p->numa_scan_period_max = task_scan_max(p);
1492 
1493 	total_faults = p->numa_faults_locality[0] +
1494 		       p->numa_faults_locality[1];
1495 	runtime = numa_get_avg_runtime(p, &period);
1496 
1497 	/* If the task is part of a group prevent parallel updates to group stats */
1498 	if (p->numa_group) {
1499 		group_lock = &p->numa_group->lock;
1500 		spin_lock(group_lock);
1501 	}
1502 
1503 	/* Find the node with the highest number of faults */
1504 	for_each_online_node(nid) {
1505 		unsigned long faults = 0, group_faults = 0;
1506 		int priv, i;
1507 
1508 		for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1509 			long diff, f_diff, f_weight;
1510 
1511 			i = task_faults_idx(nid, priv);
1512 
1513 			/* Decay existing window, copy faults since last scan */
1514 			diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1515 			fault_types[priv] += p->numa_faults_buffer_memory[i];
1516 			p->numa_faults_buffer_memory[i] = 0;
1517 
1518 			/*
1519 			 * Normalize the faults_from, so all tasks in a group
1520 			 * count according to CPU use, instead of by the raw
1521 			 * number of faults. Tasks with little runtime have
1522 			 * little over-all impact on throughput, and thus their
1523 			 * faults are less important.
1524 			 */
1525 			f_weight = div64_u64(runtime << 16, period + 1);
1526 			f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1527 				   (total_faults + 1);
1528 			f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1529 			p->numa_faults_buffer_cpu[i] = 0;
1530 
1531 			p->numa_faults_memory[i] += diff;
1532 			p->numa_faults_cpu[i] += f_diff;
1533 			faults += p->numa_faults_memory[i];
1534 			p->total_numa_faults += diff;
1535 			if (p->numa_group) {
1536 				/* safe because we can only change our own group */
1537 				p->numa_group->faults[i] += diff;
1538 				p->numa_group->faults_cpu[i] += f_diff;
1539 				p->numa_group->total_faults += diff;
1540 				group_faults += p->numa_group->faults[i];
1541 			}
1542 		}
1543 
1544 		if (faults > max_faults) {
1545 			max_faults = faults;
1546 			max_nid = nid;
1547 		}
1548 
1549 		if (group_faults > max_group_faults) {
1550 			max_group_faults = group_faults;
1551 			max_group_nid = nid;
1552 		}
1553 	}
1554 
1555 	update_task_scan_period(p, fault_types[0], fault_types[1]);
1556 
1557 	if (p->numa_group) {
1558 		update_numa_active_node_mask(p->numa_group);
1559 		/*
1560 		 * If the preferred task and group nids are different,
1561 		 * iterate over the nodes again to find the best place.
1562 		 */
1563 		if (max_nid != max_group_nid) {
1564 			unsigned long weight, max_weight = 0;
1565 
1566 			for_each_online_node(nid) {
1567 				weight = task_weight(p, nid) + group_weight(p, nid);
1568 				if (weight > max_weight) {
1569 					max_weight = weight;
1570 					max_nid = nid;
1571 				}
1572 			}
1573 		}
1574 
1575 		spin_unlock(group_lock);
1576 	}
1577 
1578 	/* Preferred node as the node with the most faults */
1579 	if (max_faults && max_nid != p->numa_preferred_nid) {
1580 		/* Update the preferred nid and migrate task if possible */
1581 		sched_setnuma(p, max_nid);
1582 		numa_migrate_preferred(p);
1583 	}
1584 }
1585 
1586 static inline int get_numa_group(struct numa_group *grp)
1587 {
1588 	return atomic_inc_not_zero(&grp->refcount);
1589 }
1590 
1591 static inline void put_numa_group(struct numa_group *grp)
1592 {
1593 	if (atomic_dec_and_test(&grp->refcount))
1594 		kfree_rcu(grp, rcu);
1595 }
1596 
1597 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1598 			int *priv)
1599 {
1600 	struct numa_group *grp, *my_grp;
1601 	struct task_struct *tsk;
1602 	bool join = false;
1603 	int cpu = cpupid_to_cpu(cpupid);
1604 	int i;
1605 
1606 	if (unlikely(!p->numa_group)) {
1607 		unsigned int size = sizeof(struct numa_group) +
1608 				    4*nr_node_ids*sizeof(unsigned long);
1609 
1610 		grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1611 		if (!grp)
1612 			return;
1613 
1614 		atomic_set(&grp->refcount, 1);
1615 		spin_lock_init(&grp->lock);
1616 		INIT_LIST_HEAD(&grp->task_list);
1617 		grp->gid = p->pid;
1618 		/* Second half of the array tracks nids where faults happen */
1619 		grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1620 						nr_node_ids;
1621 
1622 		node_set(task_node(current), grp->active_nodes);
1623 
1624 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1625 			grp->faults[i] = p->numa_faults_memory[i];
1626 
1627 		grp->total_faults = p->total_numa_faults;
1628 
1629 		list_add(&p->numa_entry, &grp->task_list);
1630 		grp->nr_tasks++;
1631 		rcu_assign_pointer(p->numa_group, grp);
1632 	}
1633 
1634 	rcu_read_lock();
1635 	tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1636 
1637 	if (!cpupid_match_pid(tsk, cpupid))
1638 		goto no_join;
1639 
1640 	grp = rcu_dereference(tsk->numa_group);
1641 	if (!grp)
1642 		goto no_join;
1643 
1644 	my_grp = p->numa_group;
1645 	if (grp == my_grp)
1646 		goto no_join;
1647 
1648 	/*
1649 	 * Only join the other group if its bigger; if we're the bigger group,
1650 	 * the other task will join us.
1651 	 */
1652 	if (my_grp->nr_tasks > grp->nr_tasks)
1653 		goto no_join;
1654 
1655 	/*
1656 	 * Tie-break on the grp address.
1657 	 */
1658 	if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1659 		goto no_join;
1660 
1661 	/* Always join threads in the same process. */
1662 	if (tsk->mm == current->mm)
1663 		join = true;
1664 
1665 	/* Simple filter to avoid false positives due to PID collisions */
1666 	if (flags & TNF_SHARED)
1667 		join = true;
1668 
1669 	/* Update priv based on whether false sharing was detected */
1670 	*priv = !join;
1671 
1672 	if (join && !get_numa_group(grp))
1673 		goto no_join;
1674 
1675 	rcu_read_unlock();
1676 
1677 	if (!join)
1678 		return;
1679 
1680 	double_lock(&my_grp->lock, &grp->lock);
1681 
1682 	for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1683 		my_grp->faults[i] -= p->numa_faults_memory[i];
1684 		grp->faults[i] += p->numa_faults_memory[i];
1685 	}
1686 	my_grp->total_faults -= p->total_numa_faults;
1687 	grp->total_faults += p->total_numa_faults;
1688 
1689 	list_move(&p->numa_entry, &grp->task_list);
1690 	my_grp->nr_tasks--;
1691 	grp->nr_tasks++;
1692 
1693 	spin_unlock(&my_grp->lock);
1694 	spin_unlock(&grp->lock);
1695 
1696 	rcu_assign_pointer(p->numa_group, grp);
1697 
1698 	put_numa_group(my_grp);
1699 	return;
1700 
1701 no_join:
1702 	rcu_read_unlock();
1703 	return;
1704 }
1705 
1706 void task_numa_free(struct task_struct *p)
1707 {
1708 	struct numa_group *grp = p->numa_group;
1709 	int i;
1710 	void *numa_faults = p->numa_faults_memory;
1711 
1712 	if (grp) {
1713 		spin_lock(&grp->lock);
1714 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1715 			grp->faults[i] -= p->numa_faults_memory[i];
1716 		grp->total_faults -= p->total_numa_faults;
1717 
1718 		list_del(&p->numa_entry);
1719 		grp->nr_tasks--;
1720 		spin_unlock(&grp->lock);
1721 		rcu_assign_pointer(p->numa_group, NULL);
1722 		put_numa_group(grp);
1723 	}
1724 
1725 	p->numa_faults_memory = NULL;
1726 	p->numa_faults_buffer_memory = NULL;
1727 	p->numa_faults_cpu= NULL;
1728 	p->numa_faults_buffer_cpu = NULL;
1729 	kfree(numa_faults);
1730 }
1731 
1732 /*
1733  * Got a PROT_NONE fault for a page on @node.
1734  */
1735 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1736 {
1737 	struct task_struct *p = current;
1738 	bool migrated = flags & TNF_MIGRATED;
1739 	int cpu_node = task_node(current);
1740 	int priv;
1741 
1742 	if (!numabalancing_enabled)
1743 		return;
1744 
1745 	/* for example, ksmd faulting in a user's mm */
1746 	if (!p->mm)
1747 		return;
1748 
1749 	/* Do not worry about placement if exiting */
1750 	if (p->state == TASK_DEAD)
1751 		return;
1752 
1753 	/* Allocate buffer to track faults on a per-node basis */
1754 	if (unlikely(!p->numa_faults_memory)) {
1755 		int size = sizeof(*p->numa_faults_memory) *
1756 			   NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1757 
1758 		p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1759 		if (!p->numa_faults_memory)
1760 			return;
1761 
1762 		BUG_ON(p->numa_faults_buffer_memory);
1763 		/*
1764 		 * The averaged statistics, shared & private, memory & cpu,
1765 		 * occupy the first half of the array. The second half of the
1766 		 * array is for current counters, which are averaged into the
1767 		 * first set by task_numa_placement.
1768 		 */
1769 		p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1770 		p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1771 		p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1772 		p->total_numa_faults = 0;
1773 		memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1774 	}
1775 
1776 	/*
1777 	 * First accesses are treated as private, otherwise consider accesses
1778 	 * to be private if the accessing pid has not changed
1779 	 */
1780 	if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1781 		priv = 1;
1782 	} else {
1783 		priv = cpupid_match_pid(p, last_cpupid);
1784 		if (!priv && !(flags & TNF_NO_GROUP))
1785 			task_numa_group(p, last_cpupid, flags, &priv);
1786 	}
1787 
1788 	task_numa_placement(p);
1789 
1790 	/*
1791 	 * Retry task to preferred node migration periodically, in case it
1792 	 * case it previously failed, or the scheduler moved us.
1793 	 */
1794 	if (time_after(jiffies, p->numa_migrate_retry))
1795 		numa_migrate_preferred(p);
1796 
1797 	if (migrated)
1798 		p->numa_pages_migrated += pages;
1799 
1800 	p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1801 	p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1802 	p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1803 }
1804 
1805 static void reset_ptenuma_scan(struct task_struct *p)
1806 {
1807 	ACCESS_ONCE(p->mm->numa_scan_seq)++;
1808 	p->mm->numa_scan_offset = 0;
1809 }
1810 
1811 /*
1812  * The expensive part of numa migration is done from task_work context.
1813  * Triggered from task_tick_numa().
1814  */
1815 void task_numa_work(struct callback_head *work)
1816 {
1817 	unsigned long migrate, next_scan, now = jiffies;
1818 	struct task_struct *p = current;
1819 	struct mm_struct *mm = p->mm;
1820 	struct vm_area_struct *vma;
1821 	unsigned long start, end;
1822 	unsigned long nr_pte_updates = 0;
1823 	long pages;
1824 
1825 	WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1826 
1827 	work->next = work; /* protect against double add */
1828 	/*
1829 	 * Who cares about NUMA placement when they're dying.
1830 	 *
1831 	 * NOTE: make sure not to dereference p->mm before this check,
1832 	 * exit_task_work() happens _after_ exit_mm() so we could be called
1833 	 * without p->mm even though we still had it when we enqueued this
1834 	 * work.
1835 	 */
1836 	if (p->flags & PF_EXITING)
1837 		return;
1838 
1839 	if (!mm->numa_next_scan) {
1840 		mm->numa_next_scan = now +
1841 			msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1842 	}
1843 
1844 	/*
1845 	 * Enforce maximal scan/migration frequency..
1846 	 */
1847 	migrate = mm->numa_next_scan;
1848 	if (time_before(now, migrate))
1849 		return;
1850 
1851 	if (p->numa_scan_period == 0) {
1852 		p->numa_scan_period_max = task_scan_max(p);
1853 		p->numa_scan_period = task_scan_min(p);
1854 	}
1855 
1856 	next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1857 	if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1858 		return;
1859 
1860 	/*
1861 	 * Delay this task enough that another task of this mm will likely win
1862 	 * the next time around.
1863 	 */
1864 	p->node_stamp += 2 * TICK_NSEC;
1865 
1866 	start = mm->numa_scan_offset;
1867 	pages = sysctl_numa_balancing_scan_size;
1868 	pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1869 	if (!pages)
1870 		return;
1871 
1872 	down_read(&mm->mmap_sem);
1873 	vma = find_vma(mm, start);
1874 	if (!vma) {
1875 		reset_ptenuma_scan(p);
1876 		start = 0;
1877 		vma = mm->mmap;
1878 	}
1879 	for (; vma; vma = vma->vm_next) {
1880 		if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1881 			continue;
1882 
1883 		/*
1884 		 * Shared library pages mapped by multiple processes are not
1885 		 * migrated as it is expected they are cache replicated. Avoid
1886 		 * hinting faults in read-only file-backed mappings or the vdso
1887 		 * as migrating the pages will be of marginal benefit.
1888 		 */
1889 		if (!vma->vm_mm ||
1890 		    (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1891 			continue;
1892 
1893 		/*
1894 		 * Skip inaccessible VMAs to avoid any confusion between
1895 		 * PROT_NONE and NUMA hinting ptes
1896 		 */
1897 		if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1898 			continue;
1899 
1900 		do {
1901 			start = max(start, vma->vm_start);
1902 			end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1903 			end = min(end, vma->vm_end);
1904 			nr_pte_updates += change_prot_numa(vma, start, end);
1905 
1906 			/*
1907 			 * Scan sysctl_numa_balancing_scan_size but ensure that
1908 			 * at least one PTE is updated so that unused virtual
1909 			 * address space is quickly skipped.
1910 			 */
1911 			if (nr_pte_updates)
1912 				pages -= (end - start) >> PAGE_SHIFT;
1913 
1914 			start = end;
1915 			if (pages <= 0)
1916 				goto out;
1917 
1918 			cond_resched();
1919 		} while (end != vma->vm_end);
1920 	}
1921 
1922 out:
1923 	/*
1924 	 * It is possible to reach the end of the VMA list but the last few
1925 	 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1926 	 * would find the !migratable VMA on the next scan but not reset the
1927 	 * scanner to the start so check it now.
1928 	 */
1929 	if (vma)
1930 		mm->numa_scan_offset = start;
1931 	else
1932 		reset_ptenuma_scan(p);
1933 	up_read(&mm->mmap_sem);
1934 }
1935 
1936 /*
1937  * Drive the periodic memory faults..
1938  */
1939 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1940 {
1941 	struct callback_head *work = &curr->numa_work;
1942 	u64 period, now;
1943 
1944 	/*
1945 	 * We don't care about NUMA placement if we don't have memory.
1946 	 */
1947 	if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1948 		return;
1949 
1950 	/*
1951 	 * Using runtime rather than walltime has the dual advantage that
1952 	 * we (mostly) drive the selection from busy threads and that the
1953 	 * task needs to have done some actual work before we bother with
1954 	 * NUMA placement.
1955 	 */
1956 	now = curr->se.sum_exec_runtime;
1957 	period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1958 
1959 	if (now - curr->node_stamp > period) {
1960 		if (!curr->node_stamp)
1961 			curr->numa_scan_period = task_scan_min(curr);
1962 		curr->node_stamp += period;
1963 
1964 		if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1965 			init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1966 			task_work_add(curr, work, true);
1967 		}
1968 	}
1969 }
1970 #else
1971 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1972 {
1973 }
1974 
1975 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1976 {
1977 }
1978 
1979 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1980 {
1981 }
1982 #endif /* CONFIG_NUMA_BALANCING */
1983 
1984 static void
1985 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1986 {
1987 	update_load_add(&cfs_rq->load, se->load.weight);
1988 	if (!parent_entity(se))
1989 		update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1990 #ifdef CONFIG_SMP
1991 	if (entity_is_task(se)) {
1992 		struct rq *rq = rq_of(cfs_rq);
1993 
1994 		account_numa_enqueue(rq, task_of(se));
1995 		list_add(&se->group_node, &rq->cfs_tasks);
1996 	}
1997 #endif
1998 	cfs_rq->nr_running++;
1999 }
2000 
2001 static void
2002 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2003 {
2004 	update_load_sub(&cfs_rq->load, se->load.weight);
2005 	if (!parent_entity(se))
2006 		update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2007 	if (entity_is_task(se)) {
2008 		account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2009 		list_del_init(&se->group_node);
2010 	}
2011 	cfs_rq->nr_running--;
2012 }
2013 
2014 #ifdef CONFIG_FAIR_GROUP_SCHED
2015 # ifdef CONFIG_SMP
2016 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2017 {
2018 	long tg_weight;
2019 
2020 	/*
2021 	 * Use this CPU's actual weight instead of the last load_contribution
2022 	 * to gain a more accurate current total weight. See
2023 	 * update_cfs_rq_load_contribution().
2024 	 */
2025 	tg_weight = atomic_long_read(&tg->load_avg);
2026 	tg_weight -= cfs_rq->tg_load_contrib;
2027 	tg_weight += cfs_rq->load.weight;
2028 
2029 	return tg_weight;
2030 }
2031 
2032 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2033 {
2034 	long tg_weight, load, shares;
2035 
2036 	tg_weight = calc_tg_weight(tg, cfs_rq);
2037 	load = cfs_rq->load.weight;
2038 
2039 	shares = (tg->shares * load);
2040 	if (tg_weight)
2041 		shares /= tg_weight;
2042 
2043 	if (shares < MIN_SHARES)
2044 		shares = MIN_SHARES;
2045 	if (shares > tg->shares)
2046 		shares = tg->shares;
2047 
2048 	return shares;
2049 }
2050 # else /* CONFIG_SMP */
2051 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2052 {
2053 	return tg->shares;
2054 }
2055 # endif /* CONFIG_SMP */
2056 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2057 			    unsigned long weight)
2058 {
2059 	if (se->on_rq) {
2060 		/* commit outstanding execution time */
2061 		if (cfs_rq->curr == se)
2062 			update_curr(cfs_rq);
2063 		account_entity_dequeue(cfs_rq, se);
2064 	}
2065 
2066 	update_load_set(&se->load, weight);
2067 
2068 	if (se->on_rq)
2069 		account_entity_enqueue(cfs_rq, se);
2070 }
2071 
2072 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2073 
2074 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2075 {
2076 	struct task_group *tg;
2077 	struct sched_entity *se;
2078 	long shares;
2079 
2080 	tg = cfs_rq->tg;
2081 	se = tg->se[cpu_of(rq_of(cfs_rq))];
2082 	if (!se || throttled_hierarchy(cfs_rq))
2083 		return;
2084 #ifndef CONFIG_SMP
2085 	if (likely(se->load.weight == tg->shares))
2086 		return;
2087 #endif
2088 	shares = calc_cfs_shares(cfs_rq, tg);
2089 
2090 	reweight_entity(cfs_rq_of(se), se, shares);
2091 }
2092 #else /* CONFIG_FAIR_GROUP_SCHED */
2093 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2094 {
2095 }
2096 #endif /* CONFIG_FAIR_GROUP_SCHED */
2097 
2098 #ifdef CONFIG_SMP
2099 /*
2100  * We choose a half-life close to 1 scheduling period.
2101  * Note: The tables below are dependent on this value.
2102  */
2103 #define LOAD_AVG_PERIOD 32
2104 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2105 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2106 
2107 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2108 static const u32 runnable_avg_yN_inv[] = {
2109 	0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2110 	0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2111 	0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2112 	0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2113 	0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2114 	0x85aac367, 0x82cd8698,
2115 };
2116 
2117 /*
2118  * Precomputed \Sum y^k { 1<=k<=n }.  These are floor(true_value) to prevent
2119  * over-estimates when re-combining.
2120  */
2121 static const u32 runnable_avg_yN_sum[] = {
2122 	    0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2123 	 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2124 	17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2125 };
2126 
2127 /*
2128  * Approximate:
2129  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2130  */
2131 static __always_inline u64 decay_load(u64 val, u64 n)
2132 {
2133 	unsigned int local_n;
2134 
2135 	if (!n)
2136 		return val;
2137 	else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2138 		return 0;
2139 
2140 	/* after bounds checking we can collapse to 32-bit */
2141 	local_n = n;
2142 
2143 	/*
2144 	 * As y^PERIOD = 1/2, we can combine
2145 	 *    y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2146 	 * With a look-up table which covers k^n (n<PERIOD)
2147 	 *
2148 	 * To achieve constant time decay_load.
2149 	 */
2150 	if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2151 		val >>= local_n / LOAD_AVG_PERIOD;
2152 		local_n %= LOAD_AVG_PERIOD;
2153 	}
2154 
2155 	val *= runnable_avg_yN_inv[local_n];
2156 	/* We don't use SRR here since we always want to round down. */
2157 	return val >> 32;
2158 }
2159 
2160 /*
2161  * For updates fully spanning n periods, the contribution to runnable
2162  * average will be: \Sum 1024*y^n
2163  *
2164  * We can compute this reasonably efficiently by combining:
2165  *   y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for  n <PERIOD}
2166  */
2167 static u32 __compute_runnable_contrib(u64 n)
2168 {
2169 	u32 contrib = 0;
2170 
2171 	if (likely(n <= LOAD_AVG_PERIOD))
2172 		return runnable_avg_yN_sum[n];
2173 	else if (unlikely(n >= LOAD_AVG_MAX_N))
2174 		return LOAD_AVG_MAX;
2175 
2176 	/* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2177 	do {
2178 		contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2179 		contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2180 
2181 		n -= LOAD_AVG_PERIOD;
2182 	} while (n > LOAD_AVG_PERIOD);
2183 
2184 	contrib = decay_load(contrib, n);
2185 	return contrib + runnable_avg_yN_sum[n];
2186 }
2187 
2188 /*
2189  * We can represent the historical contribution to runnable average as the
2190  * coefficients of a geometric series.  To do this we sub-divide our runnable
2191  * history into segments of approximately 1ms (1024us); label the segment that
2192  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2193  *
2194  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2195  *      p0            p1           p2
2196  *     (now)       (~1ms ago)  (~2ms ago)
2197  *
2198  * Let u_i denote the fraction of p_i that the entity was runnable.
2199  *
2200  * We then designate the fractions u_i as our co-efficients, yielding the
2201  * following representation of historical load:
2202  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2203  *
2204  * We choose y based on the with of a reasonably scheduling period, fixing:
2205  *   y^32 = 0.5
2206  *
2207  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2208  * approximately half as much as the contribution to load within the last ms
2209  * (u_0).
2210  *
2211  * When a period "rolls over" and we have new u_0`, multiplying the previous
2212  * sum again by y is sufficient to update:
2213  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2214  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2215  */
2216 static __always_inline int __update_entity_runnable_avg(u64 now,
2217 							struct sched_avg *sa,
2218 							int runnable)
2219 {
2220 	u64 delta, periods;
2221 	u32 runnable_contrib;
2222 	int delta_w, decayed = 0;
2223 
2224 	delta = now - sa->last_runnable_update;
2225 	/*
2226 	 * This should only happen when time goes backwards, which it
2227 	 * unfortunately does during sched clock init when we swap over to TSC.
2228 	 */
2229 	if ((s64)delta < 0) {
2230 		sa->last_runnable_update = now;
2231 		return 0;
2232 	}
2233 
2234 	/*
2235 	 * Use 1024ns as the unit of measurement since it's a reasonable
2236 	 * approximation of 1us and fast to compute.
2237 	 */
2238 	delta >>= 10;
2239 	if (!delta)
2240 		return 0;
2241 	sa->last_runnable_update = now;
2242 
2243 	/* delta_w is the amount already accumulated against our next period */
2244 	delta_w = sa->runnable_avg_period % 1024;
2245 	if (delta + delta_w >= 1024) {
2246 		/* period roll-over */
2247 		decayed = 1;
2248 
2249 		/*
2250 		 * Now that we know we're crossing a period boundary, figure
2251 		 * out how much from delta we need to complete the current
2252 		 * period and accrue it.
2253 		 */
2254 		delta_w = 1024 - delta_w;
2255 		if (runnable)
2256 			sa->runnable_avg_sum += delta_w;
2257 		sa->runnable_avg_period += delta_w;
2258 
2259 		delta -= delta_w;
2260 
2261 		/* Figure out how many additional periods this update spans */
2262 		periods = delta / 1024;
2263 		delta %= 1024;
2264 
2265 		sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2266 						  periods + 1);
2267 		sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2268 						     periods + 1);
2269 
2270 		/* Efficiently calculate \sum (1..n_period) 1024*y^i */
2271 		runnable_contrib = __compute_runnable_contrib(periods);
2272 		if (runnable)
2273 			sa->runnable_avg_sum += runnable_contrib;
2274 		sa->runnable_avg_period += runnable_contrib;
2275 	}
2276 
2277 	/* Remainder of delta accrued against u_0` */
2278 	if (runnable)
2279 		sa->runnable_avg_sum += delta;
2280 	sa->runnable_avg_period += delta;
2281 
2282 	return decayed;
2283 }
2284 
2285 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2286 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2287 {
2288 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
2289 	u64 decays = atomic64_read(&cfs_rq->decay_counter);
2290 
2291 	decays -= se->avg.decay_count;
2292 	if (!decays)
2293 		return 0;
2294 
2295 	se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2296 	se->avg.decay_count = 0;
2297 
2298 	return decays;
2299 }
2300 
2301 #ifdef CONFIG_FAIR_GROUP_SCHED
2302 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2303 						 int force_update)
2304 {
2305 	struct task_group *tg = cfs_rq->tg;
2306 	long tg_contrib;
2307 
2308 	tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2309 	tg_contrib -= cfs_rq->tg_load_contrib;
2310 
2311 	if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2312 		atomic_long_add(tg_contrib, &tg->load_avg);
2313 		cfs_rq->tg_load_contrib += tg_contrib;
2314 	}
2315 }
2316 
2317 /*
2318  * Aggregate cfs_rq runnable averages into an equivalent task_group
2319  * representation for computing load contributions.
2320  */
2321 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2322 						  struct cfs_rq *cfs_rq)
2323 {
2324 	struct task_group *tg = cfs_rq->tg;
2325 	long contrib;
2326 
2327 	/* The fraction of a cpu used by this cfs_rq */
2328 	contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2329 			  sa->runnable_avg_period + 1);
2330 	contrib -= cfs_rq->tg_runnable_contrib;
2331 
2332 	if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2333 		atomic_add(contrib, &tg->runnable_avg);
2334 		cfs_rq->tg_runnable_contrib += contrib;
2335 	}
2336 }
2337 
2338 static inline void __update_group_entity_contrib(struct sched_entity *se)
2339 {
2340 	struct cfs_rq *cfs_rq = group_cfs_rq(se);
2341 	struct task_group *tg = cfs_rq->tg;
2342 	int runnable_avg;
2343 
2344 	u64 contrib;
2345 
2346 	contrib = cfs_rq->tg_load_contrib * tg->shares;
2347 	se->avg.load_avg_contrib = div_u64(contrib,
2348 				     atomic_long_read(&tg->load_avg) + 1);
2349 
2350 	/*
2351 	 * For group entities we need to compute a correction term in the case
2352 	 * that they are consuming <1 cpu so that we would contribute the same
2353 	 * load as a task of equal weight.
2354 	 *
2355 	 * Explicitly co-ordinating this measurement would be expensive, but
2356 	 * fortunately the sum of each cpus contribution forms a usable
2357 	 * lower-bound on the true value.
2358 	 *
2359 	 * Consider the aggregate of 2 contributions.  Either they are disjoint
2360 	 * (and the sum represents true value) or they are disjoint and we are
2361 	 * understating by the aggregate of their overlap.
2362 	 *
2363 	 * Extending this to N cpus, for a given overlap, the maximum amount we
2364 	 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2365 	 * cpus that overlap for this interval and w_i is the interval width.
2366 	 *
2367 	 * On a small machine; the first term is well-bounded which bounds the
2368 	 * total error since w_i is a subset of the period.  Whereas on a
2369 	 * larger machine, while this first term can be larger, if w_i is the
2370 	 * of consequential size guaranteed to see n_i*w_i quickly converge to
2371 	 * our upper bound of 1-cpu.
2372 	 */
2373 	runnable_avg = atomic_read(&tg->runnable_avg);
2374 	if (runnable_avg < NICE_0_LOAD) {
2375 		se->avg.load_avg_contrib *= runnable_avg;
2376 		se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2377 	}
2378 }
2379 
2380 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2381 {
2382 	__update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2383 	__update_tg_runnable_avg(&rq->avg, &rq->cfs);
2384 }
2385 #else /* CONFIG_FAIR_GROUP_SCHED */
2386 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2387 						 int force_update) {}
2388 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2389 						  struct cfs_rq *cfs_rq) {}
2390 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2391 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2392 #endif /* CONFIG_FAIR_GROUP_SCHED */
2393 
2394 static inline void __update_task_entity_contrib(struct sched_entity *se)
2395 {
2396 	u32 contrib;
2397 
2398 	/* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2399 	contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2400 	contrib /= (se->avg.runnable_avg_period + 1);
2401 	se->avg.load_avg_contrib = scale_load(contrib);
2402 }
2403 
2404 /* Compute the current contribution to load_avg by se, return any delta */
2405 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2406 {
2407 	long old_contrib = se->avg.load_avg_contrib;
2408 
2409 	if (entity_is_task(se)) {
2410 		__update_task_entity_contrib(se);
2411 	} else {
2412 		__update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2413 		__update_group_entity_contrib(se);
2414 	}
2415 
2416 	return se->avg.load_avg_contrib - old_contrib;
2417 }
2418 
2419 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2420 						 long load_contrib)
2421 {
2422 	if (likely(load_contrib < cfs_rq->blocked_load_avg))
2423 		cfs_rq->blocked_load_avg -= load_contrib;
2424 	else
2425 		cfs_rq->blocked_load_avg = 0;
2426 }
2427 
2428 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2429 
2430 /* Update a sched_entity's runnable average */
2431 static inline void update_entity_load_avg(struct sched_entity *se,
2432 					  int update_cfs_rq)
2433 {
2434 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
2435 	long contrib_delta;
2436 	u64 now;
2437 
2438 	/*
2439 	 * For a group entity we need to use their owned cfs_rq_clock_task() in
2440 	 * case they are the parent of a throttled hierarchy.
2441 	 */
2442 	if (entity_is_task(se))
2443 		now = cfs_rq_clock_task(cfs_rq);
2444 	else
2445 		now = cfs_rq_clock_task(group_cfs_rq(se));
2446 
2447 	if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2448 		return;
2449 
2450 	contrib_delta = __update_entity_load_avg_contrib(se);
2451 
2452 	if (!update_cfs_rq)
2453 		return;
2454 
2455 	if (se->on_rq)
2456 		cfs_rq->runnable_load_avg += contrib_delta;
2457 	else
2458 		subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2459 }
2460 
2461 /*
2462  * Decay the load contributed by all blocked children and account this so that
2463  * their contribution may appropriately discounted when they wake up.
2464  */
2465 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2466 {
2467 	u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2468 	u64 decays;
2469 
2470 	decays = now - cfs_rq->last_decay;
2471 	if (!decays && !force_update)
2472 		return;
2473 
2474 	if (atomic_long_read(&cfs_rq->removed_load)) {
2475 		unsigned long removed_load;
2476 		removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2477 		subtract_blocked_load_contrib(cfs_rq, removed_load);
2478 	}
2479 
2480 	if (decays) {
2481 		cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2482 						      decays);
2483 		atomic64_add(decays, &cfs_rq->decay_counter);
2484 		cfs_rq->last_decay = now;
2485 	}
2486 
2487 	__update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2488 }
2489 
2490 /* Add the load generated by se into cfs_rq's child load-average */
2491 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2492 						  struct sched_entity *se,
2493 						  int wakeup)
2494 {
2495 	/*
2496 	 * We track migrations using entity decay_count <= 0, on a wake-up
2497 	 * migration we use a negative decay count to track the remote decays
2498 	 * accumulated while sleeping.
2499 	 *
2500 	 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2501 	 * are seen by enqueue_entity_load_avg() as a migration with an already
2502 	 * constructed load_avg_contrib.
2503 	 */
2504 	if (unlikely(se->avg.decay_count <= 0)) {
2505 		se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2506 		if (se->avg.decay_count) {
2507 			/*
2508 			 * In a wake-up migration we have to approximate the
2509 			 * time sleeping.  This is because we can't synchronize
2510 			 * clock_task between the two cpus, and it is not
2511 			 * guaranteed to be read-safe.  Instead, we can
2512 			 * approximate this using our carried decays, which are
2513 			 * explicitly atomically readable.
2514 			 */
2515 			se->avg.last_runnable_update -= (-se->avg.decay_count)
2516 							<< 20;
2517 			update_entity_load_avg(se, 0);
2518 			/* Indicate that we're now synchronized and on-rq */
2519 			se->avg.decay_count = 0;
2520 		}
2521 		wakeup = 0;
2522 	} else {
2523 		__synchronize_entity_decay(se);
2524 	}
2525 
2526 	/* migrated tasks did not contribute to our blocked load */
2527 	if (wakeup) {
2528 		subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2529 		update_entity_load_avg(se, 0);
2530 	}
2531 
2532 	cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2533 	/* we force update consideration on load-balancer moves */
2534 	update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2535 }
2536 
2537 /*
2538  * Remove se's load from this cfs_rq child load-average, if the entity is
2539  * transitioning to a blocked state we track its projected decay using
2540  * blocked_load_avg.
2541  */
2542 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2543 						  struct sched_entity *se,
2544 						  int sleep)
2545 {
2546 	update_entity_load_avg(se, 1);
2547 	/* we force update consideration on load-balancer moves */
2548 	update_cfs_rq_blocked_load(cfs_rq, !sleep);
2549 
2550 	cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2551 	if (sleep) {
2552 		cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2553 		se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2554 	} /* migrations, e.g. sleep=0 leave decay_count == 0 */
2555 }
2556 
2557 /*
2558  * Update the rq's load with the elapsed running time before entering
2559  * idle. if the last scheduled task is not a CFS task, idle_enter will
2560  * be the only way to update the runnable statistic.
2561  */
2562 void idle_enter_fair(struct rq *this_rq)
2563 {
2564 	update_rq_runnable_avg(this_rq, 1);
2565 }
2566 
2567 /*
2568  * Update the rq's load with the elapsed idle time before a task is
2569  * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2570  * be the only way to update the runnable statistic.
2571  */
2572 void idle_exit_fair(struct rq *this_rq)
2573 {
2574 	update_rq_runnable_avg(this_rq, 0);
2575 }
2576 
2577 static int idle_balance(struct rq *this_rq);
2578 
2579 #else /* CONFIG_SMP */
2580 
2581 static inline void update_entity_load_avg(struct sched_entity *se,
2582 					  int update_cfs_rq) {}
2583 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2584 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2585 					   struct sched_entity *se,
2586 					   int wakeup) {}
2587 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2588 					   struct sched_entity *se,
2589 					   int sleep) {}
2590 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2591 					      int force_update) {}
2592 
2593 static inline int idle_balance(struct rq *rq)
2594 {
2595 	return 0;
2596 }
2597 
2598 #endif /* CONFIG_SMP */
2599 
2600 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2601 {
2602 #ifdef CONFIG_SCHEDSTATS
2603 	struct task_struct *tsk = NULL;
2604 
2605 	if (entity_is_task(se))
2606 		tsk = task_of(se);
2607 
2608 	if (se->statistics.sleep_start) {
2609 		u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2610 
2611 		if ((s64)delta < 0)
2612 			delta = 0;
2613 
2614 		if (unlikely(delta > se->statistics.sleep_max))
2615 			se->statistics.sleep_max = delta;
2616 
2617 		se->statistics.sleep_start = 0;
2618 		se->statistics.sum_sleep_runtime += delta;
2619 
2620 		if (tsk) {
2621 			account_scheduler_latency(tsk, delta >> 10, 1);
2622 			trace_sched_stat_sleep(tsk, delta);
2623 		}
2624 	}
2625 	if (se->statistics.block_start) {
2626 		u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2627 
2628 		if ((s64)delta < 0)
2629 			delta = 0;
2630 
2631 		if (unlikely(delta > se->statistics.block_max))
2632 			se->statistics.block_max = delta;
2633 
2634 		se->statistics.block_start = 0;
2635 		se->statistics.sum_sleep_runtime += delta;
2636 
2637 		if (tsk) {
2638 			if (tsk->in_iowait) {
2639 				se->statistics.iowait_sum += delta;
2640 				se->statistics.iowait_count++;
2641 				trace_sched_stat_iowait(tsk, delta);
2642 			}
2643 
2644 			trace_sched_stat_blocked(tsk, delta);
2645 
2646 			/*
2647 			 * Blocking time is in units of nanosecs, so shift by
2648 			 * 20 to get a milliseconds-range estimation of the
2649 			 * amount of time that the task spent sleeping:
2650 			 */
2651 			if (unlikely(prof_on == SLEEP_PROFILING)) {
2652 				profile_hits(SLEEP_PROFILING,
2653 						(void *)get_wchan(tsk),
2654 						delta >> 20);
2655 			}
2656 			account_scheduler_latency(tsk, delta >> 10, 0);
2657 		}
2658 	}
2659 #endif
2660 }
2661 
2662 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2663 {
2664 #ifdef CONFIG_SCHED_DEBUG
2665 	s64 d = se->vruntime - cfs_rq->min_vruntime;
2666 
2667 	if (d < 0)
2668 		d = -d;
2669 
2670 	if (d > 3*sysctl_sched_latency)
2671 		schedstat_inc(cfs_rq, nr_spread_over);
2672 #endif
2673 }
2674 
2675 static void
2676 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2677 {
2678 	u64 vruntime = cfs_rq->min_vruntime;
2679 
2680 	/*
2681 	 * The 'current' period is already promised to the current tasks,
2682 	 * however the extra weight of the new task will slow them down a
2683 	 * little, place the new task so that it fits in the slot that
2684 	 * stays open at the end.
2685 	 */
2686 	if (initial && sched_feat(START_DEBIT))
2687 		vruntime += sched_vslice(cfs_rq, se);
2688 
2689 	/* sleeps up to a single latency don't count. */
2690 	if (!initial) {
2691 		unsigned long thresh = sysctl_sched_latency;
2692 
2693 		/*
2694 		 * Halve their sleep time's effect, to allow
2695 		 * for a gentler effect of sleepers:
2696 		 */
2697 		if (sched_feat(GENTLE_FAIR_SLEEPERS))
2698 			thresh >>= 1;
2699 
2700 		vruntime -= thresh;
2701 	}
2702 
2703 	/* ensure we never gain time by being placed backwards. */
2704 	se->vruntime = max_vruntime(se->vruntime, vruntime);
2705 }
2706 
2707 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2708 
2709 static void
2710 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2711 {
2712 	/*
2713 	 * Update the normalized vruntime before updating min_vruntime
2714 	 * through calling update_curr().
2715 	 */
2716 	if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2717 		se->vruntime += cfs_rq->min_vruntime;
2718 
2719 	/*
2720 	 * Update run-time statistics of the 'current'.
2721 	 */
2722 	update_curr(cfs_rq);
2723 	enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2724 	account_entity_enqueue(cfs_rq, se);
2725 	update_cfs_shares(cfs_rq);
2726 
2727 	if (flags & ENQUEUE_WAKEUP) {
2728 		place_entity(cfs_rq, se, 0);
2729 		enqueue_sleeper(cfs_rq, se);
2730 	}
2731 
2732 	update_stats_enqueue(cfs_rq, se);
2733 	check_spread(cfs_rq, se);
2734 	if (se != cfs_rq->curr)
2735 		__enqueue_entity(cfs_rq, se);
2736 	se->on_rq = 1;
2737 
2738 	if (cfs_rq->nr_running == 1) {
2739 		list_add_leaf_cfs_rq(cfs_rq);
2740 		check_enqueue_throttle(cfs_rq);
2741 	}
2742 }
2743 
2744 static void __clear_buddies_last(struct sched_entity *se)
2745 {
2746 	for_each_sched_entity(se) {
2747 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
2748 		if (cfs_rq->last != se)
2749 			break;
2750 
2751 		cfs_rq->last = NULL;
2752 	}
2753 }
2754 
2755 static void __clear_buddies_next(struct sched_entity *se)
2756 {
2757 	for_each_sched_entity(se) {
2758 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
2759 		if (cfs_rq->next != se)
2760 			break;
2761 
2762 		cfs_rq->next = NULL;
2763 	}
2764 }
2765 
2766 static void __clear_buddies_skip(struct sched_entity *se)
2767 {
2768 	for_each_sched_entity(se) {
2769 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
2770 		if (cfs_rq->skip != se)
2771 			break;
2772 
2773 		cfs_rq->skip = NULL;
2774 	}
2775 }
2776 
2777 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2778 {
2779 	if (cfs_rq->last == se)
2780 		__clear_buddies_last(se);
2781 
2782 	if (cfs_rq->next == se)
2783 		__clear_buddies_next(se);
2784 
2785 	if (cfs_rq->skip == se)
2786 		__clear_buddies_skip(se);
2787 }
2788 
2789 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2790 
2791 static void
2792 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2793 {
2794 	/*
2795 	 * Update run-time statistics of the 'current'.
2796 	 */
2797 	update_curr(cfs_rq);
2798 	dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2799 
2800 	update_stats_dequeue(cfs_rq, se);
2801 	if (flags & DEQUEUE_SLEEP) {
2802 #ifdef CONFIG_SCHEDSTATS
2803 		if (entity_is_task(se)) {
2804 			struct task_struct *tsk = task_of(se);
2805 
2806 			if (tsk->state & TASK_INTERRUPTIBLE)
2807 				se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2808 			if (tsk->state & TASK_UNINTERRUPTIBLE)
2809 				se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2810 		}
2811 #endif
2812 	}
2813 
2814 	clear_buddies(cfs_rq, se);
2815 
2816 	if (se != cfs_rq->curr)
2817 		__dequeue_entity(cfs_rq, se);
2818 	se->on_rq = 0;
2819 	account_entity_dequeue(cfs_rq, se);
2820 
2821 	/*
2822 	 * Normalize the entity after updating the min_vruntime because the
2823 	 * update can refer to the ->curr item and we need to reflect this
2824 	 * movement in our normalized position.
2825 	 */
2826 	if (!(flags & DEQUEUE_SLEEP))
2827 		se->vruntime -= cfs_rq->min_vruntime;
2828 
2829 	/* return excess runtime on last dequeue */
2830 	return_cfs_rq_runtime(cfs_rq);
2831 
2832 	update_min_vruntime(cfs_rq);
2833 	update_cfs_shares(cfs_rq);
2834 }
2835 
2836 /*
2837  * Preempt the current task with a newly woken task if needed:
2838  */
2839 static void
2840 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2841 {
2842 	unsigned long ideal_runtime, delta_exec;
2843 	struct sched_entity *se;
2844 	s64 delta;
2845 
2846 	ideal_runtime = sched_slice(cfs_rq, curr);
2847 	delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2848 	if (delta_exec > ideal_runtime) {
2849 		resched_task(rq_of(cfs_rq)->curr);
2850 		/*
2851 		 * The current task ran long enough, ensure it doesn't get
2852 		 * re-elected due to buddy favours.
2853 		 */
2854 		clear_buddies(cfs_rq, curr);
2855 		return;
2856 	}
2857 
2858 	/*
2859 	 * Ensure that a task that missed wakeup preemption by a
2860 	 * narrow margin doesn't have to wait for a full slice.
2861 	 * This also mitigates buddy induced latencies under load.
2862 	 */
2863 	if (delta_exec < sysctl_sched_min_granularity)
2864 		return;
2865 
2866 	se = __pick_first_entity(cfs_rq);
2867 	delta = curr->vruntime - se->vruntime;
2868 
2869 	if (delta < 0)
2870 		return;
2871 
2872 	if (delta > ideal_runtime)
2873 		resched_task(rq_of(cfs_rq)->curr);
2874 }
2875 
2876 static void
2877 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2878 {
2879 	/* 'current' is not kept within the tree. */
2880 	if (se->on_rq) {
2881 		/*
2882 		 * Any task has to be enqueued before it get to execute on
2883 		 * a CPU. So account for the time it spent waiting on the
2884 		 * runqueue.
2885 		 */
2886 		update_stats_wait_end(cfs_rq, se);
2887 		__dequeue_entity(cfs_rq, se);
2888 	}
2889 
2890 	update_stats_curr_start(cfs_rq, se);
2891 	cfs_rq->curr = se;
2892 #ifdef CONFIG_SCHEDSTATS
2893 	/*
2894 	 * Track our maximum slice length, if the CPU's load is at
2895 	 * least twice that of our own weight (i.e. dont track it
2896 	 * when there are only lesser-weight tasks around):
2897 	 */
2898 	if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2899 		se->statistics.slice_max = max(se->statistics.slice_max,
2900 			se->sum_exec_runtime - se->prev_sum_exec_runtime);
2901 	}
2902 #endif
2903 	se->prev_sum_exec_runtime = se->sum_exec_runtime;
2904 }
2905 
2906 static int
2907 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2908 
2909 /*
2910  * Pick the next process, keeping these things in mind, in this order:
2911  * 1) keep things fair between processes/task groups
2912  * 2) pick the "next" process, since someone really wants that to run
2913  * 3) pick the "last" process, for cache locality
2914  * 4) do not run the "skip" process, if something else is available
2915  */
2916 static struct sched_entity *
2917 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2918 {
2919 	struct sched_entity *left = __pick_first_entity(cfs_rq);
2920 	struct sched_entity *se;
2921 
2922 	/*
2923 	 * If curr is set we have to see if its left of the leftmost entity
2924 	 * still in the tree, provided there was anything in the tree at all.
2925 	 */
2926 	if (!left || (curr && entity_before(curr, left)))
2927 		left = curr;
2928 
2929 	se = left; /* ideally we run the leftmost entity */
2930 
2931 	/*
2932 	 * Avoid running the skip buddy, if running something else can
2933 	 * be done without getting too unfair.
2934 	 */
2935 	if (cfs_rq->skip == se) {
2936 		struct sched_entity *second;
2937 
2938 		if (se == curr) {
2939 			second = __pick_first_entity(cfs_rq);
2940 		} else {
2941 			second = __pick_next_entity(se);
2942 			if (!second || (curr && entity_before(curr, second)))
2943 				second = curr;
2944 		}
2945 
2946 		if (second && wakeup_preempt_entity(second, left) < 1)
2947 			se = second;
2948 	}
2949 
2950 	/*
2951 	 * Prefer last buddy, try to return the CPU to a preempted task.
2952 	 */
2953 	if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2954 		se = cfs_rq->last;
2955 
2956 	/*
2957 	 * Someone really wants this to run. If it's not unfair, run it.
2958 	 */
2959 	if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2960 		se = cfs_rq->next;
2961 
2962 	clear_buddies(cfs_rq, se);
2963 
2964 	return se;
2965 }
2966 
2967 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2968 
2969 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2970 {
2971 	/*
2972 	 * If still on the runqueue then deactivate_task()
2973 	 * was not called and update_curr() has to be done:
2974 	 */
2975 	if (prev->on_rq)
2976 		update_curr(cfs_rq);
2977 
2978 	/* throttle cfs_rqs exceeding runtime */
2979 	check_cfs_rq_runtime(cfs_rq);
2980 
2981 	check_spread(cfs_rq, prev);
2982 	if (prev->on_rq) {
2983 		update_stats_wait_start(cfs_rq, prev);
2984 		/* Put 'current' back into the tree. */
2985 		__enqueue_entity(cfs_rq, prev);
2986 		/* in !on_rq case, update occurred at dequeue */
2987 		update_entity_load_avg(prev, 1);
2988 	}
2989 	cfs_rq->curr = NULL;
2990 }
2991 
2992 static void
2993 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2994 {
2995 	/*
2996 	 * Update run-time statistics of the 'current'.
2997 	 */
2998 	update_curr(cfs_rq);
2999 
3000 	/*
3001 	 * Ensure that runnable average is periodically updated.
3002 	 */
3003 	update_entity_load_avg(curr, 1);
3004 	update_cfs_rq_blocked_load(cfs_rq, 1);
3005 	update_cfs_shares(cfs_rq);
3006 
3007 #ifdef CONFIG_SCHED_HRTICK
3008 	/*
3009 	 * queued ticks are scheduled to match the slice, so don't bother
3010 	 * validating it and just reschedule.
3011 	 */
3012 	if (queued) {
3013 		resched_task(rq_of(cfs_rq)->curr);
3014 		return;
3015 	}
3016 	/*
3017 	 * don't let the period tick interfere with the hrtick preemption
3018 	 */
3019 	if (!sched_feat(DOUBLE_TICK) &&
3020 			hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3021 		return;
3022 #endif
3023 
3024 	if (cfs_rq->nr_running > 1)
3025 		check_preempt_tick(cfs_rq, curr);
3026 }
3027 
3028 
3029 /**************************************************
3030  * CFS bandwidth control machinery
3031  */
3032 
3033 #ifdef CONFIG_CFS_BANDWIDTH
3034 
3035 #ifdef HAVE_JUMP_LABEL
3036 static struct static_key __cfs_bandwidth_used;
3037 
3038 static inline bool cfs_bandwidth_used(void)
3039 {
3040 	return static_key_false(&__cfs_bandwidth_used);
3041 }
3042 
3043 void cfs_bandwidth_usage_inc(void)
3044 {
3045 	static_key_slow_inc(&__cfs_bandwidth_used);
3046 }
3047 
3048 void cfs_bandwidth_usage_dec(void)
3049 {
3050 	static_key_slow_dec(&__cfs_bandwidth_used);
3051 }
3052 #else /* HAVE_JUMP_LABEL */
3053 static bool cfs_bandwidth_used(void)
3054 {
3055 	return true;
3056 }
3057 
3058 void cfs_bandwidth_usage_inc(void) {}
3059 void cfs_bandwidth_usage_dec(void) {}
3060 #endif /* HAVE_JUMP_LABEL */
3061 
3062 /*
3063  * default period for cfs group bandwidth.
3064  * default: 0.1s, units: nanoseconds
3065  */
3066 static inline u64 default_cfs_period(void)
3067 {
3068 	return 100000000ULL;
3069 }
3070 
3071 static inline u64 sched_cfs_bandwidth_slice(void)
3072 {
3073 	return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3074 }
3075 
3076 /*
3077  * Replenish runtime according to assigned quota and update expiration time.
3078  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3079  * additional synchronization around rq->lock.
3080  *
3081  * requires cfs_b->lock
3082  */
3083 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3084 {
3085 	u64 now;
3086 
3087 	if (cfs_b->quota == RUNTIME_INF)
3088 		return;
3089 
3090 	now = sched_clock_cpu(smp_processor_id());
3091 	cfs_b->runtime = cfs_b->quota;
3092 	cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3093 }
3094 
3095 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3096 {
3097 	return &tg->cfs_bandwidth;
3098 }
3099 
3100 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3101 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3102 {
3103 	if (unlikely(cfs_rq->throttle_count))
3104 		return cfs_rq->throttled_clock_task;
3105 
3106 	return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3107 }
3108 
3109 /* returns 0 on failure to allocate runtime */
3110 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3111 {
3112 	struct task_group *tg = cfs_rq->tg;
3113 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3114 	u64 amount = 0, min_amount, expires;
3115 
3116 	/* note: this is a positive sum as runtime_remaining <= 0 */
3117 	min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3118 
3119 	raw_spin_lock(&cfs_b->lock);
3120 	if (cfs_b->quota == RUNTIME_INF)
3121 		amount = min_amount;
3122 	else {
3123 		/*
3124 		 * If the bandwidth pool has become inactive, then at least one
3125 		 * period must have elapsed since the last consumption.
3126 		 * Refresh the global state and ensure bandwidth timer becomes
3127 		 * active.
3128 		 */
3129 		if (!cfs_b->timer_active) {
3130 			__refill_cfs_bandwidth_runtime(cfs_b);
3131 			__start_cfs_bandwidth(cfs_b);
3132 		}
3133 
3134 		if (cfs_b->runtime > 0) {
3135 			amount = min(cfs_b->runtime, min_amount);
3136 			cfs_b->runtime -= amount;
3137 			cfs_b->idle = 0;
3138 		}
3139 	}
3140 	expires = cfs_b->runtime_expires;
3141 	raw_spin_unlock(&cfs_b->lock);
3142 
3143 	cfs_rq->runtime_remaining += amount;
3144 	/*
3145 	 * we may have advanced our local expiration to account for allowed
3146 	 * spread between our sched_clock and the one on which runtime was
3147 	 * issued.
3148 	 */
3149 	if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3150 		cfs_rq->runtime_expires = expires;
3151 
3152 	return cfs_rq->runtime_remaining > 0;
3153 }
3154 
3155 /*
3156  * Note: This depends on the synchronization provided by sched_clock and the
3157  * fact that rq->clock snapshots this value.
3158  */
3159 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3160 {
3161 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3162 
3163 	/* if the deadline is ahead of our clock, nothing to do */
3164 	if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3165 		return;
3166 
3167 	if (cfs_rq->runtime_remaining < 0)
3168 		return;
3169 
3170 	/*
3171 	 * If the local deadline has passed we have to consider the
3172 	 * possibility that our sched_clock is 'fast' and the global deadline
3173 	 * has not truly expired.
3174 	 *
3175 	 * Fortunately we can check determine whether this the case by checking
3176 	 * whether the global deadline has advanced.
3177 	 */
3178 
3179 	if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
3180 		/* extend local deadline, drift is bounded above by 2 ticks */
3181 		cfs_rq->runtime_expires += TICK_NSEC;
3182 	} else {
3183 		/* global deadline is ahead, expiration has passed */
3184 		cfs_rq->runtime_remaining = 0;
3185 	}
3186 }
3187 
3188 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3189 {
3190 	/* dock delta_exec before expiring quota (as it could span periods) */
3191 	cfs_rq->runtime_remaining -= delta_exec;
3192 	expire_cfs_rq_runtime(cfs_rq);
3193 
3194 	if (likely(cfs_rq->runtime_remaining > 0))
3195 		return;
3196 
3197 	/*
3198 	 * if we're unable to extend our runtime we resched so that the active
3199 	 * hierarchy can be throttled
3200 	 */
3201 	if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3202 		resched_task(rq_of(cfs_rq)->curr);
3203 }
3204 
3205 static __always_inline
3206 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3207 {
3208 	if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3209 		return;
3210 
3211 	__account_cfs_rq_runtime(cfs_rq, delta_exec);
3212 }
3213 
3214 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3215 {
3216 	return cfs_bandwidth_used() && cfs_rq->throttled;
3217 }
3218 
3219 /* check whether cfs_rq, or any parent, is throttled */
3220 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3221 {
3222 	return cfs_bandwidth_used() && cfs_rq->throttle_count;
3223 }
3224 
3225 /*
3226  * Ensure that neither of the group entities corresponding to src_cpu or
3227  * dest_cpu are members of a throttled hierarchy when performing group
3228  * load-balance operations.
3229  */
3230 static inline int throttled_lb_pair(struct task_group *tg,
3231 				    int src_cpu, int dest_cpu)
3232 {
3233 	struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3234 
3235 	src_cfs_rq = tg->cfs_rq[src_cpu];
3236 	dest_cfs_rq = tg->cfs_rq[dest_cpu];
3237 
3238 	return throttled_hierarchy(src_cfs_rq) ||
3239 	       throttled_hierarchy(dest_cfs_rq);
3240 }
3241 
3242 /* updated child weight may affect parent so we have to do this bottom up */
3243 static int tg_unthrottle_up(struct task_group *tg, void *data)
3244 {
3245 	struct rq *rq = data;
3246 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3247 
3248 	cfs_rq->throttle_count--;
3249 #ifdef CONFIG_SMP
3250 	if (!cfs_rq->throttle_count) {
3251 		/* adjust cfs_rq_clock_task() */
3252 		cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3253 					     cfs_rq->throttled_clock_task;
3254 	}
3255 #endif
3256 
3257 	return 0;
3258 }
3259 
3260 static int tg_throttle_down(struct task_group *tg, void *data)
3261 {
3262 	struct rq *rq = data;
3263 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3264 
3265 	/* group is entering throttled state, stop time */
3266 	if (!cfs_rq->throttle_count)
3267 		cfs_rq->throttled_clock_task = rq_clock_task(rq);
3268 	cfs_rq->throttle_count++;
3269 
3270 	return 0;
3271 }
3272 
3273 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3274 {
3275 	struct rq *rq = rq_of(cfs_rq);
3276 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3277 	struct sched_entity *se;
3278 	long task_delta, dequeue = 1;
3279 
3280 	se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3281 
3282 	/* freeze hierarchy runnable averages while throttled */
3283 	rcu_read_lock();
3284 	walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3285 	rcu_read_unlock();
3286 
3287 	task_delta = cfs_rq->h_nr_running;
3288 	for_each_sched_entity(se) {
3289 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3290 		/* throttled entity or throttle-on-deactivate */
3291 		if (!se->on_rq)
3292 			break;
3293 
3294 		if (dequeue)
3295 			dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3296 		qcfs_rq->h_nr_running -= task_delta;
3297 
3298 		if (qcfs_rq->load.weight)
3299 			dequeue = 0;
3300 	}
3301 
3302 	if (!se)
3303 		rq->nr_running -= task_delta;
3304 
3305 	cfs_rq->throttled = 1;
3306 	cfs_rq->throttled_clock = rq_clock(rq);
3307 	raw_spin_lock(&cfs_b->lock);
3308 	list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3309 	if (!cfs_b->timer_active)
3310 		__start_cfs_bandwidth(cfs_b);
3311 	raw_spin_unlock(&cfs_b->lock);
3312 }
3313 
3314 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3315 {
3316 	struct rq *rq = rq_of(cfs_rq);
3317 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3318 	struct sched_entity *se;
3319 	int enqueue = 1;
3320 	long task_delta;
3321 
3322 	se = cfs_rq->tg->se[cpu_of(rq)];
3323 
3324 	cfs_rq->throttled = 0;
3325 
3326 	update_rq_clock(rq);
3327 
3328 	raw_spin_lock(&cfs_b->lock);
3329 	cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3330 	list_del_rcu(&cfs_rq->throttled_list);
3331 	raw_spin_unlock(&cfs_b->lock);
3332 
3333 	/* update hierarchical throttle state */
3334 	walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3335 
3336 	if (!cfs_rq->load.weight)
3337 		return;
3338 
3339 	task_delta = cfs_rq->h_nr_running;
3340 	for_each_sched_entity(se) {
3341 		if (se->on_rq)
3342 			enqueue = 0;
3343 
3344 		cfs_rq = cfs_rq_of(se);
3345 		if (enqueue)
3346 			enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3347 		cfs_rq->h_nr_running += task_delta;
3348 
3349 		if (cfs_rq_throttled(cfs_rq))
3350 			break;
3351 	}
3352 
3353 	if (!se)
3354 		rq->nr_running += task_delta;
3355 
3356 	/* determine whether we need to wake up potentially idle cpu */
3357 	if (rq->curr == rq->idle && rq->cfs.nr_running)
3358 		resched_task(rq->curr);
3359 }
3360 
3361 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3362 		u64 remaining, u64 expires)
3363 {
3364 	struct cfs_rq *cfs_rq;
3365 	u64 runtime = remaining;
3366 
3367 	rcu_read_lock();
3368 	list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3369 				throttled_list) {
3370 		struct rq *rq = rq_of(cfs_rq);
3371 
3372 		raw_spin_lock(&rq->lock);
3373 		if (!cfs_rq_throttled(cfs_rq))
3374 			goto next;
3375 
3376 		runtime = -cfs_rq->runtime_remaining + 1;
3377 		if (runtime > remaining)
3378 			runtime = remaining;
3379 		remaining -= runtime;
3380 
3381 		cfs_rq->runtime_remaining += runtime;
3382 		cfs_rq->runtime_expires = expires;
3383 
3384 		/* we check whether we're throttled above */
3385 		if (cfs_rq->runtime_remaining > 0)
3386 			unthrottle_cfs_rq(cfs_rq);
3387 
3388 next:
3389 		raw_spin_unlock(&rq->lock);
3390 
3391 		if (!remaining)
3392 			break;
3393 	}
3394 	rcu_read_unlock();
3395 
3396 	return remaining;
3397 }
3398 
3399 /*
3400  * Responsible for refilling a task_group's bandwidth and unthrottling its
3401  * cfs_rqs as appropriate. If there has been no activity within the last
3402  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3403  * used to track this state.
3404  */
3405 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3406 {
3407 	u64 runtime, runtime_expires;
3408 	int idle = 1, throttled;
3409 
3410 	raw_spin_lock(&cfs_b->lock);
3411 	/* no need to continue the timer with no bandwidth constraint */
3412 	if (cfs_b->quota == RUNTIME_INF)
3413 		goto out_unlock;
3414 
3415 	throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3416 	/* idle depends on !throttled (for the case of a large deficit) */
3417 	idle = cfs_b->idle && !throttled;
3418 	cfs_b->nr_periods += overrun;
3419 
3420 	/* if we're going inactive then everything else can be deferred */
3421 	if (idle)
3422 		goto out_unlock;
3423 
3424 	/*
3425 	 * if we have relooped after returning idle once, we need to update our
3426 	 * status as actually running, so that other cpus doing
3427 	 * __start_cfs_bandwidth will stop trying to cancel us.
3428 	 */
3429 	cfs_b->timer_active = 1;
3430 
3431 	__refill_cfs_bandwidth_runtime(cfs_b);
3432 
3433 	if (!throttled) {
3434 		/* mark as potentially idle for the upcoming period */
3435 		cfs_b->idle = 1;
3436 		goto out_unlock;
3437 	}
3438 
3439 	/* account preceding periods in which throttling occurred */
3440 	cfs_b->nr_throttled += overrun;
3441 
3442 	/*
3443 	 * There are throttled entities so we must first use the new bandwidth
3444 	 * to unthrottle them before making it generally available.  This
3445 	 * ensures that all existing debts will be paid before a new cfs_rq is
3446 	 * allowed to run.
3447 	 */
3448 	runtime = cfs_b->runtime;
3449 	runtime_expires = cfs_b->runtime_expires;
3450 	cfs_b->runtime = 0;
3451 
3452 	/*
3453 	 * This check is repeated as we are holding onto the new bandwidth
3454 	 * while we unthrottle.  This can potentially race with an unthrottled
3455 	 * group trying to acquire new bandwidth from the global pool.
3456 	 */
3457 	while (throttled && runtime > 0) {
3458 		raw_spin_unlock(&cfs_b->lock);
3459 		/* we can't nest cfs_b->lock while distributing bandwidth */
3460 		runtime = distribute_cfs_runtime(cfs_b, runtime,
3461 						 runtime_expires);
3462 		raw_spin_lock(&cfs_b->lock);
3463 
3464 		throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3465 	}
3466 
3467 	/* return (any) remaining runtime */
3468 	cfs_b->runtime = runtime;
3469 	/*
3470 	 * While we are ensured activity in the period following an
3471 	 * unthrottle, this also covers the case in which the new bandwidth is
3472 	 * insufficient to cover the existing bandwidth deficit.  (Forcing the
3473 	 * timer to remain active while there are any throttled entities.)
3474 	 */
3475 	cfs_b->idle = 0;
3476 out_unlock:
3477 	if (idle)
3478 		cfs_b->timer_active = 0;
3479 	raw_spin_unlock(&cfs_b->lock);
3480 
3481 	return idle;
3482 }
3483 
3484 /* a cfs_rq won't donate quota below this amount */
3485 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3486 /* minimum remaining period time to redistribute slack quota */
3487 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3488 /* how long we wait to gather additional slack before distributing */
3489 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3490 
3491 /*
3492  * Are we near the end of the current quota period?
3493  *
3494  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3495  * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3496  * migrate_hrtimers, base is never cleared, so we are fine.
3497  */
3498 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3499 {
3500 	struct hrtimer *refresh_timer = &cfs_b->period_timer;
3501 	u64 remaining;
3502 
3503 	/* if the call-back is running a quota refresh is already occurring */
3504 	if (hrtimer_callback_running(refresh_timer))
3505 		return 1;
3506 
3507 	/* is a quota refresh about to occur? */
3508 	remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3509 	if (remaining < min_expire)
3510 		return 1;
3511 
3512 	return 0;
3513 }
3514 
3515 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3516 {
3517 	u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3518 
3519 	/* if there's a quota refresh soon don't bother with slack */
3520 	if (runtime_refresh_within(cfs_b, min_left))
3521 		return;
3522 
3523 	start_bandwidth_timer(&cfs_b->slack_timer,
3524 				ns_to_ktime(cfs_bandwidth_slack_period));
3525 }
3526 
3527 /* we know any runtime found here is valid as update_curr() precedes return */
3528 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3529 {
3530 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3531 	s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3532 
3533 	if (slack_runtime <= 0)
3534 		return;
3535 
3536 	raw_spin_lock(&cfs_b->lock);
3537 	if (cfs_b->quota != RUNTIME_INF &&
3538 	    cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3539 		cfs_b->runtime += slack_runtime;
3540 
3541 		/* we are under rq->lock, defer unthrottling using a timer */
3542 		if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3543 		    !list_empty(&cfs_b->throttled_cfs_rq))
3544 			start_cfs_slack_bandwidth(cfs_b);
3545 	}
3546 	raw_spin_unlock(&cfs_b->lock);
3547 
3548 	/* even if it's not valid for return we don't want to try again */
3549 	cfs_rq->runtime_remaining -= slack_runtime;
3550 }
3551 
3552 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3553 {
3554 	if (!cfs_bandwidth_used())
3555 		return;
3556 
3557 	if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3558 		return;
3559 
3560 	__return_cfs_rq_runtime(cfs_rq);
3561 }
3562 
3563 /*
3564  * This is done with a timer (instead of inline with bandwidth return) since
3565  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3566  */
3567 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3568 {
3569 	u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3570 	u64 expires;
3571 
3572 	/* confirm we're still not at a refresh boundary */
3573 	raw_spin_lock(&cfs_b->lock);
3574 	if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3575 		raw_spin_unlock(&cfs_b->lock);
3576 		return;
3577 	}
3578 
3579 	if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3580 		runtime = cfs_b->runtime;
3581 		cfs_b->runtime = 0;
3582 	}
3583 	expires = cfs_b->runtime_expires;
3584 	raw_spin_unlock(&cfs_b->lock);
3585 
3586 	if (!runtime)
3587 		return;
3588 
3589 	runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3590 
3591 	raw_spin_lock(&cfs_b->lock);
3592 	if (expires == cfs_b->runtime_expires)
3593 		cfs_b->runtime = runtime;
3594 	raw_spin_unlock(&cfs_b->lock);
3595 }
3596 
3597 /*
3598  * When a group wakes up we want to make sure that its quota is not already
3599  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3600  * runtime as update_curr() throttling can not not trigger until it's on-rq.
3601  */
3602 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3603 {
3604 	if (!cfs_bandwidth_used())
3605 		return;
3606 
3607 	/* an active group must be handled by the update_curr()->put() path */
3608 	if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3609 		return;
3610 
3611 	/* ensure the group is not already throttled */
3612 	if (cfs_rq_throttled(cfs_rq))
3613 		return;
3614 
3615 	/* update runtime allocation */
3616 	account_cfs_rq_runtime(cfs_rq, 0);
3617 	if (cfs_rq->runtime_remaining <= 0)
3618 		throttle_cfs_rq(cfs_rq);
3619 }
3620 
3621 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3622 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3623 {
3624 	if (!cfs_bandwidth_used())
3625 		return false;
3626 
3627 	if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3628 		return false;
3629 
3630 	/*
3631 	 * it's possible for a throttled entity to be forced into a running
3632 	 * state (e.g. set_curr_task), in this case we're finished.
3633 	 */
3634 	if (cfs_rq_throttled(cfs_rq))
3635 		return true;
3636 
3637 	throttle_cfs_rq(cfs_rq);
3638 	return true;
3639 }
3640 
3641 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3642 {
3643 	struct cfs_bandwidth *cfs_b =
3644 		container_of(timer, struct cfs_bandwidth, slack_timer);
3645 	do_sched_cfs_slack_timer(cfs_b);
3646 
3647 	return HRTIMER_NORESTART;
3648 }
3649 
3650 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3651 {
3652 	struct cfs_bandwidth *cfs_b =
3653 		container_of(timer, struct cfs_bandwidth, period_timer);
3654 	ktime_t now;
3655 	int overrun;
3656 	int idle = 0;
3657 
3658 	for (;;) {
3659 		now = hrtimer_cb_get_time(timer);
3660 		overrun = hrtimer_forward(timer, now, cfs_b->period);
3661 
3662 		if (!overrun)
3663 			break;
3664 
3665 		idle = do_sched_cfs_period_timer(cfs_b, overrun);
3666 	}
3667 
3668 	return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3669 }
3670 
3671 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3672 {
3673 	raw_spin_lock_init(&cfs_b->lock);
3674 	cfs_b->runtime = 0;
3675 	cfs_b->quota = RUNTIME_INF;
3676 	cfs_b->period = ns_to_ktime(default_cfs_period());
3677 
3678 	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3679 	hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3680 	cfs_b->period_timer.function = sched_cfs_period_timer;
3681 	hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3682 	cfs_b->slack_timer.function = sched_cfs_slack_timer;
3683 }
3684 
3685 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3686 {
3687 	cfs_rq->runtime_enabled = 0;
3688 	INIT_LIST_HEAD(&cfs_rq->throttled_list);
3689 }
3690 
3691 /* requires cfs_b->lock, may release to reprogram timer */
3692 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3693 {
3694 	/*
3695 	 * The timer may be active because we're trying to set a new bandwidth
3696 	 * period or because we're racing with the tear-down path
3697 	 * (timer_active==0 becomes visible before the hrtimer call-back
3698 	 * terminates).  In either case we ensure that it's re-programmed
3699 	 */
3700 	while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3701 	       hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3702 		/* bounce the lock to allow do_sched_cfs_period_timer to run */
3703 		raw_spin_unlock(&cfs_b->lock);
3704 		cpu_relax();
3705 		raw_spin_lock(&cfs_b->lock);
3706 		/* if someone else restarted the timer then we're done */
3707 		if (cfs_b->timer_active)
3708 			return;
3709 	}
3710 
3711 	cfs_b->timer_active = 1;
3712 	start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3713 }
3714 
3715 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3716 {
3717 	hrtimer_cancel(&cfs_b->period_timer);
3718 	hrtimer_cancel(&cfs_b->slack_timer);
3719 }
3720 
3721 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3722 {
3723 	struct cfs_rq *cfs_rq;
3724 
3725 	for_each_leaf_cfs_rq(rq, cfs_rq) {
3726 		struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3727 
3728 		if (!cfs_rq->runtime_enabled)
3729 			continue;
3730 
3731 		/*
3732 		 * clock_task is not advancing so we just need to make sure
3733 		 * there's some valid quota amount
3734 		 */
3735 		cfs_rq->runtime_remaining = cfs_b->quota;
3736 		if (cfs_rq_throttled(cfs_rq))
3737 			unthrottle_cfs_rq(cfs_rq);
3738 	}
3739 }
3740 
3741 #else /* CONFIG_CFS_BANDWIDTH */
3742 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3743 {
3744 	return rq_clock_task(rq_of(cfs_rq));
3745 }
3746 
3747 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3748 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3749 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3750 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3751 
3752 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3753 {
3754 	return 0;
3755 }
3756 
3757 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3758 {
3759 	return 0;
3760 }
3761 
3762 static inline int throttled_lb_pair(struct task_group *tg,
3763 				    int src_cpu, int dest_cpu)
3764 {
3765 	return 0;
3766 }
3767 
3768 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3769 
3770 #ifdef CONFIG_FAIR_GROUP_SCHED
3771 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3772 #endif
3773 
3774 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3775 {
3776 	return NULL;
3777 }
3778 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3779 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3780 
3781 #endif /* CONFIG_CFS_BANDWIDTH */
3782 
3783 /**************************************************
3784  * CFS operations on tasks:
3785  */
3786 
3787 #ifdef CONFIG_SCHED_HRTICK
3788 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3789 {
3790 	struct sched_entity *se = &p->se;
3791 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
3792 
3793 	WARN_ON(task_rq(p) != rq);
3794 
3795 	if (cfs_rq->nr_running > 1) {
3796 		u64 slice = sched_slice(cfs_rq, se);
3797 		u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3798 		s64 delta = slice - ran;
3799 
3800 		if (delta < 0) {
3801 			if (rq->curr == p)
3802 				resched_task(p);
3803 			return;
3804 		}
3805 
3806 		/*
3807 		 * Don't schedule slices shorter than 10000ns, that just
3808 		 * doesn't make sense. Rely on vruntime for fairness.
3809 		 */
3810 		if (rq->curr != p)
3811 			delta = max_t(s64, 10000LL, delta);
3812 
3813 		hrtick_start(rq, delta);
3814 	}
3815 }
3816 
3817 /*
3818  * called from enqueue/dequeue and updates the hrtick when the
3819  * current task is from our class and nr_running is low enough
3820  * to matter.
3821  */
3822 static void hrtick_update(struct rq *rq)
3823 {
3824 	struct task_struct *curr = rq->curr;
3825 
3826 	if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3827 		return;
3828 
3829 	if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3830 		hrtick_start_fair(rq, curr);
3831 }
3832 #else /* !CONFIG_SCHED_HRTICK */
3833 static inline void
3834 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3835 {
3836 }
3837 
3838 static inline void hrtick_update(struct rq *rq)
3839 {
3840 }
3841 #endif
3842 
3843 /*
3844  * The enqueue_task method is called before nr_running is
3845  * increased. Here we update the fair scheduling stats and
3846  * then put the task into the rbtree:
3847  */
3848 static void
3849 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3850 {
3851 	struct cfs_rq *cfs_rq;
3852 	struct sched_entity *se = &p->se;
3853 
3854 	for_each_sched_entity(se) {
3855 		if (se->on_rq)
3856 			break;
3857 		cfs_rq = cfs_rq_of(se);
3858 		enqueue_entity(cfs_rq, se, flags);
3859 
3860 		/*
3861 		 * end evaluation on encountering a throttled cfs_rq
3862 		 *
3863 		 * note: in the case of encountering a throttled cfs_rq we will
3864 		 * post the final h_nr_running increment below.
3865 		*/
3866 		if (cfs_rq_throttled(cfs_rq))
3867 			break;
3868 		cfs_rq->h_nr_running++;
3869 
3870 		flags = ENQUEUE_WAKEUP;
3871 	}
3872 
3873 	for_each_sched_entity(se) {
3874 		cfs_rq = cfs_rq_of(se);
3875 		cfs_rq->h_nr_running++;
3876 
3877 		if (cfs_rq_throttled(cfs_rq))
3878 			break;
3879 
3880 		update_cfs_shares(cfs_rq);
3881 		update_entity_load_avg(se, 1);
3882 	}
3883 
3884 	if (!se) {
3885 		update_rq_runnable_avg(rq, rq->nr_running);
3886 		inc_nr_running(rq);
3887 	}
3888 	hrtick_update(rq);
3889 }
3890 
3891 static void set_next_buddy(struct sched_entity *se);
3892 
3893 /*
3894  * The dequeue_task method is called before nr_running is
3895  * decreased. We remove the task from the rbtree and
3896  * update the fair scheduling stats:
3897  */
3898 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3899 {
3900 	struct cfs_rq *cfs_rq;
3901 	struct sched_entity *se = &p->se;
3902 	int task_sleep = flags & DEQUEUE_SLEEP;
3903 
3904 	for_each_sched_entity(se) {
3905 		cfs_rq = cfs_rq_of(se);
3906 		dequeue_entity(cfs_rq, se, flags);
3907 
3908 		/*
3909 		 * end evaluation on encountering a throttled cfs_rq
3910 		 *
3911 		 * note: in the case of encountering a throttled cfs_rq we will
3912 		 * post the final h_nr_running decrement below.
3913 		*/
3914 		if (cfs_rq_throttled(cfs_rq))
3915 			break;
3916 		cfs_rq->h_nr_running--;
3917 
3918 		/* Don't dequeue parent if it has other entities besides us */
3919 		if (cfs_rq->load.weight) {
3920 			/*
3921 			 * Bias pick_next to pick a task from this cfs_rq, as
3922 			 * p is sleeping when it is within its sched_slice.
3923 			 */
3924 			if (task_sleep && parent_entity(se))
3925 				set_next_buddy(parent_entity(se));
3926 
3927 			/* avoid re-evaluating load for this entity */
3928 			se = parent_entity(se);
3929 			break;
3930 		}
3931 		flags |= DEQUEUE_SLEEP;
3932 	}
3933 
3934 	for_each_sched_entity(se) {
3935 		cfs_rq = cfs_rq_of(se);
3936 		cfs_rq->h_nr_running--;
3937 
3938 		if (cfs_rq_throttled(cfs_rq))
3939 			break;
3940 
3941 		update_cfs_shares(cfs_rq);
3942 		update_entity_load_avg(se, 1);
3943 	}
3944 
3945 	if (!se) {
3946 		dec_nr_running(rq);
3947 		update_rq_runnable_avg(rq, 1);
3948 	}
3949 	hrtick_update(rq);
3950 }
3951 
3952 #ifdef CONFIG_SMP
3953 /* Used instead of source_load when we know the type == 0 */
3954 static unsigned long weighted_cpuload(const int cpu)
3955 {
3956 	return cpu_rq(cpu)->cfs.runnable_load_avg;
3957 }
3958 
3959 /*
3960  * Return a low guess at the load of a migration-source cpu weighted
3961  * according to the scheduling class and "nice" value.
3962  *
3963  * We want to under-estimate the load of migration sources, to
3964  * balance conservatively.
3965  */
3966 static unsigned long source_load(int cpu, int type)
3967 {
3968 	struct rq *rq = cpu_rq(cpu);
3969 	unsigned long total = weighted_cpuload(cpu);
3970 
3971 	if (type == 0 || !sched_feat(LB_BIAS))
3972 		return total;
3973 
3974 	return min(rq->cpu_load[type-1], total);
3975 }
3976 
3977 /*
3978  * Return a high guess at the load of a migration-target cpu weighted
3979  * according to the scheduling class and "nice" value.
3980  */
3981 static unsigned long target_load(int cpu, int type)
3982 {
3983 	struct rq *rq = cpu_rq(cpu);
3984 	unsigned long total = weighted_cpuload(cpu);
3985 
3986 	if (type == 0 || !sched_feat(LB_BIAS))
3987 		return total;
3988 
3989 	return max(rq->cpu_load[type-1], total);
3990 }
3991 
3992 static unsigned long power_of(int cpu)
3993 {
3994 	return cpu_rq(cpu)->cpu_power;
3995 }
3996 
3997 static unsigned long cpu_avg_load_per_task(int cpu)
3998 {
3999 	struct rq *rq = cpu_rq(cpu);
4000 	unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4001 	unsigned long load_avg = rq->cfs.runnable_load_avg;
4002 
4003 	if (nr_running)
4004 		return load_avg / nr_running;
4005 
4006 	return 0;
4007 }
4008 
4009 static void record_wakee(struct task_struct *p)
4010 {
4011 	/*
4012 	 * Rough decay (wiping) for cost saving, don't worry
4013 	 * about the boundary, really active task won't care
4014 	 * about the loss.
4015 	 */
4016 	if (jiffies > current->wakee_flip_decay_ts + HZ) {
4017 		current->wakee_flips = 0;
4018 		current->wakee_flip_decay_ts = jiffies;
4019 	}
4020 
4021 	if (current->last_wakee != p) {
4022 		current->last_wakee = p;
4023 		current->wakee_flips++;
4024 	}
4025 }
4026 
4027 static void task_waking_fair(struct task_struct *p)
4028 {
4029 	struct sched_entity *se = &p->se;
4030 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4031 	u64 min_vruntime;
4032 
4033 #ifndef CONFIG_64BIT
4034 	u64 min_vruntime_copy;
4035 
4036 	do {
4037 		min_vruntime_copy = cfs_rq->min_vruntime_copy;
4038 		smp_rmb();
4039 		min_vruntime = cfs_rq->min_vruntime;
4040 	} while (min_vruntime != min_vruntime_copy);
4041 #else
4042 	min_vruntime = cfs_rq->min_vruntime;
4043 #endif
4044 
4045 	se->vruntime -= min_vruntime;
4046 	record_wakee(p);
4047 }
4048 
4049 #ifdef CONFIG_FAIR_GROUP_SCHED
4050 /*
4051  * effective_load() calculates the load change as seen from the root_task_group
4052  *
4053  * Adding load to a group doesn't make a group heavier, but can cause movement
4054  * of group shares between cpus. Assuming the shares were perfectly aligned one
4055  * can calculate the shift in shares.
4056  *
4057  * Calculate the effective load difference if @wl is added (subtracted) to @tg
4058  * on this @cpu and results in a total addition (subtraction) of @wg to the
4059  * total group weight.
4060  *
4061  * Given a runqueue weight distribution (rw_i) we can compute a shares
4062  * distribution (s_i) using:
4063  *
4064  *   s_i = rw_i / \Sum rw_j						(1)
4065  *
4066  * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4067  * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4068  * shares distribution (s_i):
4069  *
4070  *   rw_i = {   2,   4,   1,   0 }
4071  *   s_i  = { 2/7, 4/7, 1/7,   0 }
4072  *
4073  * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4074  * task used to run on and the CPU the waker is running on), we need to
4075  * compute the effect of waking a task on either CPU and, in case of a sync
4076  * wakeup, compute the effect of the current task going to sleep.
4077  *
4078  * So for a change of @wl to the local @cpu with an overall group weight change
4079  * of @wl we can compute the new shares distribution (s'_i) using:
4080  *
4081  *   s'_i = (rw_i + @wl) / (@wg + \Sum rw_j)				(2)
4082  *
4083  * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4084  * differences in waking a task to CPU 0. The additional task changes the
4085  * weight and shares distributions like:
4086  *
4087  *   rw'_i = {   3,   4,   1,   0 }
4088  *   s'_i  = { 3/8, 4/8, 1/8,   0 }
4089  *
4090  * We can then compute the difference in effective weight by using:
4091  *
4092  *   dw_i = S * (s'_i - s_i)						(3)
4093  *
4094  * Where 'S' is the group weight as seen by its parent.
4095  *
4096  * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4097  * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4098  * 4/7) times the weight of the group.
4099  */
4100 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4101 {
4102 	struct sched_entity *se = tg->se[cpu];
4103 
4104 	if (!tg->parent)	/* the trivial, non-cgroup case */
4105 		return wl;
4106 
4107 	for_each_sched_entity(se) {
4108 		long w, W;
4109 
4110 		tg = se->my_q->tg;
4111 
4112 		/*
4113 		 * W = @wg + \Sum rw_j
4114 		 */
4115 		W = wg + calc_tg_weight(tg, se->my_q);
4116 
4117 		/*
4118 		 * w = rw_i + @wl
4119 		 */
4120 		w = se->my_q->load.weight + wl;
4121 
4122 		/*
4123 		 * wl = S * s'_i; see (2)
4124 		 */
4125 		if (W > 0 && w < W)
4126 			wl = (w * tg->shares) / W;
4127 		else
4128 			wl = tg->shares;
4129 
4130 		/*
4131 		 * Per the above, wl is the new se->load.weight value; since
4132 		 * those are clipped to [MIN_SHARES, ...) do so now. See
4133 		 * calc_cfs_shares().
4134 		 */
4135 		if (wl < MIN_SHARES)
4136 			wl = MIN_SHARES;
4137 
4138 		/*
4139 		 * wl = dw_i = S * (s'_i - s_i); see (3)
4140 		 */
4141 		wl -= se->load.weight;
4142 
4143 		/*
4144 		 * Recursively apply this logic to all parent groups to compute
4145 		 * the final effective load change on the root group. Since
4146 		 * only the @tg group gets extra weight, all parent groups can
4147 		 * only redistribute existing shares. @wl is the shift in shares
4148 		 * resulting from this level per the above.
4149 		 */
4150 		wg = 0;
4151 	}
4152 
4153 	return wl;
4154 }
4155 #else
4156 
4157 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4158 {
4159 	return wl;
4160 }
4161 
4162 #endif
4163 
4164 static int wake_wide(struct task_struct *p)
4165 {
4166 	int factor = this_cpu_read(sd_llc_size);
4167 
4168 	/*
4169 	 * Yeah, it's the switching-frequency, could means many wakee or
4170 	 * rapidly switch, use factor here will just help to automatically
4171 	 * adjust the loose-degree, so bigger node will lead to more pull.
4172 	 */
4173 	if (p->wakee_flips > factor) {
4174 		/*
4175 		 * wakee is somewhat hot, it needs certain amount of cpu
4176 		 * resource, so if waker is far more hot, prefer to leave
4177 		 * it alone.
4178 		 */
4179 		if (current->wakee_flips > (factor * p->wakee_flips))
4180 			return 1;
4181 	}
4182 
4183 	return 0;
4184 }
4185 
4186 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4187 {
4188 	s64 this_load, load;
4189 	int idx, this_cpu, prev_cpu;
4190 	unsigned long tl_per_task;
4191 	struct task_group *tg;
4192 	unsigned long weight;
4193 	int balanced;
4194 
4195 	/*
4196 	 * If we wake multiple tasks be careful to not bounce
4197 	 * ourselves around too much.
4198 	 */
4199 	if (wake_wide(p))
4200 		return 0;
4201 
4202 	idx	  = sd->wake_idx;
4203 	this_cpu  = smp_processor_id();
4204 	prev_cpu  = task_cpu(p);
4205 	load	  = source_load(prev_cpu, idx);
4206 	this_load = target_load(this_cpu, idx);
4207 
4208 	/*
4209 	 * If sync wakeup then subtract the (maximum possible)
4210 	 * effect of the currently running task from the load
4211 	 * of the current CPU:
4212 	 */
4213 	if (sync) {
4214 		tg = task_group(current);
4215 		weight = current->se.load.weight;
4216 
4217 		this_load += effective_load(tg, this_cpu, -weight, -weight);
4218 		load += effective_load(tg, prev_cpu, 0, -weight);
4219 	}
4220 
4221 	tg = task_group(p);
4222 	weight = p->se.load.weight;
4223 
4224 	/*
4225 	 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4226 	 * due to the sync cause above having dropped this_load to 0, we'll
4227 	 * always have an imbalance, but there's really nothing you can do
4228 	 * about that, so that's good too.
4229 	 *
4230 	 * Otherwise check if either cpus are near enough in load to allow this
4231 	 * task to be woken on this_cpu.
4232 	 */
4233 	if (this_load > 0) {
4234 		s64 this_eff_load, prev_eff_load;
4235 
4236 		this_eff_load = 100;
4237 		this_eff_load *= power_of(prev_cpu);
4238 		this_eff_load *= this_load +
4239 			effective_load(tg, this_cpu, weight, weight);
4240 
4241 		prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4242 		prev_eff_load *= power_of(this_cpu);
4243 		prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4244 
4245 		balanced = this_eff_load <= prev_eff_load;
4246 	} else
4247 		balanced = true;
4248 
4249 	/*
4250 	 * If the currently running task will sleep within
4251 	 * a reasonable amount of time then attract this newly
4252 	 * woken task:
4253 	 */
4254 	if (sync && balanced)
4255 		return 1;
4256 
4257 	schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4258 	tl_per_task = cpu_avg_load_per_task(this_cpu);
4259 
4260 	if (balanced ||
4261 	    (this_load <= load &&
4262 	     this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4263 		/*
4264 		 * This domain has SD_WAKE_AFFINE and
4265 		 * p is cache cold in this domain, and
4266 		 * there is no bad imbalance.
4267 		 */
4268 		schedstat_inc(sd, ttwu_move_affine);
4269 		schedstat_inc(p, se.statistics.nr_wakeups_affine);
4270 
4271 		return 1;
4272 	}
4273 	return 0;
4274 }
4275 
4276 /*
4277  * find_idlest_group finds and returns the least busy CPU group within the
4278  * domain.
4279  */
4280 static struct sched_group *
4281 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4282 		  int this_cpu, int sd_flag)
4283 {
4284 	struct sched_group *idlest = NULL, *group = sd->groups;
4285 	unsigned long min_load = ULONG_MAX, this_load = 0;
4286 	int load_idx = sd->forkexec_idx;
4287 	int imbalance = 100 + (sd->imbalance_pct-100)/2;
4288 
4289 	if (sd_flag & SD_BALANCE_WAKE)
4290 		load_idx = sd->wake_idx;
4291 
4292 	do {
4293 		unsigned long load, avg_load;
4294 		int local_group;
4295 		int i;
4296 
4297 		/* Skip over this group if it has no CPUs allowed */
4298 		if (!cpumask_intersects(sched_group_cpus(group),
4299 					tsk_cpus_allowed(p)))
4300 			continue;
4301 
4302 		local_group = cpumask_test_cpu(this_cpu,
4303 					       sched_group_cpus(group));
4304 
4305 		/* Tally up the load of all CPUs in the group */
4306 		avg_load = 0;
4307 
4308 		for_each_cpu(i, sched_group_cpus(group)) {
4309 			/* Bias balancing toward cpus of our domain */
4310 			if (local_group)
4311 				load = source_load(i, load_idx);
4312 			else
4313 				load = target_load(i, load_idx);
4314 
4315 			avg_load += load;
4316 		}
4317 
4318 		/* Adjust by relative CPU power of the group */
4319 		avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4320 
4321 		if (local_group) {
4322 			this_load = avg_load;
4323 		} else if (avg_load < min_load) {
4324 			min_load = avg_load;
4325 			idlest = group;
4326 		}
4327 	} while (group = group->next, group != sd->groups);
4328 
4329 	if (!idlest || 100*this_load < imbalance*min_load)
4330 		return NULL;
4331 	return idlest;
4332 }
4333 
4334 /*
4335  * find_idlest_cpu - find the idlest cpu among the cpus in group.
4336  */
4337 static int
4338 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4339 {
4340 	unsigned long load, min_load = ULONG_MAX;
4341 	int idlest = -1;
4342 	int i;
4343 
4344 	/* Traverse only the allowed CPUs */
4345 	for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4346 		load = weighted_cpuload(i);
4347 
4348 		if (load < min_load || (load == min_load && i == this_cpu)) {
4349 			min_load = load;
4350 			idlest = i;
4351 		}
4352 	}
4353 
4354 	return idlest;
4355 }
4356 
4357 /*
4358  * Try and locate an idle CPU in the sched_domain.
4359  */
4360 static int select_idle_sibling(struct task_struct *p, int target)
4361 {
4362 	struct sched_domain *sd;
4363 	struct sched_group *sg;
4364 	int i = task_cpu(p);
4365 
4366 	if (idle_cpu(target))
4367 		return target;
4368 
4369 	/*
4370 	 * If the prevous cpu is cache affine and idle, don't be stupid.
4371 	 */
4372 	if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4373 		return i;
4374 
4375 	/*
4376 	 * Otherwise, iterate the domains and find an elegible idle cpu.
4377 	 */
4378 	sd = rcu_dereference(per_cpu(sd_llc, target));
4379 	for_each_lower_domain(sd) {
4380 		sg = sd->groups;
4381 		do {
4382 			if (!cpumask_intersects(sched_group_cpus(sg),
4383 						tsk_cpus_allowed(p)))
4384 				goto next;
4385 
4386 			for_each_cpu(i, sched_group_cpus(sg)) {
4387 				if (i == target || !idle_cpu(i))
4388 					goto next;
4389 			}
4390 
4391 			target = cpumask_first_and(sched_group_cpus(sg),
4392 					tsk_cpus_allowed(p));
4393 			goto done;
4394 next:
4395 			sg = sg->next;
4396 		} while (sg != sd->groups);
4397 	}
4398 done:
4399 	return target;
4400 }
4401 
4402 /*
4403  * select_task_rq_fair: Select target runqueue for the waking task in domains
4404  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4405  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4406  *
4407  * Balances load by selecting the idlest cpu in the idlest group, or under
4408  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4409  *
4410  * Returns the target cpu number.
4411  *
4412  * preempt must be disabled.
4413  */
4414 static int
4415 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4416 {
4417 	struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4418 	int cpu = smp_processor_id();
4419 	int new_cpu = cpu;
4420 	int want_affine = 0;
4421 	int sync = wake_flags & WF_SYNC;
4422 
4423 	if (p->nr_cpus_allowed == 1)
4424 		return prev_cpu;
4425 
4426 	if (sd_flag & SD_BALANCE_WAKE) {
4427 		if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4428 			want_affine = 1;
4429 		new_cpu = prev_cpu;
4430 	}
4431 
4432 	rcu_read_lock();
4433 	for_each_domain(cpu, tmp) {
4434 		if (!(tmp->flags & SD_LOAD_BALANCE))
4435 			continue;
4436 
4437 		/*
4438 		 * If both cpu and prev_cpu are part of this domain,
4439 		 * cpu is a valid SD_WAKE_AFFINE target.
4440 		 */
4441 		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4442 		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4443 			affine_sd = tmp;
4444 			break;
4445 		}
4446 
4447 		if (tmp->flags & sd_flag)
4448 			sd = tmp;
4449 	}
4450 
4451 	if (affine_sd) {
4452 		if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4453 			prev_cpu = cpu;
4454 
4455 		new_cpu = select_idle_sibling(p, prev_cpu);
4456 		goto unlock;
4457 	}
4458 
4459 	while (sd) {
4460 		struct sched_group *group;
4461 		int weight;
4462 
4463 		if (!(sd->flags & sd_flag)) {
4464 			sd = sd->child;
4465 			continue;
4466 		}
4467 
4468 		group = find_idlest_group(sd, p, cpu, sd_flag);
4469 		if (!group) {
4470 			sd = sd->child;
4471 			continue;
4472 		}
4473 
4474 		new_cpu = find_idlest_cpu(group, p, cpu);
4475 		if (new_cpu == -1 || new_cpu == cpu) {
4476 			/* Now try balancing at a lower domain level of cpu */
4477 			sd = sd->child;
4478 			continue;
4479 		}
4480 
4481 		/* Now try balancing at a lower domain level of new_cpu */
4482 		cpu = new_cpu;
4483 		weight = sd->span_weight;
4484 		sd = NULL;
4485 		for_each_domain(cpu, tmp) {
4486 			if (weight <= tmp->span_weight)
4487 				break;
4488 			if (tmp->flags & sd_flag)
4489 				sd = tmp;
4490 		}
4491 		/* while loop will break here if sd == NULL */
4492 	}
4493 unlock:
4494 	rcu_read_unlock();
4495 
4496 	return new_cpu;
4497 }
4498 
4499 /*
4500  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4501  * cfs_rq_of(p) references at time of call are still valid and identify the
4502  * previous cpu.  However, the caller only guarantees p->pi_lock is held; no
4503  * other assumptions, including the state of rq->lock, should be made.
4504  */
4505 static void
4506 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4507 {
4508 	struct sched_entity *se = &p->se;
4509 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4510 
4511 	/*
4512 	 * Load tracking: accumulate removed load so that it can be processed
4513 	 * when we next update owning cfs_rq under rq->lock.  Tasks contribute
4514 	 * to blocked load iff they have a positive decay-count.  It can never
4515 	 * be negative here since on-rq tasks have decay-count == 0.
4516 	 */
4517 	if (se->avg.decay_count) {
4518 		se->avg.decay_count = -__synchronize_entity_decay(se);
4519 		atomic_long_add(se->avg.load_avg_contrib,
4520 						&cfs_rq->removed_load);
4521 	}
4522 }
4523 #endif /* CONFIG_SMP */
4524 
4525 static unsigned long
4526 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4527 {
4528 	unsigned long gran = sysctl_sched_wakeup_granularity;
4529 
4530 	/*
4531 	 * Since its curr running now, convert the gran from real-time
4532 	 * to virtual-time in his units.
4533 	 *
4534 	 * By using 'se' instead of 'curr' we penalize light tasks, so
4535 	 * they get preempted easier. That is, if 'se' < 'curr' then
4536 	 * the resulting gran will be larger, therefore penalizing the
4537 	 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4538 	 * be smaller, again penalizing the lighter task.
4539 	 *
4540 	 * This is especially important for buddies when the leftmost
4541 	 * task is higher priority than the buddy.
4542 	 */
4543 	return calc_delta_fair(gran, se);
4544 }
4545 
4546 /*
4547  * Should 'se' preempt 'curr'.
4548  *
4549  *             |s1
4550  *        |s2
4551  *   |s3
4552  *         g
4553  *      |<--->|c
4554  *
4555  *  w(c, s1) = -1
4556  *  w(c, s2) =  0
4557  *  w(c, s3) =  1
4558  *
4559  */
4560 static int
4561 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4562 {
4563 	s64 gran, vdiff = curr->vruntime - se->vruntime;
4564 
4565 	if (vdiff <= 0)
4566 		return -1;
4567 
4568 	gran = wakeup_gran(curr, se);
4569 	if (vdiff > gran)
4570 		return 1;
4571 
4572 	return 0;
4573 }
4574 
4575 static void set_last_buddy(struct sched_entity *se)
4576 {
4577 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4578 		return;
4579 
4580 	for_each_sched_entity(se)
4581 		cfs_rq_of(se)->last = se;
4582 }
4583 
4584 static void set_next_buddy(struct sched_entity *se)
4585 {
4586 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4587 		return;
4588 
4589 	for_each_sched_entity(se)
4590 		cfs_rq_of(se)->next = se;
4591 }
4592 
4593 static void set_skip_buddy(struct sched_entity *se)
4594 {
4595 	for_each_sched_entity(se)
4596 		cfs_rq_of(se)->skip = se;
4597 }
4598 
4599 /*
4600  * Preempt the current task with a newly woken task if needed:
4601  */
4602 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4603 {
4604 	struct task_struct *curr = rq->curr;
4605 	struct sched_entity *se = &curr->se, *pse = &p->se;
4606 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4607 	int scale = cfs_rq->nr_running >= sched_nr_latency;
4608 	int next_buddy_marked = 0;
4609 
4610 	if (unlikely(se == pse))
4611 		return;
4612 
4613 	/*
4614 	 * This is possible from callers such as move_task(), in which we
4615 	 * unconditionally check_prempt_curr() after an enqueue (which may have
4616 	 * lead to a throttle).  This both saves work and prevents false
4617 	 * next-buddy nomination below.
4618 	 */
4619 	if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4620 		return;
4621 
4622 	if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4623 		set_next_buddy(pse);
4624 		next_buddy_marked = 1;
4625 	}
4626 
4627 	/*
4628 	 * We can come here with TIF_NEED_RESCHED already set from new task
4629 	 * wake up path.
4630 	 *
4631 	 * Note: this also catches the edge-case of curr being in a throttled
4632 	 * group (e.g. via set_curr_task), since update_curr() (in the
4633 	 * enqueue of curr) will have resulted in resched being set.  This
4634 	 * prevents us from potentially nominating it as a false LAST_BUDDY
4635 	 * below.
4636 	 */
4637 	if (test_tsk_need_resched(curr))
4638 		return;
4639 
4640 	/* Idle tasks are by definition preempted by non-idle tasks. */
4641 	if (unlikely(curr->policy == SCHED_IDLE) &&
4642 	    likely(p->policy != SCHED_IDLE))
4643 		goto preempt;
4644 
4645 	/*
4646 	 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4647 	 * is driven by the tick):
4648 	 */
4649 	if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4650 		return;
4651 
4652 	find_matching_se(&se, &pse);
4653 	update_curr(cfs_rq_of(se));
4654 	BUG_ON(!pse);
4655 	if (wakeup_preempt_entity(se, pse) == 1) {
4656 		/*
4657 		 * Bias pick_next to pick the sched entity that is
4658 		 * triggering this preemption.
4659 		 */
4660 		if (!next_buddy_marked)
4661 			set_next_buddy(pse);
4662 		goto preempt;
4663 	}
4664 
4665 	return;
4666 
4667 preempt:
4668 	resched_task(curr);
4669 	/*
4670 	 * Only set the backward buddy when the current task is still
4671 	 * on the rq. This can happen when a wakeup gets interleaved
4672 	 * with schedule on the ->pre_schedule() or idle_balance()
4673 	 * point, either of which can * drop the rq lock.
4674 	 *
4675 	 * Also, during early boot the idle thread is in the fair class,
4676 	 * for obvious reasons its a bad idea to schedule back to it.
4677 	 */
4678 	if (unlikely(!se->on_rq || curr == rq->idle))
4679 		return;
4680 
4681 	if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4682 		set_last_buddy(se);
4683 }
4684 
4685 static struct task_struct *
4686 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4687 {
4688 	struct cfs_rq *cfs_rq = &rq->cfs;
4689 	struct sched_entity *se;
4690 	struct task_struct *p;
4691 	int new_tasks;
4692 
4693 again:
4694 #ifdef CONFIG_FAIR_GROUP_SCHED
4695 	if (!cfs_rq->nr_running)
4696 		goto idle;
4697 
4698 	if (prev->sched_class != &fair_sched_class)
4699 		goto simple;
4700 
4701 	/*
4702 	 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4703 	 * likely that a next task is from the same cgroup as the current.
4704 	 *
4705 	 * Therefore attempt to avoid putting and setting the entire cgroup
4706 	 * hierarchy, only change the part that actually changes.
4707 	 */
4708 
4709 	do {
4710 		struct sched_entity *curr = cfs_rq->curr;
4711 
4712 		/*
4713 		 * Since we got here without doing put_prev_entity() we also
4714 		 * have to consider cfs_rq->curr. If it is still a runnable
4715 		 * entity, update_curr() will update its vruntime, otherwise
4716 		 * forget we've ever seen it.
4717 		 */
4718 		if (curr && curr->on_rq)
4719 			update_curr(cfs_rq);
4720 		else
4721 			curr = NULL;
4722 
4723 		/*
4724 		 * This call to check_cfs_rq_runtime() will do the throttle and
4725 		 * dequeue its entity in the parent(s). Therefore the 'simple'
4726 		 * nr_running test will indeed be correct.
4727 		 */
4728 		if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4729 			goto simple;
4730 
4731 		se = pick_next_entity(cfs_rq, curr);
4732 		cfs_rq = group_cfs_rq(se);
4733 	} while (cfs_rq);
4734 
4735 	p = task_of(se);
4736 
4737 	/*
4738 	 * Since we haven't yet done put_prev_entity and if the selected task
4739 	 * is a different task than we started out with, try and touch the
4740 	 * least amount of cfs_rqs.
4741 	 */
4742 	if (prev != p) {
4743 		struct sched_entity *pse = &prev->se;
4744 
4745 		while (!(cfs_rq = is_same_group(se, pse))) {
4746 			int se_depth = se->depth;
4747 			int pse_depth = pse->depth;
4748 
4749 			if (se_depth <= pse_depth) {
4750 				put_prev_entity(cfs_rq_of(pse), pse);
4751 				pse = parent_entity(pse);
4752 			}
4753 			if (se_depth >= pse_depth) {
4754 				set_next_entity(cfs_rq_of(se), se);
4755 				se = parent_entity(se);
4756 			}
4757 		}
4758 
4759 		put_prev_entity(cfs_rq, pse);
4760 		set_next_entity(cfs_rq, se);
4761 	}
4762 
4763 	if (hrtick_enabled(rq))
4764 		hrtick_start_fair(rq, p);
4765 
4766 	return p;
4767 simple:
4768 	cfs_rq = &rq->cfs;
4769 #endif
4770 
4771 	if (!cfs_rq->nr_running)
4772 		goto idle;
4773 
4774 	put_prev_task(rq, prev);
4775 
4776 	do {
4777 		se = pick_next_entity(cfs_rq, NULL);
4778 		set_next_entity(cfs_rq, se);
4779 		cfs_rq = group_cfs_rq(se);
4780 	} while (cfs_rq);
4781 
4782 	p = task_of(se);
4783 
4784 	if (hrtick_enabled(rq))
4785 		hrtick_start_fair(rq, p);
4786 
4787 	return p;
4788 
4789 idle:
4790 	new_tasks = idle_balance(rq);
4791 	/*
4792 	 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4793 	 * possible for any higher priority task to appear. In that case we
4794 	 * must re-start the pick_next_entity() loop.
4795 	 */
4796 	if (new_tasks < 0)
4797 		return RETRY_TASK;
4798 
4799 	if (new_tasks > 0)
4800 		goto again;
4801 
4802 	return NULL;
4803 }
4804 
4805 /*
4806  * Account for a descheduled task:
4807  */
4808 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4809 {
4810 	struct sched_entity *se = &prev->se;
4811 	struct cfs_rq *cfs_rq;
4812 
4813 	for_each_sched_entity(se) {
4814 		cfs_rq = cfs_rq_of(se);
4815 		put_prev_entity(cfs_rq, se);
4816 	}
4817 }
4818 
4819 /*
4820  * sched_yield() is very simple
4821  *
4822  * The magic of dealing with the ->skip buddy is in pick_next_entity.
4823  */
4824 static void yield_task_fair(struct rq *rq)
4825 {
4826 	struct task_struct *curr = rq->curr;
4827 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4828 	struct sched_entity *se = &curr->se;
4829 
4830 	/*
4831 	 * Are we the only task in the tree?
4832 	 */
4833 	if (unlikely(rq->nr_running == 1))
4834 		return;
4835 
4836 	clear_buddies(cfs_rq, se);
4837 
4838 	if (curr->policy != SCHED_BATCH) {
4839 		update_rq_clock(rq);
4840 		/*
4841 		 * Update run-time statistics of the 'current'.
4842 		 */
4843 		update_curr(cfs_rq);
4844 		/*
4845 		 * Tell update_rq_clock() that we've just updated,
4846 		 * so we don't do microscopic update in schedule()
4847 		 * and double the fastpath cost.
4848 		 */
4849 		 rq->skip_clock_update = 1;
4850 	}
4851 
4852 	set_skip_buddy(se);
4853 }
4854 
4855 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4856 {
4857 	struct sched_entity *se = &p->se;
4858 
4859 	/* throttled hierarchies are not runnable */
4860 	if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4861 		return false;
4862 
4863 	/* Tell the scheduler that we'd really like pse to run next. */
4864 	set_next_buddy(se);
4865 
4866 	yield_task_fair(rq);
4867 
4868 	return true;
4869 }
4870 
4871 #ifdef CONFIG_SMP
4872 /**************************************************
4873  * Fair scheduling class load-balancing methods.
4874  *
4875  * BASICS
4876  *
4877  * The purpose of load-balancing is to achieve the same basic fairness the
4878  * per-cpu scheduler provides, namely provide a proportional amount of compute
4879  * time to each task. This is expressed in the following equation:
4880  *
4881  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
4882  *
4883  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4884  * W_i,0 is defined as:
4885  *
4886  *   W_i,0 = \Sum_j w_i,j                                             (2)
4887  *
4888  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4889  * is derived from the nice value as per prio_to_weight[].
4890  *
4891  * The weight average is an exponential decay average of the instantaneous
4892  * weight:
4893  *
4894  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
4895  *
4896  * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4897  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4898  * can also include other factors [XXX].
4899  *
4900  * To achieve this balance we define a measure of imbalance which follows
4901  * directly from (1):
4902  *
4903  *   imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j }    (4)
4904  *
4905  * We them move tasks around to minimize the imbalance. In the continuous
4906  * function space it is obvious this converges, in the discrete case we get
4907  * a few fun cases generally called infeasible weight scenarios.
4908  *
4909  * [XXX expand on:
4910  *     - infeasible weights;
4911  *     - local vs global optima in the discrete case. ]
4912  *
4913  *
4914  * SCHED DOMAINS
4915  *
4916  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4917  * for all i,j solution, we create a tree of cpus that follows the hardware
4918  * topology where each level pairs two lower groups (or better). This results
4919  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4920  * tree to only the first of the previous level and we decrease the frequency
4921  * of load-balance at each level inv. proportional to the number of cpus in
4922  * the groups.
4923  *
4924  * This yields:
4925  *
4926  *     log_2 n     1     n
4927  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
4928  *     i = 0      2^i   2^i
4929  *                               `- size of each group
4930  *         |         |     `- number of cpus doing load-balance
4931  *         |         `- freq
4932  *         `- sum over all levels
4933  *
4934  * Coupled with a limit on how many tasks we can migrate every balance pass,
4935  * this makes (5) the runtime complexity of the balancer.
4936  *
4937  * An important property here is that each CPU is still (indirectly) connected
4938  * to every other cpu in at most O(log n) steps:
4939  *
4940  * The adjacency matrix of the resulting graph is given by:
4941  *
4942  *             log_2 n
4943  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
4944  *             k = 0
4945  *
4946  * And you'll find that:
4947  *
4948  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
4949  *
4950  * Showing there's indeed a path between every cpu in at most O(log n) steps.
4951  * The task movement gives a factor of O(m), giving a convergence complexity
4952  * of:
4953  *
4954  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
4955  *
4956  *
4957  * WORK CONSERVING
4958  *
4959  * In order to avoid CPUs going idle while there's still work to do, new idle
4960  * balancing is more aggressive and has the newly idle cpu iterate up the domain
4961  * tree itself instead of relying on other CPUs to bring it work.
4962  *
4963  * This adds some complexity to both (5) and (8) but it reduces the total idle
4964  * time.
4965  *
4966  * [XXX more?]
4967  *
4968  *
4969  * CGROUPS
4970  *
4971  * Cgroups make a horror show out of (2), instead of a simple sum we get:
4972  *
4973  *                                s_k,i
4974  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
4975  *                                 S_k
4976  *
4977  * Where
4978  *
4979  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
4980  *
4981  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4982  *
4983  * The big problem is S_k, its a global sum needed to compute a local (W_i)
4984  * property.
4985  *
4986  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4987  *      rewrite all of this once again.]
4988  */
4989 
4990 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4991 
4992 enum fbq_type { regular, remote, all };
4993 
4994 #define LBF_ALL_PINNED	0x01
4995 #define LBF_NEED_BREAK	0x02
4996 #define LBF_DST_PINNED  0x04
4997 #define LBF_SOME_PINNED	0x08
4998 
4999 struct lb_env {
5000 	struct sched_domain	*sd;
5001 
5002 	struct rq		*src_rq;
5003 	int			src_cpu;
5004 
5005 	int			dst_cpu;
5006 	struct rq		*dst_rq;
5007 
5008 	struct cpumask		*dst_grpmask;
5009 	int			new_dst_cpu;
5010 	enum cpu_idle_type	idle;
5011 	long			imbalance;
5012 	/* The set of CPUs under consideration for load-balancing */
5013 	struct cpumask		*cpus;
5014 
5015 	unsigned int		flags;
5016 
5017 	unsigned int		loop;
5018 	unsigned int		loop_break;
5019 	unsigned int		loop_max;
5020 
5021 	enum fbq_type		fbq_type;
5022 };
5023 
5024 /*
5025  * move_task - move a task from one runqueue to another runqueue.
5026  * Both runqueues must be locked.
5027  */
5028 static void move_task(struct task_struct *p, struct lb_env *env)
5029 {
5030 	deactivate_task(env->src_rq, p, 0);
5031 	set_task_cpu(p, env->dst_cpu);
5032 	activate_task(env->dst_rq, p, 0);
5033 	check_preempt_curr(env->dst_rq, p, 0);
5034 }
5035 
5036 /*
5037  * Is this task likely cache-hot:
5038  */
5039 static int
5040 task_hot(struct task_struct *p, u64 now)
5041 {
5042 	s64 delta;
5043 
5044 	if (p->sched_class != &fair_sched_class)
5045 		return 0;
5046 
5047 	if (unlikely(p->policy == SCHED_IDLE))
5048 		return 0;
5049 
5050 	/*
5051 	 * Buddy candidates are cache hot:
5052 	 */
5053 	if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
5054 			(&p->se == cfs_rq_of(&p->se)->next ||
5055 			 &p->se == cfs_rq_of(&p->se)->last))
5056 		return 1;
5057 
5058 	if (sysctl_sched_migration_cost == -1)
5059 		return 1;
5060 	if (sysctl_sched_migration_cost == 0)
5061 		return 0;
5062 
5063 	delta = now - p->se.exec_start;
5064 
5065 	return delta < (s64)sysctl_sched_migration_cost;
5066 }
5067 
5068 #ifdef CONFIG_NUMA_BALANCING
5069 /* Returns true if the destination node has incurred more faults */
5070 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5071 {
5072 	int src_nid, dst_nid;
5073 
5074 	if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5075 	    !(env->sd->flags & SD_NUMA)) {
5076 		return false;
5077 	}
5078 
5079 	src_nid = cpu_to_node(env->src_cpu);
5080 	dst_nid = cpu_to_node(env->dst_cpu);
5081 
5082 	if (src_nid == dst_nid)
5083 		return false;
5084 
5085 	/* Always encourage migration to the preferred node. */
5086 	if (dst_nid == p->numa_preferred_nid)
5087 		return true;
5088 
5089 	/* If both task and group weight improve, this move is a winner. */
5090 	if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
5091 	    group_weight(p, dst_nid) > group_weight(p, src_nid))
5092 		return true;
5093 
5094 	return false;
5095 }
5096 
5097 
5098 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5099 {
5100 	int src_nid, dst_nid;
5101 
5102 	if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5103 		return false;
5104 
5105 	if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5106 		return false;
5107 
5108 	src_nid = cpu_to_node(env->src_cpu);
5109 	dst_nid = cpu_to_node(env->dst_cpu);
5110 
5111 	if (src_nid == dst_nid)
5112 		return false;
5113 
5114 	/* Migrating away from the preferred node is always bad. */
5115 	if (src_nid == p->numa_preferred_nid)
5116 		return true;
5117 
5118 	/* If either task or group weight get worse, don't do it. */
5119 	if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
5120 	    group_weight(p, dst_nid) < group_weight(p, src_nid))
5121 		return true;
5122 
5123 	return false;
5124 }
5125 
5126 #else
5127 static inline bool migrate_improves_locality(struct task_struct *p,
5128 					     struct lb_env *env)
5129 {
5130 	return false;
5131 }
5132 
5133 static inline bool migrate_degrades_locality(struct task_struct *p,
5134 					     struct lb_env *env)
5135 {
5136 	return false;
5137 }
5138 #endif
5139 
5140 /*
5141  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5142  */
5143 static
5144 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5145 {
5146 	int tsk_cache_hot = 0;
5147 	/*
5148 	 * We do not migrate tasks that are:
5149 	 * 1) throttled_lb_pair, or
5150 	 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5151 	 * 3) running (obviously), or
5152 	 * 4) are cache-hot on their current CPU.
5153 	 */
5154 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5155 		return 0;
5156 
5157 	if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5158 		int cpu;
5159 
5160 		schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5161 
5162 		env->flags |= LBF_SOME_PINNED;
5163 
5164 		/*
5165 		 * Remember if this task can be migrated to any other cpu in
5166 		 * our sched_group. We may want to revisit it if we couldn't
5167 		 * meet load balance goals by pulling other tasks on src_cpu.
5168 		 *
5169 		 * Also avoid computing new_dst_cpu if we have already computed
5170 		 * one in current iteration.
5171 		 */
5172 		if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5173 			return 0;
5174 
5175 		/* Prevent to re-select dst_cpu via env's cpus */
5176 		for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5177 			if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5178 				env->flags |= LBF_DST_PINNED;
5179 				env->new_dst_cpu = cpu;
5180 				break;
5181 			}
5182 		}
5183 
5184 		return 0;
5185 	}
5186 
5187 	/* Record that we found atleast one task that could run on dst_cpu */
5188 	env->flags &= ~LBF_ALL_PINNED;
5189 
5190 	if (task_running(env->src_rq, p)) {
5191 		schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5192 		return 0;
5193 	}
5194 
5195 	/*
5196 	 * Aggressive migration if:
5197 	 * 1) destination numa is preferred
5198 	 * 2) task is cache cold, or
5199 	 * 3) too many balance attempts have failed.
5200 	 */
5201 	tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq));
5202 	if (!tsk_cache_hot)
5203 		tsk_cache_hot = migrate_degrades_locality(p, env);
5204 
5205 	if (migrate_improves_locality(p, env)) {
5206 #ifdef CONFIG_SCHEDSTATS
5207 		if (tsk_cache_hot) {
5208 			schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5209 			schedstat_inc(p, se.statistics.nr_forced_migrations);
5210 		}
5211 #endif
5212 		return 1;
5213 	}
5214 
5215 	if (!tsk_cache_hot ||
5216 		env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5217 
5218 		if (tsk_cache_hot) {
5219 			schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5220 			schedstat_inc(p, se.statistics.nr_forced_migrations);
5221 		}
5222 
5223 		return 1;
5224 	}
5225 
5226 	schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5227 	return 0;
5228 }
5229 
5230 /*
5231  * move_one_task tries to move exactly one task from busiest to this_rq, as
5232  * part of active balancing operations within "domain".
5233  * Returns 1 if successful and 0 otherwise.
5234  *
5235  * Called with both runqueues locked.
5236  */
5237 static int move_one_task(struct lb_env *env)
5238 {
5239 	struct task_struct *p, *n;
5240 
5241 	list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5242 		if (!can_migrate_task(p, env))
5243 			continue;
5244 
5245 		move_task(p, env);
5246 		/*
5247 		 * Right now, this is only the second place move_task()
5248 		 * is called, so we can safely collect move_task()
5249 		 * stats here rather than inside move_task().
5250 		 */
5251 		schedstat_inc(env->sd, lb_gained[env->idle]);
5252 		return 1;
5253 	}
5254 	return 0;
5255 }
5256 
5257 static const unsigned int sched_nr_migrate_break = 32;
5258 
5259 /*
5260  * move_tasks tries to move up to imbalance weighted load from busiest to
5261  * this_rq, as part of a balancing operation within domain "sd".
5262  * Returns 1 if successful and 0 otherwise.
5263  *
5264  * Called with both runqueues locked.
5265  */
5266 static int move_tasks(struct lb_env *env)
5267 {
5268 	struct list_head *tasks = &env->src_rq->cfs_tasks;
5269 	struct task_struct *p;
5270 	unsigned long load;
5271 	int pulled = 0;
5272 
5273 	if (env->imbalance <= 0)
5274 		return 0;
5275 
5276 	while (!list_empty(tasks)) {
5277 		p = list_first_entry(tasks, struct task_struct, se.group_node);
5278 
5279 		env->loop++;
5280 		/* We've more or less seen every task there is, call it quits */
5281 		if (env->loop > env->loop_max)
5282 			break;
5283 
5284 		/* take a breather every nr_migrate tasks */
5285 		if (env->loop > env->loop_break) {
5286 			env->loop_break += sched_nr_migrate_break;
5287 			env->flags |= LBF_NEED_BREAK;
5288 			break;
5289 		}
5290 
5291 		if (!can_migrate_task(p, env))
5292 			goto next;
5293 
5294 		load = task_h_load(p);
5295 
5296 		if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5297 			goto next;
5298 
5299 		if ((load / 2) > env->imbalance)
5300 			goto next;
5301 
5302 		move_task(p, env);
5303 		pulled++;
5304 		env->imbalance -= load;
5305 
5306 #ifdef CONFIG_PREEMPT
5307 		/*
5308 		 * NEWIDLE balancing is a source of latency, so preemptible
5309 		 * kernels will stop after the first task is pulled to minimize
5310 		 * the critical section.
5311 		 */
5312 		if (env->idle == CPU_NEWLY_IDLE)
5313 			break;
5314 #endif
5315 
5316 		/*
5317 		 * We only want to steal up to the prescribed amount of
5318 		 * weighted load.
5319 		 */
5320 		if (env->imbalance <= 0)
5321 			break;
5322 
5323 		continue;
5324 next:
5325 		list_move_tail(&p->se.group_node, tasks);
5326 	}
5327 
5328 	/*
5329 	 * Right now, this is one of only two places move_task() is called,
5330 	 * so we can safely collect move_task() stats here rather than
5331 	 * inside move_task().
5332 	 */
5333 	schedstat_add(env->sd, lb_gained[env->idle], pulled);
5334 
5335 	return pulled;
5336 }
5337 
5338 #ifdef CONFIG_FAIR_GROUP_SCHED
5339 /*
5340  * update tg->load_weight by folding this cpu's load_avg
5341  */
5342 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5343 {
5344 	struct sched_entity *se = tg->se[cpu];
5345 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5346 
5347 	/* throttled entities do not contribute to load */
5348 	if (throttled_hierarchy(cfs_rq))
5349 		return;
5350 
5351 	update_cfs_rq_blocked_load(cfs_rq, 1);
5352 
5353 	if (se) {
5354 		update_entity_load_avg(se, 1);
5355 		/*
5356 		 * We pivot on our runnable average having decayed to zero for
5357 		 * list removal.  This generally implies that all our children
5358 		 * have also been removed (modulo rounding error or bandwidth
5359 		 * control); however, such cases are rare and we can fix these
5360 		 * at enqueue.
5361 		 *
5362 		 * TODO: fix up out-of-order children on enqueue.
5363 		 */
5364 		if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5365 			list_del_leaf_cfs_rq(cfs_rq);
5366 	} else {
5367 		struct rq *rq = rq_of(cfs_rq);
5368 		update_rq_runnable_avg(rq, rq->nr_running);
5369 	}
5370 }
5371 
5372 static void update_blocked_averages(int cpu)
5373 {
5374 	struct rq *rq = cpu_rq(cpu);
5375 	struct cfs_rq *cfs_rq;
5376 	unsigned long flags;
5377 
5378 	raw_spin_lock_irqsave(&rq->lock, flags);
5379 	update_rq_clock(rq);
5380 	/*
5381 	 * Iterates the task_group tree in a bottom up fashion, see
5382 	 * list_add_leaf_cfs_rq() for details.
5383 	 */
5384 	for_each_leaf_cfs_rq(rq, cfs_rq) {
5385 		/*
5386 		 * Note: We may want to consider periodically releasing
5387 		 * rq->lock about these updates so that creating many task
5388 		 * groups does not result in continually extending hold time.
5389 		 */
5390 		__update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5391 	}
5392 
5393 	raw_spin_unlock_irqrestore(&rq->lock, flags);
5394 }
5395 
5396 /*
5397  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5398  * This needs to be done in a top-down fashion because the load of a child
5399  * group is a fraction of its parents load.
5400  */
5401 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5402 {
5403 	struct rq *rq = rq_of(cfs_rq);
5404 	struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5405 	unsigned long now = jiffies;
5406 	unsigned long load;
5407 
5408 	if (cfs_rq->last_h_load_update == now)
5409 		return;
5410 
5411 	cfs_rq->h_load_next = NULL;
5412 	for_each_sched_entity(se) {
5413 		cfs_rq = cfs_rq_of(se);
5414 		cfs_rq->h_load_next = se;
5415 		if (cfs_rq->last_h_load_update == now)
5416 			break;
5417 	}
5418 
5419 	if (!se) {
5420 		cfs_rq->h_load = cfs_rq->runnable_load_avg;
5421 		cfs_rq->last_h_load_update = now;
5422 	}
5423 
5424 	while ((se = cfs_rq->h_load_next) != NULL) {
5425 		load = cfs_rq->h_load;
5426 		load = div64_ul(load * se->avg.load_avg_contrib,
5427 				cfs_rq->runnable_load_avg + 1);
5428 		cfs_rq = group_cfs_rq(se);
5429 		cfs_rq->h_load = load;
5430 		cfs_rq->last_h_load_update = now;
5431 	}
5432 }
5433 
5434 static unsigned long task_h_load(struct task_struct *p)
5435 {
5436 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
5437 
5438 	update_cfs_rq_h_load(cfs_rq);
5439 	return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5440 			cfs_rq->runnable_load_avg + 1);
5441 }
5442 #else
5443 static inline void update_blocked_averages(int cpu)
5444 {
5445 }
5446 
5447 static unsigned long task_h_load(struct task_struct *p)
5448 {
5449 	return p->se.avg.load_avg_contrib;
5450 }
5451 #endif
5452 
5453 /********** Helpers for find_busiest_group ************************/
5454 /*
5455  * sg_lb_stats - stats of a sched_group required for load_balancing
5456  */
5457 struct sg_lb_stats {
5458 	unsigned long avg_load; /*Avg load across the CPUs of the group */
5459 	unsigned long group_load; /* Total load over the CPUs of the group */
5460 	unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5461 	unsigned long load_per_task;
5462 	unsigned long group_power;
5463 	unsigned int sum_nr_running; /* Nr tasks running in the group */
5464 	unsigned int group_capacity;
5465 	unsigned int idle_cpus;
5466 	unsigned int group_weight;
5467 	int group_imb; /* Is there an imbalance in the group ? */
5468 	int group_has_capacity; /* Is there extra capacity in the group? */
5469 #ifdef CONFIG_NUMA_BALANCING
5470 	unsigned int nr_numa_running;
5471 	unsigned int nr_preferred_running;
5472 #endif
5473 };
5474 
5475 /*
5476  * sd_lb_stats - Structure to store the statistics of a sched_domain
5477  *		 during load balancing.
5478  */
5479 struct sd_lb_stats {
5480 	struct sched_group *busiest;	/* Busiest group in this sd */
5481 	struct sched_group *local;	/* Local group in this sd */
5482 	unsigned long total_load;	/* Total load of all groups in sd */
5483 	unsigned long total_pwr;	/* Total power of all groups in sd */
5484 	unsigned long avg_load;	/* Average load across all groups in sd */
5485 
5486 	struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5487 	struct sg_lb_stats local_stat;	/* Statistics of the local group */
5488 };
5489 
5490 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5491 {
5492 	/*
5493 	 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5494 	 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5495 	 * We must however clear busiest_stat::avg_load because
5496 	 * update_sd_pick_busiest() reads this before assignment.
5497 	 */
5498 	*sds = (struct sd_lb_stats){
5499 		.busiest = NULL,
5500 		.local = NULL,
5501 		.total_load = 0UL,
5502 		.total_pwr = 0UL,
5503 		.busiest_stat = {
5504 			.avg_load = 0UL,
5505 		},
5506 	};
5507 }
5508 
5509 /**
5510  * get_sd_load_idx - Obtain the load index for a given sched domain.
5511  * @sd: The sched_domain whose load_idx is to be obtained.
5512  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5513  *
5514  * Return: The load index.
5515  */
5516 static inline int get_sd_load_idx(struct sched_domain *sd,
5517 					enum cpu_idle_type idle)
5518 {
5519 	int load_idx;
5520 
5521 	switch (idle) {
5522 	case CPU_NOT_IDLE:
5523 		load_idx = sd->busy_idx;
5524 		break;
5525 
5526 	case CPU_NEWLY_IDLE:
5527 		load_idx = sd->newidle_idx;
5528 		break;
5529 	default:
5530 		load_idx = sd->idle_idx;
5531 		break;
5532 	}
5533 
5534 	return load_idx;
5535 }
5536 
5537 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5538 {
5539 	return SCHED_POWER_SCALE;
5540 }
5541 
5542 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5543 {
5544 	return default_scale_freq_power(sd, cpu);
5545 }
5546 
5547 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5548 {
5549 	unsigned long weight = sd->span_weight;
5550 	unsigned long smt_gain = sd->smt_gain;
5551 
5552 	smt_gain /= weight;
5553 
5554 	return smt_gain;
5555 }
5556 
5557 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5558 {
5559 	return default_scale_smt_power(sd, cpu);
5560 }
5561 
5562 static unsigned long scale_rt_power(int cpu)
5563 {
5564 	struct rq *rq = cpu_rq(cpu);
5565 	u64 total, available, age_stamp, avg;
5566 
5567 	/*
5568 	 * Since we're reading these variables without serialization make sure
5569 	 * we read them once before doing sanity checks on them.
5570 	 */
5571 	age_stamp = ACCESS_ONCE(rq->age_stamp);
5572 	avg = ACCESS_ONCE(rq->rt_avg);
5573 
5574 	total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5575 
5576 	if (unlikely(total < avg)) {
5577 		/* Ensures that power won't end up being negative */
5578 		available = 0;
5579 	} else {
5580 		available = total - avg;
5581 	}
5582 
5583 	if (unlikely((s64)total < SCHED_POWER_SCALE))
5584 		total = SCHED_POWER_SCALE;
5585 
5586 	total >>= SCHED_POWER_SHIFT;
5587 
5588 	return div_u64(available, total);
5589 }
5590 
5591 static void update_cpu_power(struct sched_domain *sd, int cpu)
5592 {
5593 	unsigned long weight = sd->span_weight;
5594 	unsigned long power = SCHED_POWER_SCALE;
5595 	struct sched_group *sdg = sd->groups;
5596 
5597 	if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5598 		if (sched_feat(ARCH_POWER))
5599 			power *= arch_scale_smt_power(sd, cpu);
5600 		else
5601 			power *= default_scale_smt_power(sd, cpu);
5602 
5603 		power >>= SCHED_POWER_SHIFT;
5604 	}
5605 
5606 	sdg->sgp->power_orig = power;
5607 
5608 	if (sched_feat(ARCH_POWER))
5609 		power *= arch_scale_freq_power(sd, cpu);
5610 	else
5611 		power *= default_scale_freq_power(sd, cpu);
5612 
5613 	power >>= SCHED_POWER_SHIFT;
5614 
5615 	power *= scale_rt_power(cpu);
5616 	power >>= SCHED_POWER_SHIFT;
5617 
5618 	if (!power)
5619 		power = 1;
5620 
5621 	cpu_rq(cpu)->cpu_power = power;
5622 	sdg->sgp->power = power;
5623 }
5624 
5625 void update_group_power(struct sched_domain *sd, int cpu)
5626 {
5627 	struct sched_domain *child = sd->child;
5628 	struct sched_group *group, *sdg = sd->groups;
5629 	unsigned long power, power_orig;
5630 	unsigned long interval;
5631 
5632 	interval = msecs_to_jiffies(sd->balance_interval);
5633 	interval = clamp(interval, 1UL, max_load_balance_interval);
5634 	sdg->sgp->next_update = jiffies + interval;
5635 
5636 	if (!child) {
5637 		update_cpu_power(sd, cpu);
5638 		return;
5639 	}
5640 
5641 	power_orig = power = 0;
5642 
5643 	if (child->flags & SD_OVERLAP) {
5644 		/*
5645 		 * SD_OVERLAP domains cannot assume that child groups
5646 		 * span the current group.
5647 		 */
5648 
5649 		for_each_cpu(cpu, sched_group_cpus(sdg)) {
5650 			struct sched_group_power *sgp;
5651 			struct rq *rq = cpu_rq(cpu);
5652 
5653 			/*
5654 			 * build_sched_domains() -> init_sched_groups_power()
5655 			 * gets here before we've attached the domains to the
5656 			 * runqueues.
5657 			 *
5658 			 * Use power_of(), which is set irrespective of domains
5659 			 * in update_cpu_power().
5660 			 *
5661 			 * This avoids power/power_orig from being 0 and
5662 			 * causing divide-by-zero issues on boot.
5663 			 *
5664 			 * Runtime updates will correct power_orig.
5665 			 */
5666 			if (unlikely(!rq->sd)) {
5667 				power_orig += power_of(cpu);
5668 				power += power_of(cpu);
5669 				continue;
5670 			}
5671 
5672 			sgp = rq->sd->groups->sgp;
5673 			power_orig += sgp->power_orig;
5674 			power += sgp->power;
5675 		}
5676 	} else  {
5677 		/*
5678 		 * !SD_OVERLAP domains can assume that child groups
5679 		 * span the current group.
5680 		 */
5681 
5682 		group = child->groups;
5683 		do {
5684 			power_orig += group->sgp->power_orig;
5685 			power += group->sgp->power;
5686 			group = group->next;
5687 		} while (group != child->groups);
5688 	}
5689 
5690 	sdg->sgp->power_orig = power_orig;
5691 	sdg->sgp->power = power;
5692 }
5693 
5694 /*
5695  * Try and fix up capacity for tiny siblings, this is needed when
5696  * things like SD_ASYM_PACKING need f_b_g to select another sibling
5697  * which on its own isn't powerful enough.
5698  *
5699  * See update_sd_pick_busiest() and check_asym_packing().
5700  */
5701 static inline int
5702 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5703 {
5704 	/*
5705 	 * Only siblings can have significantly less than SCHED_POWER_SCALE
5706 	 */
5707 	if (!(sd->flags & SD_SHARE_CPUPOWER))
5708 		return 0;
5709 
5710 	/*
5711 	 * If ~90% of the cpu_power is still there, we're good.
5712 	 */
5713 	if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5714 		return 1;
5715 
5716 	return 0;
5717 }
5718 
5719 /*
5720  * Group imbalance indicates (and tries to solve) the problem where balancing
5721  * groups is inadequate due to tsk_cpus_allowed() constraints.
5722  *
5723  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5724  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5725  * Something like:
5726  *
5727  * 	{ 0 1 2 3 } { 4 5 6 7 }
5728  * 	        *     * * *
5729  *
5730  * If we were to balance group-wise we'd place two tasks in the first group and
5731  * two tasks in the second group. Clearly this is undesired as it will overload
5732  * cpu 3 and leave one of the cpus in the second group unused.
5733  *
5734  * The current solution to this issue is detecting the skew in the first group
5735  * by noticing the lower domain failed to reach balance and had difficulty
5736  * moving tasks due to affinity constraints.
5737  *
5738  * When this is so detected; this group becomes a candidate for busiest; see
5739  * update_sd_pick_busiest(). And calculate_imbalance() and
5740  * find_busiest_group() avoid some of the usual balance conditions to allow it
5741  * to create an effective group imbalance.
5742  *
5743  * This is a somewhat tricky proposition since the next run might not find the
5744  * group imbalance and decide the groups need to be balanced again. A most
5745  * subtle and fragile situation.
5746  */
5747 
5748 static inline int sg_imbalanced(struct sched_group *group)
5749 {
5750 	return group->sgp->imbalance;
5751 }
5752 
5753 /*
5754  * Compute the group capacity.
5755  *
5756  * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5757  * first dividing out the smt factor and computing the actual number of cores
5758  * and limit power unit capacity with that.
5759  */
5760 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5761 {
5762 	unsigned int capacity, smt, cpus;
5763 	unsigned int power, power_orig;
5764 
5765 	power = group->sgp->power;
5766 	power_orig = group->sgp->power_orig;
5767 	cpus = group->group_weight;
5768 
5769 	/* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5770 	smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5771 	capacity = cpus / smt; /* cores */
5772 
5773 	capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5774 	if (!capacity)
5775 		capacity = fix_small_capacity(env->sd, group);
5776 
5777 	return capacity;
5778 }
5779 
5780 /**
5781  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5782  * @env: The load balancing environment.
5783  * @group: sched_group whose statistics are to be updated.
5784  * @load_idx: Load index of sched_domain of this_cpu for load calc.
5785  * @local_group: Does group contain this_cpu.
5786  * @sgs: variable to hold the statistics for this group.
5787  */
5788 static inline void update_sg_lb_stats(struct lb_env *env,
5789 			struct sched_group *group, int load_idx,
5790 			int local_group, struct sg_lb_stats *sgs)
5791 {
5792 	unsigned long load;
5793 	int i;
5794 
5795 	memset(sgs, 0, sizeof(*sgs));
5796 
5797 	for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5798 		struct rq *rq = cpu_rq(i);
5799 
5800 		/* Bias balancing toward cpus of our domain */
5801 		if (local_group)
5802 			load = target_load(i, load_idx);
5803 		else
5804 			load = source_load(i, load_idx);
5805 
5806 		sgs->group_load += load;
5807 		sgs->sum_nr_running += rq->nr_running;
5808 #ifdef CONFIG_NUMA_BALANCING
5809 		sgs->nr_numa_running += rq->nr_numa_running;
5810 		sgs->nr_preferred_running += rq->nr_preferred_running;
5811 #endif
5812 		sgs->sum_weighted_load += weighted_cpuload(i);
5813 		if (idle_cpu(i))
5814 			sgs->idle_cpus++;
5815 	}
5816 
5817 	/* Adjust by relative CPU power of the group */
5818 	sgs->group_power = group->sgp->power;
5819 	sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5820 
5821 	if (sgs->sum_nr_running)
5822 		sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5823 
5824 	sgs->group_weight = group->group_weight;
5825 
5826 	sgs->group_imb = sg_imbalanced(group);
5827 	sgs->group_capacity = sg_capacity(env, group);
5828 
5829 	if (sgs->group_capacity > sgs->sum_nr_running)
5830 		sgs->group_has_capacity = 1;
5831 }
5832 
5833 /**
5834  * update_sd_pick_busiest - return 1 on busiest group
5835  * @env: The load balancing environment.
5836  * @sds: sched_domain statistics
5837  * @sg: sched_group candidate to be checked for being the busiest
5838  * @sgs: sched_group statistics
5839  *
5840  * Determine if @sg is a busier group than the previously selected
5841  * busiest group.
5842  *
5843  * Return: %true if @sg is a busier group than the previously selected
5844  * busiest group. %false otherwise.
5845  */
5846 static bool update_sd_pick_busiest(struct lb_env *env,
5847 				   struct sd_lb_stats *sds,
5848 				   struct sched_group *sg,
5849 				   struct sg_lb_stats *sgs)
5850 {
5851 	if (sgs->avg_load <= sds->busiest_stat.avg_load)
5852 		return false;
5853 
5854 	if (sgs->sum_nr_running > sgs->group_capacity)
5855 		return true;
5856 
5857 	if (sgs->group_imb)
5858 		return true;
5859 
5860 	/*
5861 	 * ASYM_PACKING needs to move all the work to the lowest
5862 	 * numbered CPUs in the group, therefore mark all groups
5863 	 * higher than ourself as busy.
5864 	 */
5865 	if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5866 	    env->dst_cpu < group_first_cpu(sg)) {
5867 		if (!sds->busiest)
5868 			return true;
5869 
5870 		if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5871 			return true;
5872 	}
5873 
5874 	return false;
5875 }
5876 
5877 #ifdef CONFIG_NUMA_BALANCING
5878 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5879 {
5880 	if (sgs->sum_nr_running > sgs->nr_numa_running)
5881 		return regular;
5882 	if (sgs->sum_nr_running > sgs->nr_preferred_running)
5883 		return remote;
5884 	return all;
5885 }
5886 
5887 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5888 {
5889 	if (rq->nr_running > rq->nr_numa_running)
5890 		return regular;
5891 	if (rq->nr_running > rq->nr_preferred_running)
5892 		return remote;
5893 	return all;
5894 }
5895 #else
5896 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5897 {
5898 	return all;
5899 }
5900 
5901 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5902 {
5903 	return regular;
5904 }
5905 #endif /* CONFIG_NUMA_BALANCING */
5906 
5907 /**
5908  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5909  * @env: The load balancing environment.
5910  * @sds: variable to hold the statistics for this sched_domain.
5911  */
5912 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5913 {
5914 	struct sched_domain *child = env->sd->child;
5915 	struct sched_group *sg = env->sd->groups;
5916 	struct sg_lb_stats tmp_sgs;
5917 	int load_idx, prefer_sibling = 0;
5918 
5919 	if (child && child->flags & SD_PREFER_SIBLING)
5920 		prefer_sibling = 1;
5921 
5922 	load_idx = get_sd_load_idx(env->sd, env->idle);
5923 
5924 	do {
5925 		struct sg_lb_stats *sgs = &tmp_sgs;
5926 		int local_group;
5927 
5928 		local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5929 		if (local_group) {
5930 			sds->local = sg;
5931 			sgs = &sds->local_stat;
5932 
5933 			if (env->idle != CPU_NEWLY_IDLE ||
5934 			    time_after_eq(jiffies, sg->sgp->next_update))
5935 				update_group_power(env->sd, env->dst_cpu);
5936 		}
5937 
5938 		update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5939 
5940 		if (local_group)
5941 			goto next_group;
5942 
5943 		/*
5944 		 * In case the child domain prefers tasks go to siblings
5945 		 * first, lower the sg capacity to one so that we'll try
5946 		 * and move all the excess tasks away. We lower the capacity
5947 		 * of a group only if the local group has the capacity to fit
5948 		 * these excess tasks, i.e. nr_running < group_capacity. The
5949 		 * extra check prevents the case where you always pull from the
5950 		 * heaviest group when it is already under-utilized (possible
5951 		 * with a large weight task outweighs the tasks on the system).
5952 		 */
5953 		if (prefer_sibling && sds->local &&
5954 		    sds->local_stat.group_has_capacity)
5955 			sgs->group_capacity = min(sgs->group_capacity, 1U);
5956 
5957 		if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5958 			sds->busiest = sg;
5959 			sds->busiest_stat = *sgs;
5960 		}
5961 
5962 next_group:
5963 		/* Now, start updating sd_lb_stats */
5964 		sds->total_load += sgs->group_load;
5965 		sds->total_pwr += sgs->group_power;
5966 
5967 		sg = sg->next;
5968 	} while (sg != env->sd->groups);
5969 
5970 	if (env->sd->flags & SD_NUMA)
5971 		env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5972 }
5973 
5974 /**
5975  * check_asym_packing - Check to see if the group is packed into the
5976  *			sched doman.
5977  *
5978  * This is primarily intended to used at the sibling level.  Some
5979  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
5980  * case of POWER7, it can move to lower SMT modes only when higher
5981  * threads are idle.  When in lower SMT modes, the threads will
5982  * perform better since they share less core resources.  Hence when we
5983  * have idle threads, we want them to be the higher ones.
5984  *
5985  * This packing function is run on idle threads.  It checks to see if
5986  * the busiest CPU in this domain (core in the P7 case) has a higher
5987  * CPU number than the packing function is being run on.  Here we are
5988  * assuming lower CPU number will be equivalent to lower a SMT thread
5989  * number.
5990  *
5991  * Return: 1 when packing is required and a task should be moved to
5992  * this CPU.  The amount of the imbalance is returned in *imbalance.
5993  *
5994  * @env: The load balancing environment.
5995  * @sds: Statistics of the sched_domain which is to be packed
5996  */
5997 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5998 {
5999 	int busiest_cpu;
6000 
6001 	if (!(env->sd->flags & SD_ASYM_PACKING))
6002 		return 0;
6003 
6004 	if (!sds->busiest)
6005 		return 0;
6006 
6007 	busiest_cpu = group_first_cpu(sds->busiest);
6008 	if (env->dst_cpu > busiest_cpu)
6009 		return 0;
6010 
6011 	env->imbalance = DIV_ROUND_CLOSEST(
6012 		sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
6013 		SCHED_POWER_SCALE);
6014 
6015 	return 1;
6016 }
6017 
6018 /**
6019  * fix_small_imbalance - Calculate the minor imbalance that exists
6020  *			amongst the groups of a sched_domain, during
6021  *			load balancing.
6022  * @env: The load balancing environment.
6023  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6024  */
6025 static inline
6026 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6027 {
6028 	unsigned long tmp, pwr_now = 0, pwr_move = 0;
6029 	unsigned int imbn = 2;
6030 	unsigned long scaled_busy_load_per_task;
6031 	struct sg_lb_stats *local, *busiest;
6032 
6033 	local = &sds->local_stat;
6034 	busiest = &sds->busiest_stat;
6035 
6036 	if (!local->sum_nr_running)
6037 		local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6038 	else if (busiest->load_per_task > local->load_per_task)
6039 		imbn = 1;
6040 
6041 	scaled_busy_load_per_task =
6042 		(busiest->load_per_task * SCHED_POWER_SCALE) /
6043 		busiest->group_power;
6044 
6045 	if (busiest->avg_load + scaled_busy_load_per_task >=
6046 	    local->avg_load + (scaled_busy_load_per_task * imbn)) {
6047 		env->imbalance = busiest->load_per_task;
6048 		return;
6049 	}
6050 
6051 	/*
6052 	 * OK, we don't have enough imbalance to justify moving tasks,
6053 	 * however we may be able to increase total CPU power used by
6054 	 * moving them.
6055 	 */
6056 
6057 	pwr_now += busiest->group_power *
6058 			min(busiest->load_per_task, busiest->avg_load);
6059 	pwr_now += local->group_power *
6060 			min(local->load_per_task, local->avg_load);
6061 	pwr_now /= SCHED_POWER_SCALE;
6062 
6063 	/* Amount of load we'd subtract */
6064 	if (busiest->avg_load > scaled_busy_load_per_task) {
6065 		pwr_move += busiest->group_power *
6066 			    min(busiest->load_per_task,
6067 				busiest->avg_load - scaled_busy_load_per_task);
6068 	}
6069 
6070 	/* Amount of load we'd add */
6071 	if (busiest->avg_load * busiest->group_power <
6072 	    busiest->load_per_task * SCHED_POWER_SCALE) {
6073 		tmp = (busiest->avg_load * busiest->group_power) /
6074 		      local->group_power;
6075 	} else {
6076 		tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
6077 		      local->group_power;
6078 	}
6079 	pwr_move += local->group_power *
6080 		    min(local->load_per_task, local->avg_load + tmp);
6081 	pwr_move /= SCHED_POWER_SCALE;
6082 
6083 	/* Move if we gain throughput */
6084 	if (pwr_move > pwr_now)
6085 		env->imbalance = busiest->load_per_task;
6086 }
6087 
6088 /**
6089  * calculate_imbalance - Calculate the amount of imbalance present within the
6090  *			 groups of a given sched_domain during load balance.
6091  * @env: load balance environment
6092  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6093  */
6094 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6095 {
6096 	unsigned long max_pull, load_above_capacity = ~0UL;
6097 	struct sg_lb_stats *local, *busiest;
6098 
6099 	local = &sds->local_stat;
6100 	busiest = &sds->busiest_stat;
6101 
6102 	if (busiest->group_imb) {
6103 		/*
6104 		 * In the group_imb case we cannot rely on group-wide averages
6105 		 * to ensure cpu-load equilibrium, look at wider averages. XXX
6106 		 */
6107 		busiest->load_per_task =
6108 			min(busiest->load_per_task, sds->avg_load);
6109 	}
6110 
6111 	/*
6112 	 * In the presence of smp nice balancing, certain scenarios can have
6113 	 * max load less than avg load(as we skip the groups at or below
6114 	 * its cpu_power, while calculating max_load..)
6115 	 */
6116 	if (busiest->avg_load <= sds->avg_load ||
6117 	    local->avg_load >= sds->avg_load) {
6118 		env->imbalance = 0;
6119 		return fix_small_imbalance(env, sds);
6120 	}
6121 
6122 	if (!busiest->group_imb) {
6123 		/*
6124 		 * Don't want to pull so many tasks that a group would go idle.
6125 		 * Except of course for the group_imb case, since then we might
6126 		 * have to drop below capacity to reach cpu-load equilibrium.
6127 		 */
6128 		load_above_capacity =
6129 			(busiest->sum_nr_running - busiest->group_capacity);
6130 
6131 		load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
6132 		load_above_capacity /= busiest->group_power;
6133 	}
6134 
6135 	/*
6136 	 * We're trying to get all the cpus to the average_load, so we don't
6137 	 * want to push ourselves above the average load, nor do we wish to
6138 	 * reduce the max loaded cpu below the average load. At the same time,
6139 	 * we also don't want to reduce the group load below the group capacity
6140 	 * (so that we can implement power-savings policies etc). Thus we look
6141 	 * for the minimum possible imbalance.
6142 	 */
6143 	max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6144 
6145 	/* How much load to actually move to equalise the imbalance */
6146 	env->imbalance = min(
6147 		max_pull * busiest->group_power,
6148 		(sds->avg_load - local->avg_load) * local->group_power
6149 	) / SCHED_POWER_SCALE;
6150 
6151 	/*
6152 	 * if *imbalance is less than the average load per runnable task
6153 	 * there is no guarantee that any tasks will be moved so we'll have
6154 	 * a think about bumping its value to force at least one task to be
6155 	 * moved
6156 	 */
6157 	if (env->imbalance < busiest->load_per_task)
6158 		return fix_small_imbalance(env, sds);
6159 }
6160 
6161 /******* find_busiest_group() helpers end here *********************/
6162 
6163 /**
6164  * find_busiest_group - Returns the busiest group within the sched_domain
6165  * if there is an imbalance. If there isn't an imbalance, and
6166  * the user has opted for power-savings, it returns a group whose
6167  * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6168  * such a group exists.
6169  *
6170  * Also calculates the amount of weighted load which should be moved
6171  * to restore balance.
6172  *
6173  * @env: The load balancing environment.
6174  *
6175  * Return:	- The busiest group if imbalance exists.
6176  *		- If no imbalance and user has opted for power-savings balance,
6177  *		   return the least loaded group whose CPUs can be
6178  *		   put to idle by rebalancing its tasks onto our group.
6179  */
6180 static struct sched_group *find_busiest_group(struct lb_env *env)
6181 {
6182 	struct sg_lb_stats *local, *busiest;
6183 	struct sd_lb_stats sds;
6184 
6185 	init_sd_lb_stats(&sds);
6186 
6187 	/*
6188 	 * Compute the various statistics relavent for load balancing at
6189 	 * this level.
6190 	 */
6191 	update_sd_lb_stats(env, &sds);
6192 	local = &sds.local_stat;
6193 	busiest = &sds.busiest_stat;
6194 
6195 	if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6196 	    check_asym_packing(env, &sds))
6197 		return sds.busiest;
6198 
6199 	/* There is no busy sibling group to pull tasks from */
6200 	if (!sds.busiest || busiest->sum_nr_running == 0)
6201 		goto out_balanced;
6202 
6203 	sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
6204 
6205 	/*
6206 	 * If the busiest group is imbalanced the below checks don't
6207 	 * work because they assume all things are equal, which typically
6208 	 * isn't true due to cpus_allowed constraints and the like.
6209 	 */
6210 	if (busiest->group_imb)
6211 		goto force_balance;
6212 
6213 	/* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6214 	if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
6215 	    !busiest->group_has_capacity)
6216 		goto force_balance;
6217 
6218 	/*
6219 	 * If the local group is more busy than the selected busiest group
6220 	 * don't try and pull any tasks.
6221 	 */
6222 	if (local->avg_load >= busiest->avg_load)
6223 		goto out_balanced;
6224 
6225 	/*
6226 	 * Don't pull any tasks if this group is already above the domain
6227 	 * average load.
6228 	 */
6229 	if (local->avg_load >= sds.avg_load)
6230 		goto out_balanced;
6231 
6232 	if (env->idle == CPU_IDLE) {
6233 		/*
6234 		 * This cpu is idle. If the busiest group load doesn't
6235 		 * have more tasks than the number of available cpu's and
6236 		 * there is no imbalance between this and busiest group
6237 		 * wrt to idle cpu's, it is balanced.
6238 		 */
6239 		if ((local->idle_cpus < busiest->idle_cpus) &&
6240 		    busiest->sum_nr_running <= busiest->group_weight)
6241 			goto out_balanced;
6242 	} else {
6243 		/*
6244 		 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6245 		 * imbalance_pct to be conservative.
6246 		 */
6247 		if (100 * busiest->avg_load <=
6248 				env->sd->imbalance_pct * local->avg_load)
6249 			goto out_balanced;
6250 	}
6251 
6252 force_balance:
6253 	/* Looks like there is an imbalance. Compute it */
6254 	calculate_imbalance(env, &sds);
6255 	return sds.busiest;
6256 
6257 out_balanced:
6258 	env->imbalance = 0;
6259 	return NULL;
6260 }
6261 
6262 /*
6263  * find_busiest_queue - find the busiest runqueue among the cpus in group.
6264  */
6265 static struct rq *find_busiest_queue(struct lb_env *env,
6266 				     struct sched_group *group)
6267 {
6268 	struct rq *busiest = NULL, *rq;
6269 	unsigned long busiest_load = 0, busiest_power = 1;
6270 	int i;
6271 
6272 	for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6273 		unsigned long power, capacity, wl;
6274 		enum fbq_type rt;
6275 
6276 		rq = cpu_rq(i);
6277 		rt = fbq_classify_rq(rq);
6278 
6279 		/*
6280 		 * We classify groups/runqueues into three groups:
6281 		 *  - regular: there are !numa tasks
6282 		 *  - remote:  there are numa tasks that run on the 'wrong' node
6283 		 *  - all:     there is no distinction
6284 		 *
6285 		 * In order to avoid migrating ideally placed numa tasks,
6286 		 * ignore those when there's better options.
6287 		 *
6288 		 * If we ignore the actual busiest queue to migrate another
6289 		 * task, the next balance pass can still reduce the busiest
6290 		 * queue by moving tasks around inside the node.
6291 		 *
6292 		 * If we cannot move enough load due to this classification
6293 		 * the next pass will adjust the group classification and
6294 		 * allow migration of more tasks.
6295 		 *
6296 		 * Both cases only affect the total convergence complexity.
6297 		 */
6298 		if (rt > env->fbq_type)
6299 			continue;
6300 
6301 		power = power_of(i);
6302 		capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6303 		if (!capacity)
6304 			capacity = fix_small_capacity(env->sd, group);
6305 
6306 		wl = weighted_cpuload(i);
6307 
6308 		/*
6309 		 * When comparing with imbalance, use weighted_cpuload()
6310 		 * which is not scaled with the cpu power.
6311 		 */
6312 		if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6313 			continue;
6314 
6315 		/*
6316 		 * For the load comparisons with the other cpu's, consider
6317 		 * the weighted_cpuload() scaled with the cpu power, so that
6318 		 * the load can be moved away from the cpu that is potentially
6319 		 * running at a lower capacity.
6320 		 *
6321 		 * Thus we're looking for max(wl_i / power_i), crosswise
6322 		 * multiplication to rid ourselves of the division works out
6323 		 * to: wl_i * power_j > wl_j * power_i;  where j is our
6324 		 * previous maximum.
6325 		 */
6326 		if (wl * busiest_power > busiest_load * power) {
6327 			busiest_load = wl;
6328 			busiest_power = power;
6329 			busiest = rq;
6330 		}
6331 	}
6332 
6333 	return busiest;
6334 }
6335 
6336 /*
6337  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6338  * so long as it is large enough.
6339  */
6340 #define MAX_PINNED_INTERVAL	512
6341 
6342 /* Working cpumask for load_balance and load_balance_newidle. */
6343 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6344 
6345 static int need_active_balance(struct lb_env *env)
6346 {
6347 	struct sched_domain *sd = env->sd;
6348 
6349 	if (env->idle == CPU_NEWLY_IDLE) {
6350 
6351 		/*
6352 		 * ASYM_PACKING needs to force migrate tasks from busy but
6353 		 * higher numbered CPUs in order to pack all tasks in the
6354 		 * lowest numbered CPUs.
6355 		 */
6356 		if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6357 			return 1;
6358 	}
6359 
6360 	return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6361 }
6362 
6363 static int active_load_balance_cpu_stop(void *data);
6364 
6365 static int should_we_balance(struct lb_env *env)
6366 {
6367 	struct sched_group *sg = env->sd->groups;
6368 	struct cpumask *sg_cpus, *sg_mask;
6369 	int cpu, balance_cpu = -1;
6370 
6371 	/*
6372 	 * In the newly idle case, we will allow all the cpu's
6373 	 * to do the newly idle load balance.
6374 	 */
6375 	if (env->idle == CPU_NEWLY_IDLE)
6376 		return 1;
6377 
6378 	sg_cpus = sched_group_cpus(sg);
6379 	sg_mask = sched_group_mask(sg);
6380 	/* Try to find first idle cpu */
6381 	for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6382 		if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6383 			continue;
6384 
6385 		balance_cpu = cpu;
6386 		break;
6387 	}
6388 
6389 	if (balance_cpu == -1)
6390 		balance_cpu = group_balance_cpu(sg);
6391 
6392 	/*
6393 	 * First idle cpu or the first cpu(busiest) in this sched group
6394 	 * is eligible for doing load balancing at this and above domains.
6395 	 */
6396 	return balance_cpu == env->dst_cpu;
6397 }
6398 
6399 /*
6400  * Check this_cpu to ensure it is balanced within domain. Attempt to move
6401  * tasks if there is an imbalance.
6402  */
6403 static int load_balance(int this_cpu, struct rq *this_rq,
6404 			struct sched_domain *sd, enum cpu_idle_type idle,
6405 			int *continue_balancing)
6406 {
6407 	int ld_moved, cur_ld_moved, active_balance = 0;
6408 	struct sched_domain *sd_parent = sd->parent;
6409 	struct sched_group *group;
6410 	struct rq *busiest;
6411 	unsigned long flags;
6412 	struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6413 
6414 	struct lb_env env = {
6415 		.sd		= sd,
6416 		.dst_cpu	= this_cpu,
6417 		.dst_rq		= this_rq,
6418 		.dst_grpmask    = sched_group_cpus(sd->groups),
6419 		.idle		= idle,
6420 		.loop_break	= sched_nr_migrate_break,
6421 		.cpus		= cpus,
6422 		.fbq_type	= all,
6423 	};
6424 
6425 	/*
6426 	 * For NEWLY_IDLE load_balancing, we don't need to consider
6427 	 * other cpus in our group
6428 	 */
6429 	if (idle == CPU_NEWLY_IDLE)
6430 		env.dst_grpmask = NULL;
6431 
6432 	cpumask_copy(cpus, cpu_active_mask);
6433 
6434 	schedstat_inc(sd, lb_count[idle]);
6435 
6436 redo:
6437 	if (!should_we_balance(&env)) {
6438 		*continue_balancing = 0;
6439 		goto out_balanced;
6440 	}
6441 
6442 	group = find_busiest_group(&env);
6443 	if (!group) {
6444 		schedstat_inc(sd, lb_nobusyg[idle]);
6445 		goto out_balanced;
6446 	}
6447 
6448 	busiest = find_busiest_queue(&env, group);
6449 	if (!busiest) {
6450 		schedstat_inc(sd, lb_nobusyq[idle]);
6451 		goto out_balanced;
6452 	}
6453 
6454 	BUG_ON(busiest == env.dst_rq);
6455 
6456 	schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6457 
6458 	ld_moved = 0;
6459 	if (busiest->nr_running > 1) {
6460 		/*
6461 		 * Attempt to move tasks. If find_busiest_group has found
6462 		 * an imbalance but busiest->nr_running <= 1, the group is
6463 		 * still unbalanced. ld_moved simply stays zero, so it is
6464 		 * correctly treated as an imbalance.
6465 		 */
6466 		env.flags |= LBF_ALL_PINNED;
6467 		env.src_cpu   = busiest->cpu;
6468 		env.src_rq    = busiest;
6469 		env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
6470 
6471 more_balance:
6472 		local_irq_save(flags);
6473 		double_rq_lock(env.dst_rq, busiest);
6474 
6475 		/*
6476 		 * cur_ld_moved - load moved in current iteration
6477 		 * ld_moved     - cumulative load moved across iterations
6478 		 */
6479 		cur_ld_moved = move_tasks(&env);
6480 		ld_moved += cur_ld_moved;
6481 		double_rq_unlock(env.dst_rq, busiest);
6482 		local_irq_restore(flags);
6483 
6484 		/*
6485 		 * some other cpu did the load balance for us.
6486 		 */
6487 		if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6488 			resched_cpu(env.dst_cpu);
6489 
6490 		if (env.flags & LBF_NEED_BREAK) {
6491 			env.flags &= ~LBF_NEED_BREAK;
6492 			goto more_balance;
6493 		}
6494 
6495 		/*
6496 		 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6497 		 * us and move them to an alternate dst_cpu in our sched_group
6498 		 * where they can run. The upper limit on how many times we
6499 		 * iterate on same src_cpu is dependent on number of cpus in our
6500 		 * sched_group.
6501 		 *
6502 		 * This changes load balance semantics a bit on who can move
6503 		 * load to a given_cpu. In addition to the given_cpu itself
6504 		 * (or a ilb_cpu acting on its behalf where given_cpu is
6505 		 * nohz-idle), we now have balance_cpu in a position to move
6506 		 * load to given_cpu. In rare situations, this may cause
6507 		 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6508 		 * _independently_ and at _same_ time to move some load to
6509 		 * given_cpu) causing exceess load to be moved to given_cpu.
6510 		 * This however should not happen so much in practice and
6511 		 * moreover subsequent load balance cycles should correct the
6512 		 * excess load moved.
6513 		 */
6514 		if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6515 
6516 			/* Prevent to re-select dst_cpu via env's cpus */
6517 			cpumask_clear_cpu(env.dst_cpu, env.cpus);
6518 
6519 			env.dst_rq	 = cpu_rq(env.new_dst_cpu);
6520 			env.dst_cpu	 = env.new_dst_cpu;
6521 			env.flags	&= ~LBF_DST_PINNED;
6522 			env.loop	 = 0;
6523 			env.loop_break	 = sched_nr_migrate_break;
6524 
6525 			/*
6526 			 * Go back to "more_balance" rather than "redo" since we
6527 			 * need to continue with same src_cpu.
6528 			 */
6529 			goto more_balance;
6530 		}
6531 
6532 		/*
6533 		 * We failed to reach balance because of affinity.
6534 		 */
6535 		if (sd_parent) {
6536 			int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6537 
6538 			if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6539 				*group_imbalance = 1;
6540 			} else if (*group_imbalance)
6541 				*group_imbalance = 0;
6542 		}
6543 
6544 		/* All tasks on this runqueue were pinned by CPU affinity */
6545 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
6546 			cpumask_clear_cpu(cpu_of(busiest), cpus);
6547 			if (!cpumask_empty(cpus)) {
6548 				env.loop = 0;
6549 				env.loop_break = sched_nr_migrate_break;
6550 				goto redo;
6551 			}
6552 			goto out_balanced;
6553 		}
6554 	}
6555 
6556 	if (!ld_moved) {
6557 		schedstat_inc(sd, lb_failed[idle]);
6558 		/*
6559 		 * Increment the failure counter only on periodic balance.
6560 		 * We do not want newidle balance, which can be very
6561 		 * frequent, pollute the failure counter causing
6562 		 * excessive cache_hot migrations and active balances.
6563 		 */
6564 		if (idle != CPU_NEWLY_IDLE)
6565 			sd->nr_balance_failed++;
6566 
6567 		if (need_active_balance(&env)) {
6568 			raw_spin_lock_irqsave(&busiest->lock, flags);
6569 
6570 			/* don't kick the active_load_balance_cpu_stop,
6571 			 * if the curr task on busiest cpu can't be
6572 			 * moved to this_cpu
6573 			 */
6574 			if (!cpumask_test_cpu(this_cpu,
6575 					tsk_cpus_allowed(busiest->curr))) {
6576 				raw_spin_unlock_irqrestore(&busiest->lock,
6577 							    flags);
6578 				env.flags |= LBF_ALL_PINNED;
6579 				goto out_one_pinned;
6580 			}
6581 
6582 			/*
6583 			 * ->active_balance synchronizes accesses to
6584 			 * ->active_balance_work.  Once set, it's cleared
6585 			 * only after active load balance is finished.
6586 			 */
6587 			if (!busiest->active_balance) {
6588 				busiest->active_balance = 1;
6589 				busiest->push_cpu = this_cpu;
6590 				active_balance = 1;
6591 			}
6592 			raw_spin_unlock_irqrestore(&busiest->lock, flags);
6593 
6594 			if (active_balance) {
6595 				stop_one_cpu_nowait(cpu_of(busiest),
6596 					active_load_balance_cpu_stop, busiest,
6597 					&busiest->active_balance_work);
6598 			}
6599 
6600 			/*
6601 			 * We've kicked active balancing, reset the failure
6602 			 * counter.
6603 			 */
6604 			sd->nr_balance_failed = sd->cache_nice_tries+1;
6605 		}
6606 	} else
6607 		sd->nr_balance_failed = 0;
6608 
6609 	if (likely(!active_balance)) {
6610 		/* We were unbalanced, so reset the balancing interval */
6611 		sd->balance_interval = sd->min_interval;
6612 	} else {
6613 		/*
6614 		 * If we've begun active balancing, start to back off. This
6615 		 * case may not be covered by the all_pinned logic if there
6616 		 * is only 1 task on the busy runqueue (because we don't call
6617 		 * move_tasks).
6618 		 */
6619 		if (sd->balance_interval < sd->max_interval)
6620 			sd->balance_interval *= 2;
6621 	}
6622 
6623 	goto out;
6624 
6625 out_balanced:
6626 	schedstat_inc(sd, lb_balanced[idle]);
6627 
6628 	sd->nr_balance_failed = 0;
6629 
6630 out_one_pinned:
6631 	/* tune up the balancing interval */
6632 	if (((env.flags & LBF_ALL_PINNED) &&
6633 			sd->balance_interval < MAX_PINNED_INTERVAL) ||
6634 			(sd->balance_interval < sd->max_interval))
6635 		sd->balance_interval *= 2;
6636 
6637 	ld_moved = 0;
6638 out:
6639 	return ld_moved;
6640 }
6641 
6642 /*
6643  * idle_balance is called by schedule() if this_cpu is about to become
6644  * idle. Attempts to pull tasks from other CPUs.
6645  */
6646 static int idle_balance(struct rq *this_rq)
6647 {
6648 	struct sched_domain *sd;
6649 	int pulled_task = 0;
6650 	unsigned long next_balance = jiffies + HZ;
6651 	u64 curr_cost = 0;
6652 	int this_cpu = this_rq->cpu;
6653 
6654 	idle_enter_fair(this_rq);
6655 	/*
6656 	 * We must set idle_stamp _before_ calling idle_balance(), such that we
6657 	 * measure the duration of idle_balance() as idle time.
6658 	 */
6659 	this_rq->idle_stamp = rq_clock(this_rq);
6660 
6661 	if (this_rq->avg_idle < sysctl_sched_migration_cost)
6662 		goto out;
6663 
6664 	/*
6665 	 * Drop the rq->lock, but keep IRQ/preempt disabled.
6666 	 */
6667 	raw_spin_unlock(&this_rq->lock);
6668 
6669 	update_blocked_averages(this_cpu);
6670 	rcu_read_lock();
6671 	for_each_domain(this_cpu, sd) {
6672 		unsigned long interval;
6673 		int continue_balancing = 1;
6674 		u64 t0, domain_cost;
6675 
6676 		if (!(sd->flags & SD_LOAD_BALANCE))
6677 			continue;
6678 
6679 		if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6680 			break;
6681 
6682 		if (sd->flags & SD_BALANCE_NEWIDLE) {
6683 			t0 = sched_clock_cpu(this_cpu);
6684 
6685 			/* If we've pulled tasks over stop searching: */
6686 			pulled_task = load_balance(this_cpu, this_rq,
6687 						   sd, CPU_NEWLY_IDLE,
6688 						   &continue_balancing);
6689 
6690 			domain_cost = sched_clock_cpu(this_cpu) - t0;
6691 			if (domain_cost > sd->max_newidle_lb_cost)
6692 				sd->max_newidle_lb_cost = domain_cost;
6693 
6694 			curr_cost += domain_cost;
6695 		}
6696 
6697 		interval = msecs_to_jiffies(sd->balance_interval);
6698 		if (time_after(next_balance, sd->last_balance + interval))
6699 			next_balance = sd->last_balance + interval;
6700 		if (pulled_task)
6701 			break;
6702 	}
6703 	rcu_read_unlock();
6704 
6705 	raw_spin_lock(&this_rq->lock);
6706 
6707 	/*
6708 	 * While browsing the domains, we released the rq lock.
6709 	 * A task could have be enqueued in the meantime
6710 	 */
6711 	if (this_rq->cfs.h_nr_running && !pulled_task) {
6712 		pulled_task = 1;
6713 		goto out;
6714 	}
6715 
6716 	if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6717 		/*
6718 		 * We are going idle. next_balance may be set based on
6719 		 * a busy processor. So reset next_balance.
6720 		 */
6721 		this_rq->next_balance = next_balance;
6722 	}
6723 
6724 	if (curr_cost > this_rq->max_idle_balance_cost)
6725 		this_rq->max_idle_balance_cost = curr_cost;
6726 
6727 out:
6728 	/* Is there a task of a high priority class? */
6729 	if (this_rq->nr_running != this_rq->cfs.h_nr_running &&
6730 	    (this_rq->dl.dl_nr_running ||
6731 	     (this_rq->rt.rt_nr_running && !rt_rq_throttled(&this_rq->rt))))
6732 		pulled_task = -1;
6733 
6734 	if (pulled_task) {
6735 		idle_exit_fair(this_rq);
6736 		this_rq->idle_stamp = 0;
6737 	}
6738 
6739 	return pulled_task;
6740 }
6741 
6742 /*
6743  * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6744  * running tasks off the busiest CPU onto idle CPUs. It requires at
6745  * least 1 task to be running on each physical CPU where possible, and
6746  * avoids physical / logical imbalances.
6747  */
6748 static int active_load_balance_cpu_stop(void *data)
6749 {
6750 	struct rq *busiest_rq = data;
6751 	int busiest_cpu = cpu_of(busiest_rq);
6752 	int target_cpu = busiest_rq->push_cpu;
6753 	struct rq *target_rq = cpu_rq(target_cpu);
6754 	struct sched_domain *sd;
6755 
6756 	raw_spin_lock_irq(&busiest_rq->lock);
6757 
6758 	/* make sure the requested cpu hasn't gone down in the meantime */
6759 	if (unlikely(busiest_cpu != smp_processor_id() ||
6760 		     !busiest_rq->active_balance))
6761 		goto out_unlock;
6762 
6763 	/* Is there any task to move? */
6764 	if (busiest_rq->nr_running <= 1)
6765 		goto out_unlock;
6766 
6767 	/*
6768 	 * This condition is "impossible", if it occurs
6769 	 * we need to fix it. Originally reported by
6770 	 * Bjorn Helgaas on a 128-cpu setup.
6771 	 */
6772 	BUG_ON(busiest_rq == target_rq);
6773 
6774 	/* move a task from busiest_rq to target_rq */
6775 	double_lock_balance(busiest_rq, target_rq);
6776 
6777 	/* Search for an sd spanning us and the target CPU. */
6778 	rcu_read_lock();
6779 	for_each_domain(target_cpu, sd) {
6780 		if ((sd->flags & SD_LOAD_BALANCE) &&
6781 		    cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6782 				break;
6783 	}
6784 
6785 	if (likely(sd)) {
6786 		struct lb_env env = {
6787 			.sd		= sd,
6788 			.dst_cpu	= target_cpu,
6789 			.dst_rq		= target_rq,
6790 			.src_cpu	= busiest_rq->cpu,
6791 			.src_rq		= busiest_rq,
6792 			.idle		= CPU_IDLE,
6793 		};
6794 
6795 		schedstat_inc(sd, alb_count);
6796 
6797 		if (move_one_task(&env))
6798 			schedstat_inc(sd, alb_pushed);
6799 		else
6800 			schedstat_inc(sd, alb_failed);
6801 	}
6802 	rcu_read_unlock();
6803 	double_unlock_balance(busiest_rq, target_rq);
6804 out_unlock:
6805 	busiest_rq->active_balance = 0;
6806 	raw_spin_unlock_irq(&busiest_rq->lock);
6807 	return 0;
6808 }
6809 
6810 static inline int on_null_domain(struct rq *rq)
6811 {
6812 	return unlikely(!rcu_dereference_sched(rq->sd));
6813 }
6814 
6815 #ifdef CONFIG_NO_HZ_COMMON
6816 /*
6817  * idle load balancing details
6818  * - When one of the busy CPUs notice that there may be an idle rebalancing
6819  *   needed, they will kick the idle load balancer, which then does idle
6820  *   load balancing for all the idle CPUs.
6821  */
6822 static struct {
6823 	cpumask_var_t idle_cpus_mask;
6824 	atomic_t nr_cpus;
6825 	unsigned long next_balance;     /* in jiffy units */
6826 } nohz ____cacheline_aligned;
6827 
6828 static inline int find_new_ilb(void)
6829 {
6830 	int ilb = cpumask_first(nohz.idle_cpus_mask);
6831 
6832 	if (ilb < nr_cpu_ids && idle_cpu(ilb))
6833 		return ilb;
6834 
6835 	return nr_cpu_ids;
6836 }
6837 
6838 /*
6839  * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6840  * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6841  * CPU (if there is one).
6842  */
6843 static void nohz_balancer_kick(void)
6844 {
6845 	int ilb_cpu;
6846 
6847 	nohz.next_balance++;
6848 
6849 	ilb_cpu = find_new_ilb();
6850 
6851 	if (ilb_cpu >= nr_cpu_ids)
6852 		return;
6853 
6854 	if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6855 		return;
6856 	/*
6857 	 * Use smp_send_reschedule() instead of resched_cpu().
6858 	 * This way we generate a sched IPI on the target cpu which
6859 	 * is idle. And the softirq performing nohz idle load balance
6860 	 * will be run before returning from the IPI.
6861 	 */
6862 	smp_send_reschedule(ilb_cpu);
6863 	return;
6864 }
6865 
6866 static inline void nohz_balance_exit_idle(int cpu)
6867 {
6868 	if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6869 		/*
6870 		 * Completely isolated CPUs don't ever set, so we must test.
6871 		 */
6872 		if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
6873 			cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6874 			atomic_dec(&nohz.nr_cpus);
6875 		}
6876 		clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6877 	}
6878 }
6879 
6880 static inline void set_cpu_sd_state_busy(void)
6881 {
6882 	struct sched_domain *sd;
6883 	int cpu = smp_processor_id();
6884 
6885 	rcu_read_lock();
6886 	sd = rcu_dereference(per_cpu(sd_busy, cpu));
6887 
6888 	if (!sd || !sd->nohz_idle)
6889 		goto unlock;
6890 	sd->nohz_idle = 0;
6891 
6892 	atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6893 unlock:
6894 	rcu_read_unlock();
6895 }
6896 
6897 void set_cpu_sd_state_idle(void)
6898 {
6899 	struct sched_domain *sd;
6900 	int cpu = smp_processor_id();
6901 
6902 	rcu_read_lock();
6903 	sd = rcu_dereference(per_cpu(sd_busy, cpu));
6904 
6905 	if (!sd || sd->nohz_idle)
6906 		goto unlock;
6907 	sd->nohz_idle = 1;
6908 
6909 	atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6910 unlock:
6911 	rcu_read_unlock();
6912 }
6913 
6914 /*
6915  * This routine will record that the cpu is going idle with tick stopped.
6916  * This info will be used in performing idle load balancing in the future.
6917  */
6918 void nohz_balance_enter_idle(int cpu)
6919 {
6920 	/*
6921 	 * If this cpu is going down, then nothing needs to be done.
6922 	 */
6923 	if (!cpu_active(cpu))
6924 		return;
6925 
6926 	if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6927 		return;
6928 
6929 	/*
6930 	 * If we're a completely isolated CPU, we don't play.
6931 	 */
6932 	if (on_null_domain(cpu_rq(cpu)))
6933 		return;
6934 
6935 	cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6936 	atomic_inc(&nohz.nr_cpus);
6937 	set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6938 }
6939 
6940 static int sched_ilb_notifier(struct notifier_block *nfb,
6941 					unsigned long action, void *hcpu)
6942 {
6943 	switch (action & ~CPU_TASKS_FROZEN) {
6944 	case CPU_DYING:
6945 		nohz_balance_exit_idle(smp_processor_id());
6946 		return NOTIFY_OK;
6947 	default:
6948 		return NOTIFY_DONE;
6949 	}
6950 }
6951 #endif
6952 
6953 static DEFINE_SPINLOCK(balancing);
6954 
6955 /*
6956  * Scale the max load_balance interval with the number of CPUs in the system.
6957  * This trades load-balance latency on larger machines for less cross talk.
6958  */
6959 void update_max_interval(void)
6960 {
6961 	max_load_balance_interval = HZ*num_online_cpus()/10;
6962 }
6963 
6964 /*
6965  * It checks each scheduling domain to see if it is due to be balanced,
6966  * and initiates a balancing operation if so.
6967  *
6968  * Balancing parameters are set up in init_sched_domains.
6969  */
6970 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
6971 {
6972 	int continue_balancing = 1;
6973 	int cpu = rq->cpu;
6974 	unsigned long interval;
6975 	struct sched_domain *sd;
6976 	/* Earliest time when we have to do rebalance again */
6977 	unsigned long next_balance = jiffies + 60*HZ;
6978 	int update_next_balance = 0;
6979 	int need_serialize, need_decay = 0;
6980 	u64 max_cost = 0;
6981 
6982 	update_blocked_averages(cpu);
6983 
6984 	rcu_read_lock();
6985 	for_each_domain(cpu, sd) {
6986 		/*
6987 		 * Decay the newidle max times here because this is a regular
6988 		 * visit to all the domains. Decay ~1% per second.
6989 		 */
6990 		if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6991 			sd->max_newidle_lb_cost =
6992 				(sd->max_newidle_lb_cost * 253) / 256;
6993 			sd->next_decay_max_lb_cost = jiffies + HZ;
6994 			need_decay = 1;
6995 		}
6996 		max_cost += sd->max_newidle_lb_cost;
6997 
6998 		if (!(sd->flags & SD_LOAD_BALANCE))
6999 			continue;
7000 
7001 		/*
7002 		 * Stop the load balance at this level. There is another
7003 		 * CPU in our sched group which is doing load balancing more
7004 		 * actively.
7005 		 */
7006 		if (!continue_balancing) {
7007 			if (need_decay)
7008 				continue;
7009 			break;
7010 		}
7011 
7012 		interval = sd->balance_interval;
7013 		if (idle != CPU_IDLE)
7014 			interval *= sd->busy_factor;
7015 
7016 		/* scale ms to jiffies */
7017 		interval = msecs_to_jiffies(interval);
7018 		interval = clamp(interval, 1UL, max_load_balance_interval);
7019 
7020 		need_serialize = sd->flags & SD_SERIALIZE;
7021 
7022 		if (need_serialize) {
7023 			if (!spin_trylock(&balancing))
7024 				goto out;
7025 		}
7026 
7027 		if (time_after_eq(jiffies, sd->last_balance + interval)) {
7028 			if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7029 				/*
7030 				 * The LBF_DST_PINNED logic could have changed
7031 				 * env->dst_cpu, so we can't know our idle
7032 				 * state even if we migrated tasks. Update it.
7033 				 */
7034 				idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7035 			}
7036 			sd->last_balance = jiffies;
7037 		}
7038 		if (need_serialize)
7039 			spin_unlock(&balancing);
7040 out:
7041 		if (time_after(next_balance, sd->last_balance + interval)) {
7042 			next_balance = sd->last_balance + interval;
7043 			update_next_balance = 1;
7044 		}
7045 	}
7046 	if (need_decay) {
7047 		/*
7048 		 * Ensure the rq-wide value also decays but keep it at a
7049 		 * reasonable floor to avoid funnies with rq->avg_idle.
7050 		 */
7051 		rq->max_idle_balance_cost =
7052 			max((u64)sysctl_sched_migration_cost, max_cost);
7053 	}
7054 	rcu_read_unlock();
7055 
7056 	/*
7057 	 * next_balance will be updated only when there is a need.
7058 	 * When the cpu is attached to null domain for ex, it will not be
7059 	 * updated.
7060 	 */
7061 	if (likely(update_next_balance))
7062 		rq->next_balance = next_balance;
7063 }
7064 
7065 #ifdef CONFIG_NO_HZ_COMMON
7066 /*
7067  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7068  * rebalancing for all the cpus for whom scheduler ticks are stopped.
7069  */
7070 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7071 {
7072 	int this_cpu = this_rq->cpu;
7073 	struct rq *rq;
7074 	int balance_cpu;
7075 
7076 	if (idle != CPU_IDLE ||
7077 	    !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7078 		goto end;
7079 
7080 	for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7081 		if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7082 			continue;
7083 
7084 		/*
7085 		 * If this cpu gets work to do, stop the load balancing
7086 		 * work being done for other cpus. Next load
7087 		 * balancing owner will pick it up.
7088 		 */
7089 		if (need_resched())
7090 			break;
7091 
7092 		rq = cpu_rq(balance_cpu);
7093 
7094 		raw_spin_lock_irq(&rq->lock);
7095 		update_rq_clock(rq);
7096 		update_idle_cpu_load(rq);
7097 		raw_spin_unlock_irq(&rq->lock);
7098 
7099 		rebalance_domains(rq, CPU_IDLE);
7100 
7101 		if (time_after(this_rq->next_balance, rq->next_balance))
7102 			this_rq->next_balance = rq->next_balance;
7103 	}
7104 	nohz.next_balance = this_rq->next_balance;
7105 end:
7106 	clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7107 }
7108 
7109 /*
7110  * Current heuristic for kicking the idle load balancer in the presence
7111  * of an idle cpu is the system.
7112  *   - This rq has more than one task.
7113  *   - At any scheduler domain level, this cpu's scheduler group has multiple
7114  *     busy cpu's exceeding the group's power.
7115  *   - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7116  *     domain span are idle.
7117  */
7118 static inline int nohz_kick_needed(struct rq *rq)
7119 {
7120 	unsigned long now = jiffies;
7121 	struct sched_domain *sd;
7122 	struct sched_group_power *sgp;
7123 	int nr_busy, cpu = rq->cpu;
7124 
7125 	if (unlikely(rq->idle_balance))
7126 		return 0;
7127 
7128        /*
7129 	* We may be recently in ticked or tickless idle mode. At the first
7130 	* busy tick after returning from idle, we will update the busy stats.
7131 	*/
7132 	set_cpu_sd_state_busy();
7133 	nohz_balance_exit_idle(cpu);
7134 
7135 	/*
7136 	 * None are in tickless mode and hence no need for NOHZ idle load
7137 	 * balancing.
7138 	 */
7139 	if (likely(!atomic_read(&nohz.nr_cpus)))
7140 		return 0;
7141 
7142 	if (time_before(now, nohz.next_balance))
7143 		return 0;
7144 
7145 	if (rq->nr_running >= 2)
7146 		goto need_kick;
7147 
7148 	rcu_read_lock();
7149 	sd = rcu_dereference(per_cpu(sd_busy, cpu));
7150 
7151 	if (sd) {
7152 		sgp = sd->groups->sgp;
7153 		nr_busy = atomic_read(&sgp->nr_busy_cpus);
7154 
7155 		if (nr_busy > 1)
7156 			goto need_kick_unlock;
7157 	}
7158 
7159 	sd = rcu_dereference(per_cpu(sd_asym, cpu));
7160 
7161 	if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7162 				  sched_domain_span(sd)) < cpu))
7163 		goto need_kick_unlock;
7164 
7165 	rcu_read_unlock();
7166 	return 0;
7167 
7168 need_kick_unlock:
7169 	rcu_read_unlock();
7170 need_kick:
7171 	return 1;
7172 }
7173 #else
7174 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7175 #endif
7176 
7177 /*
7178  * run_rebalance_domains is triggered when needed from the scheduler tick.
7179  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7180  */
7181 static void run_rebalance_domains(struct softirq_action *h)
7182 {
7183 	struct rq *this_rq = this_rq();
7184 	enum cpu_idle_type idle = this_rq->idle_balance ?
7185 						CPU_IDLE : CPU_NOT_IDLE;
7186 
7187 	rebalance_domains(this_rq, idle);
7188 
7189 	/*
7190 	 * If this cpu has a pending nohz_balance_kick, then do the
7191 	 * balancing on behalf of the other idle cpus whose ticks are
7192 	 * stopped.
7193 	 */
7194 	nohz_idle_balance(this_rq, idle);
7195 }
7196 
7197 /*
7198  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7199  */
7200 void trigger_load_balance(struct rq *rq)
7201 {
7202 	/* Don't need to rebalance while attached to NULL domain */
7203 	if (unlikely(on_null_domain(rq)))
7204 		return;
7205 
7206 	if (time_after_eq(jiffies, rq->next_balance))
7207 		raise_softirq(SCHED_SOFTIRQ);
7208 #ifdef CONFIG_NO_HZ_COMMON
7209 	if (nohz_kick_needed(rq))
7210 		nohz_balancer_kick();
7211 #endif
7212 }
7213 
7214 static void rq_online_fair(struct rq *rq)
7215 {
7216 	update_sysctl();
7217 }
7218 
7219 static void rq_offline_fair(struct rq *rq)
7220 {
7221 	update_sysctl();
7222 
7223 	/* Ensure any throttled groups are reachable by pick_next_task */
7224 	unthrottle_offline_cfs_rqs(rq);
7225 }
7226 
7227 #endif /* CONFIG_SMP */
7228 
7229 /*
7230  * scheduler tick hitting a task of our scheduling class:
7231  */
7232 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7233 {
7234 	struct cfs_rq *cfs_rq;
7235 	struct sched_entity *se = &curr->se;
7236 
7237 	for_each_sched_entity(se) {
7238 		cfs_rq = cfs_rq_of(se);
7239 		entity_tick(cfs_rq, se, queued);
7240 	}
7241 
7242 	if (numabalancing_enabled)
7243 		task_tick_numa(rq, curr);
7244 
7245 	update_rq_runnable_avg(rq, 1);
7246 }
7247 
7248 /*
7249  * called on fork with the child task as argument from the parent's context
7250  *  - child not yet on the tasklist
7251  *  - preemption disabled
7252  */
7253 static void task_fork_fair(struct task_struct *p)
7254 {
7255 	struct cfs_rq *cfs_rq;
7256 	struct sched_entity *se = &p->se, *curr;
7257 	int this_cpu = smp_processor_id();
7258 	struct rq *rq = this_rq();
7259 	unsigned long flags;
7260 
7261 	raw_spin_lock_irqsave(&rq->lock, flags);
7262 
7263 	update_rq_clock(rq);
7264 
7265 	cfs_rq = task_cfs_rq(current);
7266 	curr = cfs_rq->curr;
7267 
7268 	/*
7269 	 * Not only the cpu but also the task_group of the parent might have
7270 	 * been changed after parent->se.parent,cfs_rq were copied to
7271 	 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7272 	 * of child point to valid ones.
7273 	 */
7274 	rcu_read_lock();
7275 	__set_task_cpu(p, this_cpu);
7276 	rcu_read_unlock();
7277 
7278 	update_curr(cfs_rq);
7279 
7280 	if (curr)
7281 		se->vruntime = curr->vruntime;
7282 	place_entity(cfs_rq, se, 1);
7283 
7284 	if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7285 		/*
7286 		 * Upon rescheduling, sched_class::put_prev_task() will place
7287 		 * 'current' within the tree based on its new key value.
7288 		 */
7289 		swap(curr->vruntime, se->vruntime);
7290 		resched_task(rq->curr);
7291 	}
7292 
7293 	se->vruntime -= cfs_rq->min_vruntime;
7294 
7295 	raw_spin_unlock_irqrestore(&rq->lock, flags);
7296 }
7297 
7298 /*
7299  * Priority of the task has changed. Check to see if we preempt
7300  * the current task.
7301  */
7302 static void
7303 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7304 {
7305 	if (!p->se.on_rq)
7306 		return;
7307 
7308 	/*
7309 	 * Reschedule if we are currently running on this runqueue and
7310 	 * our priority decreased, or if we are not currently running on
7311 	 * this runqueue and our priority is higher than the current's
7312 	 */
7313 	if (rq->curr == p) {
7314 		if (p->prio > oldprio)
7315 			resched_task(rq->curr);
7316 	} else
7317 		check_preempt_curr(rq, p, 0);
7318 }
7319 
7320 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7321 {
7322 	struct sched_entity *se = &p->se;
7323 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
7324 
7325 	/*
7326 	 * Ensure the task's vruntime is normalized, so that when it's
7327 	 * switched back to the fair class the enqueue_entity(.flags=0) will
7328 	 * do the right thing.
7329 	 *
7330 	 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7331 	 * have normalized the vruntime, if it's !on_rq, then only when
7332 	 * the task is sleeping will it still have non-normalized vruntime.
7333 	 */
7334 	if (!p->on_rq && p->state != TASK_RUNNING) {
7335 		/*
7336 		 * Fix up our vruntime so that the current sleep doesn't
7337 		 * cause 'unlimited' sleep bonus.
7338 		 */
7339 		place_entity(cfs_rq, se, 0);
7340 		se->vruntime -= cfs_rq->min_vruntime;
7341 	}
7342 
7343 #ifdef CONFIG_SMP
7344 	/*
7345 	* Remove our load from contribution when we leave sched_fair
7346 	* and ensure we don't carry in an old decay_count if we
7347 	* switch back.
7348 	*/
7349 	if (se->avg.decay_count) {
7350 		__synchronize_entity_decay(se);
7351 		subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7352 	}
7353 #endif
7354 }
7355 
7356 /*
7357  * We switched to the sched_fair class.
7358  */
7359 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7360 {
7361 	struct sched_entity *se = &p->se;
7362 #ifdef CONFIG_FAIR_GROUP_SCHED
7363 	/*
7364 	 * Since the real-depth could have been changed (only FAIR
7365 	 * class maintain depth value), reset depth properly.
7366 	 */
7367 	se->depth = se->parent ? se->parent->depth + 1 : 0;
7368 #endif
7369 	if (!se->on_rq)
7370 		return;
7371 
7372 	/*
7373 	 * We were most likely switched from sched_rt, so
7374 	 * kick off the schedule if running, otherwise just see
7375 	 * if we can still preempt the current task.
7376 	 */
7377 	if (rq->curr == p)
7378 		resched_task(rq->curr);
7379 	else
7380 		check_preempt_curr(rq, p, 0);
7381 }
7382 
7383 /* Account for a task changing its policy or group.
7384  *
7385  * This routine is mostly called to set cfs_rq->curr field when a task
7386  * migrates between groups/classes.
7387  */
7388 static void set_curr_task_fair(struct rq *rq)
7389 {
7390 	struct sched_entity *se = &rq->curr->se;
7391 
7392 	for_each_sched_entity(se) {
7393 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
7394 
7395 		set_next_entity(cfs_rq, se);
7396 		/* ensure bandwidth has been allocated on our new cfs_rq */
7397 		account_cfs_rq_runtime(cfs_rq, 0);
7398 	}
7399 }
7400 
7401 void init_cfs_rq(struct cfs_rq *cfs_rq)
7402 {
7403 	cfs_rq->tasks_timeline = RB_ROOT;
7404 	cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7405 #ifndef CONFIG_64BIT
7406 	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7407 #endif
7408 #ifdef CONFIG_SMP
7409 	atomic64_set(&cfs_rq->decay_counter, 1);
7410 	atomic_long_set(&cfs_rq->removed_load, 0);
7411 #endif
7412 }
7413 
7414 #ifdef CONFIG_FAIR_GROUP_SCHED
7415 static void task_move_group_fair(struct task_struct *p, int on_rq)
7416 {
7417 	struct sched_entity *se = &p->se;
7418 	struct cfs_rq *cfs_rq;
7419 
7420 	/*
7421 	 * If the task was not on the rq at the time of this cgroup movement
7422 	 * it must have been asleep, sleeping tasks keep their ->vruntime
7423 	 * absolute on their old rq until wakeup (needed for the fair sleeper
7424 	 * bonus in place_entity()).
7425 	 *
7426 	 * If it was on the rq, we've just 'preempted' it, which does convert
7427 	 * ->vruntime to a relative base.
7428 	 *
7429 	 * Make sure both cases convert their relative position when migrating
7430 	 * to another cgroup's rq. This does somewhat interfere with the
7431 	 * fair sleeper stuff for the first placement, but who cares.
7432 	 */
7433 	/*
7434 	 * When !on_rq, vruntime of the task has usually NOT been normalized.
7435 	 * But there are some cases where it has already been normalized:
7436 	 *
7437 	 * - Moving a forked child which is waiting for being woken up by
7438 	 *   wake_up_new_task().
7439 	 * - Moving a task which has been woken up by try_to_wake_up() and
7440 	 *   waiting for actually being woken up by sched_ttwu_pending().
7441 	 *
7442 	 * To prevent boost or penalty in the new cfs_rq caused by delta
7443 	 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7444 	 */
7445 	if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7446 		on_rq = 1;
7447 
7448 	if (!on_rq)
7449 		se->vruntime -= cfs_rq_of(se)->min_vruntime;
7450 	set_task_rq(p, task_cpu(p));
7451 	se->depth = se->parent ? se->parent->depth + 1 : 0;
7452 	if (!on_rq) {
7453 		cfs_rq = cfs_rq_of(se);
7454 		se->vruntime += cfs_rq->min_vruntime;
7455 #ifdef CONFIG_SMP
7456 		/*
7457 		 * migrate_task_rq_fair() will have removed our previous
7458 		 * contribution, but we must synchronize for ongoing future
7459 		 * decay.
7460 		 */
7461 		se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7462 		cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7463 #endif
7464 	}
7465 }
7466 
7467 void free_fair_sched_group(struct task_group *tg)
7468 {
7469 	int i;
7470 
7471 	destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7472 
7473 	for_each_possible_cpu(i) {
7474 		if (tg->cfs_rq)
7475 			kfree(tg->cfs_rq[i]);
7476 		if (tg->se)
7477 			kfree(tg->se[i]);
7478 	}
7479 
7480 	kfree(tg->cfs_rq);
7481 	kfree(tg->se);
7482 }
7483 
7484 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7485 {
7486 	struct cfs_rq *cfs_rq;
7487 	struct sched_entity *se;
7488 	int i;
7489 
7490 	tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7491 	if (!tg->cfs_rq)
7492 		goto err;
7493 	tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7494 	if (!tg->se)
7495 		goto err;
7496 
7497 	tg->shares = NICE_0_LOAD;
7498 
7499 	init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7500 
7501 	for_each_possible_cpu(i) {
7502 		cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7503 				      GFP_KERNEL, cpu_to_node(i));
7504 		if (!cfs_rq)
7505 			goto err;
7506 
7507 		se = kzalloc_node(sizeof(struct sched_entity),
7508 				  GFP_KERNEL, cpu_to_node(i));
7509 		if (!se)
7510 			goto err_free_rq;
7511 
7512 		init_cfs_rq(cfs_rq);
7513 		init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7514 	}
7515 
7516 	return 1;
7517 
7518 err_free_rq:
7519 	kfree(cfs_rq);
7520 err:
7521 	return 0;
7522 }
7523 
7524 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7525 {
7526 	struct rq *rq = cpu_rq(cpu);
7527 	unsigned long flags;
7528 
7529 	/*
7530 	* Only empty task groups can be destroyed; so we can speculatively
7531 	* check on_list without danger of it being re-added.
7532 	*/
7533 	if (!tg->cfs_rq[cpu]->on_list)
7534 		return;
7535 
7536 	raw_spin_lock_irqsave(&rq->lock, flags);
7537 	list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7538 	raw_spin_unlock_irqrestore(&rq->lock, flags);
7539 }
7540 
7541 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7542 			struct sched_entity *se, int cpu,
7543 			struct sched_entity *parent)
7544 {
7545 	struct rq *rq = cpu_rq(cpu);
7546 
7547 	cfs_rq->tg = tg;
7548 	cfs_rq->rq = rq;
7549 	init_cfs_rq_runtime(cfs_rq);
7550 
7551 	tg->cfs_rq[cpu] = cfs_rq;
7552 	tg->se[cpu] = se;
7553 
7554 	/* se could be NULL for root_task_group */
7555 	if (!se)
7556 		return;
7557 
7558 	if (!parent) {
7559 		se->cfs_rq = &rq->cfs;
7560 		se->depth = 0;
7561 	} else {
7562 		se->cfs_rq = parent->my_q;
7563 		se->depth = parent->depth + 1;
7564 	}
7565 
7566 	se->my_q = cfs_rq;
7567 	/* guarantee group entities always have weight */
7568 	update_load_set(&se->load, NICE_0_LOAD);
7569 	se->parent = parent;
7570 }
7571 
7572 static DEFINE_MUTEX(shares_mutex);
7573 
7574 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7575 {
7576 	int i;
7577 	unsigned long flags;
7578 
7579 	/*
7580 	 * We can't change the weight of the root cgroup.
7581 	 */
7582 	if (!tg->se[0])
7583 		return -EINVAL;
7584 
7585 	shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7586 
7587 	mutex_lock(&shares_mutex);
7588 	if (tg->shares == shares)
7589 		goto done;
7590 
7591 	tg->shares = shares;
7592 	for_each_possible_cpu(i) {
7593 		struct rq *rq = cpu_rq(i);
7594 		struct sched_entity *se;
7595 
7596 		se = tg->se[i];
7597 		/* Propagate contribution to hierarchy */
7598 		raw_spin_lock_irqsave(&rq->lock, flags);
7599 
7600 		/* Possible calls to update_curr() need rq clock */
7601 		update_rq_clock(rq);
7602 		for_each_sched_entity(se)
7603 			update_cfs_shares(group_cfs_rq(se));
7604 		raw_spin_unlock_irqrestore(&rq->lock, flags);
7605 	}
7606 
7607 done:
7608 	mutex_unlock(&shares_mutex);
7609 	return 0;
7610 }
7611 #else /* CONFIG_FAIR_GROUP_SCHED */
7612 
7613 void free_fair_sched_group(struct task_group *tg) { }
7614 
7615 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7616 {
7617 	return 1;
7618 }
7619 
7620 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7621 
7622 #endif /* CONFIG_FAIR_GROUP_SCHED */
7623 
7624 
7625 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7626 {
7627 	struct sched_entity *se = &task->se;
7628 	unsigned int rr_interval = 0;
7629 
7630 	/*
7631 	 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7632 	 * idle runqueue:
7633 	 */
7634 	if (rq->cfs.load.weight)
7635 		rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7636 
7637 	return rr_interval;
7638 }
7639 
7640 /*
7641  * All the scheduling class methods:
7642  */
7643 const struct sched_class fair_sched_class = {
7644 	.next			= &idle_sched_class,
7645 	.enqueue_task		= enqueue_task_fair,
7646 	.dequeue_task		= dequeue_task_fair,
7647 	.yield_task		= yield_task_fair,
7648 	.yield_to_task		= yield_to_task_fair,
7649 
7650 	.check_preempt_curr	= check_preempt_wakeup,
7651 
7652 	.pick_next_task		= pick_next_task_fair,
7653 	.put_prev_task		= put_prev_task_fair,
7654 
7655 #ifdef CONFIG_SMP
7656 	.select_task_rq		= select_task_rq_fair,
7657 	.migrate_task_rq	= migrate_task_rq_fair,
7658 
7659 	.rq_online		= rq_online_fair,
7660 	.rq_offline		= rq_offline_fair,
7661 
7662 	.task_waking		= task_waking_fair,
7663 #endif
7664 
7665 	.set_curr_task          = set_curr_task_fair,
7666 	.task_tick		= task_tick_fair,
7667 	.task_fork		= task_fork_fair,
7668 
7669 	.prio_changed		= prio_changed_fair,
7670 	.switched_from		= switched_from_fair,
7671 	.switched_to		= switched_to_fair,
7672 
7673 	.get_rr_interval	= get_rr_interval_fair,
7674 
7675 #ifdef CONFIG_FAIR_GROUP_SCHED
7676 	.task_move_group	= task_move_group_fair,
7677 #endif
7678 };
7679 
7680 #ifdef CONFIG_SCHED_DEBUG
7681 void print_cfs_stats(struct seq_file *m, int cpu)
7682 {
7683 	struct cfs_rq *cfs_rq;
7684 
7685 	rcu_read_lock();
7686 	for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7687 		print_cfs_rq(m, cpu, cfs_rq);
7688 	rcu_read_unlock();
7689 }
7690 #endif
7691 
7692 __init void init_sched_fair_class(void)
7693 {
7694 #ifdef CONFIG_SMP
7695 	open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7696 
7697 #ifdef CONFIG_NO_HZ_COMMON
7698 	nohz.next_balance = jiffies;
7699 	zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7700 	cpu_notifier(sched_ilb_notifier, 0);
7701 #endif
7702 #endif /* SMP */
7703 
7704 }
7705