xref: /linux/kernel/sched/fair.c (revision 78b47fb995b5f3e7e54f0bde365063b530299726)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4  *
5  *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6  *
7  *  Interactivity improvements by Mike Galbraith
8  *  (C) 2007 Mike Galbraith <efault@gmx.de>
9  *
10  *  Various enhancements by Dmitry Adamushko.
11  *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12  *
13  *  Group scheduling enhancements by Srivatsa Vaddagiri
14  *  Copyright IBM Corporation, 2007
15  *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16  *
17  *  Scaled math optimizations by Thomas Gleixner
18  *  Copyright (C) 2007, Linutronix GmbH, Thomas Gleixner <tglx@kernel.org>
19  *
20  *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
22  */
23 #include <linux/energy_model.h>
24 #include <linux/mmap_lock.h>
25 #include <linux/jiffies.h>
26 #include <linux/math.h>
27 #include <linux/mm_api.h>
28 #include <linux/highmem.h>
29 #include <linux/hrtimer.h>
30 #include <linux/hrtimer_bases.h>
31 #include <linux/spinlock_api.h>
32 #include <linux/cpumask_api.h>
33 #include <linux/lockdep_api.h>
34 #include <linux/softirq.h>
35 #include <linux/refcount_api.h>
36 #include <linux/topology.h>
37 #include <linux/sched/clock.h>
38 #include <linux/sched/cond_resched.h>
39 #include <linux/sched/cputime.h>
40 #include <linux/sched/isolation.h>
41 #include <linux/sched/nohz.h>
42 #include <linux/sched/prio.h>
43 #include <linux/static_call.h>
44 
45 #include <linux/cpuidle.h>
46 #include <linux/interrupt.h>
47 #include <linux/memory-tiers.h>
48 #include <linux/mempolicy.h>
49 #include <linux/mutex_api.h>
50 #include <linux/profile.h>
51 #include <linux/psi.h>
52 #include <linux/ratelimit.h>
53 #include <linux/task_work.h>
54 #include <linux/rbtree_augmented.h>
55 
56 #include <asm/switch_to.h>
57 
58 #include <uapi/linux/sched/types.h>
59 
60 #include "sched.h"
61 #include "stats.h"
62 #include "autogroup.h"
63 
64 /*
65  * The initial- and re-scaling of tunables is configurable
66  *
67  * Options are:
68  *
69  *   SCHED_TUNABLESCALING_NONE - unscaled, always *1
70  *   SCHED_TUNABLESCALING_LOG - scaled logarithmically, *1+ilog(ncpus)
71  *   SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
72  *
73  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
74  */
75 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
76 
77 /*
78  * Default base time slice (request size r_i) for SCHED_NORMAL/SCHED_BATCH:
79  *
80  * Under EEVDF this is the request size used to compute the virtual
81  * deadline; see update_deadline().
82  *
83  * (default: 0.70 msec * (1 + ilog(ncpus)), units: nanoseconds)
84  */
85 unsigned int sysctl_sched_base_slice			= 700000ULL;
86 static unsigned int normalized_sysctl_sched_base_slice	= 700000ULL;
87 
88 __read_mostly unsigned int sysctl_sched_migration_cost	= 500000UL;
89 
90 static int __init setup_sched_thermal_decay_shift(char *str)
91 {
92 	pr_warn("Ignoring the deprecated sched_thermal_decay_shift= option\n");
93 	return 1;
94 }
95 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
96 
97 /*
98  * For asym packing, by default the lower numbered CPU has higher priority.
99  */
100 int __weak arch_asym_cpu_priority(int cpu)
101 {
102 	return -cpu;
103 }
104 
105 /*
106  * The margin used when comparing utilization with CPU capacity.
107  *
108  * (default: ~20%)
109  */
110 #define fits_capacity(cap, max)	((cap) * 1280 < (max) * 1024)
111 
112 /*
113  * The margin used when comparing CPU capacities.
114  * is 'cap1' noticeably greater than 'cap2'
115  *
116  * (default: ~5%)
117  */
118 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
119 
120 #ifdef CONFIG_CFS_BANDWIDTH
121 /*
122  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
123  * each time a cfs_rq requests quota.
124  *
125  * Note: in the case that the slice exceeds the runtime remaining (either due
126  * to consumption or the quota being specified to be smaller than the slice)
127  * we will always only issue the remaining available time.
128  *
129  * (default: 5 msec, units: microseconds)
130  */
131 static unsigned int sysctl_sched_cfs_bandwidth_slice		= 5000UL;
132 #endif
133 
134 #ifdef CONFIG_NUMA_BALANCING
135 /* Restrict the NUMA promotion throughput (MB/s) for each target node. */
136 static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536;
137 #endif
138 
139 #ifdef CONFIG_SYSCTL
140 static const struct ctl_table sched_fair_sysctls[] = {
141 #ifdef CONFIG_CFS_BANDWIDTH
142 	{
143 		.procname       = "sched_cfs_bandwidth_slice_us",
144 		.data           = &sysctl_sched_cfs_bandwidth_slice,
145 		.maxlen         = sizeof(unsigned int),
146 		.mode           = 0644,
147 		.proc_handler   = proc_dointvec_minmax,
148 		.extra1         = SYSCTL_ONE,
149 	},
150 #endif
151 #ifdef CONFIG_NUMA_BALANCING
152 	{
153 		.procname	= "numa_balancing_promote_rate_limit_MBps",
154 		.data		= &sysctl_numa_balancing_promote_rate_limit,
155 		.maxlen		= sizeof(unsigned int),
156 		.mode		= 0644,
157 		.proc_handler	= proc_dointvec_minmax,
158 		.extra1		= SYSCTL_ZERO,
159 	},
160 #endif /* CONFIG_NUMA_BALANCING */
161 };
162 
163 static int __init sched_fair_sysctl_init(void)
164 {
165 	register_sysctl_init("kernel", sched_fair_sysctls);
166 	return 0;
167 }
168 late_initcall(sched_fair_sysctl_init);
169 #endif /* CONFIG_SYSCTL */
170 
171 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
172 {
173 	lw->weight += inc;
174 	lw->inv_weight = 0;
175 }
176 
177 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
178 {
179 	lw->weight -= dec;
180 	lw->inv_weight = 0;
181 }
182 
183 static inline void update_load_set(struct load_weight *lw, unsigned long w)
184 {
185 	lw->weight = w;
186 	lw->inv_weight = 0;
187 }
188 
189 /*
190  * Increase the granularity value when there are more CPUs,
191  * because with more CPUs the 'effective latency' as visible
192  * to users decreases. But the relationship is not linear,
193  * so pick a second-best guess by going with the log2 of the
194  * number of CPUs.
195  *
196  * This idea comes from the SD scheduler of Con Kolivas:
197  */
198 static unsigned int get_update_sysctl_factor(void)
199 {
200 	unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
201 	unsigned int factor;
202 
203 	switch (sysctl_sched_tunable_scaling) {
204 	case SCHED_TUNABLESCALING_NONE:
205 		factor = 1;
206 		break;
207 	case SCHED_TUNABLESCALING_LINEAR:
208 		factor = cpus;
209 		break;
210 	case SCHED_TUNABLESCALING_LOG:
211 	default:
212 		factor = 1 + ilog2(cpus);
213 		break;
214 	}
215 
216 	return factor;
217 }
218 
219 static void update_sysctl(void)
220 {
221 	unsigned int factor = get_update_sysctl_factor();
222 
223 #define SET_SYSCTL(name) \
224 	(sysctl_##name = (factor) * normalized_sysctl_##name)
225 	SET_SYSCTL(sched_base_slice);
226 #undef SET_SYSCTL
227 }
228 
229 void __init sched_init_granularity(void)
230 {
231 	update_sysctl();
232 }
233 
234 #ifndef CONFIG_64BIT
235 #define WMULT_CONST	(~0U)
236 #define WMULT_SHIFT	32
237 
238 static void __update_inv_weight(struct load_weight *lw)
239 {
240 	unsigned long w;
241 
242 	if (likely(lw->inv_weight))
243 		return;
244 
245 	w = scale_load_down(lw->weight);
246 
247 	if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
248 		lw->inv_weight = 1;
249 	else if (unlikely(!w))
250 		lw->inv_weight = WMULT_CONST;
251 	else
252 		lw->inv_weight = WMULT_CONST / w;
253 }
254 
255 /*
256  * delta_exec * weight / lw.weight
257  *   OR
258  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
259  *
260  * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
261  * we're guaranteed shift stays positive because inv_weight is guaranteed to
262  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
263  *
264  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
265  * weight/lw.weight <= 1, and therefore our shift will also be positive.
266  */
267 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
268 {
269 	u64 fact = scale_load_down(weight);
270 	u32 fact_hi = (u32)(fact >> 32);
271 	int shift = WMULT_SHIFT;
272 	int fs;
273 
274 	__update_inv_weight(lw);
275 
276 	if (unlikely(fact_hi)) {
277 		fs = fls(fact_hi);
278 		shift -= fs;
279 		fact >>= fs;
280 	}
281 
282 	fact = mul_u32_u32(fact, lw->inv_weight);
283 
284 	fact_hi = (u32)(fact >> 32);
285 	if (fact_hi) {
286 		fs = fls(fact_hi);
287 		shift -= fs;
288 		fact >>= fs;
289 	}
290 
291 	return mul_u64_u32_shr(delta_exec, fact, shift);
292 }
293 #else
294 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
295 {
296 	return (delta_exec * weight) / lw->weight;
297 }
298 #endif
299 
300 /*
301  * delta /= w
302  */
303 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
304 {
305 	if (se->h_load.weight != NICE_0_LOAD)
306 		delta = __calc_delta(delta, NICE_0_LOAD, &se->h_load);
307 
308 	return delta;
309 }
310 
311 const struct sched_class fair_sched_class;
312 
313 /**************************************************************
314  * CFS operations on generic schedulable entities:
315  */
316 
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 
319 /* Walk up scheduling entities hierarchy */
320 #define for_each_sched_entity(se) \
321 		for (; se; se = se->parent)
322 
323 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
324 {
325 	struct rq *rq = rq_of(cfs_rq);
326 	int cpu = cpu_of(rq);
327 
328 	if (cfs_rq->on_list)
329 		return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
330 
331 	cfs_rq->on_list = 1;
332 
333 	/*
334 	 * Ensure we either appear before our parent (if already
335 	 * enqueued) or force our parent to appear after us when it is
336 	 * enqueued. The fact that we always enqueue bottom-up
337 	 * reduces this to two cases and a special case for the root
338 	 * cfs_rq. Furthermore, it also means that we will always reset
339 	 * tmp_alone_branch either when the branch is connected
340 	 * to a tree or when we reach the top of the tree
341 	 */
342 	if (cfs_rq->tg->parent &&
343 	    tg_cfs_rq(cfs_rq->tg->parent, cpu)->on_list) {
344 		/*
345 		 * If parent is already on the list, we add the child
346 		 * just before. Thanks to circular linked property of
347 		 * the list, this means to put the child at the tail
348 		 * of the list that starts by parent.
349 		 */
350 		list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
351 			&(tg_cfs_rq(cfs_rq->tg->parent, cpu)->leaf_cfs_rq_list));
352 		/*
353 		 * The branch is now connected to its tree so we can
354 		 * reset tmp_alone_branch to the beginning of the
355 		 * list.
356 		 */
357 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
358 		return true;
359 	}
360 
361 	if (!cfs_rq->tg->parent) {
362 		/*
363 		 * cfs rq without parent should be put
364 		 * at the tail of the list.
365 		 */
366 		list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
367 			&rq->leaf_cfs_rq_list);
368 		/*
369 		 * We have reach the top of a tree so we can reset
370 		 * tmp_alone_branch to the beginning of the list.
371 		 */
372 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
373 		return true;
374 	}
375 
376 	/*
377 	 * The parent has not already been added so we want to
378 	 * make sure that it will be put after us.
379 	 * tmp_alone_branch points to the begin of the branch
380 	 * where we will add parent.
381 	 */
382 	list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
383 	/*
384 	 * update tmp_alone_branch to points to the new begin
385 	 * of the branch
386 	 */
387 	rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
388 	return false;
389 }
390 
391 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
392 {
393 	if (cfs_rq->on_list) {
394 		struct rq *rq = rq_of(cfs_rq);
395 
396 		/*
397 		 * With cfs_rq being unthrottled/throttled during an enqueue,
398 		 * it can happen the tmp_alone_branch points to the leaf that
399 		 * we finally want to delete. In this case, tmp_alone_branch moves
400 		 * to the prev element but it will point to rq->leaf_cfs_rq_list
401 		 * at the end of the enqueue.
402 		 */
403 		if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
404 			rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
405 
406 		list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
407 		cfs_rq->on_list = 0;
408 	}
409 }
410 
411 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
412 {
413 	WARN_ON_ONCE(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
414 }
415 
416 /* Iterate through all leaf cfs_rq's on a runqueue */
417 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)			\
418 	list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list,	\
419 				 leaf_cfs_rq_list)
420 
421 /* Do the two (enqueued) entities belong to the same group ? */
422 static inline struct cfs_rq *
423 is_same_group(struct sched_entity *se, struct sched_entity *pse)
424 {
425 	if (se->cfs_rq == pse->cfs_rq)
426 		return se->cfs_rq;
427 
428 	return NULL;
429 }
430 
431 static inline struct sched_entity *parent_entity(const struct sched_entity *se)
432 {
433 	return se->parent;
434 }
435 
436 static int tg_is_idle(struct task_group *tg)
437 {
438 	return tg->idle > 0;
439 }
440 
441 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
442 {
443 	return cfs_rq->idle > 0;
444 }
445 
446 static int se_is_idle(struct sched_entity *se)
447 {
448 	if (entity_is_task(se))
449 		return task_has_idle_policy(task_of(se));
450 	return cfs_rq_is_idle(group_cfs_rq(se));
451 }
452 
453 #else /* !CONFIG_FAIR_GROUP_SCHED: */
454 
455 #define for_each_sched_entity(se) \
456 		for (; se; se = NULL)
457 
458 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
459 {
460 	return true;
461 }
462 
463 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
464 {
465 }
466 
467 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
468 {
469 }
470 
471 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)	\
472 		for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
473 
474 static inline struct sched_entity *parent_entity(struct sched_entity *se)
475 {
476 	return NULL;
477 }
478 
479 static inline int tg_is_idle(struct task_group *tg)
480 {
481 	return 0;
482 }
483 
484 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
485 {
486 	return 0;
487 }
488 
489 static int se_is_idle(struct sched_entity *se)
490 {
491 	return task_has_idle_policy(task_of(se));
492 }
493 
494 #endif /* !CONFIG_FAIR_GROUP_SCHED */
495 
496 static __always_inline
497 bool account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
498 
499 /**************************************************************
500  * Scheduling class tree data structure manipulation methods:
501  */
502 
503 extern void __BUILD_BUG_vruntime_cmp(void);
504 
505 /* Use __builtin_strcmp() because of __HAVE_ARCH_STRCMP: */
506 
507 #define vruntime_cmp(A, CMP_STR, B) ({				\
508 	int __res = 0;						\
509 								\
510 	if (!__builtin_strcmp(CMP_STR, "<")) {			\
511 		__res = ((s64)((A)-(B)) < 0);			\
512 	} else if (!__builtin_strcmp(CMP_STR, "<=")) {		\
513 		__res = ((s64)((A)-(B)) <= 0);			\
514 	} else if (!__builtin_strcmp(CMP_STR, ">")) {		\
515 		__res = ((s64)((A)-(B)) > 0);			\
516 	} else if (!__builtin_strcmp(CMP_STR, ">=")) {		\
517 		__res = ((s64)((A)-(B)) >= 0);			\
518 	} else {						\
519 		/* Unknown operator throws linker error: */	\
520 		__BUILD_BUG_vruntime_cmp();			\
521 	}							\
522 								\
523 	__res;							\
524 })
525 
526 extern void __BUILD_BUG_vruntime_op(void);
527 
528 #define vruntime_op(A, OP_STR, B) ({				\
529 	s64 __res = 0;						\
530 								\
531 	if (!__builtin_strcmp(OP_STR, "-")) {			\
532 		__res = (s64)((A)-(B));				\
533 	} else {						\
534 		/* Unknown operator throws linker error: */	\
535 		__BUILD_BUG_vruntime_op();			\
536 	}							\
537 								\
538 	__res;						\
539 })
540 
541 
542 static inline __maybe_unused u64 max_vruntime(u64 max_vruntime, u64 vruntime)
543 {
544 	if (vruntime_cmp(vruntime, ">", max_vruntime))
545 		max_vruntime = vruntime;
546 
547 	return max_vruntime;
548 }
549 
550 static inline __maybe_unused u64 min_vruntime(u64 min_vruntime, u64 vruntime)
551 {
552 	if (vruntime_cmp(vruntime, "<", min_vruntime))
553 		min_vruntime = vruntime;
554 
555 	return min_vruntime;
556 }
557 
558 static inline bool entity_before(const struct sched_entity *a,
559 				 const struct sched_entity *b)
560 {
561 	/*
562 	 * Tiebreak on vruntime seems unnecessary since it can
563 	 * hardly happen.
564 	 */
565 	return vruntime_cmp(a->deadline, "<", b->deadline);
566 }
567 
568 /*
569  * Per avg_vruntime() below, cfs_rq::zero_vruntime is only slightly stale
570  * and this value should be no more than two lag bounds. Which puts it in the
571  * general order of:
572  *
573  *	(slice + TICK_NSEC) << NICE_0_LOAD_SHIFT
574  *
575  * which is around 44 bits in size (on 64bit); that is 20 for
576  * NICE_0_LOAD_SHIFT, another 20 for NSEC_PER_MSEC and then a handful for
577  * however many msec the actual slice+tick ends up begin.
578  *
579  * (disregarding the actual divide-by-weight part makes for the worst case
580  * weight of 2, which nicely cancels vs the fuzz in zero_vruntime not actually
581  * being the zero-lag point).
582  */
583 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
584 {
585 	return vruntime_op(se->vruntime, "-", cfs_rq->zero_vruntime);
586 }
587 
588 #define __node_2_se(node) \
589 	rb_entry((node), struct sched_entity, run_node)
590 
591 /*
592  * Compute virtual time from the per-task service numbers:
593  *
594  * Fair schedulers conserve lag:
595  *
596  *   \Sum lag_i = 0
597  *
598  * Where lag_i is given by:
599  *
600  *   lag_i = S - s_i = w_i * (V - v_i)
601  *
602  * Where S is the ideal service time and V is it's virtual time counterpart.
603  * Therefore:
604  *
605  *   \Sum lag_i = 0
606  *   \Sum w_i * (V - v_i) = 0
607  *   \Sum (w_i * V - w_i * v_i) = 0
608  *
609  * From which we can solve an expression for V in v_i (which we have in
610  * se->vruntime):
611  *
612  *       \Sum v_i * w_i   \Sum v_i * w_i
613  *   V = -------------- = --------------
614  *          \Sum w_i            W
615  *
616  * Specifically, this is the weighted average of all entity virtual runtimes.
617  *
618  * [[ NOTE: this is only equal to the ideal scheduler under the condition
619  *          that join/leave operations happen at lag_i = 0, otherwise the
620  *          virtual time has non-contiguous motion equivalent to:
621  *
622  *	      V +-= lag_i / W
623  *
624  *	    Also see the comment in place_entity() that deals with this. ]]
625  *
626  * However, since v_i is u64, and the multiplication could easily overflow
627  * transform it into a relative form that uses smaller quantities:
628  *
629  * Substitute: v_i == (v_i - v0) + v0
630  *
631  *     \Sum ((v_i - v0) + v0) * w_i   \Sum (v_i - v0) * w_i
632  * V = ---------------------------- = --------------------- + v0
633  *                  W                            W
634  *
635  * Which we track using:
636  *
637  *                    v0 := cfs_rq->zero_vruntime
638  * \Sum (v_i - v0) * w_i := cfs_rq->sum_w_vruntime
639  *              \Sum w_i := cfs_rq->sum_weight
640  *
641  * Since zero_vruntime closely tracks the per-task service, these
642  * deltas: (v_i - v0), will be in the order of the maximal (virtual) lag
643  * induced in the system due to quantisation.
644  */
645 static inline unsigned long avg_vruntime_weight(struct cfs_rq *cfs_rq, unsigned long w)
646 {
647 #ifdef CONFIG_64BIT
648 	if (cfs_rq->sum_shift)
649 		w = max(2UL, w >> cfs_rq->sum_shift);
650 #endif
651 	return w;
652 }
653 
654 static inline void
655 __sum_w_vruntime_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
656 {
657 	unsigned long weight = avg_vruntime_weight(cfs_rq, se->h_load.weight);
658 	s64 w_vruntime, key = entity_key(cfs_rq, se);
659 
660 	w_vruntime = key * weight;
661 	WARN_ON_ONCE((w_vruntime >> 63) != (w_vruntime >> 62));
662 
663 	cfs_rq->sum_w_vruntime += w_vruntime;
664 	cfs_rq->sum_weight += weight;
665 }
666 
667 static void
668 sum_w_vruntime_add_paranoid(struct cfs_rq *cfs_rq, struct sched_entity *se)
669 {
670 	unsigned long weight;
671 	s64 key, tmp;
672 
673 again:
674 	weight = avg_vruntime_weight(cfs_rq, se->h_load.weight);
675 	key = entity_key(cfs_rq, se);
676 
677 	if (check_mul_overflow(key, weight, &key))
678 		goto overflow;
679 
680 	if (check_add_overflow(cfs_rq->sum_w_vruntime, key, &tmp))
681 		goto overflow;
682 
683 	cfs_rq->sum_w_vruntime = tmp;
684 	cfs_rq->sum_weight += weight;
685 	return;
686 
687 overflow:
688 	/*
689 	 * There's gotta be a limit -- if we're still failing at this point
690 	 * there's really nothing much to be done about things.
691 	 */
692 	BUG_ON(cfs_rq->sum_shift >= 10);
693 	cfs_rq->sum_shift++;
694 
695 	/*
696 	 * Note: \Sum (k_i * (w_i >> 1)) != (\Sum (k_i * w_i)) >> 1
697 	 */
698 	cfs_rq->sum_w_vruntime = 0;
699 	cfs_rq->sum_weight = 0;
700 
701 	for (struct rb_node *node = cfs_rq->tasks_timeline.rb_leftmost;
702 	     node; node = rb_next(node))
703 		__sum_w_vruntime_add(cfs_rq, __node_2_se(node));
704 
705 	goto again;
706 }
707 
708 static void
709 sum_w_vruntime_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
710 {
711 	if (sched_feat(PARANOID_AVG))
712 		return sum_w_vruntime_add_paranoid(cfs_rq, se);
713 
714 	__sum_w_vruntime_add(cfs_rq, se);
715 }
716 
717 static void
718 sum_w_vruntime_sub(struct cfs_rq *cfs_rq, struct sched_entity *se)
719 {
720 	unsigned long weight = avg_vruntime_weight(cfs_rq, se->h_load.weight);
721 	s64 key = entity_key(cfs_rq, se);
722 
723 	cfs_rq->sum_w_vruntime -= key * weight;
724 	cfs_rq->sum_weight -= weight;
725 }
726 
727 static inline
728 void update_zero_vruntime(struct cfs_rq *cfs_rq, s64 delta)
729 {
730 	/*
731 	 * v' = v + d ==> sum_w_vruntime' = sum_w_vruntime - d*sum_weight
732 	 */
733 	cfs_rq->sum_w_vruntime -= cfs_rq->sum_weight * delta;
734 	cfs_rq->zero_vruntime += delta;
735 }
736 
737 /*
738  * Specifically: avg_vruntime() + 0 must result in entity_eligible() := true
739  * For this to be so, the result of this function must have a left bias.
740  *
741  * Called in:
742  *  - place_entity()      -- before enqueue
743  *  - update_entity_lag() -- before dequeue
744  *  - update_deadline()   -- slice expiration
745  *
746  * This means it is one entry 'behind' but that puts it close enough to where
747  * the bound on entity_key() is at most two lag bounds.
748  */
749 u64 avg_vruntime(struct cfs_rq *cfs_rq)
750 {
751 	struct sched_entity *curr = cfs_rq->curr;
752 	long weight = cfs_rq->sum_weight;
753 	s64 delta = 0;
754 
755 	if (curr && !curr->on_rq)
756 		curr = NULL;
757 
758 	if (weight) {
759 		s64 runtime = cfs_rq->sum_w_vruntime;
760 
761 		if (curr) {
762 			unsigned long w = avg_vruntime_weight(cfs_rq, curr->h_load.weight);
763 
764 			runtime += entity_key(cfs_rq, curr) * w;
765 			weight += w;
766 		}
767 
768 		/* sign flips effective floor / ceiling */
769 		if (runtime < 0)
770 			runtime -= (weight - 1);
771 
772 		delta = div64_long(runtime, weight);
773 	} else if (curr) {
774 		/*
775 		 * When there is but one element, it is the average.
776 		 */
777 		delta = curr->vruntime - cfs_rq->zero_vruntime;
778 	}
779 
780 	update_zero_vruntime(cfs_rq, delta);
781 
782 	return cfs_rq->zero_vruntime;
783 }
784 
785 /*
786  *     \Sum (v_i - v0)*w_i
787  * V = ------------------- + v0
788  *          \Sum w_i
789  *
790  * Let W = \Sum w_i, and move v_j such that 'v_j == V', thus:
791  *
792  * V = 1/W * {(v_j - v0)*w_j + \Sum_i!=j (v_i - v0)*w_i} + v0
793  *
794  * v_j = 1/W * {(v_j - v0)*w_j + \Sum_i!=j (v_i - v0)*w_i} + v0
795  *
796  * v_j = 1/W * (v_j - v0)*w_j + 1/W * \Sum_i!=j (v_i - v0)*w_i + v0
797  *
798  * v_j - 1/W * (v_j - v0)*w_j = 1/W * \Sum_i!=j (v_i - v0)*w_i + v0
799  *
800  * v_j*W - (v_j - v0)*w_j = \Sum_i!=j (v_i - v0)*w_i + v0*W
801  *
802  * v_j*(W - w_j) + v0*w_j = \Sum_i!=j (v_i - v0)*w_i + v0*W
803  *
804  * v_j*(W - w_j) = \Sum_i!=j (v_i - v0)*w_i + v0*(W - w_j)
805  *
806  *       \Sum_i!=j (v_i - v0)*w_i
807  * v_j = ------------------------ + v0
808  *               W - w_j
809  *
810  * When v_j happens to be curr, then '\Sum_i!=j (v_i - v0)*w_i'
811  * is cfs_rq->sum_w_runtime, and 'W - w_j' is cfs_rq->sum_weight, since curr
812  * is not included in the sum.
813  */
814 static u64 ineligible_vruntime(struct cfs_rq *cfs_rq)
815 {
816 	struct sched_entity *curr = cfs_rq->curr;
817 	long weight = cfs_rq->sum_weight;
818 	s64 delta = 0;
819 
820 	if (curr && !curr->on_rq)
821 		curr = NULL;
822 
823 	if (weight) {
824 		s64 runtime = cfs_rq->sum_w_vruntime;
825 
826 		/*
827 		 * Do not add @curr to obtain the effective '- w_j' terms.
828 		 */
829 
830 		/* sign flips effective floor / ceiling */
831 		if (runtime < 0)
832 			runtime -= (weight - 1);
833 
834 		delta = div64_long(runtime, weight);
835 	}
836 
837 	return cfs_rq->zero_vruntime + delta + 1;
838 }
839 
840 static inline u64 cfs_rq_max_slice(struct cfs_rq *cfs_rq);
841 
842 /*
843  * lag_i = S - s_i = w_i * (V - v_i)
844  *
845  * However, since V is approximated by the weighted average of all entities it
846  * is possible -- by addition/removal/reweight to the tree -- to move V around
847  * and end up with a larger lag than we started with.
848  *
849  * Limit this to either double the slice length with a minimum of TICK_NSEC
850  * since that is the timing granularity.
851  *
852  * EEVDF gives the following limit for a steady state system:
853  *
854  *   -r_max < lag < max(r_max, q)
855  */
856 static s64 entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 avruntime)
857 {
858 	u64 max_slice = cfs_rq_max_slice(cfs_rq) + TICK_NSEC;
859 	s64 vlag, limit;
860 
861 	vlag = avruntime - se->vruntime;
862 	limit = calc_delta_fair(max_slice, se);
863 
864 	return clamp(vlag, -limit, limit);
865 }
866 
867 /*
868  * Delayed dequeue aims to reduce the negative lag of a dequeued task. While
869  * updating the lag of an entity, check that negative lag didn't increase
870  * during the delayed dequeue period which would be unfair.
871  * Similarly, check that the entity didn't gain positive lag when DELAY_ZERO
872  * is set.
873  *
874  * Return true if the vlag has been modified. Specifically:
875  *
876  *   se->vlag != avg_vruntime() - se->vruntime
877  *
878  * This can be due to clamping in entity_lag() or clamping due to
879  * sched_delayed. Either way, when vlag is modified and the entity is
880  * retained, the tree needs to be adjusted.
881  */
882 static __always_inline
883 bool update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
884 {
885 	u64 avruntime = avg_vruntime(cfs_rq);
886 	s64 vlag = entity_lag(cfs_rq, se, avruntime);
887 
888 	if (se->sched_delayed) {
889 		/* previous vlag < 0 otherwise se would not be delayed */
890 		vlag = max(vlag, se->vlag);
891 		if (sched_feat(DELAY_ZERO))
892 			vlag = min(vlag, 0);
893 	}
894 	se->vlag = vlag;
895 
896 	return avruntime - vlag != se->vruntime;
897 }
898 
899 /*
900  * Entity is eligible once it received less service than it ought to have,
901  * eg. lag >= 0.
902  *
903  * lag_i = S - s_i = w_i*(V - v_i)
904  *
905  * lag_i >= 0 -> V >= v_i
906  *
907  *     \Sum (v_i - v0)*w_i
908  * V = ------------------- + v0
909  *          \Sum w_i
910  *
911  * lag_i >= 0 -> \Sum (v_i - v0)*w_i >= (v_i - v0)*(\Sum w_i)
912  *
913  * Note: using 'avg_vruntime() > se->vruntime' is inaccurate due
914  *       to the loss in precision caused by the division.
915  */
916 static int vruntime_eligible(struct cfs_rq *cfs_rq, u64 vruntime)
917 {
918 	struct sched_entity *curr = cfs_rq->curr;
919 	s64 key, avg = cfs_rq->sum_w_vruntime;
920 	long load = cfs_rq->sum_weight;
921 
922 	if (curr && curr->on_rq) {
923 		unsigned long weight = avg_vruntime_weight(cfs_rq, curr->h_load.weight);
924 
925 		avg += entity_key(cfs_rq, curr) * weight;
926 		load += weight;
927 	}
928 
929 	key = vruntime_op(vruntime, "-", cfs_rq->zero_vruntime);
930 
931 	/*
932 	 * The worst case term for @key includes 'NSEC_TICK * NICE_0_LOAD'
933 	 * and @load obviously includes NICE_0_LOAD. NSEC_TICK is around 24
934 	 * bits, while NICE_0_LOAD is 20 on 64bit and 10 otherwise.
935 	 *
936 	 * This gives that on 64bit the product will be at least 64bit which
937 	 * overflows s64, while on 32bit it will only be 44bits and should fit
938 	 * comfortably.
939 	 */
940 #ifdef CONFIG_64BIT
941 #ifdef CONFIG_ARCH_SUPPORTS_INT128
942 	/* This often results in simpler code than __builtin_mul_overflow(). */
943 	return avg >= (__int128)key * load;
944 #else
945 	s64 rhs;
946 	/*
947 	 * On overflow, the sign of key tells us the correct answer: a large
948 	 * positive key means vruntime >> V, so not eligible; a large negative
949 	 * key means vruntime << V, so eligible.
950 	 */
951 	if (check_mul_overflow(key, load, &rhs))
952 		return key <= 0;
953 
954 	return avg >= rhs;
955 #endif
956 #else /* 32bit */
957 	return avg >= key * load;
958 #endif
959 }
960 
961 int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se)
962 {
963 	return vruntime_eligible(cfs_rq, se->vruntime);
964 }
965 
966 static inline u64 cfs_rq_min_slice(struct cfs_rq *cfs_rq)
967 {
968 	struct sched_entity *root = __pick_root_entity(cfs_rq);
969 	struct sched_entity *curr = cfs_rq->curr;
970 	u64 min_slice = ~0ULL;
971 
972 	if (curr && curr->on_rq)
973 		min_slice = curr->slice;
974 
975 	if (root)
976 		min_slice = min(min_slice, root->min_slice);
977 
978 	return min_slice;
979 }
980 
981 static inline u64 cfs_rq_max_slice(struct cfs_rq *cfs_rq)
982 {
983 	struct sched_entity *root = __pick_root_entity(cfs_rq);
984 	struct sched_entity *curr = cfs_rq->curr;
985 	u64 max_slice = 0ULL;
986 
987 	if (curr && curr->on_rq)
988 		max_slice = curr->slice;
989 
990 	if (root)
991 		max_slice = max(max_slice, root->max_slice);
992 
993 	return max_slice;
994 }
995 
996 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
997 {
998 	return entity_before(__node_2_se(a), __node_2_se(b));
999 }
1000 
1001 /* min() for wrapping vruntimes */
1002 static inline u64 __min_vruntime(u64 a, u64 b)
1003 {
1004 	return vruntime_cmp(a, "<", b) ? a : b;
1005 }
1006 
1007 RB_DECLARE_CALLBACKS(static, min_vruntime_cb,
1008 		     struct sched_entity, run_node,
1009 		     RB_AUG(vruntime, min_vruntime, __min_vruntime),
1010 		     RB_AUG(slice, min_slice, min),
1011 		     RB_AUG(slice, max_slice, max));
1012 
1013 /*
1014  * Enqueue an entity into the rb-tree:
1015  */
1016 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1017 {
1018 	WARN_ON_ONCE(&rq_of(cfs_rq)->cfs != cfs_rq);
1019 	WARN_ON_ONCE(!entity_is_task(se));
1020 
1021 	sum_w_vruntime_add(cfs_rq, se);
1022 	se->min_vruntime = se->vruntime;
1023 	se->min_slice = se->slice;
1024 	se->max_slice = se->slice;
1025 
1026 	rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
1027 				__entity_less, &min_vruntime_cb);
1028 }
1029 
1030 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1031 {
1032 	WARN_ON_ONCE(&rq_of(cfs_rq)->cfs != cfs_rq);
1033 	WARN_ON_ONCE(!entity_is_task(se));
1034 
1035 	rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
1036 				  &min_vruntime_cb);
1037 	sum_w_vruntime_sub(cfs_rq, se);
1038 }
1039 
1040 struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq)
1041 {
1042 	struct rb_node *root = cfs_rq->tasks_timeline.rb_root.rb_node;
1043 
1044 	if (!root)
1045 		return NULL;
1046 
1047 	return __node_2_se(root);
1048 }
1049 
1050 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
1051 {
1052 	struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
1053 
1054 	if (!left)
1055 		return NULL;
1056 
1057 	return __node_2_se(left);
1058 }
1059 
1060 /*
1061  * Set the vruntime up to which an entity can run before looking
1062  * for another entity to pick.
1063  * In case of run to parity, we use the shortest slice of the enqueued
1064  * entities to set the protected period.
1065  * When run to parity is disabled, we give a minimum quantum to the running
1066  * entity to ensure progress.
1067  */
1068 static inline void set_protect_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
1069 {
1070 	u64 slice = normalized_sysctl_sched_base_slice;
1071 	u64 vprot = se->deadline;
1072 
1073 	if (sched_feat(RUN_TO_PARITY))
1074 		slice = cfs_rq_min_slice(cfs_rq);
1075 
1076 	slice = min(slice, se->slice);
1077 
1078 	/* If there are shorter slices than se's one */
1079 	if (slice != se->slice) {
1080 		vprot = min_vruntime(vprot, se->vruntime + calc_delta_fair(slice, se));
1081 		if (sched_feat(PREEMPT_SHORT))
1082 			vprot = min_vruntime(vprot, ineligible_vruntime(cfs_rq));
1083 	}
1084 
1085 	se->vprot = vprot;
1086 }
1087 
1088 static inline void update_protect_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
1089 {
1090 	u64 vruntime = min_vruntime(se->vruntime, avg_vruntime(cfs_rq));
1091 	u64 slice = cfs_rq_min_slice(cfs_rq);
1092 	u64 vprot = min_vruntime(se->vprot, vruntime + calc_delta_fair(slice, se));
1093 
1094 	if (sched_feat(PREEMPT_SHORT) && slice != se->slice)
1095 		vprot = min_vruntime(vprot, ineligible_vruntime(cfs_rq));
1096 
1097 	se->vprot = vprot;
1098 }
1099 
1100 static inline bool protect_slice(struct sched_entity *se)
1101 {
1102 	return vruntime_cmp(se->vruntime, "<", se->vprot);
1103 }
1104 
1105 static inline void cancel_protect_slice(struct sched_entity *se)
1106 {
1107 	if (protect_slice(se))
1108 		se->vprot = se->vruntime;
1109 }
1110 
1111 /*
1112  * Earliest Eligible Virtual Deadline First
1113  *
1114  * In order to provide latency guarantees for different request sizes
1115  * EEVDF selects the best runnable task from two criteria:
1116  *
1117  *  1) the task must be eligible (must be owed service)
1118  *
1119  *  2) from those tasks that meet 1), we select the one
1120  *     with the earliest virtual deadline.
1121  *
1122  * We can do this in O(log n) time due to an augmented RB-tree. The
1123  * tree keeps the entries sorted on deadline, but also functions as a
1124  * heap based on the vruntime by keeping:
1125  *
1126  *  se->min_vruntime = min(se->vruntime, se->{left,right}->min_vruntime)
1127  *
1128  * Which allows tree pruning through eligibility.
1129  */
1130 static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq, bool protect)
1131 {
1132 	struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node;
1133 	struct sched_entity *se = __pick_first_entity(cfs_rq);
1134 	struct sched_entity *curr = cfs_rq->curr;
1135 	struct sched_entity *best = NULL;
1136 
1137 	/*
1138 	 * We can safely skip eligibility check if there is only one entity
1139 	 * in this cfs_rq, saving some cycles.
1140 	 */
1141 	if (cfs_rq->h_nr_queued == 1)
1142 		return curr && curr->on_rq ? curr : se;
1143 
1144 	/*
1145 	 * Picking the ->next buddy will affect latency but not fairness.
1146 	 */
1147 	if (sched_feat(PICK_BUDDY) && protect &&
1148 	    cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next)) {
1149 		/* ->next will never be delayed */
1150 		WARN_ON_ONCE(cfs_rq->next->sched_delayed);
1151 		return cfs_rq->next;
1152 	}
1153 
1154 	if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))
1155 		curr = NULL;
1156 
1157 	if (curr && protect && protect_slice(curr))
1158 		return curr;
1159 
1160 	/* Pick the leftmost entity if it's eligible */
1161 	if (se && entity_eligible(cfs_rq, se)) {
1162 		best = se;
1163 		goto found;
1164 	}
1165 
1166 	/* Heap search for the EEVD entity */
1167 	while (node) {
1168 		struct rb_node *left = node->rb_left;
1169 
1170 		/*
1171 		 * Eligible entities in left subtree are always better
1172 		 * choices, since they have earlier deadlines.
1173 		 */
1174 		if (left && vruntime_eligible(cfs_rq,
1175 					__node_2_se(left)->min_vruntime)) {
1176 			node = left;
1177 			continue;
1178 		}
1179 
1180 		se = __node_2_se(node);
1181 
1182 		/*
1183 		 * The left subtree either is empty or has no eligible
1184 		 * entity, so check the current node since it is the one
1185 		 * with earliest deadline that might be eligible.
1186 		 */
1187 		if (entity_eligible(cfs_rq, se)) {
1188 			best = se;
1189 			break;
1190 		}
1191 
1192 		node = node->rb_right;
1193 	}
1194 found:
1195 	if (!best || (curr && entity_before(curr, best)))
1196 		best = curr;
1197 
1198 	return best;
1199 }
1200 
1201 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
1202 {
1203 	struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
1204 
1205 	if (!last)
1206 		return NULL;
1207 
1208 	return __node_2_se(last);
1209 }
1210 
1211 /**************************************************************
1212  * Scheduling class statistics methods:
1213  */
1214 int sched_update_scaling(void)
1215 {
1216 	unsigned int factor = get_update_sysctl_factor();
1217 
1218 #define WRT_SYSCTL(name) \
1219 	(normalized_sysctl_##name = sysctl_##name / (factor))
1220 	WRT_SYSCTL(sched_base_slice);
1221 #undef WRT_SYSCTL
1222 
1223 	return 0;
1224 }
1225 
1226 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se);
1227 
1228 /*
1229  * XXX: strictly: vd_i += N*r_i/w_i such that: vd_i > ve_i
1230  * this is probably good enough.
1231  */
1232 static bool update_deadline(struct cfs_rq *cfs_rq, struct sched_entity *se)
1233 {
1234 	if (vruntime_cmp(se->vruntime, "<", se->deadline))
1235 		return false;
1236 
1237 	/*
1238 	 * For EEVDF the virtual time slope is determined by w_i (iow.
1239 	 * nice) while the request time r_i is determined by
1240 	 * sysctl_sched_base_slice.
1241 	 */
1242 	if (!se->custom_slice)
1243 		se->slice = sysctl_sched_base_slice;
1244 
1245 	/*
1246 	 * EEVDF: vd_i = ve_i + r_i / w_i
1247 	 */
1248 	se->deadline = se->vruntime + calc_delta_fair(se->slice, se);
1249 	avg_vruntime(cfs_rq);
1250 
1251 	/*
1252 	 * The task has consumed its request, reschedule.
1253 	 */
1254 	return true;
1255 }
1256 
1257 #include "pelt.h"
1258 
1259 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
1260 static unsigned long task_h_load(struct task_struct *p);
1261 static unsigned long capacity_of(int cpu);
1262 
1263 /* Give new sched_entity start runnable values to heavy its load in infant time */
1264 void init_entity_runnable_average(struct sched_entity *se)
1265 {
1266 	struct sched_avg *sa = &se->avg;
1267 
1268 	memset(sa, 0, sizeof(*sa));
1269 
1270 	/*
1271 	 * Tasks are initialized with full load to be seen as heavy tasks until
1272 	 * they get a chance to stabilize to their real load level.
1273 	 * Group entities are initialized with zero load to reflect the fact that
1274 	 * nothing has been attached to the task group yet.
1275 	 */
1276 	if (entity_is_task(se))
1277 		sa->load_avg = scale_load_down(se->load.weight);
1278 
1279 	/* when this task is enqueued, it will contribute to its cfs_rq's load_avg */
1280 }
1281 
1282 /*
1283  * With new tasks being created, their initial util_avgs are extrapolated
1284  * based on the cfs_rq's current util_avg:
1285  *
1286  *   util_avg = cfs_rq->avg.util_avg / (cfs_rq->avg.load_avg + 1)
1287  *		* se_weight(se)
1288  *
1289  * However, in many cases, the above util_avg does not give a desired
1290  * value. Moreover, the sum of the util_avgs may be divergent, such
1291  * as when the series is a harmonic series.
1292  *
1293  * To solve this problem, we also cap the util_avg of successive tasks to
1294  * only 1/2 of the left utilization budget:
1295  *
1296  *   util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
1297  *
1298  * where n denotes the nth task and cpu_scale the CPU capacity.
1299  *
1300  * For example, for a CPU with 1024 of capacity, a simplest series from
1301  * the beginning would be like:
1302  *
1303  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
1304  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
1305  *
1306  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
1307  * if util_avg > util_avg_cap.
1308  */
1309 void post_init_entity_util_avg(struct task_struct *p)
1310 {
1311 	struct sched_entity *se = &p->se;
1312 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
1313 	struct sched_avg *sa = &se->avg;
1314 	long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
1315 	long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
1316 
1317 	if (p->sched_class != &fair_sched_class) {
1318 		/*
1319 		 * For !fair tasks do:
1320 		 *
1321 		update_cfs_rq_load_avg(now, cfs_rq);
1322 		attach_entity_load_avg(cfs_rq, se);
1323 		switched_from_fair(rq, p);
1324 		 *
1325 		 * such that the next switched_to_fair() has the
1326 		 * expected state.
1327 		 */
1328 		se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
1329 		return;
1330 	}
1331 
1332 	if (cap > 0) {
1333 		if (cfs_rq->avg.util_avg != 0) {
1334 			sa->util_avg  = cfs_rq->avg.util_avg * se_weight(se);
1335 			sa->util_avg /= (cfs_rq->avg.load_avg + 1);
1336 
1337 			if (sa->util_avg > cap)
1338 				sa->util_avg = cap;
1339 		} else {
1340 			sa->util_avg = cap;
1341 		}
1342 	}
1343 
1344 	sa->runnable_avg = sa->util_avg;
1345 }
1346 
1347 static inline void account_mm_sched(struct rq *rq, struct task_struct *p, s64 delta_exec);
1348 
1349 static s64 update_se(struct rq *rq, struct sched_entity *se)
1350 {
1351 	u64 now = rq_clock_task(rq);
1352 	s64 delta_exec;
1353 
1354 	delta_exec = now - se->exec_start;
1355 	if (unlikely(delta_exec <= 0))
1356 		return delta_exec;
1357 
1358 	se->exec_start = now;
1359 	if (entity_is_task(se)) {
1360 		struct task_struct *running = rq->curr;
1361 		/*
1362 		 * If se is a task, we account the time against the running
1363 		 * task, as w/ proxy-exec they may not be the same.
1364 		 */
1365 		running->se.exec_start = now;
1366 		running->se.sum_exec_runtime += delta_exec;
1367 
1368 		trace_sched_stat_runtime(running, delta_exec);
1369 		account_group_exec_runtime(running, delta_exec);
1370 		account_mm_sched(rq, running, delta_exec);
1371 
1372 		cgroup_account_cputime(running, delta_exec);
1373 	} else {
1374 		/* If not task, account the time against donor se  */
1375 		se->sum_exec_runtime += delta_exec;
1376 	}
1377 
1378 	if (schedstat_enabled()) {
1379 		struct sched_statistics *stats;
1380 
1381 		stats = __schedstats_from_se(se);
1382 		__schedstat_set(stats->exec_max,
1383 				max(delta_exec, stats->exec_max));
1384 	}
1385 
1386 	return delta_exec;
1387 }
1388 
1389 #ifdef CONFIG_SCHED_CACHE
1390 
1391 /*
1392  * XXX numbers come from a place the sun don't shine -- probably wants to be SD
1393  * tunable or so.
1394  */
1395 #define EPOCH_PERIOD	(HZ / 100)	/* 10 ms */
1396 #define EPOCH_LLC_AFFINITY_TIMEOUT	5	/* 50 ms */
1397 __read_mostly unsigned int llc_aggr_tolerance	= 1;
1398 __read_mostly unsigned int llc_epoch_period	= EPOCH_PERIOD;
1399 __read_mostly unsigned int llc_epoch_affinity_timeout = EPOCH_LLC_AFFINITY_TIMEOUT;
1400 __read_mostly unsigned int llc_imb_pct		= 20;
1401 __read_mostly unsigned int llc_overaggr_pct	= 50;
1402 
1403 static int llc_id(int cpu)
1404 {
1405 	if (cpu < 0)
1406 		return -1;
1407 
1408 	return per_cpu(sd_llc_id, cpu);
1409 }
1410 
1411 static inline int get_sched_cache_scale(int mul)
1412 {
1413 	unsigned int tol = READ_ONCE(llc_aggr_tolerance);
1414 
1415 	if (!tol)
1416 		return 0;
1417 
1418 	if (tol >= 100)
1419 		return INT_MAX;
1420 
1421 	return (1 + (tol - 1) * mul);
1422 }
1423 
1424 static bool exceed_llc_capacity(struct sched_cache_group *grp, int cpu)
1425 {
1426 #ifdef CONFIG_NUMA_BALANCING
1427 	unsigned long llc, footprint;
1428 	struct sched_domain *sd;
1429 	int scale;
1430 
1431 	guard(rcu)();
1432 
1433 	sd = rcu_dereference_sched_domain(cpu_rq(cpu)->sd);
1434 	if (!sd)
1435 		return true;
1436 
1437 	if (static_branch_likely(&sched_numa_balancing)) {
1438 		/*
1439 		 * TBD: RDT exclusive LLC ways reserved should be
1440 		 * excluded.
1441 		 */
1442 		llc = sd->llc_bytes;
1443 		footprint = READ_ONCE(grp->footprint);
1444 
1445 		/*
1446 		 * Scale the LLC size by 256*llc_aggr_tolerance
1447 		 * and compare it to the task's footprint.
1448 		 *
1449 		 * Suppose the L3 size is 32MB. If the
1450 		 * llc_aggr_tolerance is 1:
1451 		 * When the footprint is larger than 32MB, the
1452 		 * process is regarded as exceeding the LLC
1453 		 * capacity. If the llc_aggr_tolerance is 99:
1454 		 * When the footprint is larger than 784GB, the
1455 		 * process is regarded as exceeding the LLC
1456 		 * capacity:
1457 		 * 784GB = (1 + (99 - 1) * 256) * 32MB
1458 		 * If the llc_aggr_tolerance is 100:
1459 		 * ignore the footprint and do the aggregation
1460 		 * anyway.
1461 		 */
1462 		scale = get_sched_cache_scale(256);
1463 		if (scale == INT_MAX)
1464 			return false;
1465 
1466 		return ((llc * (u64)scale) < (footprint * PAGE_SIZE));
1467 	}
1468 #endif
1469 	return false;
1470 }
1471 
1472 /* Are the running threads (in SMT siblings) below ~80% of the LLC's CPUs? */
1473 static inline bool fits_llc_nr(u64 nr_threads, unsigned int nr_cpus)
1474 {
1475 	return nr_threads * 5 < (u64)nr_cpus * 4;
1476 }
1477 
1478 static bool invalid_llc_nr(struct sched_cache_group *grp, struct task_struct *p,
1479 			   int cpu)
1480 {
1481 	int scale;
1482 
1483 	if (get_nr_threads(p) <= 1)
1484 		return true;
1485 
1486 	/*
1487 	 * Scale the number of 'cores' in a LLC by llc_aggr_tolerance
1488 	 * and compare it to the task's active threads.
1489 	 */
1490 	scale = get_sched_cache_scale(1);
1491 	if (scale == INT_MAX)
1492 		return false;
1493 
1494 	return !fits_llc_nr(READ_ONCE(grp->nr_running_avg) * cpu_smt_num_threads,
1495 			    scale * per_cpu(sd_llc_size, cpu));
1496 }
1497 
1498 /*
1499  * A task counts in nr_pref_llc_running while it is queued on its preferred
1500  * LLC (pref_llc_queued) and runnable (!sched_delayed), keeping the counter in
1501  * the runnable domain so alb_break_llc() can compare it with h_nr_runnable.
1502  */
1503 static bool task_pref_llc_runnable(struct task_struct *p)
1504 {
1505 	return p->pref_llc_queued && !p->se.sched_delayed;
1506 }
1507 
1508 static void pref_llc_running_inc(struct rq *rq, struct task_struct *p)
1509 {
1510 	if (task_pref_llc_runnable(p))
1511 		rq->nr_pref_llc_running++;
1512 }
1513 
1514 static void pref_llc_running_dec(struct rq *rq, struct task_struct *p)
1515 {
1516 	if (task_pref_llc_runnable(p))
1517 		rq->nr_pref_llc_running--;
1518 }
1519 
1520 static void account_llc_enqueue(struct rq *rq, struct task_struct *p)
1521 {
1522 	int pref_llc, pref_llc_queued;
1523 	struct sched_domain *sd;
1524 
1525 	pref_llc = p->preferred_llc;
1526 	if (pref_llc < 0)
1527 		return;
1528 
1529 	pref_llc_queued = (pref_llc == task_llc(p));
1530 	rq->nr_llc_running++;
1531 
1532 	/*
1533 	 * Record whether p is enqueued on its preferred
1534 	 * LLC, in order to pair with account_llc_dequeue()
1535 	 * to maintain a consistent nr_pref_llc_running per
1536 	 * runqueue.
1537 	 * This is necessary because a race condition exists:
1538 	 * after a task is enqueued on a runqueue, task_llc(p)
1539 	 * may change due to CPU hotplug. Therefore, checking
1540 	 * task_llc(p) to determine whether the task is being
1541 	 * dequeued from its preferred LLC is unreliable and
1542 	 * can cause inconsistent values - checking the
1543 	 * p->pref_llc_queued in account_llc_dequeue() would
1544 	 * be reliable.
1545 	 */
1546 	p->pref_llc_queued = pref_llc_queued;
1547 
1548 	/* Skipped while delayed; clear_delayed() adds it back on wake. */
1549 	pref_llc_running_inc(rq, p);
1550 
1551 	sd = rcu_dereference_all(rq->sd);
1552 	if (sd && (unsigned int)pref_llc < sd->llc_max)
1553 		sd->llc_counts[pref_llc]++;
1554 }
1555 
1556 static void account_llc_dequeue(struct rq *rq, struct task_struct *p)
1557 {
1558 	struct sched_domain *sd;
1559 	int pref_llc;
1560 
1561 	pref_llc = p->preferred_llc;
1562 	if (pref_llc < 0)
1563 		return;
1564 
1565 	rq->nr_llc_running--;
1566 	if (p->pref_llc_queued) {
1567 		/*
1568 		 * Skipped if still delayed (set_delayed() already removed it);
1569 		 * clearing pref_llc_queued below also stops clear_delayed()
1570 		 * from re-adding it.
1571 		 */
1572 		pref_llc_running_dec(rq, p);
1573 		/*
1574 		 * Update the status in case
1575 		 * other logic might query
1576 		 * this.
1577 		 */
1578 		p->pref_llc_queued = 0;
1579 	}
1580 
1581 	sd = rcu_dereference_all(rq->sd);
1582 	if (sd && (unsigned int)pref_llc < sd->llc_max) {
1583 		/*
1584 		 * There is a race condition between dequeue
1585 		 * and CPU hotplug. After a task has been enqueued
1586 		 * on CPUx, a CPU hotplug event occurs, and all online
1587 		 * CPUs (including CPUx) rebuild their sched_domains
1588 		 * and reset statistics to zero(including sd->llc_counts).
1589 		 * This can cause temporary undercount and we have to
1590 		 * check for such underflow in sd->llc_counts.
1591 		 *
1592 		 * This undercount is temporary and accurate accounting
1593 		 * will resume once the rq has a chance to be idle.
1594 		 */
1595 		if (sd->llc_counts[pref_llc])
1596 			sd->llc_counts[pref_llc]--;
1597 	}
1598 }
1599 
1600 int mm_init_sched(struct mm_struct *mm,
1601 		  struct sched_cache_time __percpu *_pcpu_sched)
1602 {
1603 	struct sched_cache_group *grp;
1604 	unsigned long epoch = 0;
1605 	int i;
1606 
1607 	grp = kzalloc_obj(*grp);
1608 	if (!grp) {
1609 		free_percpu(_pcpu_sched);
1610 		mm->sched_cache_grp = NULL;
1611 		return -ENOMEM;
1612 	}
1613 
1614 	for_each_possible_cpu(i) {
1615 		struct sched_cache_time *pcpu_sched = per_cpu_ptr(_pcpu_sched, i);
1616 		struct rq *rq = cpu_rq(i);
1617 
1618 		pcpu_sched->runtime = 0;
1619 		/* a slightly stale cpu epoch is acceptible */
1620 		pcpu_sched->epoch = rq->cpu_epoch;
1621 		epoch = rq->cpu_epoch;
1622 	}
1623 
1624 	raw_spin_lock_init(&grp->lock);
1625 	grp->epoch = epoch;
1626 	grp->cpu = -1;
1627 	grp->next_scan = jiffies;
1628 	grp->nr_running_avg = 0;
1629 	grp->footprint = 0;
1630 	refcount_set(&grp->refcnt, 1);
1631 	/*
1632 	 * The update to grp->pcpu_sched should not be reordered
1633 	 * before initialization to grp's other fields, in case
1634 	 * the readers may get invalid mm_sched_epoch, etc.
1635 	 */
1636 	smp_store_release(&grp->pcpu_sched, _pcpu_sched);
1637 	/*
1638 	 * Publish the group last.  Not every reader qualifies it by
1639 	 * grp->pcpu_sched - can_migrate_llc_task() only checks that the
1640 	 * pointer is non-NULL before reading grp->footprint and
1641 	 * grp->nr_running_avg - so a reachable group must already be
1642 	 * fully initialized.
1643 	 */
1644 	smp_store_release(&mm->sched_cache_grp, grp);
1645 	return 0;
1646 }
1647 
1648 static void sched_cache_group_free_rcu(struct rcu_head *rcu)
1649 {
1650 	struct sched_cache_group *grp =
1651 		container_of(rcu, struct sched_cache_group, rcu);
1652 
1653 	free_percpu(grp->pcpu_sched);
1654 	kfree(grp);
1655 }
1656 
1657 static void sched_cache_group_put(struct sched_cache_group *grp)
1658 {
1659 	if (!grp || !refcount_dec_and_test(&grp->refcnt))
1660 		return;
1661 
1662 	call_rcu(&grp->rcu, sched_cache_group_free_rcu);
1663 }
1664 
1665 DEFINE_FREE(sched_cache_group_put, struct sched_cache_group *,
1666 	    sched_cache_group_put(_T));
1667 
1668 #define rcu_deref_sched_cache_grp(tsk) \
1669 	rcu_dereference_check((tsk)->sched_cache_grp, (tsk) == current)
1670 
1671 static struct sched_cache_group *sched_cache_replace_grp(struct task_struct *p,
1672 							 struct sched_cache_group *new)
1673 {
1674 	struct sched_cache_group *old;
1675 
1676 	old = rcu_deref_sched_cache_grp(p);
1677 	rcu_assign_pointer(p->sched_cache_grp, new);
1678 
1679 	return old;
1680 }
1681 
1682 struct sched_cache_group *sched_cache_group_get(struct sched_cache_group *grp)
1683 {
1684 	/*
1685 	 * refcount_inc_not_zero() is the acquire primitive for lockless
1686 	 * (RCU) lookups; plain refcount_inc() would scribble the count if
1687 	 * it already reached zero. Return NULL in that case.
1688 	 */
1689 	if (grp && !refcount_inc_not_zero(&grp->refcnt))
1690 		grp = NULL;
1691 
1692 	return grp;
1693 }
1694 
1695 struct sched_cache_group *task_cache_group_get(struct task_struct *p)
1696 {
1697 	guard(rcu)();
1698 	return sched_cache_group_get(rcu_dereference(p->sched_cache_grp));
1699 }
1700 
1701 void sched_cache_fork(struct task_struct *p)
1702 {
1703 	/*
1704 	 * The child takes its own reference on the mm's cache group, separate
1705 	 * from the reference held by the mm. @p is not yet visible to readers,
1706 	 * so a plain initializing store is enough.
1707 	 */
1708 	RCU_INIT_POINTER(p->sched_cache_grp,
1709 			 sched_cache_group_get(p->mm->sched_cache_grp));
1710 }
1711 
1712 void sched_cache_fork_cleanup(struct task_struct *p)
1713 {
1714 	/*
1715 	 * A fork that fails after sched_cache_fork() never reaches exit_mm(),
1716 	 * so drop the reference here. @p never became visible, so there are no
1717 	 * concurrent readers and the reference we hold keeps the group alive.
1718 	 */
1719 	sched_cache_group_put(rcu_access_pointer(p->sched_cache_grp));
1720 	RCU_INIT_POINTER(p->sched_cache_grp, NULL);
1721 }
1722 
1723 void sched_cache_exec_mmap(struct task_struct *p, struct mm_struct *mm)
1724 {
1725 	struct sched_cache_group *old;
1726 
1727 	/*
1728 	 * Acquire the new reference before publishing the pointer, then drop
1729 	 * the old one. @p is current and the only writer of its own pointer.
1730 	 */
1731 	old = sched_cache_replace_grp(p, sched_cache_group_get(mm->sched_cache_grp));
1732 	sched_cache_group_put(old);
1733 }
1734 
1735 void sched_cache_exit_mm(struct task_struct *p)
1736 {
1737 	struct sched_cache_group *grp = sched_cache_replace_grp(p, NULL);
1738 
1739 #ifdef CONFIG_NUMA_BALANCING
1740 	/*
1741 	 * Subtract this task's footprint from the group before dropping the
1742 	 * reference, so the group footprint converges as its threads exit.
1743 	 * Unlocked for performance; clamp to avoid underflow.
1744 	 */
1745 	if (grp && p->total_numa_faults) {
1746 		unsigned long fp = READ_ONCE(grp->footprint);
1747 		unsigned long sub = min(fp, p->total_numa_faults);
1748 
1749 		WRITE_ONCE(grp->footprint, fp - sub);
1750 	}
1751 #endif
1752 	sched_cache_group_put(grp);
1753 }
1754 
1755 void mm_destroy_sched(struct mm_struct *mm)
1756 {
1757 	sched_cache_group_put(mm->sched_cache_grp);
1758 	mm->sched_cache_grp = NULL;
1759 }
1760 
1761 /* because why would C be fully specified */
1762 static __always_inline void __shr_u64(u64 *val, unsigned int n)
1763 {
1764 	if (n >= 64) {
1765 		*val = 0;
1766 		return;
1767 	}
1768 	*val >>= n;
1769 }
1770 
1771 static inline void __update_mm_sched(struct rq *rq,
1772 				     struct sched_cache_time *pcpu_sched)
1773 {
1774 	lockdep_assert_held(&rq->cpu_epoch_lock);
1775 
1776 	unsigned int period = max(READ_ONCE(llc_epoch_period), 1U);
1777 	unsigned long n, now = jiffies;
1778 	long delta = now - rq->cpu_epoch_next;
1779 
1780 	if (delta > 0) {
1781 		n = (delta + period - 1) / period;
1782 		rq->cpu_epoch += n;
1783 		rq->cpu_epoch_next += n * period;
1784 		__shr_u64(&rq->cpu_runtime, n);
1785 	}
1786 
1787 	n = rq->cpu_epoch - pcpu_sched->epoch;
1788 	if (n) {
1789 		pcpu_sched->epoch += n;
1790 		__shr_u64(&pcpu_sched->runtime, n);
1791 	}
1792 }
1793 
1794 static unsigned long fraction_mm_sched(struct rq *rq,
1795 				       struct sched_cache_time *pcpu_sched)
1796 {
1797 	guard(raw_spinlock_irqsave)(&rq->cpu_epoch_lock);
1798 
1799 	__update_mm_sched(rq, pcpu_sched);
1800 
1801 	/*
1802 	 * Runtime is a geometric series (r=0.5) and as such will sum to twice
1803 	 * the accumulation period, this means the multiplcation here should
1804 	 * not overflow.
1805 	 */
1806 	return div64_u64(NICE_0_LOAD * pcpu_sched->runtime, rq->cpu_runtime + 1);
1807 }
1808 
1809 static int get_pref_llc(struct task_struct *p, struct sched_cache_group *grp)
1810 {
1811 	int mm_sched_llc = -1, mm_sched_cpu;
1812 
1813 	if (!grp)
1814 		return -1;
1815 
1816 	mm_sched_cpu = READ_ONCE(grp->cpu);
1817 	if (mm_sched_cpu != -1) {
1818 		mm_sched_llc = llc_id(mm_sched_cpu);
1819 
1820 #ifdef CONFIG_NUMA_BALANCING
1821 		/*
1822 		 * Don't assign preferred LLC if it
1823 		 * conflicts with NUMA balancing.
1824 		 * This can happen when sched_setnuma() gets
1825 		 * called, however it is not much of an issue
1826 		 * because we expect account_mm_sched() to get
1827 		 * called fairly regularly -- at a higher rate
1828 		 * than sched_setnuma() at least -- and thus the
1829 		 * conflict only exists for a short period of time.
1830 		 */
1831 		if (static_branch_likely(&sched_numa_balancing) &&
1832 		    p->numa_preferred_nid >= 0 &&
1833 		    cpu_to_node(mm_sched_cpu) != p->numa_preferred_nid)
1834 			mm_sched_llc = -1;
1835 #endif
1836 	}
1837 
1838 	return mm_sched_llc;
1839 }
1840 
1841 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p);
1842 
1843 static inline
1844 void account_mm_sched(struct rq *rq, struct task_struct *p, s64 delta_exec)
1845 {
1846 	struct sched_cache_group *grp = rcu_dereference_all(p->sched_cache_grp);
1847 	struct sched_cache_time *pcpu_sched;
1848 	int mm_sched_llc = -1;
1849 	unsigned long epoch;
1850 
1851 	if (!sched_cache_enabled())
1852 		return;
1853 
1854 	if (p->sched_class != &fair_sched_class)
1855 		return;
1856 	/*
1857 	 * init_task, kthreads and user thread created
1858 	 * by user_mode_thread() don't have a cache group.
1859 	 * In theory a kernel thread does not have any valid
1860 	 * cache group, because sched_cache_fork() is not
1861 	 * invoked for a kernel thread - !grp should gate the
1862 	 * kernel thread. Use the PF_KTHREAD check explicitly
1863 	 * here for safety reasons, to guard against future
1864 	 * modifications and to pair with task_tick_cache().
1865 	 */
1866 	if (p->flags & PF_KTHREAD || !grp || !grp->pcpu_sched)
1867 		return;
1868 
1869 	pcpu_sched = per_cpu_ptr(grp->pcpu_sched, cpu_of(rq));
1870 
1871 	scoped_guard (raw_spinlock, &rq->cpu_epoch_lock) {
1872 		__update_mm_sched(rq, pcpu_sched);
1873 		pcpu_sched->runtime += delta_exec;
1874 		rq->cpu_runtime += delta_exec;
1875 		epoch = rq->cpu_epoch;
1876 	}
1877 
1878 	/*
1879 	 * If this process hasn't hit task_cache_work() for a while invalidate
1880 	 * its preferred state.
1881 	 */
1882 	if ((long)(epoch - READ_ONCE(grp->epoch)) > llc_epoch_affinity_timeout ||
1883 	    invalid_llc_nr(grp, p, cpu_of(rq)) ||
1884 	    exceed_llc_capacity(grp, cpu_of(rq))) {
1885 		if (READ_ONCE(grp->cpu) != -1)
1886 			WRITE_ONCE(grp->cpu, -1);
1887 	}
1888 
1889 	mm_sched_llc = get_pref_llc(p, grp);
1890 
1891 	/* task not on rq accounted later in account_entity_enqueue() */
1892 	if (task_running_on_cpu(rq->cpu, p) &&
1893 	    READ_ONCE(p->preferred_llc) != mm_sched_llc) {
1894 		account_llc_dequeue(rq, p);
1895 		WRITE_ONCE(p->preferred_llc, mm_sched_llc);
1896 		account_llc_enqueue(rq, p);
1897 	}
1898 }
1899 
1900 static void task_tick_cache(struct rq *rq, struct task_struct *p)
1901 {
1902 	struct sched_cache_group *grp = rcu_dereference_all(p->sched_cache_grp);
1903 	struct callback_head *work = &p->cache_work;
1904 	unsigned long epoch;
1905 
1906 	if (!sched_cache_enabled())
1907 		return;
1908 
1909 	if (!grp || p->flags & PF_KTHREAD ||
1910 	    !grp->pcpu_sched)
1911 		return;
1912 
1913 	epoch = rq->cpu_epoch;
1914 	/* avoid moving backwards */
1915 	if (time_after_eq(grp->epoch, epoch))
1916 		return;
1917 
1918 	guard(raw_spinlock)(&grp->lock);
1919 
1920 	if (work->next == work) {
1921 		task_work_add(p, work, TWA_RESUME);
1922 		WRITE_ONCE(grp->epoch, epoch);
1923 	}
1924 }
1925 
1926 static void get_scan_cpumasks(cpumask_var_t cpus, struct task_struct *p,
1927 			      struct sched_cache_group *grp)
1928 {
1929 #ifdef CONFIG_NUMA_BALANCING
1930 	int cpu, curr_cpu, nid, pref_nid;
1931 
1932 	if (!static_branch_likely(&sched_numa_balancing))
1933 		goto out;
1934 
1935 	cpu = READ_ONCE(grp->cpu);
1936 	if (cpu != -1)
1937 		nid = cpu_to_node(cpu);
1938 	curr_cpu = task_cpu(p);
1939 
1940 	/*
1941 	 * Scanning in the preferred NUMA node is ideal. However, the NUMA
1942 	 * preferred node is per-task rather than per-process. It is possible
1943 	 * for different threads of the process to have distinct preferred
1944 	 * nodes; consequently, the process-wide preferred LLC may bounce
1945 	 * between different nodes. As a workaround, maintain the scan
1946 	 * CPU mask to also cover the process's current preferred LLC and the
1947 	 * current running node to mitigate the bouncing risk.
1948 	 * TBD: numa_group should be considered during task aggregation.
1949 	 */
1950 	pref_nid = p->numa_preferred_nid;
1951 	/* honor the task's preferred node */
1952 	if (pref_nid == NUMA_NO_NODE)
1953 		goto out;
1954 
1955 	cpumask_or(cpus, cpus, cpumask_of_node(pref_nid));
1956 
1957 	/* honor the task's preferred LLC CPU */
1958 	if (cpu != -1 && !cpumask_test_cpu(cpu, cpus) && nid != NUMA_NO_NODE)
1959 		cpumask_or(cpus, cpus, cpumask_of_node(nid));
1960 
1961 	/* make sure the task's current running node is included */
1962 	if (!cpumask_test_cpu(curr_cpu, cpus))
1963 		cpumask_or(cpus, cpus, cpumask_of_node(cpu_to_node(curr_cpu)));
1964 
1965 	return;
1966 
1967 out:
1968 #endif
1969 	cpumask_copy(cpus, cpu_online_mask);
1970 }
1971 
1972 static inline void update_avg_scale(u64 *avg, u64 sample)
1973 {
1974 	int factor = per_cpu(sd_llc_size, raw_smp_processor_id());
1975 	s64 diff = sample - *avg;
1976 	u32 divisor;
1977 
1978 	/*
1979 	 * Scale the divisor based on the number of CPUs contained
1980 	 * in the LLC. This scaling ensures smaller LLC domains use
1981 	 * a smaller divisor to achieve more precise sensitivity to
1982 	 * changes in nr_running, while larger LLC domains are capped
1983 	 * at a maximum divisor of 8 which is the default smoothing
1984 	 * factor of EWMA in update_avg().
1985 	 */
1986 	divisor = clamp_t(u32, (factor >> 2), 2, 8);
1987 	*avg += div64_s64(diff, divisor);
1988 }
1989 
1990 static void task_cache_work(struct callback_head *work)
1991 {
1992 	struct sched_cache_group *grp __free(sched_cache_group_put) = NULL;
1993 	cpumask_var_t cpus __free(free_cpumask_var) = CPUMASK_VAR_NULL;
1994 	int cpu, m_a_cpu = -1, nr_running = 0, curr_cpu;
1995 	unsigned long next_scan, now = jiffies;
1996 	struct task_struct *p = current, *cur;
1997 	unsigned long curr_m_a_occ = 0;
1998 	unsigned long m_a_occ = 0;
1999 
2000 	WARN_ON_ONCE(work != &p->cache_work);
2001 
2002 	work->next = work;
2003 
2004 	if (p->flags & PF_EXITING)
2005 		return;
2006 
2007 	/*
2008 	 * A reference makes sure grp is not released by others. The rcu
2009 	 * lock can not be held till after zalloc_cpumask_var() below,
2010 	 * because the latter might sleep.
2011 	 */
2012 	grp = task_cache_group_get(p);
2013 	if (!grp)
2014 		return;
2015 
2016 	next_scan = READ_ONCE(grp->next_scan);
2017 	if (time_before(now, next_scan))
2018 		return;
2019 
2020 	/* only 1 thread is allowed to scan */
2021 	if (!try_cmpxchg(&grp->next_scan, &next_scan,
2022 			 now + max_t(unsigned long,
2023 				     READ_ONCE(llc_epoch_period), 1)))
2024 		return;
2025 
2026 	curr_cpu = task_cpu(p);
2027 	if (invalid_llc_nr(grp, p, curr_cpu) ||
2028 	    exceed_llc_capacity(grp, curr_cpu)) {
2029 		if (READ_ONCE(grp->cpu) != -1)
2030 			WRITE_ONCE(grp->cpu, -1);
2031 
2032 		return;
2033 	}
2034 
2035 	if (!zalloc_cpumask_var(&cpus, GFP_KERNEL))
2036 		return;
2037 
2038 	scoped_guard (cpus_read_lock) {
2039 		guard(rcu)();
2040 
2041 		get_scan_cpumasks(cpus, p, grp);
2042 
2043 		for_each_cpu(cpu, cpus) {
2044 			/* XXX sched_cluster_active */
2045 			struct sched_domain *sd = rcu_dereference_all(per_cpu(sd_llc, cpu));
2046 			unsigned long occ, m_occ = 0, a_occ = 0;
2047 			int m_cpu = -1, i;
2048 
2049 			if (!sd)
2050 				continue;
2051 
2052 			for_each_cpu(i, sched_domain_span(sd)) {
2053 				occ = fraction_mm_sched(cpu_rq(i),
2054 							per_cpu_ptr(grp->pcpu_sched, i));
2055 				a_occ += occ;
2056 				if (occ > m_occ) {
2057 					m_occ = occ;
2058 					m_cpu = i;
2059 				}
2060 
2061 				/*
2062 				 * rcu_access_pointer() is used because the
2063 				 * pointer is only compared, never dereferenced.
2064 				 */
2065 				cur = rcu_dereference_all(cpu_rq(i)->curr);
2066 				if (cur && !(cur->flags & (PF_EXITING | PF_KTHREAD)) &&
2067 				    rcu_access_pointer(cur->sched_cache_grp) == grp)
2068 					nr_running++;
2069 			}
2070 
2071 			/*
2072 			 * Compare the accumulated occupancy of each LLC. The
2073 			 * reason for using accumulated occupancy rather than average
2074 			 * per CPU occupancy is that it works better in asymmetric LLC
2075 			 * scenarios.
2076 			 * For example, if there are 2 threads in a 4CPU LLC and 3
2077 			 * threads in an 8CPU LLC, it might be better to choose the one
2078 			 * with 3 threads. However, this would not be the case if the
2079 			 * occupancy is divided by the number of CPUs in an LLC (i.e.,
2080 			 * if average per CPU occupancy is used).
2081 			 * Besides, NUMA balancing fault statistics behave similarly:
2082 			 * the total number of faults per node is compared rather than
2083 			 * the average number of faults per CPU. This strategy is also
2084 			 * followed here.
2085 			 */
2086 			if (a_occ > m_a_occ) {
2087 				m_a_occ = a_occ;
2088 				m_a_cpu = m_cpu;
2089 			}
2090 
2091 			if (llc_id(cpu) == llc_id(READ_ONCE(grp->cpu)))
2092 				curr_m_a_occ = a_occ;
2093 
2094 			cpumask_andnot(cpus, cpus, sched_domain_span(sd));
2095 		}
2096 	}
2097 
2098 	if (m_a_occ > (2 * curr_m_a_occ)) {
2099 		/*
2100 		 * Avoid switching sched_cache_grp->cpu too fast.
2101 		 * The reason to choose 2X is because:
2102 		 * 1. It is better to keep the preferred LLC stable,
2103 		 *    rather than changing it frequently and cause migrations
2104 		 * 2. 2X means the new preferred LLC has at least 1 more
2105 		 *    busy CPU than the old one(200% vs 100%, eg)
2106 		 * 3. 2X is chosen based on test results, as it delivers
2107 		 *    the optimal performance gain so far.
2108 		 */
2109 		WRITE_ONCE(grp->cpu, m_a_cpu);
2110 	}
2111 
2112 	update_avg_scale(&grp->nr_running_avg, nr_running);
2113 }
2114 
2115 void init_sched_mm(struct task_struct *p)
2116 {
2117 	struct callback_head *work = &p->cache_work;
2118 
2119 	init_task_work(work, task_cache_work);
2120 	work->next = work;
2121 	/*
2122 	 * dup_task_struct() copies the parent's task_struct, including its
2123 	 * sched_cache_grp, for which the child holds no reference.  Clear it
2124 	 * here - before copy_mm() runs - so the child never carries a
2125 	 * borrowed pointer that the fork error path would put.
2126 	 */
2127 	RCU_INIT_POINTER(p->sched_cache_grp, NULL);
2128 	/*
2129 	 * Reset new task's preference to avoid
2130 	 * polluting account_llc_enqueue().
2131 	 */
2132 	p->preferred_llc = -1;
2133 	p->pref_llc_queued = 0;
2134 }
2135 
2136 #else /* CONFIG_SCHED_CACHE */
2137 
2138 static inline void account_mm_sched(struct rq *rq, struct task_struct *p,
2139 				    s64 delta_exec) { }
2140 
2141 void init_sched_mm(struct task_struct *p) { }
2142 
2143 static void task_tick_cache(struct rq *rq, struct task_struct *p) { }
2144 
2145 static inline int get_pref_llc(struct task_struct *p,
2146 			       struct sched_cache_group *grp)
2147 {
2148 	return -1;
2149 }
2150 
2151 static void account_llc_enqueue(struct rq *rq, struct task_struct *p) {}
2152 
2153 static void account_llc_dequeue(struct rq *rq, struct task_struct *p) {}
2154 
2155 static void pref_llc_running_inc(struct rq *rq, struct task_struct *p) {}
2156 
2157 static void pref_llc_running_dec(struct rq *rq, struct task_struct *p) {}
2158 
2159 #endif /* CONFIG_SCHED_CACHE */
2160 
2161 /*
2162  * Used by other classes to account runtime.
2163  */
2164 s64 update_curr_common(struct rq *rq)
2165 {
2166 	return update_se(rq, &rq->donor->se);
2167 }
2168 
2169 /*
2170  * Update the current task's runtime statistics.
2171  */
2172 static void update_curr(struct cfs_rq *cfs_rq)
2173 {
2174 	/*
2175 	 * Note: cfs_rq->curr corresponds to the task picked to
2176 	 * run (ie: rq->donor.se) which due to proxy-exec may
2177 	 * not necessarily be the actual task running
2178 	 * (rq->curr.se). This is easy to confuse!
2179 	 */
2180 	struct sched_entity *curr = cfs_rq->h_curr;
2181 	struct rq *rq = rq_of(cfs_rq);
2182 	s64 delta_exec;
2183 	bool resched;
2184 
2185 	if (unlikely(!curr))
2186 		return;
2187 
2188 	delta_exec = update_se(rq, curr);
2189 	if (unlikely(delta_exec <= 0))
2190 		return;
2191 
2192 	account_cfs_rq_runtime(cfs_rq, delta_exec);
2193 
2194 	if (!entity_is_task(curr))
2195 		return;
2196 
2197 	cfs_rq = &rq->cfs;
2198 
2199 	curr->vruntime += calc_delta_fair(delta_exec, curr);
2200 	resched = update_deadline(cfs_rq, curr);
2201 
2202 	/*
2203 	 * If the fair_server is active, we need to account for the
2204 	 * fair_server time whether or not the task is running on
2205 	 * behalf of fair_server or not:
2206 	 *  - If the task is running on behalf of fair_server, we need
2207 	 *    to limit its time based on the assigned runtime.
2208 	 *  - Fair task that runs outside of fair_server should account
2209 	 *    against fair_server such that it can account for this time
2210 	 *    and possibly avoid running this period.
2211 	 */
2212 	dl_server_update(&rq->fair_server, delta_exec);
2213 
2214 	if (cfs_rq->h_nr_queued == 1)
2215 		return;
2216 
2217 	if (resched || !protect_slice(curr)) {
2218 		resched_curr_lazy(rq);
2219 		clear_buddies(cfs_rq, curr);
2220 	}
2221 }
2222 
2223 static void update_curr_fair(struct rq *rq)
2224 {
2225 	struct sched_entity *se = &rq->donor->se;
2226 
2227 	for_each_sched_entity(se)
2228 		update_curr(cfs_rq_of(se));
2229 }
2230 
2231 static inline void
2232 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
2233 {
2234 	struct sched_statistics *stats;
2235 	struct task_struct *p = NULL;
2236 
2237 	if (!schedstat_enabled())
2238 		return;
2239 
2240 	stats = __schedstats_from_se(se);
2241 
2242 	if (entity_is_task(se))
2243 		p = task_of(se);
2244 
2245 	__update_stats_wait_start(rq_of(cfs_rq), p, stats);
2246 }
2247 
2248 static inline void
2249 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
2250 {
2251 	struct sched_statistics *stats;
2252 	struct task_struct *p = NULL;
2253 
2254 	if (!schedstat_enabled())
2255 		return;
2256 
2257 	stats = __schedstats_from_se(se);
2258 
2259 	/*
2260 	 * When the sched_schedstat changes from 0 to 1, some sched se
2261 	 * maybe already in the runqueue, the se->statistics.wait_start
2262 	 * will be 0.So it will let the delta wrong. We need to avoid this
2263 	 * scenario.
2264 	 */
2265 	if (unlikely(!schedstat_val(stats->wait_start)))
2266 		return;
2267 
2268 	if (entity_is_task(se))
2269 		p = task_of(se);
2270 
2271 	__update_stats_wait_end(rq_of(cfs_rq), p, stats);
2272 }
2273 
2274 static inline void
2275 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
2276 {
2277 	struct sched_statistics *stats;
2278 	struct task_struct *tsk = NULL;
2279 
2280 	if (!schedstat_enabled())
2281 		return;
2282 
2283 	stats = __schedstats_from_se(se);
2284 
2285 	if (entity_is_task(se))
2286 		tsk = task_of(se);
2287 
2288 	__update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
2289 }
2290 
2291 /*
2292  * Task is being enqueued - update stats:
2293  */
2294 static inline void
2295 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2296 {
2297 	if (!schedstat_enabled())
2298 		return;
2299 
2300 	/*
2301 	 * Are we enqueueing a waiting task? (for current tasks
2302 	 * a dequeue/enqueue event is a NOP)
2303 	 */
2304 	if (se != cfs_rq->h_curr)
2305 		update_stats_wait_start_fair(cfs_rq, se);
2306 
2307 	if (flags & ENQUEUE_WAKEUP)
2308 		update_stats_enqueue_sleeper_fair(cfs_rq, se);
2309 }
2310 
2311 static inline void
2312 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2313 {
2314 
2315 	if (!schedstat_enabled())
2316 		return;
2317 
2318 	/*
2319 	 * Mark the end of the wait period if dequeueing a
2320 	 * waiting task:
2321 	 */
2322 	if (se != cfs_rq->h_curr)
2323 		update_stats_wait_end_fair(cfs_rq, se);
2324 
2325 	if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
2326 		struct task_struct *tsk = task_of(se);
2327 		unsigned int state;
2328 
2329 		/* XXX racy against TTWU */
2330 		state = READ_ONCE(tsk->__state);
2331 		if (state & TASK_INTERRUPTIBLE)
2332 			__schedstat_set(tsk->stats.sleep_start,
2333 				      rq_clock(rq_of(cfs_rq)));
2334 		if (state & TASK_UNINTERRUPTIBLE)
2335 			__schedstat_set(tsk->stats.block_start,
2336 				      rq_clock(rq_of(cfs_rq)));
2337 	}
2338 }
2339 
2340 /*
2341  * We are picking a new current task - update its stats:
2342  */
2343 static inline void
2344 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
2345 {
2346 	/*
2347 	 * We are starting a new run period:
2348 	 */
2349 	se->exec_start = rq_clock_task(rq_of(cfs_rq));
2350 }
2351 
2352 /* Check sched_smt_active before calling this to avoid overheads in fastpaths */
2353 static inline bool is_core_idle(int cpu)
2354 {
2355 	int sibling;
2356 
2357 	for_each_cpu(sibling, cpu_smt_mask(cpu)) {
2358 		if (cpu == sibling)
2359 			continue;
2360 
2361 		if (!idle_cpu(sibling))
2362 			return false;
2363 	}
2364 
2365 	return true;
2366 }
2367 
2368 #ifdef CONFIG_NUMA
2369 #define NUMA_IMBALANCE_MIN 2
2370 
2371 static inline long
2372 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
2373 {
2374 	/*
2375 	 * Allow a NUMA imbalance if busy CPUs is less than the maximum
2376 	 * threshold. Above this threshold, individual tasks may be contending
2377 	 * for both memory bandwidth and any shared HT resources.  This is an
2378 	 * approximation as the number of running tasks may not be related to
2379 	 * the number of busy CPUs due to sched_setaffinity.
2380 	 */
2381 	if (dst_running > imb_numa_nr)
2382 		return imbalance;
2383 
2384 	/*
2385 	 * Allow a small imbalance based on a simple pair of communicating
2386 	 * tasks that remain local when the destination is lightly loaded.
2387 	 */
2388 	if (imbalance <= NUMA_IMBALANCE_MIN)
2389 		return 0;
2390 
2391 	return imbalance;
2392 }
2393 #endif /* CONFIG_NUMA */
2394 
2395 #ifdef CONFIG_NUMA_BALANCING
2396 /*
2397  * Approximate time to scan a full NUMA task in ms. The task scan period is
2398  * calculated based on the tasks virtual memory size and
2399  * numa_balancing_scan_size.
2400  */
2401 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
2402 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
2403 
2404 /* Portion of address space to scan in MB */
2405 unsigned int sysctl_numa_balancing_scan_size = 256;
2406 
2407 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
2408 unsigned int sysctl_numa_balancing_scan_delay = 1000;
2409 
2410 /* The page with hint page fault latency < threshold in ms is considered hot */
2411 unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
2412 
2413 struct numa_group {
2414 	refcount_t refcount;
2415 
2416 	spinlock_t lock; /* nr_tasks, tasks */
2417 	int nr_tasks;
2418 	pid_t gid;
2419 	int active_nodes;
2420 
2421 	struct rcu_head rcu;
2422 	unsigned long total_faults;
2423 	unsigned long max_faults_cpu;
2424 	/*
2425 	 * faults[] array is split into two regions: faults_mem and faults_cpu.
2426 	 *
2427 	 * Faults_cpu is used to decide whether memory should move
2428 	 * towards the CPU. As a consequence, these stats are weighted
2429 	 * more by CPU use than by memory faults.
2430 	 */
2431 	unsigned long faults[];
2432 };
2433 
2434 /*
2435  * For functions that can be called in multiple contexts that permit reading
2436  * ->numa_group (see struct task_struct for locking rules).
2437  */
2438 static struct numa_group *deref_task_numa_group(struct task_struct *p)
2439 {
2440 	return rcu_dereference_check(p->numa_group, p == current ||
2441 		(lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
2442 }
2443 
2444 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
2445 {
2446 	return rcu_dereference_protected(p->numa_group, p == current);
2447 }
2448 
2449 static inline unsigned long group_faults_priv(struct numa_group *ng);
2450 static inline unsigned long group_faults_shared(struct numa_group *ng);
2451 
2452 static unsigned int task_nr_scan_windows(struct task_struct *p)
2453 {
2454 	unsigned long rss = 0;
2455 	unsigned long nr_scan_pages;
2456 
2457 	/*
2458 	 * Calculations based on RSS as non-present and empty pages are skipped
2459 	 * by the PTE scanner and NUMA hinting faults should be trapped based
2460 	 * on resident pages
2461 	 */
2462 	nr_scan_pages = MB_TO_PAGES(sysctl_numa_balancing_scan_size);
2463 	rss = get_mm_rss(p->mm);
2464 	if (!rss)
2465 		rss = nr_scan_pages;
2466 
2467 	rss = round_up(rss, nr_scan_pages);
2468 	return rss / nr_scan_pages;
2469 }
2470 
2471 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
2472 #define MAX_SCAN_WINDOW 2560
2473 
2474 static unsigned int task_scan_min(struct task_struct *p)
2475 {
2476 	unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
2477 	unsigned int scan, floor;
2478 	unsigned int windows = 1;
2479 
2480 	if (scan_size < MAX_SCAN_WINDOW)
2481 		windows = MAX_SCAN_WINDOW / scan_size;
2482 	floor = 1000 / windows;
2483 
2484 	scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
2485 	return max_t(unsigned int, floor, scan);
2486 }
2487 
2488 static unsigned int task_scan_start(struct task_struct *p)
2489 {
2490 	unsigned long smin = task_scan_min(p);
2491 	unsigned long period = smin;
2492 	struct numa_group *ng;
2493 
2494 	/* Scale the maximum scan period with the amount of shared memory. */
2495 	rcu_read_lock();
2496 	ng = rcu_dereference_all(p->numa_group);
2497 	if (ng) {
2498 		unsigned long shared = group_faults_shared(ng);
2499 		unsigned long private = group_faults_priv(ng);
2500 
2501 		period *= refcount_read(&ng->refcount);
2502 		period *= shared + 1;
2503 		period /= private + shared + 1;
2504 	}
2505 	rcu_read_unlock();
2506 
2507 	return max(smin, period);
2508 }
2509 
2510 static unsigned int task_scan_max(struct task_struct *p)
2511 {
2512 	unsigned long smin = task_scan_min(p);
2513 	unsigned long smax;
2514 	struct numa_group *ng;
2515 
2516 	/* Watch for min being lower than max due to floor calculations */
2517 	smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
2518 
2519 	/* Scale the maximum scan period with the amount of shared memory. */
2520 	ng = deref_curr_numa_group(p);
2521 	if (ng) {
2522 		unsigned long shared = group_faults_shared(ng);
2523 		unsigned long private = group_faults_priv(ng);
2524 		unsigned long period = smax;
2525 
2526 		period *= refcount_read(&ng->refcount);
2527 		period *= shared + 1;
2528 		period /= private + shared + 1;
2529 
2530 		smax = max(smax, period);
2531 	}
2532 
2533 	return max(smin, smax);
2534 }
2535 
2536 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2537 {
2538 	rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
2539 	rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
2540 }
2541 
2542 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2543 {
2544 	rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
2545 	rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
2546 }
2547 
2548 /* Shared or private faults. */
2549 #define NR_NUMA_HINT_FAULT_TYPES 2
2550 
2551 /* Memory and CPU locality */
2552 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
2553 
2554 /* Averaged statistics, and temporary buffers. */
2555 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
2556 
2557 pid_t task_numa_group_id(struct task_struct *p)
2558 {
2559 	struct numa_group *ng;
2560 	pid_t gid = 0;
2561 
2562 	rcu_read_lock();
2563 	ng = rcu_dereference_all(p->numa_group);
2564 	if (ng)
2565 		gid = ng->gid;
2566 	rcu_read_unlock();
2567 
2568 	return gid;
2569 }
2570 
2571 /*
2572  * The averaged statistics, shared & private, memory & CPU,
2573  * occupy the first half of the array. The second half of the
2574  * array is for current counters, which are averaged into the
2575  * first set by task_numa_placement.
2576  */
2577 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
2578 {
2579 	return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
2580 }
2581 
2582 static inline unsigned long task_faults(struct task_struct *p, int nid)
2583 {
2584 	if (!p->numa_faults)
2585 		return 0;
2586 
2587 	return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
2588 		p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
2589 }
2590 
2591 static inline unsigned long group_faults(struct task_struct *p, int nid)
2592 {
2593 	struct numa_group *ng = deref_task_numa_group(p);
2594 
2595 	if (!ng)
2596 		return 0;
2597 
2598 	return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
2599 		ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
2600 }
2601 
2602 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
2603 {
2604 	return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
2605 		group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
2606 }
2607 
2608 static inline unsigned long group_faults_priv(struct numa_group *ng)
2609 {
2610 	unsigned long faults = 0;
2611 	int node;
2612 
2613 	for_each_online_node(node) {
2614 		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
2615 	}
2616 
2617 	return faults;
2618 }
2619 
2620 static inline unsigned long group_faults_shared(struct numa_group *ng)
2621 {
2622 	unsigned long faults = 0;
2623 	int node;
2624 
2625 	for_each_online_node(node) {
2626 		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
2627 	}
2628 
2629 	return faults;
2630 }
2631 
2632 /*
2633  * A node triggering more than 1/3 as many NUMA faults as the maximum is
2634  * considered part of a numa group's pseudo-interleaving set. Migrations
2635  * between these nodes are slowed down, to allow things to settle down.
2636  */
2637 #define ACTIVE_NODE_FRACTION 3
2638 
2639 static bool numa_is_active_node(int nid, struct numa_group *ng)
2640 {
2641 	return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
2642 }
2643 
2644 /* Handle placement on systems where not all nodes are directly connected. */
2645 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
2646 					int lim_dist, bool task)
2647 {
2648 	unsigned long score = 0;
2649 	int node, max_dist;
2650 
2651 	/*
2652 	 * All nodes are directly connected, and the same distance
2653 	 * from each other. No need for fancy placement algorithms.
2654 	 */
2655 	if (sched_numa_topology_type == NUMA_DIRECT)
2656 		return 0;
2657 
2658 	/* sched_max_numa_distance may be changed in parallel. */
2659 	max_dist = READ_ONCE(sched_max_numa_distance);
2660 	/*
2661 	 * This code is called for each node, introducing N^2 complexity,
2662 	 * which should be OK given the number of nodes rarely exceeds 8.
2663 	 */
2664 	for_each_online_node(node) {
2665 		unsigned long faults;
2666 		int dist = node_distance(nid, node);
2667 
2668 		/*
2669 		 * The furthest away nodes in the system are not interesting
2670 		 * for placement; nid was already counted.
2671 		 */
2672 		if (dist >= max_dist || node == nid)
2673 			continue;
2674 
2675 		/*
2676 		 * On systems with a backplane NUMA topology, compare groups
2677 		 * of nodes, and move tasks towards the group with the most
2678 		 * memory accesses. When comparing two nodes at distance
2679 		 * "hoplimit", only nodes closer by than "hoplimit" are part
2680 		 * of each group. Skip other nodes.
2681 		 */
2682 		if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
2683 			continue;
2684 
2685 		/* Add up the faults from nearby nodes. */
2686 		if (task)
2687 			faults = task_faults(p, node);
2688 		else
2689 			faults = group_faults(p, node);
2690 
2691 		/*
2692 		 * On systems with a glueless mesh NUMA topology, there are
2693 		 * no fixed "groups of nodes". Instead, nodes that are not
2694 		 * directly connected bounce traffic through intermediate
2695 		 * nodes; a numa_group can occupy any set of nodes.
2696 		 * The further away a node is, the less the faults count.
2697 		 * This seems to result in good task placement.
2698 		 */
2699 		if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2700 			faults *= (max_dist - dist);
2701 			faults /= (max_dist - LOCAL_DISTANCE);
2702 		}
2703 
2704 		score += faults;
2705 	}
2706 
2707 	return score;
2708 }
2709 
2710 /*
2711  * These return the fraction of accesses done by a particular task, or
2712  * task group, on a particular numa node.  The group weight is given a
2713  * larger multiplier, in order to group tasks together that are almost
2714  * evenly spread out between numa nodes.
2715  */
2716 static inline unsigned long task_weight(struct task_struct *p, int nid,
2717 					int dist)
2718 {
2719 	unsigned long faults, total_faults;
2720 
2721 	if (!p->numa_faults)
2722 		return 0;
2723 
2724 	total_faults = p->total_numa_faults;
2725 
2726 	if (!total_faults)
2727 		return 0;
2728 
2729 	faults = task_faults(p, nid);
2730 	faults += score_nearby_nodes(p, nid, dist, true);
2731 
2732 	return 1000 * faults / total_faults;
2733 }
2734 
2735 static inline unsigned long group_weight(struct task_struct *p, int nid,
2736 					 int dist)
2737 {
2738 	struct numa_group *ng = deref_task_numa_group(p);
2739 	unsigned long faults, total_faults;
2740 
2741 	if (!ng)
2742 		return 0;
2743 
2744 	total_faults = ng->total_faults;
2745 
2746 	if (!total_faults)
2747 		return 0;
2748 
2749 	faults = group_faults(p, nid);
2750 	faults += score_nearby_nodes(p, nid, dist, false);
2751 
2752 	return 1000 * faults / total_faults;
2753 }
2754 
2755 /*
2756  * If memory tiering mode is enabled, cpupid of slow memory page is
2757  * used to record scan time instead of CPU and PID.  When tiering mode
2758  * is disabled at run time, the scan time (in cpupid) will be
2759  * interpreted as CPU and PID.  So CPU needs to be checked to avoid to
2760  * access out of array bound.
2761  */
2762 static inline bool cpupid_valid(int cpupid)
2763 {
2764 	return cpupid_to_cpu(cpupid) < nr_cpu_ids;
2765 }
2766 
2767 /*
2768  * For memory tiering mode, if there are enough free pages (more than
2769  * enough watermark defined here) in fast memory node, to take full
2770  * advantage of fast memory capacity, all recently accessed slow
2771  * memory pages will be migrated to fast memory node without
2772  * considering hot threshold.
2773  */
2774 static bool pgdat_free_space_enough(struct pglist_data *pgdat)
2775 {
2776 	int z;
2777 	unsigned long enough_wmark;
2778 
2779 	enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT,
2780 			   pgdat->node_present_pages >> 4);
2781 	for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2782 		struct zone *zone = pgdat->node_zones + z;
2783 
2784 		if (!populated_zone(zone))
2785 			continue;
2786 
2787 		if (zone_watermark_ok(zone, 0,
2788 				      promo_wmark_pages(zone) + enough_wmark,
2789 				      ZONE_MOVABLE, 0))
2790 			return true;
2791 	}
2792 	return false;
2793 }
2794 
2795 /*
2796  * For memory tiering mode, when page tables are scanned, the scan
2797  * time will be recorded in struct page in addition to make page
2798  * PROT_NONE for slow memory page.  So when the page is accessed, in
2799  * hint page fault handler, the hint page fault latency is calculated
2800  * via,
2801  *
2802  *	hint page fault latency = hint page fault time - scan time
2803  *
2804  * The smaller the hint page fault latency, the higher the possibility
2805  * for the page to be hot.
2806  */
2807 static int numa_hint_fault_latency(struct folio *folio)
2808 {
2809 	int last_time, time;
2810 
2811 	time = jiffies_to_msecs(jiffies);
2812 	last_time = folio_xchg_access_time(folio, time);
2813 
2814 	return (time - last_time) & PAGE_ACCESS_TIME_MASK;
2815 }
2816 
2817 /*
2818  * For memory tiering mode, too high promotion/demotion throughput may
2819  * hurt application latency.  So we provide a mechanism to rate limit
2820  * the number of pages that are tried to be promoted.
2821  */
2822 static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
2823 				      unsigned long rate_limit, int nr)
2824 {
2825 	unsigned long nr_cand;
2826 	unsigned int now, start;
2827 
2828 	now = jiffies_to_msecs(jiffies);
2829 	mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr);
2830 	nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
2831 	start = pgdat->nbp_rl_start;
2832 	if (now - start > MSEC_PER_SEC &&
2833 	    cmpxchg(&pgdat->nbp_rl_start, start, now) == start)
2834 		pgdat->nbp_rl_nr_cand = nr_cand;
2835 	if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)
2836 		return true;
2837 	return false;
2838 }
2839 
2840 #define NUMA_MIGRATION_ADJUST_STEPS	16
2841 
2842 static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
2843 					    unsigned long rate_limit,
2844 					    unsigned int ref_th)
2845 {
2846 	unsigned int now, start, th_period, unit_th, th;
2847 	unsigned long nr_cand, ref_cand, diff_cand;
2848 
2849 	now = jiffies_to_msecs(jiffies);
2850 	th_period = sysctl_numa_balancing_scan_period_max;
2851 	start = pgdat->nbp_th_start;
2852 	if (now - start > th_period &&
2853 	    cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
2854 		ref_cand = rate_limit *
2855 			sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
2856 		nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
2857 		diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
2858 		unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
2859 		th = pgdat->nbp_threshold ? : ref_th;
2860 		if (diff_cand > ref_cand * 11 / 10)
2861 			th = max(th - unit_th, unit_th);
2862 		else if (diff_cand < ref_cand * 9 / 10)
2863 			th = min(th + unit_th, ref_th * 2);
2864 		pgdat->nbp_th_nr_cand = nr_cand;
2865 		pgdat->nbp_threshold = th;
2866 	}
2867 }
2868 
2869 bool should_numa_migrate_memory(struct task_struct *p, struct folio *folio,
2870 				int src_nid, int dst_cpu)
2871 {
2872 	struct numa_group *ng = deref_curr_numa_group(p);
2873 	int dst_nid = cpu_to_node(dst_cpu);
2874 	int last_cpupid, this_cpupid;
2875 
2876 	/*
2877 	 * Cannot migrate to memoryless nodes.
2878 	 */
2879 	if (!node_state(dst_nid, N_MEMORY))
2880 		return false;
2881 
2882 	/*
2883 	 * The pages in slow memory node should be migrated according
2884 	 * to hot/cold instead of private/shared.
2885 	 */
2886 	if (folio_use_access_time(folio)) {
2887 		struct pglist_data *pgdat;
2888 		unsigned long rate_limit;
2889 		unsigned int latency, th, def_th;
2890 		long nr = folio_nr_pages(folio);
2891 
2892 		pgdat = NODE_DATA(dst_nid);
2893 		if (pgdat_free_space_enough(pgdat)) {
2894 			/* workload changed, reset hot threshold */
2895 			pgdat->nbp_threshold = 0;
2896 			mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE_NRL, nr);
2897 			return true;
2898 		}
2899 
2900 		def_th = sysctl_numa_balancing_hot_threshold;
2901 		rate_limit = MB_TO_PAGES(sysctl_numa_balancing_promote_rate_limit);
2902 		numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
2903 
2904 		th = pgdat->nbp_threshold ? : def_th;
2905 		latency = numa_hint_fault_latency(folio);
2906 		if (latency >= th)
2907 			return false;
2908 
2909 		return !numa_promotion_rate_limit(pgdat, rate_limit, nr);
2910 	}
2911 
2912 	this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
2913 	last_cpupid = folio_xchg_last_cpupid(folio, this_cpupid);
2914 
2915 	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
2916 	    !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
2917 		return false;
2918 
2919 	/*
2920 	 * Allow first faults or private faults to migrate immediately early in
2921 	 * the lifetime of a task. The magic number 4 is based on waiting for
2922 	 * two full passes of the "multi-stage node selection" test that is
2923 	 * executed below.
2924 	 */
2925 	if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
2926 	    (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
2927 		return true;
2928 
2929 	/*
2930 	 * Multi-stage node selection is used in conjunction with a periodic
2931 	 * migration fault to build a temporal task<->page relation. By using
2932 	 * a two-stage filter we remove short/unlikely relations.
2933 	 *
2934 	 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
2935 	 * a task's usage of a particular page (n_p) per total usage of this
2936 	 * page (n_t) (in a given time-span) to a probability.
2937 	 *
2938 	 * Our periodic faults will sample this probability and getting the
2939 	 * same result twice in a row, given these samples are fully
2940 	 * independent, is then given by P(n)^2, provided our sample period
2941 	 * is sufficiently short compared to the usage pattern.
2942 	 *
2943 	 * This quadric squishes small probabilities, making it less likely we
2944 	 * act on an unlikely task<->page relation.
2945 	 */
2946 	if (!cpupid_pid_unset(last_cpupid) &&
2947 				cpupid_to_nid(last_cpupid) != dst_nid)
2948 		return false;
2949 
2950 	/* Always allow migrate on private faults */
2951 	if (cpupid_match_pid(p, last_cpupid))
2952 		return true;
2953 
2954 	/* A shared fault, but p->numa_group has not been set up yet. */
2955 	if (!ng)
2956 		return true;
2957 
2958 	/*
2959 	 * Destination node is much more heavily used than the source
2960 	 * node? Allow migration.
2961 	 */
2962 	if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
2963 					ACTIVE_NODE_FRACTION)
2964 		return true;
2965 
2966 	/*
2967 	 * Distribute memory according to CPU & memory use on each node,
2968 	 * with 3/4 hysteresis to avoid unnecessary memory migrations:
2969 	 *
2970 	 * faults_cpu(dst)   3   faults_cpu(src)
2971 	 * --------------- * - > ---------------
2972 	 * faults_mem(dst)   4   faults_mem(src)
2973 	 */
2974 	return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
2975 	       group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
2976 }
2977 
2978 /*
2979  * 'numa_type' describes the node at the moment of load balancing.
2980  */
2981 enum numa_type {
2982 	/* The node has spare capacity that can be used to run more tasks.  */
2983 	node_has_spare = 0,
2984 	/*
2985 	 * The node is fully used and the tasks don't compete for more CPU
2986 	 * cycles. Nevertheless, some tasks might wait before running.
2987 	 */
2988 	node_fully_busy,
2989 	/*
2990 	 * The node is overloaded and can't provide expected CPU cycles to all
2991 	 * tasks.
2992 	 */
2993 	node_overloaded
2994 };
2995 
2996 /* Cached statistics for all CPUs within a node */
2997 struct numa_stats {
2998 	unsigned long load;
2999 	unsigned long runnable;
3000 	unsigned long util;
3001 	/* Total compute capacity of CPUs on a node */
3002 	unsigned long compute_capacity;
3003 	unsigned int nr_running;
3004 	unsigned int weight;
3005 	enum numa_type node_type;
3006 	int idle_cpu;
3007 };
3008 
3009 struct task_numa_env {
3010 	struct task_struct *p;
3011 
3012 	int src_cpu, src_nid;
3013 	int dst_cpu, dst_nid;
3014 	int imb_numa_nr;
3015 
3016 	struct numa_stats src_stats, dst_stats;
3017 
3018 	int imbalance_pct;
3019 	int dist;
3020 
3021 	struct task_struct *best_task;
3022 	long best_imp;
3023 	int best_cpu;
3024 };
3025 
3026 static unsigned long cpu_load(struct rq *rq);
3027 static unsigned long cpu_runnable(struct rq *rq);
3028 
3029 static inline enum
3030 numa_type numa_classify(unsigned int imbalance_pct,
3031 			 struct numa_stats *ns)
3032 {
3033 	if ((ns->nr_running > ns->weight) &&
3034 	    (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
3035 	     ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
3036 		return node_overloaded;
3037 
3038 	if ((ns->nr_running < ns->weight) ||
3039 	    (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
3040 	     ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
3041 		return node_has_spare;
3042 
3043 	return node_fully_busy;
3044 }
3045 
3046 /* Forward declarations of select_idle_sibling helpers */
3047 static inline bool test_idle_cores(int cpu);
3048 static inline int numa_idle_core(int idle_core, int cpu)
3049 {
3050 	if (!sched_smt_active() ||
3051 	    idle_core >= 0 || !test_idle_cores(cpu))
3052 		return idle_core;
3053 
3054 	/*
3055 	 * Prefer cores instead of packing HT siblings
3056 	 * and triggering future load balancing.
3057 	 */
3058 	if (is_core_idle(cpu))
3059 		idle_core = cpu;
3060 
3061 	return idle_core;
3062 }
3063 
3064 /*
3065  * Gather all necessary information to make NUMA balancing placement
3066  * decisions that are compatible with standard load balancer. This
3067  * borrows code and logic from update_sg_lb_stats but sharing a
3068  * common implementation is impractical.
3069  */
3070 static void update_numa_stats(struct task_numa_env *env,
3071 			      struct numa_stats *ns, int nid,
3072 			      bool find_idle)
3073 {
3074 	int cpu, idle_core = -1;
3075 
3076 	memset(ns, 0, sizeof(*ns));
3077 	ns->idle_cpu = -1;
3078 
3079 	rcu_read_lock();
3080 	for_each_cpu(cpu, cpumask_of_node(nid)) {
3081 		struct rq *rq = cpu_rq(cpu);
3082 
3083 		ns->load += cpu_load(rq);
3084 		ns->runnable += cpu_runnable(rq);
3085 		ns->util += cpu_util_cfs(cpu);
3086 		ns->nr_running += rq->cfs.h_nr_runnable;
3087 		ns->compute_capacity += capacity_of(cpu);
3088 
3089 		if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) {
3090 			if (READ_ONCE(rq->numa_migrate_on) ||
3091 			    !cpumask_test_cpu(cpu, env->p->cpus_ptr))
3092 				continue;
3093 
3094 			if (ns->idle_cpu == -1)
3095 				ns->idle_cpu = cpu;
3096 
3097 			idle_core = numa_idle_core(idle_core, cpu);
3098 		}
3099 	}
3100 	rcu_read_unlock();
3101 
3102 	ns->weight = cpumask_weight(cpumask_of_node(nid));
3103 
3104 	ns->node_type = numa_classify(env->imbalance_pct, ns);
3105 
3106 	if (idle_core >= 0)
3107 		ns->idle_cpu = idle_core;
3108 }
3109 
3110 static void task_numa_assign(struct task_numa_env *env,
3111 			     struct task_struct *p, long imp)
3112 {
3113 	struct rq *rq = cpu_rq(env->dst_cpu);
3114 
3115 	/* Check if run-queue part of active NUMA balance. */
3116 	if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
3117 		int cpu;
3118 		int start = env->dst_cpu;
3119 
3120 		/* Find alternative idle CPU. */
3121 		for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) {
3122 			if (cpu == env->best_cpu || !idle_cpu(cpu) ||
3123 			    !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
3124 				continue;
3125 			}
3126 
3127 			env->dst_cpu = cpu;
3128 			rq = cpu_rq(env->dst_cpu);
3129 			if (!xchg(&rq->numa_migrate_on, 1))
3130 				goto assign;
3131 		}
3132 
3133 		/* Failed to find an alternative idle CPU */
3134 		return;
3135 	}
3136 
3137 assign:
3138 	/*
3139 	 * Clear previous best_cpu/rq numa-migrate flag, since task now
3140 	 * found a better CPU to move/swap.
3141 	 */
3142 	if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
3143 		rq = cpu_rq(env->best_cpu);
3144 		WRITE_ONCE(rq->numa_migrate_on, 0);
3145 	}
3146 
3147 	if (env->best_task)
3148 		put_task_struct(env->best_task);
3149 	if (p)
3150 		get_task_struct(p);
3151 
3152 	env->best_task = p;
3153 	env->best_imp = imp;
3154 	env->best_cpu = env->dst_cpu;
3155 }
3156 
3157 static bool load_too_imbalanced(long src_load, long dst_load,
3158 				struct task_numa_env *env)
3159 {
3160 	long imb, old_imb;
3161 	long orig_src_load, orig_dst_load;
3162 	long src_capacity, dst_capacity;
3163 
3164 	/*
3165 	 * The load is corrected for the CPU capacity available on each node.
3166 	 *
3167 	 * src_load        dst_load
3168 	 * ------------ vs ---------
3169 	 * src_capacity    dst_capacity
3170 	 */
3171 	src_capacity = env->src_stats.compute_capacity;
3172 	dst_capacity = env->dst_stats.compute_capacity;
3173 
3174 	imb = abs(dst_load * src_capacity - src_load * dst_capacity);
3175 
3176 	orig_src_load = env->src_stats.load;
3177 	orig_dst_load = env->dst_stats.load;
3178 
3179 	old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
3180 
3181 	/* Would this change make things worse? */
3182 	return (imb > old_imb);
3183 }
3184 
3185 /*
3186  * Maximum NUMA importance can be 1998 (2*999);
3187  * SMALLIMP @ 30 would be close to 1998/64.
3188  * Used to deter task migration.
3189  */
3190 #define SMALLIMP	30
3191 
3192 /*
3193  * This checks if the overall compute and NUMA accesses of the system would
3194  * be improved if the source tasks was migrated to the target dst_cpu taking
3195  * into account that it might be best if task running on the dst_cpu should
3196  * be exchanged with the source task
3197  */
3198 static bool task_numa_compare(struct task_numa_env *env,
3199 			      long taskimp, long groupimp, bool maymove)
3200 {
3201 	struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
3202 	struct rq *dst_rq = cpu_rq(env->dst_cpu);
3203 	long imp = p_ng ? groupimp : taskimp;
3204 	struct task_struct *cur;
3205 	long src_load, dst_load;
3206 	int dist = env->dist;
3207 	long moveimp = imp;
3208 	long load;
3209 	bool stopsearch = false;
3210 
3211 	if (READ_ONCE(dst_rq->numa_migrate_on))
3212 		return false;
3213 
3214 	rcu_read_lock();
3215 	cur = rcu_dereference_all(dst_rq->curr);
3216 	if (cur && ((cur->flags & (PF_EXITING | PF_KTHREAD)) ||
3217 		    !cur->mm))
3218 		cur = NULL;
3219 
3220 	/*
3221 	 * Because we have preemption enabled we can get migrated around and
3222 	 * end try selecting ourselves (current == env->p) as a swap candidate.
3223 	 */
3224 	if (cur == env->p) {
3225 		stopsearch = true;
3226 		goto unlock;
3227 	}
3228 
3229 	if (!cur) {
3230 		if (maymove && moveimp >= env->best_imp)
3231 			goto assign;
3232 		else
3233 			goto unlock;
3234 	}
3235 
3236 	/* Skip this swap candidate if cannot move to the source cpu. */
3237 	if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
3238 		goto unlock;
3239 
3240 	/*
3241 	 * Skip this swap candidate if it is not moving to its preferred
3242 	 * node and the best task is.
3243 	 */
3244 	if (env->best_task &&
3245 	    env->best_task->numa_preferred_nid == env->src_nid &&
3246 	    cur->numa_preferred_nid != env->src_nid) {
3247 		goto unlock;
3248 	}
3249 
3250 	/*
3251 	 * "imp" is the fault differential for the source task between the
3252 	 * source and destination node. Calculate the total differential for
3253 	 * the source task and potential destination task. The more negative
3254 	 * the value is, the more remote accesses that would be expected to
3255 	 * be incurred if the tasks were swapped.
3256 	 *
3257 	 * If dst and source tasks are in the same NUMA group, or not
3258 	 * in any group then look only at task weights.
3259 	 */
3260 	cur_ng = rcu_dereference_all(cur->numa_group);
3261 	if (cur_ng == p_ng) {
3262 		/*
3263 		 * Do not swap within a group or between tasks that have
3264 		 * no group if there is spare capacity. Swapping does
3265 		 * not address the load imbalance and helps one task at
3266 		 * the cost of punishing another.
3267 		 */
3268 		if (env->dst_stats.node_type == node_has_spare)
3269 			goto unlock;
3270 
3271 		imp = taskimp + task_weight(cur, env->src_nid, dist) -
3272 		      task_weight(cur, env->dst_nid, dist);
3273 		/*
3274 		 * Add some hysteresis to prevent swapping the
3275 		 * tasks within a group over tiny differences.
3276 		 */
3277 		if (cur_ng)
3278 			imp -= imp / 16;
3279 	} else {
3280 		/*
3281 		 * Compare the group weights. If a task is all by itself
3282 		 * (not part of a group), use the task weight instead.
3283 		 */
3284 		if (cur_ng && p_ng)
3285 			imp += group_weight(cur, env->src_nid, dist) -
3286 			       group_weight(cur, env->dst_nid, dist);
3287 		else
3288 			imp += task_weight(cur, env->src_nid, dist) -
3289 			       task_weight(cur, env->dst_nid, dist);
3290 	}
3291 
3292 	/* Discourage picking a task already on its preferred node */
3293 	if (cur->numa_preferred_nid == env->dst_nid)
3294 		imp -= imp / 16;
3295 
3296 	/*
3297 	 * Encourage picking a task that moves to its preferred node.
3298 	 * This potentially makes imp larger than it's maximum of
3299 	 * 1998 (see SMALLIMP and task_weight for why) but in this
3300 	 * case, it does not matter.
3301 	 */
3302 	if (cur->numa_preferred_nid == env->src_nid)
3303 		imp += imp / 8;
3304 
3305 	if (maymove && moveimp > imp && moveimp > env->best_imp) {
3306 		imp = moveimp;
3307 		cur = NULL;
3308 		goto assign;
3309 	}
3310 
3311 	/*
3312 	 * Prefer swapping with a task moving to its preferred node over a
3313 	 * task that is not.
3314 	 */
3315 	if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
3316 	    env->best_task->numa_preferred_nid != env->src_nid) {
3317 		goto assign;
3318 	}
3319 
3320 	/*
3321 	 * If the NUMA importance is less than SMALLIMP,
3322 	 * task migration might only result in ping pong
3323 	 * of tasks and also hurt performance due to cache
3324 	 * misses.
3325 	 */
3326 	if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
3327 		goto unlock;
3328 
3329 	/*
3330 	 * In the overloaded case, try and keep the load balanced.
3331 	 */
3332 	load = task_h_load(env->p) - task_h_load(cur);
3333 	if (!load)
3334 		goto assign;
3335 
3336 	dst_load = env->dst_stats.load + load;
3337 	src_load = env->src_stats.load - load;
3338 
3339 	if (load_too_imbalanced(src_load, dst_load, env))
3340 		goto unlock;
3341 
3342 assign:
3343 	/* Evaluate an idle CPU for a task numa move. */
3344 	if (!cur) {
3345 		int cpu = env->dst_stats.idle_cpu;
3346 
3347 		/* Nothing cached so current CPU went idle since the search. */
3348 		if (cpu < 0)
3349 			cpu = env->dst_cpu;
3350 
3351 		/*
3352 		 * If the CPU is no longer truly idle and the previous best CPU
3353 		 * is, keep using it.
3354 		 */
3355 		if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
3356 		    idle_cpu(env->best_cpu)) {
3357 			cpu = env->best_cpu;
3358 		}
3359 
3360 		env->dst_cpu = cpu;
3361 	}
3362 
3363 	task_numa_assign(env, cur, imp);
3364 
3365 	/*
3366 	 * If a move to idle is allowed because there is capacity or load
3367 	 * balance improves then stop the search. While a better swap
3368 	 * candidate may exist, a search is not free.
3369 	 */
3370 	if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
3371 		stopsearch = true;
3372 
3373 	/*
3374 	 * If a swap candidate must be identified and the current best task
3375 	 * moves its preferred node then stop the search.
3376 	 */
3377 	if (!maymove && env->best_task &&
3378 	    env->best_task->numa_preferred_nid == env->src_nid) {
3379 		stopsearch = true;
3380 	}
3381 unlock:
3382 	rcu_read_unlock();
3383 
3384 	return stopsearch;
3385 }
3386 
3387 static void task_numa_find_cpu(struct task_numa_env *env,
3388 				long taskimp, long groupimp)
3389 {
3390 	bool maymove = false;
3391 	int cpu;
3392 
3393 	/*
3394 	 * If dst node has spare capacity, then check if there is an
3395 	 * imbalance that would be overruled by the load balancer.
3396 	 */
3397 	if (env->dst_stats.node_type == node_has_spare) {
3398 		unsigned int imbalance;
3399 		int src_running, dst_running;
3400 
3401 		/*
3402 		 * Would movement cause an imbalance? Note that if src has
3403 		 * more running tasks that the imbalance is ignored as the
3404 		 * move improves the imbalance from the perspective of the
3405 		 * CPU load balancer.
3406 		 * */
3407 		src_running = env->src_stats.nr_running - 1;
3408 		dst_running = env->dst_stats.nr_running + 1;
3409 		imbalance = max(0, dst_running - src_running);
3410 		imbalance = adjust_numa_imbalance(imbalance, dst_running,
3411 						  env->imb_numa_nr);
3412 
3413 		/* Use idle CPU if there is no imbalance */
3414 		if (!imbalance) {
3415 			maymove = true;
3416 			if (env->dst_stats.idle_cpu >= 0) {
3417 				env->dst_cpu = env->dst_stats.idle_cpu;
3418 				task_numa_assign(env, NULL, 0);
3419 				return;
3420 			}
3421 		}
3422 	} else {
3423 		long src_load, dst_load, load;
3424 		/*
3425 		 * If the improvement from just moving env->p direction is better
3426 		 * than swapping tasks around, check if a move is possible.
3427 		 */
3428 		load = task_h_load(env->p);
3429 		dst_load = env->dst_stats.load + load;
3430 		src_load = env->src_stats.load - load;
3431 		maymove = !load_too_imbalanced(src_load, dst_load, env);
3432 	}
3433 
3434 	/* Skip CPUs if the source task cannot migrate */
3435 	for_each_cpu_and(cpu, cpumask_of_node(env->dst_nid), env->p->cpus_ptr) {
3436 		env->dst_cpu = cpu;
3437 		if (task_numa_compare(env, taskimp, groupimp, maymove))
3438 			break;
3439 	}
3440 }
3441 
3442 static int task_numa_migrate(struct task_struct *p)
3443 {
3444 	struct task_numa_env env = {
3445 		.p = p,
3446 
3447 		.src_cpu = task_cpu(p),
3448 		.src_nid = task_node(p),
3449 
3450 		.imbalance_pct = 112,
3451 
3452 		.best_task = NULL,
3453 		.best_imp = 0,
3454 		.best_cpu = -1,
3455 	};
3456 	unsigned long taskweight, groupweight;
3457 	struct sched_domain *sd;
3458 	long taskimp, groupimp;
3459 	struct numa_group *ng;
3460 	struct rq *best_rq;
3461 	int nid, ret, dist;
3462 
3463 	/*
3464 	 * Pick the lowest SD_NUMA domain, as that would have the smallest
3465 	 * imbalance and would be the first to start moving tasks about.
3466 	 *
3467 	 * And we want to avoid any moving of tasks about, as that would create
3468 	 * random movement of tasks -- counter the numa conditions we're trying
3469 	 * to satisfy here.
3470 	 */
3471 	rcu_read_lock();
3472 	sd = rcu_dereference_all(per_cpu(sd_numa, env.src_cpu));
3473 	if (sd) {
3474 		env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
3475 		env.imb_numa_nr = sd->imb_numa_nr;
3476 	}
3477 	rcu_read_unlock();
3478 
3479 	/*
3480 	 * Cpusets can break the scheduler domain tree into smaller
3481 	 * balance domains, some of which do not cross NUMA boundaries.
3482 	 * Tasks that are "trapped" in such domains cannot be migrated
3483 	 * elsewhere, so there is no point in (re)trying.
3484 	 */
3485 	if (unlikely(!sd)) {
3486 		sched_setnuma(p, task_node(p));
3487 		return -EINVAL;
3488 	}
3489 
3490 	env.dst_nid = p->numa_preferred_nid;
3491 	dist = env.dist = node_distance(env.src_nid, env.dst_nid);
3492 	taskweight = task_weight(p, env.src_nid, dist);
3493 	groupweight = group_weight(p, env.src_nid, dist);
3494 	update_numa_stats(&env, &env.src_stats, env.src_nid, false);
3495 	taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
3496 	groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
3497 	update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
3498 
3499 	/* Try to find a spot on the preferred nid. */
3500 	task_numa_find_cpu(&env, taskimp, groupimp);
3501 
3502 	/*
3503 	 * Look at other nodes in these cases:
3504 	 * - there is no space available on the preferred_nid
3505 	 * - the task is part of a numa_group that is interleaved across
3506 	 *   multiple NUMA nodes; in order to better consolidate the group,
3507 	 *   we need to check other locations.
3508 	 */
3509 	ng = deref_curr_numa_group(p);
3510 	if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
3511 		for_each_node_state(nid, N_CPU) {
3512 			if (nid == env.src_nid || nid == p->numa_preferred_nid)
3513 				continue;
3514 
3515 			dist = node_distance(env.src_nid, env.dst_nid);
3516 			if (sched_numa_topology_type == NUMA_BACKPLANE &&
3517 						dist != env.dist) {
3518 				taskweight = task_weight(p, env.src_nid, dist);
3519 				groupweight = group_weight(p, env.src_nid, dist);
3520 			}
3521 
3522 			/* Only consider nodes where both task and groups benefit */
3523 			taskimp = task_weight(p, nid, dist) - taskweight;
3524 			groupimp = group_weight(p, nid, dist) - groupweight;
3525 			if (taskimp < 0 && groupimp < 0)
3526 				continue;
3527 
3528 			env.dist = dist;
3529 			env.dst_nid = nid;
3530 			update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
3531 			task_numa_find_cpu(&env, taskimp, groupimp);
3532 		}
3533 	}
3534 
3535 	/*
3536 	 * If the task is part of a workload that spans multiple NUMA nodes,
3537 	 * and is migrating into one of the workload's active nodes, remember
3538 	 * this node as the task's preferred numa node, so the workload can
3539 	 * settle down.
3540 	 * A task that migrated to a second choice node will be better off
3541 	 * trying for a better one later. Do not set the preferred node here.
3542 	 */
3543 	if (ng) {
3544 		if (env.best_cpu == -1)
3545 			nid = env.src_nid;
3546 		else
3547 			nid = cpu_to_node(env.best_cpu);
3548 
3549 		if (nid != p->numa_preferred_nid)
3550 			sched_setnuma(p, nid);
3551 	}
3552 
3553 	/* No better CPU than the current one was found. */
3554 	if (env.best_cpu == -1) {
3555 		trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
3556 		return -EAGAIN;
3557 	}
3558 
3559 	best_rq = cpu_rq(env.best_cpu);
3560 	if (env.best_task == NULL) {
3561 		ret = migrate_task_to(p, env.best_cpu);
3562 		WRITE_ONCE(best_rq->numa_migrate_on, 0);
3563 		if (ret != 0)
3564 			trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
3565 		return ret;
3566 	}
3567 
3568 	ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
3569 	WRITE_ONCE(best_rq->numa_migrate_on, 0);
3570 
3571 	if (ret != 0)
3572 		trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
3573 	put_task_struct(env.best_task);
3574 	return ret;
3575 }
3576 
3577 /* Attempt to migrate a task to a CPU on the preferred node. */
3578 static void numa_migrate_preferred(struct task_struct *p)
3579 {
3580 	unsigned long interval = HZ;
3581 
3582 	/* This task has no NUMA fault statistics yet */
3583 	if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
3584 		return;
3585 
3586 	/* Periodically retry migrating the task to the preferred node */
3587 	interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
3588 	p->numa_migrate_retry = jiffies + interval;
3589 
3590 	/* Success if task is already running on preferred CPU */
3591 	if (task_node(p) == p->numa_preferred_nid)
3592 		return;
3593 
3594 	/* Otherwise, try migrate to a CPU on the preferred node */
3595 	task_numa_migrate(p);
3596 }
3597 
3598 /*
3599  * Find out how many nodes the workload is actively running on. Do this by
3600  * tracking the nodes from which NUMA hinting faults are triggered. This can
3601  * be different from the set of nodes where the workload's memory is currently
3602  * located.
3603  */
3604 static void numa_group_count_active_nodes(struct numa_group *numa_group)
3605 {
3606 	unsigned long faults, max_faults = 0;
3607 	int nid, active_nodes = 0;
3608 
3609 	for_each_node_state(nid, N_CPU) {
3610 		faults = group_faults_cpu(numa_group, nid);
3611 		if (faults > max_faults)
3612 			max_faults = faults;
3613 	}
3614 
3615 	for_each_node_state(nid, N_CPU) {
3616 		faults = group_faults_cpu(numa_group, nid);
3617 		if (faults * ACTIVE_NODE_FRACTION > max_faults)
3618 			active_nodes++;
3619 	}
3620 
3621 	numa_group->max_faults_cpu = max_faults;
3622 	numa_group->active_nodes = active_nodes;
3623 }
3624 
3625 /*
3626  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
3627  * increments. The more local the fault statistics are, the higher the scan
3628  * period will be for the next scan window. If local/(local+remote) ratio is
3629  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
3630  * the scan period will decrease. Aim for 70% local accesses.
3631  */
3632 #define NUMA_PERIOD_SLOTS 10
3633 #define NUMA_PERIOD_THRESHOLD 7
3634 
3635 /*
3636  * Increase the scan period (slow down scanning) if the majority of
3637  * our memory is already on our local node, or if the majority of
3638  * the page accesses are shared with other processes.
3639  * Otherwise, decrease the scan period.
3640  */
3641 static void update_task_scan_period(struct task_struct *p,
3642 			unsigned long shared, unsigned long private)
3643 {
3644 	unsigned int period_slot;
3645 	int lr_ratio, ps_ratio;
3646 	int diff;
3647 
3648 	unsigned long remote = p->numa_faults_locality[0];
3649 	unsigned long local = p->numa_faults_locality[1];
3650 
3651 	/*
3652 	 * If there were no record hinting faults then either the task is
3653 	 * completely idle or all activity is in areas that are not of interest
3654 	 * to automatic numa balancing. Related to that, if there were failed
3655 	 * migration then it implies we are migrating too quickly or the local
3656 	 * node is overloaded. In either case, scan slower
3657 	 */
3658 	if (local + shared == 0 || p->numa_faults_locality[2]) {
3659 		p->numa_scan_period = min(p->numa_scan_period_max,
3660 			p->numa_scan_period << 1);
3661 
3662 		p->mm->numa_next_scan = jiffies +
3663 			msecs_to_jiffies(p->numa_scan_period);
3664 
3665 		goto out;
3666 	}
3667 
3668 	/*
3669 	 * Prepare to scale scan period relative to the current period.
3670 	 *	 == NUMA_PERIOD_THRESHOLD scan period stays the same
3671 	 *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
3672 	 *	 >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
3673 	 */
3674 	period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
3675 	lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
3676 	ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
3677 
3678 	if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
3679 		/*
3680 		 * Most memory accesses are local. There is no need to
3681 		 * do fast NUMA scanning, since memory is already local.
3682 		 */
3683 		int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
3684 		if (!slot)
3685 			slot = 1;
3686 		diff = slot * period_slot;
3687 	} else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
3688 		/*
3689 		 * Most memory accesses are shared with other tasks.
3690 		 * There is no point in continuing fast NUMA scanning,
3691 		 * since other tasks may just move the memory elsewhere.
3692 		 */
3693 		int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
3694 		if (!slot)
3695 			slot = 1;
3696 		diff = slot * period_slot;
3697 	} else {
3698 		/*
3699 		 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
3700 		 * yet they are not on the local NUMA node. Speed up
3701 		 * NUMA scanning to get the memory moved over.
3702 		 */
3703 		int ratio = max(lr_ratio, ps_ratio);
3704 		diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
3705 	}
3706 
3707 	p->numa_scan_period = clamp(p->numa_scan_period + diff,
3708 			task_scan_min(p), task_scan_max(p));
3709 
3710 out:
3711 	memset(p->numa_faults_locality, 0,
3712 	       sizeof(p->numa_faults_locality));
3713 }
3714 
3715 /*
3716  * Get the fraction of time the task has been running since the last
3717  * NUMA placement cycle. The scheduler keeps similar statistics, but
3718  * decays those on a 32ms period, which is orders of magnitude off
3719  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
3720  * stats only if the task is so new there are no NUMA statistics yet.
3721  */
3722 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
3723 {
3724 	u64 runtime, delta, now;
3725 	/* Use the start of this time slice to avoid calculations. */
3726 	now = p->se.exec_start;
3727 	runtime = p->se.sum_exec_runtime;
3728 
3729 	if (p->last_task_numa_placement) {
3730 		delta = runtime - p->last_sum_exec_runtime;
3731 		*period = now - p->last_task_numa_placement;
3732 
3733 		/* Avoid time going backwards, prevent potential divide error: */
3734 		if (unlikely((s64)*period < 0))
3735 			*period = 0;
3736 	} else {
3737 		delta = p->se.avg.load_sum;
3738 		*period = LOAD_AVG_MAX;
3739 	}
3740 
3741 	p->last_sum_exec_runtime = runtime;
3742 	p->last_task_numa_placement = now;
3743 
3744 	return delta;
3745 }
3746 
3747 /*
3748  * Determine the preferred nid for a task in a numa_group. This needs to
3749  * be done in a way that produces consistent results with group_weight,
3750  * otherwise workloads might not converge.
3751  */
3752 static int preferred_group_nid(struct task_struct *p, int nid)
3753 {
3754 	nodemask_t nodes;
3755 	int dist;
3756 
3757 	/* Direct connections between all NUMA nodes. */
3758 	if (sched_numa_topology_type == NUMA_DIRECT)
3759 		return nid;
3760 
3761 	/*
3762 	 * On a system with glueless mesh NUMA topology, group_weight
3763 	 * scores nodes according to the number of NUMA hinting faults on
3764 	 * both the node itself, and on nearby nodes.
3765 	 */
3766 	if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
3767 		unsigned long score, max_score = 0;
3768 		int node, max_node = nid;
3769 
3770 		dist = sched_max_numa_distance;
3771 
3772 		for_each_node_state(node, N_CPU) {
3773 			score = group_weight(p, node, dist);
3774 			if (score > max_score) {
3775 				max_score = score;
3776 				max_node = node;
3777 			}
3778 		}
3779 		return max_node;
3780 	}
3781 
3782 	/*
3783 	 * Finding the preferred nid in a system with NUMA backplane
3784 	 * interconnect topology is more involved. The goal is to locate
3785 	 * tasks from numa_groups near each other in the system, and
3786 	 * untangle workloads from different sides of the system. This requires
3787 	 * searching down the hierarchy of node groups, recursively searching
3788 	 * inside the highest scoring group of nodes. The nodemask tricks
3789 	 * keep the complexity of the search down.
3790 	 */
3791 	nodes = node_states[N_CPU];
3792 	for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
3793 		unsigned long max_faults = 0;
3794 		nodemask_t max_group = NODE_MASK_NONE;
3795 		int a, b;
3796 
3797 		/* Are there nodes at this distance from each other? */
3798 		if (!find_numa_distance(dist))
3799 			continue;
3800 
3801 		for_each_node_mask(a, nodes) {
3802 			unsigned long faults = 0;
3803 			nodemask_t this_group;
3804 			nodes_clear(this_group);
3805 
3806 			/* Sum group's NUMA faults; includes a==b case. */
3807 			for_each_node_mask(b, nodes) {
3808 				if (node_distance(a, b) < dist) {
3809 					faults += group_faults(p, b);
3810 					node_set(b, this_group);
3811 					node_clear(b, nodes);
3812 				}
3813 			}
3814 
3815 			/* Remember the top group. */
3816 			if (faults > max_faults) {
3817 				max_faults = faults;
3818 				max_group = this_group;
3819 				/*
3820 				 * subtle: at the smallest distance there is
3821 				 * just one node left in each "group", the
3822 				 * winner is the preferred nid.
3823 				 */
3824 				nid = a;
3825 			}
3826 		}
3827 		/* Next round, evaluate the nodes within max_group. */
3828 		if (!max_faults)
3829 			break;
3830 		nodes = max_group;
3831 	}
3832 	return nid;
3833 }
3834 
3835 static void task_numa_placement(struct task_struct *p)
3836 	__context_unsafe(/* conditional locking */)
3837 {
3838 	struct sched_cache_group __maybe_unused *grp;
3839 	int seq, nid, max_nid = NUMA_NO_NODE;
3840 	unsigned long max_faults = 0;
3841 	unsigned long fault_types[2] = { 0, 0 };
3842 	unsigned long total_faults;
3843 	u64 runtime, period;
3844 	spinlock_t *group_lock = NULL;
3845 	long __maybe_unused new_fp;
3846 	struct numa_group *ng;
3847 
3848 	/*
3849 	 * The p->mm->numa_scan_seq field gets updated without
3850 	 * exclusive access. Use READ_ONCE() here to ensure
3851 	 * that the field is read in a single access:
3852 	 */
3853 	seq = READ_ONCE(p->mm->numa_scan_seq);
3854 	if (p->numa_scan_seq == seq)
3855 		return;
3856 	p->numa_scan_seq = seq;
3857 	p->numa_scan_period_max = task_scan_max(p);
3858 
3859 	total_faults = p->numa_faults_locality[0] +
3860 		       p->numa_faults_locality[1];
3861 	runtime = numa_get_avg_runtime(p, &period);
3862 
3863 	/* If the task is part of a group prevent parallel updates to group stats */
3864 	ng = deref_curr_numa_group(p);
3865 	if (ng) {
3866 		group_lock = &ng->lock;
3867 		spin_lock_irq(group_lock);
3868 	}
3869 
3870 	/* Find the node with the highest number of faults */
3871 	for_each_online_node(nid) {
3872 		/* Keep track of the offsets in numa_faults array */
3873 		int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
3874 		unsigned long faults = 0, group_faults = 0;
3875 		int priv;
3876 
3877 		for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
3878 			long diff, f_diff, f_weight;
3879 
3880 			mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
3881 			membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
3882 			cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
3883 			cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
3884 
3885 			/* Decay existing window, copy faults since last scan */
3886 			diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
3887 			fault_types[priv] += p->numa_faults[membuf_idx];
3888 			p->numa_faults[membuf_idx] = 0;
3889 
3890 			/*
3891 			 * Normalize the faults_from, so all tasks in a group
3892 			 * count according to CPU use, instead of by the raw
3893 			 * number of faults. Tasks with little runtime have
3894 			 * little over-all impact on throughput, and thus their
3895 			 * faults are less important.
3896 			 */
3897 			f_weight = div64_u64(runtime << 16, period + 1);
3898 			f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
3899 				   (total_faults + 1);
3900 			f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
3901 			p->numa_faults[cpubuf_idx] = 0;
3902 
3903 			p->numa_faults[mem_idx] += diff;
3904 			p->numa_faults[cpu_idx] += f_diff;
3905 			faults += p->numa_faults[mem_idx];
3906 			p->total_numa_faults += diff;
3907 			if (ng) {
3908 				/*
3909 				 * safe because we can only change our own group
3910 				 *
3911 				 * mem_idx represents the offset for a given
3912 				 * nid and priv in a specific region because it
3913 				 * is at the beginning of the numa_faults array.
3914 				 */
3915 				ng->faults[mem_idx] += diff;
3916 				ng->faults[cpu_idx] += f_diff;
3917 				ng->total_faults += diff;
3918 				group_faults += ng->faults[mem_idx];
3919 			}
3920 #ifdef CONFIG_SCHED_CACHE
3921 			/*
3922 			 * Per task p->numa_faults[mem_idx] converges,
3923 			 * so the accumulation of each task's faults
3924 			 * converges too - Given the number of threads,
3925 			 * it cannot overflow an unsigned long.
3926 			 * Racy with concurrent updates from other threads
3927 			 * sharing this mm. Acceptable since footprint is a
3928 			 * heuristic and occasional lost updates are tolerable.
3929 			 *
3930 			 * If a task exits, its corresponding footprint must
3931 			 * be subtracted from p->sched_cache_grp->footprint,
3932 			 * otherwise the footprint will not converge: the
3933 			 * exiting thread's footprint remains unchanged/undecayed.
3934 			 * See exit_mm().
3935 			 *
3936 			 * Lost updates and unsynchronized subtraction
3937 			 * in exit_mm() can cause footprint + diff to
3938 			 * go negative. Clamp to zero to prevent the
3939 			 * unsigned footprint from wrapping.
3940 			 */
3941 			scoped_guard(rcu) {
3942 				grp = rcu_dereference(p->sched_cache_grp);
3943 
3944 				if (grp) {
3945 					new_fp = (long)READ_ONCE(grp->footprint) + diff;
3946 					WRITE_ONCE(grp->footprint, max(new_fp, 0L));
3947 				}
3948 			}
3949 #endif
3950 		}
3951 
3952 		if (!ng) {
3953 			if (faults > max_faults) {
3954 				max_faults = faults;
3955 				max_nid = nid;
3956 			}
3957 		} else if (group_faults > max_faults) {
3958 			max_faults = group_faults;
3959 			max_nid = nid;
3960 		}
3961 	}
3962 
3963 	/* Cannot migrate task to CPU-less node */
3964 	max_nid = numa_nearest_node(max_nid, N_CPU);
3965 
3966 	if (ng) {
3967 		numa_group_count_active_nodes(ng);
3968 		spin_unlock_irq(group_lock);
3969 		max_nid = preferred_group_nid(p, max_nid);
3970 	}
3971 
3972 	if (max_faults) {
3973 		/* Set the new preferred node */
3974 		if (max_nid != p->numa_preferred_nid)
3975 			sched_setnuma(p, max_nid);
3976 	}
3977 
3978 	update_task_scan_period(p, fault_types[0], fault_types[1]);
3979 }
3980 
3981 static inline int get_numa_group(struct numa_group *grp)
3982 {
3983 	return refcount_inc_not_zero(&grp->refcount);
3984 }
3985 
3986 static inline void put_numa_group(struct numa_group *grp)
3987 {
3988 	if (refcount_dec_and_test(&grp->refcount))
3989 		kfree_rcu(grp, rcu);
3990 }
3991 
3992 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
3993 			int *priv)
3994 {
3995 	struct numa_group *grp, *my_grp;
3996 	struct task_struct *tsk;
3997 	bool join = false;
3998 	int cpu = cpupid_to_cpu(cpupid);
3999 	int i;
4000 
4001 	if (unlikely(!deref_curr_numa_group(p))) {
4002 		unsigned int size = sizeof(struct numa_group) +
4003 				    NR_NUMA_HINT_FAULT_STATS *
4004 				    nr_node_ids * sizeof(unsigned long);
4005 
4006 		grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
4007 		if (!grp)
4008 			return;
4009 
4010 		refcount_set(&grp->refcount, 1);
4011 		grp->active_nodes = 1;
4012 		grp->max_faults_cpu = 0;
4013 		spin_lock_init(&grp->lock);
4014 		grp->gid = p->pid;
4015 
4016 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
4017 			grp->faults[i] = p->numa_faults[i];
4018 
4019 		grp->total_faults = p->total_numa_faults;
4020 
4021 		grp->nr_tasks++;
4022 		rcu_assign_pointer(p->numa_group, grp);
4023 	}
4024 
4025 	rcu_read_lock();
4026 	tsk = READ_ONCE(cpu_rq(cpu)->curr);
4027 
4028 	if (!cpupid_match_pid(tsk, cpupid))
4029 		goto no_join;
4030 
4031 	grp = rcu_dereference_all(tsk->numa_group);
4032 	if (!grp)
4033 		goto no_join;
4034 
4035 	my_grp = deref_curr_numa_group(p);
4036 	if (grp == my_grp)
4037 		goto no_join;
4038 
4039 	/*
4040 	 * Only join the other group if its bigger; if we're the bigger group,
4041 	 * the other task will join us.
4042 	 */
4043 	if (my_grp->nr_tasks > grp->nr_tasks)
4044 		goto no_join;
4045 
4046 	/*
4047 	 * Tie-break on the grp address.
4048 	 */
4049 	if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
4050 		goto no_join;
4051 
4052 	/* Always join threads in the same process. */
4053 	if (tsk->mm == current->mm)
4054 		join = true;
4055 
4056 	/* Simple filter to avoid false positives due to PID collisions */
4057 	if (flags & TNF_SHARED)
4058 		join = true;
4059 
4060 	/* Update priv based on whether false sharing was detected */
4061 	*priv = !join;
4062 
4063 	if (join && !get_numa_group(grp))
4064 		goto no_join;
4065 
4066 	rcu_read_unlock();
4067 
4068 	if (!join)
4069 		return;
4070 
4071 	WARN_ON_ONCE(irqs_disabled());
4072 	double_lock_irq(&my_grp->lock, &grp->lock);
4073 
4074 	for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
4075 		my_grp->faults[i] -= p->numa_faults[i];
4076 		grp->faults[i] += p->numa_faults[i];
4077 	}
4078 	my_grp->total_faults -= p->total_numa_faults;
4079 	grp->total_faults += p->total_numa_faults;
4080 
4081 	my_grp->nr_tasks--;
4082 	grp->nr_tasks++;
4083 
4084 	spin_unlock(&my_grp->lock);
4085 	spin_unlock_irq(&grp->lock);
4086 
4087 	rcu_assign_pointer(p->numa_group, grp);
4088 
4089 	put_numa_group(my_grp);
4090 	return;
4091 
4092 no_join:
4093 	rcu_read_unlock();
4094 	return;
4095 }
4096 
4097 /*
4098  * Get rid of NUMA statistics associated with a task (either current or dead).
4099  * If @final is set, the task is dead and has reached refcount zero, so we can
4100  * safely free all relevant data structures. Otherwise, there might be
4101  * concurrent reads from places like load balancing and procfs, and we should
4102  * reset the data back to default state without freeing ->numa_faults.
4103  */
4104 void task_numa_free(struct task_struct *p, bool final)
4105 {
4106 	/* safe: p either is current or is being freed by current */
4107 	struct numa_group *grp = rcu_dereference_raw(p->numa_group);
4108 	unsigned long *numa_faults = p->numa_faults;
4109 	unsigned long flags;
4110 	int i;
4111 
4112 	if (!numa_faults)
4113 		return;
4114 
4115 	if (grp) {
4116 		spin_lock_irqsave(&grp->lock, flags);
4117 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
4118 			grp->faults[i] -= p->numa_faults[i];
4119 		grp->total_faults -= p->total_numa_faults;
4120 
4121 		grp->nr_tasks--;
4122 		spin_unlock_irqrestore(&grp->lock, flags);
4123 		RCU_INIT_POINTER(p->numa_group, NULL);
4124 		put_numa_group(grp);
4125 	}
4126 
4127 	if (final) {
4128 		p->numa_faults = NULL;
4129 		kfree(numa_faults);
4130 	} else {
4131 		p->total_numa_faults = 0;
4132 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
4133 			numa_faults[i] = 0;
4134 	}
4135 }
4136 
4137 /*
4138  * Got a PROT_NONE fault for a page on @node.
4139  */
4140 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
4141 {
4142 	struct task_struct *p = current;
4143 	bool migrated = flags & TNF_MIGRATED;
4144 	int cpu_node = task_node(current);
4145 	int local = !!(flags & TNF_FAULT_LOCAL);
4146 	struct numa_group *ng;
4147 	int priv;
4148 
4149 	if (!static_branch_likely(&sched_numa_balancing))
4150 		return;
4151 
4152 	/* for example, ksmd faulting in a user's mm */
4153 	if (!p->mm)
4154 		return;
4155 
4156 	/*
4157 	 * NUMA faults statistics are unnecessary for the slow memory
4158 	 * node for memory tiering mode.
4159 	 */
4160 	if (!node_is_toptier(mem_node) &&
4161 	    (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
4162 	     !cpupid_valid(last_cpupid)))
4163 		return;
4164 
4165 	/* Allocate buffer to track faults on a per-node basis */
4166 	if (unlikely(!p->numa_faults)) {
4167 		int size = sizeof(*p->numa_faults) *
4168 			   NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
4169 
4170 		p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
4171 		if (!p->numa_faults)
4172 			return;
4173 
4174 		p->total_numa_faults = 0;
4175 		memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
4176 	}
4177 
4178 	/*
4179 	 * First accesses are treated as private, otherwise consider accesses
4180 	 * to be private if the accessing pid has not changed
4181 	 */
4182 	if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
4183 		priv = 1;
4184 	} else {
4185 		priv = cpupid_match_pid(p, last_cpupid);
4186 		if (!priv && !(flags & TNF_NO_GROUP))
4187 			task_numa_group(p, last_cpupid, flags, &priv);
4188 	}
4189 
4190 	/*
4191 	 * If a workload spans multiple NUMA nodes, a shared fault that
4192 	 * occurs wholly within the set of nodes that the workload is
4193 	 * actively using should be counted as local. This allows the
4194 	 * scan rate to slow down when a workload has settled down.
4195 	 */
4196 	ng = deref_curr_numa_group(p);
4197 	if (!priv && !local && ng && ng->active_nodes > 1 &&
4198 				numa_is_active_node(cpu_node, ng) &&
4199 				numa_is_active_node(mem_node, ng))
4200 		local = 1;
4201 
4202 	/*
4203 	 * Retry to migrate task to preferred node periodically, in case it
4204 	 * previously failed, or the scheduler moved us.
4205 	 */
4206 	if (time_after(jiffies, p->numa_migrate_retry)) {
4207 		task_numa_placement(p);
4208 		numa_migrate_preferred(p);
4209 	}
4210 
4211 	if (migrated)
4212 		p->numa_pages_migrated += pages;
4213 	if (flags & TNF_MIGRATE_FAIL)
4214 		p->numa_faults_locality[2] += pages;
4215 
4216 	p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
4217 	p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
4218 	p->numa_faults_locality[local] += pages;
4219 }
4220 
4221 static void reset_ptenuma_scan(struct task_struct *p)
4222 {
4223 	/*
4224 	 * We only did a read acquisition of the mmap sem, so
4225 	 * p->mm->numa_scan_seq is written to without exclusive access
4226 	 * and the update is not guaranteed to be atomic. That's not
4227 	 * much of an issue though, since this is just used for
4228 	 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
4229 	 * expensive, to avoid any form of compiler optimizations:
4230 	 */
4231 	WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
4232 	p->mm->numa_scan_offset = 0;
4233 }
4234 
4235 static bool vma_is_accessed(struct mm_struct *mm, struct vm_area_struct *vma)
4236 {
4237 	unsigned long pids;
4238 	/*
4239 	 * Allow unconditional access first two times, so that all the (pages)
4240 	 * of VMAs get prot_none fault introduced irrespective of accesses.
4241 	 * This is also done to avoid any side effect of task scanning
4242 	 * amplifying the unfairness of disjoint set of VMAs' access.
4243 	 */
4244 	if ((READ_ONCE(current->mm->numa_scan_seq) - vma->numab_state->start_scan_seq) < 2)
4245 		return true;
4246 
4247 	pids = vma->numab_state->pids_active[0] | vma->numab_state->pids_active[1];
4248 	if (test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids))
4249 		return true;
4250 
4251 	/*
4252 	 * Complete a scan that has already started regardless of PID access, or
4253 	 * some VMAs may never be scanned in multi-threaded applications:
4254 	 */
4255 	if (mm->numa_scan_offset > vma->vm_start) {
4256 		trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_IGNORE_PID);
4257 		return true;
4258 	}
4259 
4260 	/*
4261 	 * This vma has not been accessed for a while, and if the number
4262 	 * the threads in the same process is low, which means no other
4263 	 * threads can help scan this vma, force a vma scan.
4264 	 */
4265 	if (READ_ONCE(mm->numa_scan_seq) >
4266 	   (vma->numab_state->prev_scan_seq + get_nr_threads(current)))
4267 		return true;
4268 
4269 	return false;
4270 }
4271 
4272 #define VMA_PID_RESET_PERIOD (4 * sysctl_numa_balancing_scan_delay)
4273 
4274 /*
4275  * The expensive part of numa migration is done from task_work context.
4276  * Triggered from task_tick_numa().
4277  */
4278 static void task_numa_work(struct callback_head *work)
4279 {
4280 	unsigned long migrate, next_scan, now = jiffies;
4281 	struct task_struct *p = current;
4282 	struct mm_struct *mm = p->mm;
4283 	u64 runtime = p->se.sum_exec_runtime;
4284 	struct vm_area_struct *vma;
4285 	unsigned long start, end;
4286 	unsigned long nr_pte_updates = 0;
4287 	long pages, virtpages;
4288 	struct vma_iterator vmi;
4289 	bool vma_pids_skipped;
4290 	bool vma_pids_forced = false;
4291 
4292 	WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
4293 
4294 	work->next = work;
4295 	/*
4296 	 * Who cares about NUMA placement when they're dying.
4297 	 *
4298 	 * NOTE: make sure not to dereference p->mm before this check,
4299 	 * exit_task_work() happens _after_ exit_mm() so we could be called
4300 	 * without p->mm even though we still had it when we enqueued this
4301 	 * work.
4302 	 */
4303 	if (p->flags & PF_EXITING)
4304 		return;
4305 
4306 	/*
4307 	 * Memory is pinned to only one NUMA node via cpuset.mems, naturally
4308 	 * no page can be migrated.
4309 	 */
4310 	if (cpusets_enabled() && nodes_weight(cpuset_current_mems_allowed) == 1) {
4311 		trace_sched_skip_cpuset_numa(current, &cpuset_current_mems_allowed);
4312 		return;
4313 	}
4314 
4315 	if (!mm->numa_next_scan) {
4316 		mm->numa_next_scan = now +
4317 			msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
4318 	}
4319 
4320 	/*
4321 	 * Enforce maximal scan/migration frequency..
4322 	 */
4323 	migrate = mm->numa_next_scan;
4324 	if (time_before(now, migrate))
4325 		return;
4326 
4327 	if (p->numa_scan_period == 0) {
4328 		p->numa_scan_period_max = task_scan_max(p);
4329 		p->numa_scan_period = task_scan_start(p);
4330 	}
4331 
4332 	next_scan = now + msecs_to_jiffies(p->numa_scan_period);
4333 	if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
4334 		return;
4335 
4336 	/*
4337 	 * Delay this task enough that another task of this mm will likely win
4338 	 * the next time around.
4339 	 */
4340 	p->node_stamp += 2 * TICK_NSEC;
4341 
4342 	pages = sysctl_numa_balancing_scan_size;
4343 	pages <<= 20 - PAGE_SHIFT; /* MB in pages */
4344 	virtpages = pages * 8;	   /* Scan up to this much virtual space */
4345 	if (!pages)
4346 		return;
4347 
4348 
4349 	if (!mmap_read_trylock(mm))
4350 		return;
4351 
4352 	/*
4353 	 * VMAs are skipped if the current PID has not trapped a fault within
4354 	 * the VMA recently. Allow scanning to be forced if there is no
4355 	 * suitable VMA remaining.
4356 	 */
4357 	vma_pids_skipped = false;
4358 
4359 retry_pids:
4360 	start = mm->numa_scan_offset;
4361 	vma_iter_init(&vmi, mm, start);
4362 	vma = vma_next(&vmi);
4363 	if (!vma) {
4364 		reset_ptenuma_scan(p);
4365 		start = 0;
4366 		vma_iter_set(&vmi, start);
4367 		vma = vma_next(&vmi);
4368 	}
4369 
4370 	for (; vma; vma = vma_next(&vmi)) {
4371 		if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
4372 			vma_is_hugetlb(vma) || !vma_is_mm_managed(vma)) {
4373 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_UNSUITABLE);
4374 			continue;
4375 		}
4376 
4377 		/*
4378 		 * Shared library pages mapped by multiple processes are not
4379 		 * migrated as it is expected they are cache replicated. Avoid
4380 		 * hinting faults in read-only file-backed mappings or the vDSO
4381 		 * as migrating the pages will be of marginal benefit.
4382 		 */
4383 		if (!vma->vm_mm ||
4384 		    (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) {
4385 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SHARED_RO);
4386 			continue;
4387 		}
4388 
4389 		/*
4390 		 * Skip inaccessible VMAs to avoid any confusion between
4391 		 * PROT_NONE and NUMA hinting PTEs
4392 		 */
4393 		if (!vma_is_accessible(vma)) {
4394 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_INACCESSIBLE);
4395 			continue;
4396 		}
4397 
4398 		/* Initialise new per-VMA NUMAB state. */
4399 		if (!vma->numab_state) {
4400 			struct vma_numab_state *ptr;
4401 
4402 			ptr = kzalloc_obj(*ptr);
4403 			if (!ptr)
4404 				continue;
4405 
4406 			if (cmpxchg(&vma->numab_state, NULL, ptr)) {
4407 				kfree(ptr);
4408 				continue;
4409 			}
4410 
4411 			vma->numab_state->start_scan_seq = mm->numa_scan_seq;
4412 
4413 			vma->numab_state->next_scan = now +
4414 				msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
4415 
4416 			/* Reset happens after 4 times scan delay of scan start */
4417 			vma->numab_state->pids_active_reset =  vma->numab_state->next_scan +
4418 				msecs_to_jiffies(VMA_PID_RESET_PERIOD);
4419 
4420 			/*
4421 			 * Ensure prev_scan_seq does not match numa_scan_seq,
4422 			 * to prevent VMAs being skipped prematurely on the
4423 			 * first scan:
4424 			 */
4425 			 vma->numab_state->prev_scan_seq = mm->numa_scan_seq - 1;
4426 		}
4427 
4428 		/*
4429 		 * Scanning the VMAs of short lived tasks add more overhead. So
4430 		 * delay the scan for new VMAs.
4431 		 */
4432 		if (mm->numa_scan_seq && time_before(jiffies,
4433 						vma->numab_state->next_scan)) {
4434 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SCAN_DELAY);
4435 			continue;
4436 		}
4437 
4438 		/* RESET access PIDs regularly for old VMAs. */
4439 		if (mm->numa_scan_seq &&
4440 				time_after(jiffies, vma->numab_state->pids_active_reset)) {
4441 			vma->numab_state->pids_active_reset = vma->numab_state->pids_active_reset +
4442 				msecs_to_jiffies(VMA_PID_RESET_PERIOD);
4443 			vma->numab_state->pids_active[0] = READ_ONCE(vma->numab_state->pids_active[1]);
4444 			vma->numab_state->pids_active[1] = 0;
4445 		}
4446 
4447 		/* Do not rescan VMAs twice within the same sequence. */
4448 		if (vma->numab_state->prev_scan_seq == mm->numa_scan_seq) {
4449 			mm->numa_scan_offset = vma->vm_end;
4450 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SEQ_COMPLETED);
4451 			continue;
4452 		}
4453 
4454 		/*
4455 		 * Do not scan the VMA if task has not accessed it, unless no other
4456 		 * VMA candidate exists.
4457 		 */
4458 		if (!vma_pids_forced && !vma_is_accessed(mm, vma)) {
4459 			vma_pids_skipped = true;
4460 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_PID_INACTIVE);
4461 			continue;
4462 		}
4463 
4464 		do {
4465 			start = max(start, vma->vm_start);
4466 			end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
4467 			end = min(end, vma->vm_end);
4468 			nr_pte_updates = change_prot_numa(vma, start, end);
4469 
4470 			/*
4471 			 * Try to scan sysctl_numa_balancing_size worth of
4472 			 * hpages that have at least one present PTE that
4473 			 * is not already PTE-numa. If the VMA contains
4474 			 * areas that are unused or already full of prot_numa
4475 			 * PTEs, scan up to virtpages, to skip through those
4476 			 * areas faster.
4477 			 */
4478 			if (nr_pte_updates)
4479 				pages -= (end - start) >> PAGE_SHIFT;
4480 			virtpages -= (end - start) >> PAGE_SHIFT;
4481 
4482 			start = end;
4483 			if (pages <= 0 || virtpages <= 0)
4484 				goto out;
4485 
4486 			cond_resched();
4487 		} while (end != vma->vm_end);
4488 
4489 		/* VMA scan is complete, do not scan until next sequence. */
4490 		vma->numab_state->prev_scan_seq = mm->numa_scan_seq;
4491 
4492 		/*
4493 		 * Only force scan within one VMA at a time, to limit the
4494 		 * cost of scanning a potentially uninteresting VMA.
4495 		 */
4496 		if (vma_pids_forced)
4497 			break;
4498 	}
4499 
4500 	/*
4501 	 * If no VMAs are remaining and VMAs were skipped due to the PID
4502 	 * not accessing the VMA previously, then force a scan to ensure
4503 	 * forward progress:
4504 	 */
4505 	if (!vma && !vma_pids_forced && vma_pids_skipped) {
4506 		vma_pids_forced = true;
4507 		goto retry_pids;
4508 	}
4509 
4510 out:
4511 	/*
4512 	 * It is possible to reach the end of the VMA list but the last few
4513 	 * VMAs are not guaranteed to the vma_migratable. If they are not, we
4514 	 * would find the !migratable VMA on the next scan but not reset the
4515 	 * scanner to the start so check it now.
4516 	 */
4517 	if (vma)
4518 		mm->numa_scan_offset = start;
4519 	else
4520 		reset_ptenuma_scan(p);
4521 	mmap_read_unlock(mm);
4522 
4523 	/*
4524 	 * Make sure tasks use at least 32x as much time to run other code
4525 	 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
4526 	 * Usually update_task_scan_period slows down scanning enough; on an
4527 	 * overloaded system we need to limit overhead on a per task basis.
4528 	 */
4529 	if (unlikely(p->se.sum_exec_runtime != runtime)) {
4530 		u64 diff = p->se.sum_exec_runtime - runtime;
4531 		p->node_stamp += 32 * diff;
4532 	}
4533 }
4534 
4535 void init_numa_balancing(u64 clone_flags, struct task_struct *p)
4536 {
4537 	int mm_users = 0;
4538 	struct mm_struct *mm = p->mm;
4539 
4540 	if (mm) {
4541 		mm_users = atomic_read(&mm->mm_users);
4542 		if (mm_users == 1) {
4543 			mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
4544 			mm->numa_scan_seq = 0;
4545 		}
4546 	}
4547 	p->node_stamp			= 0;
4548 	p->numa_scan_seq		= mm ? mm->numa_scan_seq : 0;
4549 	p->numa_scan_period		= sysctl_numa_balancing_scan_delay;
4550 	p->numa_migrate_retry		= 0;
4551 	/* Protect against double add, see task_tick_numa and task_numa_work */
4552 	p->numa_work.next		= &p->numa_work;
4553 	p->numa_faults			= NULL;
4554 	p->numa_pages_migrated		= 0;
4555 	p->total_numa_faults		= 0;
4556 	RCU_INIT_POINTER(p->numa_group, NULL);
4557 	p->last_task_numa_placement	= 0;
4558 	p->last_sum_exec_runtime	= 0;
4559 
4560 	init_task_work(&p->numa_work, task_numa_work);
4561 
4562 	/* New address space, reset the preferred nid */
4563 	if (!(clone_flags & CLONE_VM)) {
4564 		p->numa_preferred_nid = NUMA_NO_NODE;
4565 		return;
4566 	}
4567 
4568 	/*
4569 	 * New thread, keep existing numa_preferred_nid which should be copied
4570 	 * already by arch_dup_task_struct but stagger when scans start.
4571 	 */
4572 	if (mm) {
4573 		unsigned int delay;
4574 
4575 		delay = min_t(unsigned int, task_scan_max(current),
4576 			current->numa_scan_period * mm_users * NSEC_PER_MSEC);
4577 		delay += 2 * TICK_NSEC;
4578 		p->node_stamp = delay;
4579 	}
4580 }
4581 
4582 /*
4583  * Drive the periodic memory faults..
4584  */
4585 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
4586 {
4587 	struct callback_head *work = &curr->numa_work;
4588 	u64 period, now;
4589 
4590 	/*
4591 	 * We don't care about NUMA placement if we don't have memory.
4592 	 */
4593 	if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
4594 		return;
4595 
4596 	/*
4597 	 * Using runtime rather than walltime has the dual advantage that
4598 	 * we (mostly) drive the selection from busy threads and that the
4599 	 * task needs to have done some actual work before we bother with
4600 	 * NUMA placement.
4601 	 */
4602 	now = curr->se.sum_exec_runtime;
4603 	period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
4604 
4605 	if (now > curr->node_stamp + period) {
4606 		if (!curr->node_stamp)
4607 			curr->numa_scan_period = task_scan_start(curr);
4608 		curr->node_stamp += period;
4609 
4610 		if (!time_before(jiffies, curr->mm->numa_next_scan))
4611 			task_work_add(curr, work, TWA_RESUME);
4612 	}
4613 }
4614 
4615 static void update_scan_period(struct task_struct *p, int new_cpu)
4616 {
4617 	int src_nid = cpu_to_node(task_cpu(p));
4618 	int dst_nid = cpu_to_node(new_cpu);
4619 
4620 	if (!static_branch_likely(&sched_numa_balancing))
4621 		return;
4622 
4623 	if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
4624 		return;
4625 
4626 	if (src_nid == dst_nid)
4627 		return;
4628 
4629 	/*
4630 	 * Allow resets if faults have been trapped before one scan
4631 	 * has completed. This is most likely due to a new task that
4632 	 * is pulled cross-node due to wakeups or load balancing.
4633 	 */
4634 	if (p->numa_scan_seq) {
4635 		/*
4636 		 * Avoid scan adjustments if moving to the preferred
4637 		 * node or if the task was not previously running on
4638 		 * the preferred node.
4639 		 */
4640 		if (dst_nid == p->numa_preferred_nid ||
4641 		    (p->numa_preferred_nid != NUMA_NO_NODE &&
4642 			src_nid != p->numa_preferred_nid))
4643 			return;
4644 	}
4645 
4646 	p->numa_scan_period = task_scan_start(p);
4647 }
4648 
4649 #else /* !CONFIG_NUMA_BALANCING: */
4650 
4651 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
4652 {
4653 }
4654 
4655 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
4656 {
4657 }
4658 
4659 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
4660 {
4661 }
4662 
4663 static inline void update_scan_period(struct task_struct *p, int new_cpu)
4664 {
4665 }
4666 
4667 #endif /* !CONFIG_NUMA_BALANCING */
4668 
4669 static void
4670 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
4671 {
4672 	WARN_ON_ONCE(cfs_rq != cfs_rq_of(se));
4673 	update_load_add(&cfs_rq->load, se->load.weight);
4674 	if (entity_is_task(se)) {
4675 		struct task_struct *p = task_of(se);
4676 		struct rq *rq = rq_of(cfs_rq);
4677 
4678 		account_numa_enqueue(rq, p);
4679 		account_llc_enqueue(rq, p);
4680 		list_add(&se->group_node, &rq->cfs_tasks);
4681 	}
4682 	cfs_rq->nr_queued++;
4683 }
4684 
4685 static void
4686 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
4687 {
4688 	WARN_ON_ONCE(cfs_rq != cfs_rq_of(se));
4689 	update_load_sub(&cfs_rq->load, se->load.weight);
4690 	if (entity_is_task(se)) {
4691 		struct task_struct *p = task_of(se);
4692 		struct rq *rq = rq_of(cfs_rq);
4693 
4694 		account_numa_dequeue(rq, p);
4695 		account_llc_dequeue(rq, p);
4696 		list_del_init(&se->group_node);
4697 	}
4698 	cfs_rq->nr_queued--;
4699 }
4700 
4701 /*
4702  * Signed add and clamp on underflow.
4703  *
4704  * Explicitly do a load-store to ensure the intermediate value never hits
4705  * memory. This allows lockless observations without ever seeing the negative
4706  * values.
4707  */
4708 #define add_positive(_ptr, _val) do {                           \
4709 	typeof(_ptr) ptr = (_ptr);                              \
4710 	__signed_scalar_typeof(*ptr) val = (_val);              \
4711 	typeof(*ptr) res, var = READ_ONCE(*ptr);                \
4712 								\
4713 	res = var + val;                                        \
4714 								\
4715 	if (val < 0 && res > var)                               \
4716 		res = 0;                                        \
4717 								\
4718 	WRITE_ONCE(*ptr, res);                                  \
4719 } while (0)
4720 
4721 /*
4722  * Remove and clamp on negative, from a local variable.
4723  *
4724  * A variant of sub_positive(), which does not use explicit load-store
4725  * and is thus optimized for local variable updates.
4726  */
4727 #define lsub_positive(_ptr, _val) do {				\
4728 	typeof(_ptr) ptr = (_ptr);				\
4729 	*ptr -= min_t(typeof(*ptr), *ptr, _val);		\
4730 } while (0)
4731 
4732 
4733 /*
4734  * Because of rounding, se->util_sum might ends up being +1 more than
4735  * cfs->util_sum. Although this is not a problem by itself, detaching
4736  * a lot of tasks with the rounding problem between 2 updates of
4737  * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4738  * cfs_util_avg is not.
4739  *
4740  * Check that util_sum is still above its lower bound for the new
4741  * util_avg. Given that period_contrib might have moved since the last
4742  * sync, we are only sure that util_sum must be above or equal to
4743  *    util_avg * minimum possible divider
4744  */
4745 #define __update_sa(sa, name, delta_avg, delta_sum) do {	\
4746 	add_positive(&(sa)->name##_avg, delta_avg);		\
4747 	add_positive(&(sa)->name##_sum, delta_sum);		\
4748 	(sa)->name##_sum = max_t(typeof((sa)->name##_sum),	\
4749 			       (sa)->name##_sum,		\
4750 			       (sa)->name##_avg * PELT_MIN_DIVIDER); \
4751 } while (0)
4752 
4753 static inline void
4754 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4755 {
4756 	__update_sa(&cfs_rq->avg, load, se->avg.load_avg,
4757 		    se_weight(se) * se->avg.load_sum);
4758 }
4759 
4760 static inline void
4761 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4762 {
4763 	__update_sa(&cfs_rq->avg, load, -se->avg.load_avg,
4764 		    se_weight(se) * -se->avg.load_sum);
4765 }
4766 
4767 static void
4768 rescale_entity(struct sched_entity *se, unsigned long weight, bool rel_vprot)
4769 {
4770 	long old_weight = se->h_load.weight;
4771 
4772 	/*
4773 	 * VRUNTIME
4774 	 * --------
4775 	 *
4776 	 * COROLLARY #1: The virtual runtime of the entity needs to be
4777 	 * adjusted if re-weight at !0-lag point.
4778 	 *
4779 	 * Proof: For contradiction assume this is not true, so we can
4780 	 * re-weight without changing vruntime at !0-lag point.
4781 	 *
4782 	 *             Weight	VRuntime   Avg-VRuntime
4783 	 *     before    w          v            V
4784 	 *      after    w'         v'           V'
4785 	 *
4786 	 * Since lag needs to be preserved through re-weight:
4787 	 *
4788 	 *	lag = (V - v)*w = (V'- v')*w', where v = v'
4789 	 *	==>	V' = (V - v)*w/w' + v		(1)
4790 	 *
4791 	 * Let W be the total weight of the entities before reweight,
4792 	 * since V' is the new weighted average of entities:
4793 	 *
4794 	 *	V' = (WV + w'v - wv) / (W + w' - w)	(2)
4795 	 *
4796 	 * by using (1) & (2) we obtain:
4797 	 *
4798 	 *	(WV + w'v - wv) / (W + w' - w) = (V - v)*w/w' + v
4799 	 *	==> (WV-Wv+Wv+w'v-wv)/(W+w'-w) = (V - v)*w/w' + v
4800 	 *	==> (WV - Wv)/(W + w' - w) + v = (V - v)*w/w' + v
4801 	 *	==>	(V - v)*W/(W + w' - w) = (V - v)*w/w' (3)
4802 	 *
4803 	 * Since we are doing at !0-lag point which means V != v, we
4804 	 * can simplify (3):
4805 	 *
4806 	 *	==>	W / (W + w' - w) = w / w'
4807 	 *	==>	Ww' = Ww + ww' - ww
4808 	 *	==>	W * (w' - w) = w * (w' - w)
4809 	 *	==>	W = w	(re-weight indicates w' != w)
4810 	 *
4811 	 * So the cfs_rq contains only one entity, hence vruntime of
4812 	 * the entity @v should always equal to the cfs_rq's weighted
4813 	 * average vruntime @V, which means we will always re-weight
4814 	 * at 0-lag point, thus breach assumption. Proof completed.
4815 	 *
4816 	 *
4817 	 * COROLLARY #2: Re-weight does NOT affect weighted average
4818 	 * vruntime of all the entities.
4819 	 *
4820 	 * Proof: According to corollary #1, Eq. (1) should be:
4821 	 *
4822 	 *	(V - v)*w = (V' - v')*w'
4823 	 *	==>    v' = V' - (V - v)*w/w'		(4)
4824 	 *
4825 	 * According to the weighted average formula, we have:
4826 	 *
4827 	 *	V' = (WV - wv + w'v') / (W - w + w')
4828 	 *	   = (WV - wv + w'(V' - (V - v)w/w')) / (W - w + w')
4829 	 *	   = (WV - wv + w'V' - Vw + wv) / (W - w + w')
4830 	 *	   = (WV + w'V' - Vw) / (W - w + w')
4831 	 *
4832 	 *	==>  V'*(W - w + w') = WV + w'V' - Vw
4833 	 *	==>	V' * (W - w) = (W - w) * V	(5)
4834 	 *
4835 	 * If the entity is the only one in the cfs_rq, then reweight
4836 	 * always occurs at 0-lag point, so V won't change. Or else
4837 	 * there are other entities, hence W != w, then Eq. (5) turns
4838 	 * into V' = V. So V won't change in either case, proof done.
4839 	 *
4840 	 *
4841 	 * So according to corollary #1 & #2, the effect of re-weight
4842 	 * on vruntime should be:
4843 	 *
4844 	 *	v' = V' - (V - v) * w / w'		(4)
4845 	 *	   = V  - (V - v) * w / w'
4846 	 *	   = V  - vl * w / w'
4847 	 *	   = V  - vl'
4848 	 */
4849 	se->vlag = div64_long(se->vlag * old_weight, weight);
4850 
4851 	/*
4852 	 * DEADLINE
4853 	 * --------
4854 	 *
4855 	 * When the weight changes, the virtual time slope changes and
4856 	 * we should adjust the relative virtual deadline accordingly.
4857 	 *
4858 	 *	d' = v' + (d - v)*w/w'
4859 	 *	   = V' - (V - v)*w/w' + (d - v)*w/w'
4860 	 *	   = V  - (V - v)*w/w' + (d - v)*w/w'
4861 	 *	   = V  + (d - V)*w/w'
4862 	 */
4863 	if (se->rel_deadline)
4864 		se->deadline = div64_long(se->deadline * old_weight, weight);
4865 
4866 	if (rel_vprot)
4867 		se->vprot = div64_long(se->vprot * old_weight, weight);
4868 }
4869 
4870 static void reweight_eevdf(struct cfs_rq *cfs_rq, struct sched_entity *se,
4871 			   unsigned long weight, bool on_rq)
4872 {
4873 	bool curr = cfs_rq->curr == se;
4874 	bool rel_vprot = false;
4875 	u64 avruntime = 0;
4876 
4877 	if (se->h_load.weight == weight)
4878 		return;
4879 
4880 	if (on_rq) {
4881 		avruntime = avg_vruntime(cfs_rq);
4882 		se->vlag = entity_lag(cfs_rq, se, avruntime);
4883 		se->deadline -= avruntime;
4884 		se->rel_deadline = 1;
4885 		if (curr && protect_slice(se)) {
4886 			se->vprot -= avruntime;
4887 			rel_vprot = true;
4888 		}
4889 
4890 		cfs_rq->h_nr_queued--;
4891 		if (!curr)
4892 			__dequeue_entity(cfs_rq, se);
4893 	}
4894 
4895 	rescale_entity(se, weight, rel_vprot);
4896 
4897 	update_load_set(&se->h_load, weight);
4898 
4899 	if (on_rq) {
4900 		if (rel_vprot)
4901 			se->vprot += avruntime;
4902 		se->deadline += avruntime;
4903 		se->rel_deadline = 0;
4904 		se->vruntime = avruntime - se->vlag;
4905 
4906 		if (!curr)
4907 			__enqueue_entity(cfs_rq, se);
4908 		cfs_rq->h_nr_queued++;
4909 	}
4910 }
4911 
4912 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
4913 			    unsigned long weight)
4914 {
4915 	if (se->load.weight == weight)
4916 		return;
4917 
4918 	if (se->on_rq) {
4919 		WARN_ON_ONCE(cfs_rq != cfs_rq_of(se));
4920 		update_load_sub(&cfs_rq->load, se->load.weight);
4921 	}
4922 	dequeue_load_avg(cfs_rq, se);
4923 
4924 	update_load_set(&se->load, weight);
4925 
4926 	do {
4927 		u32 divider = get_pelt_divider(&se->avg);
4928 		se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
4929 	} while (0);
4930 
4931 	enqueue_load_avg(cfs_rq, se);
4932 
4933 	if (se->on_rq)
4934 		update_load_add(&cfs_rq->load, se->load.weight);
4935 }
4936 
4937 /*
4938  * weight = NICE_0_LOAD;
4939  * for_each_entity_se(se)
4940  *   weight = __calc_prop_weight(cfs_rq_of(se), se, weight);
4941  */
4942 static __always_inline
4943 unsigned long __calc_prop_weight(struct cfs_rq *cfs_rq, struct sched_entity *se,
4944 				 unsigned long weight)
4945 {
4946 	weight *= se->load.weight;
4947 	if (parent_entity(se))
4948 		weight /= cfs_rq->load.weight;
4949 	else
4950 		weight /= NICE_0_LOAD;
4951 
4952 	return max(weight, MIN_SHARES);
4953 }
4954 
4955 static void reweight_task_fair(struct rq *rq, struct task_struct *p,
4956 			       const struct load_weight *lw)
4957 {
4958 	struct sched_entity *se = &p->se;
4959 	unsigned long weight = NICE_0_LOAD;
4960 
4961 	if (se->on_rq)
4962 		update_curr_fair(rq);
4963 
4964 	reweight_entity(cfs_rq_of(se), se, lw->weight);
4965 	se->load.inv_weight = lw->inv_weight;
4966 
4967 	if (!se->on_rq)
4968 		return;
4969 
4970 	for_each_sched_entity(se)
4971 		weight = __calc_prop_weight(cfs_rq_of(se), se, weight);
4972 
4973 	reweight_eevdf(&rq->cfs, &p->se, weight, p->se.on_rq);
4974 }
4975 
4976 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
4977 
4978 #ifdef CONFIG_FAIR_GROUP_SCHED
4979 /*
4980  * All this does is approximate the hierarchical proportion which includes that
4981  * global sum we all love to hate.
4982  *
4983  * That is, the weight of a group entity, is the proportional share of the
4984  * group weight based on the group runqueue weights. That is:
4985  *
4986  *                     tg->weight * grq->load.weight
4987  *   ge->load.weight = -----------------------------               (1)
4988  *                       \Sum grq->load.weight
4989  *
4990  * Now, because computing that sum is prohibitively expensive to compute (been
4991  * there, done that) we approximate it with this average stuff. The average
4992  * moves slower and therefore the approximation is cheaper and more stable.
4993  *
4994  * So instead of the above, we substitute:
4995  *
4996  *   grq->load.weight -> grq->avg.load_avg                         (2)
4997  *
4998  * which yields the following:
4999  *
5000  *                     tg->weight * grq->avg.load_avg
5001  *   ge->load.weight = ------------------------------              (3)
5002  *                             tg->load_avg
5003  *
5004  * Where: tg->load_avg ~= \Sum grq->avg.load_avg
5005  *
5006  * That is shares_avg, and it is right (given the approximation (2)).
5007  *
5008  * The problem with it is that because the average is slow -- it was designed
5009  * to be exactly that of course -- this leads to transients in boundary
5010  * conditions. In specific, the case where the group was idle and we start the
5011  * one task. It takes time for our CPU's grq->avg.load_avg to build up,
5012  * yielding bad latency etc..
5013  *
5014  * Now, in that special case (1) reduces to:
5015  *
5016  *                     tg->weight * grq->load.weight
5017  *   ge->load.weight = ----------------------------- = tg->weight   (4)
5018  *                         grp->load.weight
5019  *
5020  * That is, the sum collapses because all other CPUs are idle; the UP scenario.
5021  *
5022  * So what we do is modify our approximation (3) to approach (4) in the (near)
5023  * UP case, like:
5024  *
5025  *   ge->load.weight =
5026  *
5027  *              tg->weight * grq->load.weight
5028  *     ---------------------------------------------------         (5)
5029  *     tg->load_avg - grq->avg.load_avg + grq->load.weight
5030  *
5031  * But because grq->load.weight can drop to 0, resulting in a divide by zero,
5032  * we need to use grq->avg.load_avg as its lower bound, which then gives:
5033  *
5034  *
5035  *                     tg->weight * grq->load.weight
5036  *   ge->load.weight = -----------------------------		   (6)
5037  *                             tg_load_avg'
5038  *
5039  * Where:
5040  *
5041  *   tg_load_avg' = tg->load_avg - grq->avg.load_avg +
5042  *                  max(grq->load.weight, grq->avg.load_avg)
5043  *
5044  * And that is shares_weight and is icky. In the (near) UP case it approaches
5045  * (4) while in the normal case it approaches (3). It consistently
5046  * overestimates the ge->load.weight and therefore:
5047  *
5048  *   \Sum ge->load.weight >= tg->weight
5049  *
5050  * hence icky!
5051  */
5052 static long __calc_smp_shares(struct cfs_rq *cfs_rq, long tg_shares, long shares_max)
5053 {
5054 	struct task_group *tg = cfs_rq->tg;
5055 	long tg_weight, load, shares;
5056 
5057 	load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
5058 
5059 	tg_weight = atomic_long_read(&tg->load_avg);
5060 
5061 	/* Ensure tg_weight >= load */
5062 	tg_weight -= cfs_rq->tg_load_avg_contrib;
5063 	tg_weight += load;
5064 
5065 	shares = (tg_shares * load);
5066 	if (tg_weight)
5067 		shares /= tg_weight;
5068 
5069 	/*
5070 	 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
5071 	 * of a group with small tg->shares value. It is a floor value which is
5072 	 * assigned as a minimum load.weight to the sched_entity representing
5073 	 * the group on a CPU.
5074 	 *
5075 	 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
5076 	 * on an 8-core system with 8 tasks each runnable on one CPU shares has
5077 	 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
5078 	 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
5079 	 * instead of 0.
5080 	 */
5081 	return clamp_t(long, shares, MIN_SHARES, shares_max);
5082 }
5083 
5084 static int tg_cpus(struct task_group *tg)
5085 {
5086 	int nr = num_online_cpus();
5087 
5088 	if (cpusets_enabled()) {
5089 		struct cgroup *cgrp = tg->css.cgroup;
5090 		if (cgrp)
5091 			nr = cpuset_num_cpus(cgrp);
5092 	}
5093 
5094 	/*
5095 	 * An empty cpuset would propagate a 0 shares_max into
5096 	 * __calc_smp_shares(), where clamp() yields hi when hi < lo and so
5097 	 * defeats the MIN_SHARES floor. Match tg_tasks(), which floors at 1.
5098 	 */
5099 	return max(nr, 1);
5100 }
5101 
5102 static inline int tg_tasks(struct task_group *tg)
5103 {
5104 	return max(1, atomic_long_read(&tg->runnable_avg) >> SCHED_CAPACITY_SHIFT);
5105 }
5106 
5107 /*
5108  * Func: fraction(nr_tasks * tg->shares)
5109  *
5110  * Scale tg->shares by the number of tasks.
5111  */
5112 static long calc_tasks_shares(struct cfs_rq *cfs_rq)
5113 {
5114 	struct task_group *tg = cfs_rq->tg;
5115 	int nr = tg_tasks(tg);
5116 	long tg_shares = READ_ONCE(tg->shares);
5117 	return __calc_smp_shares(cfs_rq, nr * tg_shares, nr * tg_shares);
5118 }
5119 
5120 /*
5121  * Func: min(fraction(nr_cpus * tg->shares), nice -20)
5122  *
5123  * Scale tg->shares by the maximal number of CPUs; but clip the max shares at
5124  * nice -20, otherwise a single spinner on a 512 CPU machine would result in
5125  * 512*NICE_0_LOAD, which is also crazy.
5126  */
5127 static long calc_max_shares(struct cfs_rq *cfs_rq)
5128 {
5129 	struct task_group *tg = cfs_rq->tg;
5130 	int nr = tg_cpus(tg);
5131 	long tg_shares = READ_ONCE(tg->shares);
5132 	long max_shares = scale_load(sched_prio_to_weight[0]);
5133 	return __calc_smp_shares(cfs_rq, tg_shares * nr, max_shares);
5134 }
5135 
5136 /*
5137  * Func: fraction(nr * tg->shares); nr = min(nr_tasks, nr_cpus)
5138  *
5139  * Scales between "smp" and "max" in a natural way. No longer needs clipping
5140  * since there are no unnatural inflations like with "max".
5141  */
5142 static long calc_concur_shares(struct cfs_rq *cfs_rq)
5143 {
5144 	struct task_group *tg = cfs_rq->tg;
5145 	int nr = min(tg_tasks(tg), tg_cpus(tg));
5146 	long tg_shares = READ_ONCE(tg->shares);
5147 	return __calc_smp_shares(cfs_rq, nr * tg_shares, nr * tg_shares);
5148 }
5149 
5150 /*
5151  * Func: fraction(tg->shares)
5152  *
5153  * This infamously results in tiny shares when you have many CPUs.
5154  */
5155 static long calc_smp_shares(struct cfs_rq *cfs_rq)
5156 {
5157 	struct task_group *tg = cfs_rq->tg;
5158 	long tg_shares = READ_ONCE(tg->shares);
5159 	return __calc_smp_shares(cfs_rq, tg_shares, tg_shares);
5160 }
5161 
5162 /*
5163  * Ignore this pesky SMP stuff, use (4).
5164  */
5165 static long calc_up_shares(struct cfs_rq *cfs_rq)
5166 {
5167 	struct task_group *tg = cfs_rq->tg;
5168 	return READ_ONCE(tg->shares);
5169 }
5170 
5171 DEFINE_STATIC_CALL(calc_group_shares, calc_concur_shares);
5172 
5173 void __sched_cgroup_mode_update(int mode)
5174 {
5175 	long (*func)(struct cfs_rq *);
5176 	switch (mode) {
5177 	case 0:
5178 		func = &calc_up_shares;
5179 		break;
5180 	case 1:
5181 		func = &calc_smp_shares;
5182 		break;
5183 	case 2:
5184 	default:
5185 		func = &calc_concur_shares;
5186 		break;
5187 	case 3:
5188 		func = &calc_max_shares;
5189 		break;
5190 	case 4:
5191 		func = &calc_tasks_shares;
5192 		break;
5193 	}
5194 	static_call_update(calc_group_shares, func);
5195 }
5196 
5197 /*
5198  * Recomputes the group entity based on the current state of its group
5199  * runqueue.
5200  */
5201 static void update_cfs_group(struct sched_entity *se)
5202 {
5203 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
5204 	long shares;
5205 
5206 	/*
5207 	 * When a group becomes empty, preserve its weight. This matters for
5208 	 * DELAY_DEQUEUE.
5209 	 */
5210 	if (!gcfs_rq || !gcfs_rq->load.weight)
5211 		return;
5212 
5213 	shares = static_call(calc_group_shares)(gcfs_rq);
5214 	reweight_entity(cfs_rq_of(se), se, shares);
5215 }
5216 
5217 #else /* !CONFIG_FAIR_GROUP_SCHED: */
5218 static inline void update_cfs_group(struct sched_entity *se)
5219 {
5220 }
5221 #endif /* !CONFIG_FAIR_GROUP_SCHED */
5222 
5223 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
5224 {
5225 	struct rq *rq = rq_of(cfs_rq);
5226 
5227 	if (&rq->cfs == cfs_rq) {
5228 		/*
5229 		 * There are a few boundary cases this might miss but it should
5230 		 * get called often enough that that should (hopefully) not be
5231 		 * a real problem.
5232 		 *
5233 		 * It will not get called when we go idle, because the idle
5234 		 * thread is a different class (!fair), nor will the utilization
5235 		 * number include things like RT tasks.
5236 		 *
5237 		 * As is, the util number is not freq-invariant (we'd have to
5238 		 * implement arch_scale_freq_capacity() for that).
5239 		 *
5240 		 * See cpu_util_cfs().
5241 		 */
5242 		cpufreq_update_util(rq, flags);
5243 	}
5244 }
5245 
5246 static inline bool load_avg_is_decayed(struct sched_avg *sa)
5247 {
5248 	if (sa->load_sum)
5249 		return false;
5250 
5251 	if (sa->util_sum)
5252 		return false;
5253 
5254 	if (sa->runnable_sum)
5255 		return false;
5256 
5257 	/*
5258 	 * _avg must be null when _sum are null because _avg = _sum / divider
5259 	 * Make sure that rounding and/or propagation of PELT values never
5260 	 * break this.
5261 	 */
5262 	WARN_ON_ONCE(sa->load_avg ||
5263 		      sa->util_avg ||
5264 		      sa->runnable_avg);
5265 
5266 	return true;
5267 }
5268 
5269 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
5270 {
5271 	return u64_u32_load_copy(cfs_rq->avg.last_update_time,
5272 				 cfs_rq->last_update_time_copy);
5273 }
5274 #ifdef CONFIG_FAIR_GROUP_SCHED
5275 /*
5276  * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
5277  * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
5278  * bottom-up, we only have to test whether the cfs_rq before us on the list
5279  * is our child.
5280  * If cfs_rq is not on the list, test whether a child needs its to be added to
5281  * connect a branch to the tree  * (see list_add_leaf_cfs_rq() for details).
5282  */
5283 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
5284 {
5285 	struct cfs_rq *prev_cfs_rq;
5286 	struct list_head *prev;
5287 	struct rq *rq = rq_of(cfs_rq);
5288 
5289 	if (cfs_rq->on_list) {
5290 		prev = cfs_rq->leaf_cfs_rq_list.prev;
5291 	} else {
5292 		prev = rq->tmp_alone_branch;
5293 	}
5294 
5295 	if (prev == &rq->leaf_cfs_rq_list)
5296 		return false;
5297 
5298 	prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
5299 
5300 	return (prev_cfs_rq->tg->parent == cfs_rq->tg);
5301 }
5302 
5303 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
5304 {
5305 	if (cfs_rq->load.weight)
5306 		return false;
5307 
5308 	if (!load_avg_is_decayed(&cfs_rq->avg))
5309 		return false;
5310 
5311 	if (child_cfs_rq_on_list(cfs_rq))
5312 		return false;
5313 
5314 	if (cfs_rq->tg_load_avg_contrib)
5315 		return false;
5316 
5317 	return true;
5318 }
5319 
5320 /**
5321  * update_tg_load_avg - update the tg's load avg
5322  * @cfs_rq: the cfs_rq whose avg changed
5323  *
5324  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
5325  * However, because tg->load_avg is a global value there are performance
5326  * considerations.
5327  *
5328  * In order to avoid having to look at the other cfs_rq's, we use a
5329  * differential update where we store the last value we propagated. This in
5330  * turn allows skipping updates if the differential is 'small'.
5331  *
5332  * Updating tg's load_avg is necessary before update_cfs_group().
5333  */
5334 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
5335 {
5336 	long dl, dr;
5337 	u64 now;
5338 
5339 	/*
5340 	 * No need to update load_avg for root_task_group as it is not used.
5341 	 */
5342 	if (cfs_rq->tg == &root_task_group)
5343 		return;
5344 
5345 	/* rq has been offline and doesn't contribute to the share anymore: */
5346 	if (!cpu_active(cpu_of(rq_of(cfs_rq))))
5347 		return;
5348 
5349 	/*
5350 	 * For migration heavy workloads, access to tg->load_avg can be
5351 	 * unbound. Limit the update rate to at most once per ms.
5352 	 */
5353 	now = rq_clock(rq_of(cfs_rq));
5354 	if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC)
5355 		return;
5356 
5357 	dl = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
5358 	dr = cfs_rq->avg.runnable_avg - cfs_rq->tg_runnable_avg_contrib;
5359 	if (abs(dl) > cfs_rq->tg_load_avg_contrib / 64 ||
5360 	    abs(dr) > cfs_rq->tg_runnable_avg_contrib / 64) {
5361 		atomic_long_add(dl, &cfs_rq->tg->load_avg);
5362 		atomic_long_add(dr, &cfs_rq->tg->runnable_avg);
5363 		cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
5364 		cfs_rq->tg_runnable_avg_contrib = cfs_rq->avg.runnable_avg;
5365 		cfs_rq->last_update_tg_load_avg = now;
5366 	}
5367 }
5368 
5369 static inline void clear_tg_load_avg(struct cfs_rq *cfs_rq)
5370 {
5371 	long dl, dr;
5372 	u64 now;
5373 
5374 	/*
5375 	 * No need to update load_avg for root_task_group, as it is not used.
5376 	 */
5377 	if (cfs_rq->tg == &root_task_group)
5378 		return;
5379 
5380 	now = rq_clock(rq_of(cfs_rq));
5381 	dl = 0 - cfs_rq->tg_load_avg_contrib;
5382 	dr = 0 - cfs_rq->tg_runnable_avg_contrib;
5383 	atomic_long_add(dl, &cfs_rq->tg->load_avg);
5384 	atomic_long_add(dr, &cfs_rq->tg->runnable_avg);
5385 	cfs_rq->tg_load_avg_contrib = 0;
5386 	cfs_rq->tg_runnable_avg_contrib = 0;
5387 	cfs_rq->last_update_tg_load_avg = now;
5388 }
5389 
5390 /* CPU offline callback: */
5391 static void __maybe_unused clear_tg_offline_cfs_rqs(struct rq *rq)
5392 {
5393 	struct task_group *tg;
5394 
5395 	lockdep_assert_rq_held(rq);
5396 
5397 	/*
5398 	 * The rq clock has already been updated in
5399 	 * set_rq_offline(), so we should skip updating
5400 	 * the rq clock again in unthrottle_cfs_rq().
5401 	 */
5402 	rq_clock_start_loop_update(rq);
5403 
5404 	guard(rcu)();
5405 
5406 	list_for_each_entry_rcu(tg, &task_groups, list) {
5407 		struct cfs_rq *cfs_rq = tg_cfs_rq(tg, cpu_of(rq));
5408 
5409 		clear_tg_load_avg(cfs_rq);
5410 	}
5411 
5412 	rq_clock_stop_loop_update(rq);
5413 }
5414 
5415 /*
5416  * Called within set_task_rq() right before setting a task's CPU. The
5417  * caller only guarantees p->pi_lock is held; no other assumptions,
5418  * including the state of rq->lock, should be made.
5419  */
5420 void set_task_rq_fair(struct sched_entity *se,
5421 		      struct cfs_rq *prev, struct cfs_rq *next)
5422 {
5423 	u64 p_last_update_time;
5424 	u64 n_last_update_time;
5425 
5426 	if (!sched_feat(ATTACH_AGE_LOAD))
5427 		return;
5428 
5429 	/*
5430 	 * We are supposed to update the task to "current" time, then its up to
5431 	 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
5432 	 * getting what current time is, so simply throw away the out-of-date
5433 	 * time. This will result in the wakee task is less decayed, but giving
5434 	 * the wakee more load sounds not bad.
5435 	 */
5436 	if (!(se->avg.last_update_time && prev))
5437 		return;
5438 
5439 	p_last_update_time = cfs_rq_last_update_time(prev);
5440 	n_last_update_time = cfs_rq_last_update_time(next);
5441 
5442 	__update_load_avg_blocked_se(p_last_update_time, se);
5443 	se->avg.last_update_time = n_last_update_time;
5444 }
5445 
5446 /*
5447  * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
5448  * propagate its contribution. The key to this propagation is the invariant
5449  * that for each group:
5450  *
5451  *   ge->avg == grq->avg						(1)
5452  *
5453  * _IFF_ we look at the pure running and runnable sums. Because they
5454  * represent the very same entity, just at different points in the hierarchy.
5455  *
5456  * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
5457  * and simply copies the running/runnable sum over (but still wrong, because
5458  * the group entity and group rq do not have their PELT windows aligned).
5459  *
5460  * However, update_tg_cfs_load() is more complex. So we have:
5461  *
5462  *   ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg		(2)
5463  *
5464  * And since, like util, the runnable part should be directly transferable,
5465  * the following would _appear_ to be the straight forward approach:
5466  *
5467  *   grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg	(3)
5468  *
5469  * And per (1) we have:
5470  *
5471  *   ge->avg.runnable_avg == grq->avg.runnable_avg
5472  *
5473  * Which gives:
5474  *
5475  *                      ge->load.weight * grq->avg.load_avg
5476  *   ge->avg.load_avg = -----------------------------------		(4)
5477  *                               grq->load.weight
5478  *
5479  * Except that is wrong!
5480  *
5481  * Because while for entities historical weight is not important and we
5482  * really only care about our future and therefore can consider a pure
5483  * runnable sum, runqueues can NOT do this.
5484  *
5485  * We specifically want runqueues to have a load_avg that includes
5486  * historical weights. Those represent the blocked load, the load we expect
5487  * to (shortly) return to us. This only works by keeping the weights as
5488  * integral part of the sum. We therefore cannot decompose as per (3).
5489  *
5490  * Another reason this doesn't work is that runnable isn't a 0-sum entity.
5491  * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
5492  * rq itself is runnable anywhere between 2/3 and 1 depending on how the
5493  * runnable section of these tasks overlap (or not). If they were to perfectly
5494  * align the rq as a whole would be runnable 2/3 of the time. If however we
5495  * always have at least 1 runnable task, the rq as a whole is always runnable.
5496  *
5497  * So we'll have to approximate.. :/
5498  *
5499  * Given the constraint:
5500  *
5501  *   ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
5502  *
5503  * We can construct a rule that adds runnable to a rq by assuming minimal
5504  * overlap.
5505  *
5506  * On removal, we'll assume each task is equally runnable; which yields:
5507  *
5508  *   grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
5509  *
5510  * XXX: only do this for the part of runnable > running ?
5511  *
5512  */
5513 static inline void
5514 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
5515 {
5516 	long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
5517 	u32 new_sum, divider;
5518 
5519 	/* Nothing to update */
5520 	if (!delta_avg)
5521 		return;
5522 
5523 	/*
5524 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
5525 	 * See ___update_load_avg() for details.
5526 	 */
5527 	divider = get_pelt_divider(&cfs_rq->avg);
5528 
5529 	/* Set new sched_entity's utilization */
5530 	se->avg.util_avg = gcfs_rq->avg.util_avg;
5531 	new_sum = se->avg.util_avg * divider;
5532 	delta_sum = (long)new_sum - (long)se->avg.util_sum;
5533 	se->avg.util_sum = new_sum;
5534 
5535 	/* Update parent cfs_rq utilization */
5536 	__update_sa(&cfs_rq->avg, util, delta_avg, delta_sum);
5537 }
5538 
5539 static inline void
5540 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
5541 {
5542 	long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
5543 	u64 new_sum;
5544 	u32 divider;
5545 
5546 	/* Nothing to update */
5547 	if (!delta_avg)
5548 		return;
5549 
5550 	/*
5551 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
5552 	 * See ___update_load_avg() for details.
5553 	 */
5554 	divider = get_pelt_divider(&cfs_rq->avg);
5555 
5556 	/* Set new sched_entity's runnable */
5557 	se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
5558 	new_sum = (u64)se->avg.runnable_avg * divider;
5559 	delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
5560 	se->avg.runnable_sum = new_sum;
5561 
5562 	/* Update parent cfs_rq runnable */
5563 	__update_sa(&cfs_rq->avg, runnable, delta_avg, delta_sum);
5564 }
5565 
5566 static inline void
5567 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
5568 {
5569 	long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
5570 	unsigned long load_avg;
5571 	u64 load_sum = 0;
5572 	s64 delta_sum;
5573 	u32 divider;
5574 
5575 	if (!runnable_sum)
5576 		return;
5577 
5578 	gcfs_rq->prop_runnable_sum = 0;
5579 
5580 	/*
5581 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
5582 	 * See ___update_load_avg() for details.
5583 	 */
5584 	divider = get_pelt_divider(&cfs_rq->avg);
5585 
5586 	if (runnable_sum >= 0) {
5587 		/*
5588 		 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
5589 		 * the CPU is saturated running == runnable.
5590 		 */
5591 		runnable_sum += se->avg.load_sum;
5592 		runnable_sum = min_t(long, runnable_sum, divider);
5593 	} else {
5594 		/*
5595 		 * Estimate the new unweighted runnable_sum of the gcfs_rq by
5596 		 * assuming all tasks are equally runnable.
5597 		 */
5598 		if (scale_load_down(gcfs_rq->load.weight)) {
5599 			load_sum = div_u64(gcfs_rq->avg.load_sum,
5600 				scale_load_down(gcfs_rq->load.weight));
5601 		}
5602 
5603 		/* But make sure to not inflate se's runnable */
5604 		runnable_sum = min(se->avg.load_sum, load_sum);
5605 	}
5606 
5607 	/*
5608 	 * runnable_sum can't be lower than running_sum
5609 	 * Rescale running sum to be in the same range as runnable sum
5610 	 * running_sum is in [0 : LOAD_AVG_MAX <<  SCHED_CAPACITY_SHIFT]
5611 	 * runnable_sum is in [0 : LOAD_AVG_MAX]
5612 	 */
5613 	running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
5614 	runnable_sum = max(runnable_sum, running_sum);
5615 
5616 	load_sum = se_weight(se) * runnable_sum;
5617 	load_avg = div_u64(load_sum, divider);
5618 
5619 	delta_avg = load_avg - se->avg.load_avg;
5620 	if (!delta_avg)
5621 		return;
5622 
5623 	delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
5624 
5625 	se->avg.load_sum = runnable_sum;
5626 	se->avg.load_avg = load_avg;
5627 	__update_sa(&cfs_rq->avg, load, delta_avg, delta_sum);
5628 }
5629 
5630 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
5631 {
5632 	cfs_rq->propagate = 1;
5633 	cfs_rq->prop_runnable_sum += runnable_sum;
5634 }
5635 
5636 /* Update task and its cfs_rq load average */
5637 static inline int propagate_entity_load_avg(struct sched_entity *se)
5638 {
5639 	struct cfs_rq *cfs_rq, *gcfs_rq;
5640 
5641 	if (entity_is_task(se))
5642 		return 0;
5643 
5644 	gcfs_rq = group_cfs_rq(se);
5645 	if (!gcfs_rq->propagate)
5646 		return 0;
5647 
5648 	gcfs_rq->propagate = 0;
5649 
5650 	cfs_rq = cfs_rq_of(se);
5651 
5652 	add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
5653 
5654 	update_tg_cfs_util(cfs_rq, se, gcfs_rq);
5655 	update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
5656 	update_tg_cfs_load(cfs_rq, se, gcfs_rq);
5657 
5658 	trace_pelt_cfs_tp(cfs_rq);
5659 	trace_pelt_se_tp(se);
5660 
5661 	return 1;
5662 }
5663 
5664 /*
5665  * Check if we need to update the load and the utilization of a blocked
5666  * group_entity:
5667  */
5668 static inline bool skip_blocked_update(struct sched_entity *se)
5669 {
5670 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
5671 
5672 	/*
5673 	 * If sched_entity still have not zero load or utilization, we have to
5674 	 * decay it:
5675 	 */
5676 	if (se->avg.load_avg || se->avg.util_avg)
5677 		return false;
5678 
5679 	/*
5680 	 * If there is a pending propagation, we have to update the load and
5681 	 * the utilization of the sched_entity:
5682 	 */
5683 	if (gcfs_rq->propagate)
5684 		return false;
5685 
5686 	/*
5687 	 * Otherwise, the load and the utilization of the sched_entity is
5688 	 * already zero and there is no pending propagation, so it will be a
5689 	 * waste of time to try to decay it:
5690 	 */
5691 	return true;
5692 }
5693 
5694 #else /* !CONFIG_FAIR_GROUP_SCHED: */
5695 
5696 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
5697 
5698 static inline void clear_tg_offline_cfs_rqs(struct rq *rq) {}
5699 
5700 static inline int propagate_entity_load_avg(struct sched_entity *se)
5701 {
5702 	return 0;
5703 }
5704 
5705 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
5706 
5707 #endif /* !CONFIG_FAIR_GROUP_SCHED */
5708 
5709 #ifdef CONFIG_NO_HZ_COMMON
5710 static inline void migrate_se_pelt_lag(struct sched_entity *se)
5711 {
5712 	u64 throttled = 0, now, lut;
5713 	struct cfs_rq *cfs_rq;
5714 	struct rq *rq;
5715 	bool is_idle;
5716 
5717 	if (load_avg_is_decayed(&se->avg))
5718 		return;
5719 
5720 	cfs_rq = cfs_rq_of(se);
5721 	rq = rq_of(cfs_rq);
5722 
5723 	rcu_read_lock();
5724 	is_idle = is_idle_task(rcu_dereference_all(rq->curr));
5725 	rcu_read_unlock();
5726 
5727 	/*
5728 	 * The lag estimation comes with a cost we don't want to pay all the
5729 	 * time. Hence, limiting to the case where the source CPU is idle and
5730 	 * we know we are at the greatest risk to have an outdated clock.
5731 	 */
5732 	if (!is_idle)
5733 		return;
5734 
5735 	/*
5736 	 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
5737 	 *
5738 	 *   last_update_time (the cfs_rq's last_update_time)
5739 	 *	= cfs_rq_clock_pelt()@cfs_rq_idle
5740 	 *      = rq_clock_pelt()@cfs_rq_idle
5741 	 *        - cfs->throttled_clock_pelt_time@cfs_rq_idle
5742 	 *
5743 	 *   cfs_idle_lag (delta between rq's update and cfs_rq's update)
5744 	 *      = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
5745 	 *
5746 	 *   rq_idle_lag (delta between now and rq's update)
5747 	 *      = sched_clock_cpu() - rq_clock()@rq_idle
5748 	 *
5749 	 * We can then write:
5750 	 *
5751 	 *    now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
5752 	 *          sched_clock_cpu() - rq_clock()@rq_idle
5753 	 * Where:
5754 	 *      rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
5755 	 *      rq_clock()@rq_idle      is rq->clock_idle
5756 	 *      cfs->throttled_clock_pelt_time@cfs_rq_idle
5757 	 *                              is cfs_rq->throttled_pelt_idle
5758 	 */
5759 
5760 #ifdef CONFIG_CFS_BANDWIDTH
5761 	throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
5762 	/* The clock has been stopped for throttling */
5763 	if (throttled == U64_MAX)
5764 		return;
5765 #endif
5766 	now = u64_u32_load(rq->clock_pelt_idle);
5767 	/*
5768 	 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
5769 	 * is observed the old clock_pelt_idle value and the new clock_idle,
5770 	 * which lead to an underestimation. The opposite would lead to an
5771 	 * overestimation.
5772 	 */
5773 	smp_rmb();
5774 	lut = cfs_rq_last_update_time(cfs_rq);
5775 
5776 	now -= throttled;
5777 	if (now < lut)
5778 		/*
5779 		 * cfs_rq->avg.last_update_time is more recent than our
5780 		 * estimation, let's use it.
5781 		 */
5782 		now = lut;
5783 	else
5784 		now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
5785 
5786 	__update_load_avg_blocked_se(now, se);
5787 }
5788 #else /* !CONFIG_NO_HZ_COMMON: */
5789 static void migrate_se_pelt_lag(struct sched_entity *se) {}
5790 #endif /* !CONFIG_NO_HZ_COMMON */
5791 
5792 /**
5793  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
5794  * @now: current time, as per cfs_rq_clock_pelt()
5795  * @cfs_rq: cfs_rq to update
5796  *
5797  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
5798  * avg. The immediate corollary is that all (fair) tasks must be attached.
5799  *
5800  * cfs_rq->avg is used for task_h_load() and update_cfs_group() for example.
5801  *
5802  * Return: true if the load decayed or we removed load.
5803  *
5804  * Since both these conditions indicate a changed cfs_rq->avg.load we should
5805  * call update_tg_load_avg() when this function returns true.
5806  */
5807 static inline int
5808 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
5809 {
5810 	unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
5811 	struct sched_avg *sa = &cfs_rq->avg;
5812 	int decayed = 0;
5813 
5814 	if (cfs_rq->removed.nr) {
5815 		unsigned long r;
5816 		u32 divider = get_pelt_divider(&cfs_rq->avg);
5817 
5818 		raw_spin_lock(&cfs_rq->removed.lock);
5819 		swap(cfs_rq->removed.util_avg, removed_util);
5820 		swap(cfs_rq->removed.load_avg, removed_load);
5821 		swap(cfs_rq->removed.runnable_avg, removed_runnable);
5822 		cfs_rq->removed.nr = 0;
5823 		raw_spin_unlock(&cfs_rq->removed.lock);
5824 
5825 		r = removed_load;
5826 		__update_sa(sa, load, -r, -r*divider);
5827 
5828 		r = removed_util;
5829 		__update_sa(sa, util, -r, -r*divider);
5830 
5831 		r = removed_runnable;
5832 		__update_sa(sa, runnable, -r, -r*divider);
5833 
5834 		/*
5835 		 * removed_runnable is the unweighted version of removed_load so we
5836 		 * can use it to estimate removed_load_sum.
5837 		 */
5838 		add_tg_cfs_propagate(cfs_rq,
5839 			-(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
5840 
5841 		decayed = 1;
5842 	}
5843 
5844 	decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
5845 	u64_u32_store_copy(sa->last_update_time,
5846 			   cfs_rq->last_update_time_copy,
5847 			   sa->last_update_time);
5848 	return decayed;
5849 }
5850 
5851 /**
5852  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
5853  * @cfs_rq: cfs_rq to attach to
5854  * @se: sched_entity to attach
5855  *
5856  * Must call update_cfs_rq_load_avg() before this, since we rely on
5857  * cfs_rq->avg.last_update_time being current.
5858  */
5859 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
5860 {
5861 	/*
5862 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
5863 	 * See ___update_load_avg() for details.
5864 	 */
5865 	u32 divider = get_pelt_divider(&cfs_rq->avg);
5866 
5867 	/*
5868 	 * When we attach the @se to the @cfs_rq, we must align the decay
5869 	 * window because without that, really weird and wonderful things can
5870 	 * happen.
5871 	 *
5872 	 * XXX illustrate
5873 	 */
5874 	se->avg.last_update_time = cfs_rq->avg.last_update_time;
5875 	se->avg.period_contrib = cfs_rq->avg.period_contrib;
5876 
5877 	/*
5878 	 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
5879 	 * period_contrib. This isn't strictly correct, but since we're
5880 	 * entirely outside of the PELT hierarchy, nobody cares if we truncate
5881 	 * _sum a little.
5882 	 */
5883 	se->avg.util_sum = se->avg.util_avg * divider;
5884 
5885 	se->avg.runnable_sum = se->avg.runnable_avg * divider;
5886 
5887 	se->avg.load_sum = se->avg.load_avg * divider;
5888 	if (se_weight(se) < se->avg.load_sum)
5889 		se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
5890 	else
5891 		se->avg.load_sum = 1;
5892 
5893 	enqueue_load_avg(cfs_rq, se);
5894 	cfs_rq->avg.util_avg += se->avg.util_avg;
5895 	cfs_rq->avg.util_sum += se->avg.util_sum;
5896 	cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
5897 	cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
5898 
5899 	add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
5900 
5901 	cfs_rq_util_change(cfs_rq, 0);
5902 
5903 	trace_pelt_cfs_tp(cfs_rq);
5904 }
5905 
5906 /**
5907  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
5908  * @cfs_rq: cfs_rq to detach from
5909  * @se: sched_entity to detach
5910  *
5911  * Must call update_cfs_rq_load_avg() before this, since we rely on
5912  * cfs_rq->avg.last_update_time being current.
5913  */
5914 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
5915 {
5916 	dequeue_load_avg(cfs_rq, se);
5917 	__update_sa(&cfs_rq->avg, util, -se->avg.util_avg, -se->avg.util_sum);
5918 	__update_sa(&cfs_rq->avg, runnable, -se->avg.runnable_avg, -se->avg.runnable_sum);
5919 
5920 	add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
5921 
5922 	cfs_rq_util_change(cfs_rq, 0);
5923 
5924 	trace_pelt_cfs_tp(cfs_rq);
5925 }
5926 
5927 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
5928 
5929 static inline void util_est_update(struct sched_entity *se)
5930 {
5931 	unsigned int ewma, dequeued, last_ewma_diff;
5932 
5933 	if (!sched_feat(UTIL_EST))
5934 		return;
5935 
5936 	/* Get current estimate of utilization */
5937 	ewma = READ_ONCE(se->avg.util_est);
5938 
5939 	/*
5940 	 * If the PELT values haven't changed since enqueue time,
5941 	 * skip the util_est update.
5942 	 */
5943 	if (ewma & UTIL_AVG_UNCHANGED)
5944 		return;
5945 
5946 	/* Get utilization at dequeue */
5947 	dequeued = READ_ONCE(se->avg.util_avg);
5948 
5949 	/*
5950 	 * Reset EWMA on utilization increases, the moving average is used only
5951 	 * to smooth utilization decreases.
5952 	 */
5953 	if (ewma <= dequeued) {
5954 		ewma = dequeued;
5955 		goto done;
5956 	}
5957 
5958 	/*
5959 	 * Skip update of task's estimated utilization when its members are
5960 	 * already ~1% close to its last activation value.
5961 	 */
5962 	last_ewma_diff = ewma - dequeued;
5963 	if (last_ewma_diff < UTIL_EST_MARGIN)
5964 		goto done;
5965 
5966 	/*
5967 	 * To avoid underestimate of task utilization, skip updates of EWMA if
5968 	 * we cannot grant that thread got all CPU time it wanted.
5969 	 */
5970 	if ((dequeued + UTIL_EST_MARGIN) < READ_ONCE(se->avg.runnable_avg))
5971 		goto done;
5972 
5973 	/*
5974 	 * Update Task's estimated utilization
5975 	 *
5976 	 * When *p completes an activation we can consolidate another sample
5977 	 * of the task size. This is done by using this value to update the
5978 	 * Exponential Weighted Moving Average (EWMA):
5979 	 *
5980 	 *  ewma(t) = w *  task_util(p) + (1-w) * ewma(t-1)
5981 	 *          = w *  task_util(p) +         ewma(t-1)  - w * ewma(t-1)
5982 	 *          = w * (task_util(p) -         ewma(t-1)) +     ewma(t-1)
5983 	 *          = w * (      -last_ewma_diff           ) +     ewma(t-1)
5984 	 *          = w * (-last_ewma_diff +  ewma(t-1) / w)
5985 	 *
5986 	 * Where 'w' is the weight of new samples, which is configured to be
5987 	 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
5988 	 */
5989 	ewma <<= UTIL_EST_WEIGHT_SHIFT;
5990 	ewma  -= last_ewma_diff;
5991 	ewma >>= UTIL_EST_WEIGHT_SHIFT;
5992 done:
5993 	ewma |= UTIL_AVG_UNCHANGED;
5994 	WRITE_ONCE(se->avg.util_est, ewma);
5995 
5996 	trace_sched_util_est_se_tp(se);
5997 }
5998 
5999 /*
6000  * Optional action to be done while updating the load average
6001  */
6002 #define UPDATE_TG	0x01
6003 #define SKIP_AGE_LOAD	0x02
6004 #define DO_ATTACH	0x04
6005 #define DO_DETACH	0x08
6006 #define UPDATE_UTIL_EST	0x10
6007 
6008 /* Update task and its cfs_rq load average */
6009 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
6010 {
6011 	u64 now = cfs_rq_clock_pelt(cfs_rq);
6012 	int decayed;
6013 
6014 	/*
6015 	 * Track task load average for carrying it to new CPU after migrated, and
6016 	 * track group sched_entity load average for task_h_load calculation in migration
6017 	 */
6018 	if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
6019 		__update_load_avg_se(now, cfs_rq, se);
6020 
6021 	decayed  = update_cfs_rq_load_avg(now, cfs_rq);
6022 	decayed |= propagate_entity_load_avg(se);
6023 
6024 	if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
6025 
6026 		/*
6027 		 * DO_ATTACH means we're here from enqueue_entity().
6028 		 * !last_update_time means we've passed through
6029 		 * migrate_task_rq_fair() indicating we migrated.
6030 		 *
6031 		 * IOW we're enqueueing a task on a new CPU.
6032 		 */
6033 		attach_entity_load_avg(cfs_rq, se);
6034 		update_tg_load_avg(cfs_rq);
6035 
6036 	} else if (flags & DO_DETACH) {
6037 		/*
6038 		 * DO_DETACH means we're here from dequeue_entity()
6039 		 * and we are migrating task out of the CPU.
6040 		 */
6041 		detach_entity_load_avg(cfs_rq, se);
6042 		update_tg_load_avg(cfs_rq);
6043 	} else if (decayed) {
6044 		cfs_rq_util_change(cfs_rq, 0);
6045 
6046 		if (flags & UPDATE_TG)
6047 			update_tg_load_avg(cfs_rq);
6048 	}
6049 
6050 	if (flags & UPDATE_UTIL_EST)
6051 		util_est_update(se);
6052 }
6053 
6054 /*
6055  * Synchronize entity load avg of dequeued entity without locking
6056  * the previous rq.
6057  */
6058 static void sync_entity_load_avg(struct sched_entity *se)
6059 {
6060 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
6061 	u64 last_update_time;
6062 
6063 	last_update_time = cfs_rq_last_update_time(cfs_rq);
6064 	__update_load_avg_blocked_se(last_update_time, se);
6065 }
6066 
6067 /*
6068  * Task first catches up with cfs_rq, and then subtract
6069  * itself from the cfs_rq (task must be off the queue now).
6070  */
6071 static void remove_entity_load_avg(struct sched_entity *se)
6072 {
6073 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
6074 	unsigned long flags;
6075 
6076 	/*
6077 	 * tasks cannot exit without having gone through wake_up_new_task() ->
6078 	 * enqueue_task_fair() which will have added things to the cfs_rq,
6079 	 * so we can remove unconditionally.
6080 	 */
6081 
6082 	sync_entity_load_avg(se);
6083 
6084 	raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
6085 	++cfs_rq->removed.nr;
6086 	cfs_rq->removed.util_avg	+= se->avg.util_avg;
6087 	cfs_rq->removed.load_avg	+= se->avg.load_avg;
6088 	cfs_rq->removed.runnable_avg	+= se->avg.runnable_avg;
6089 	raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
6090 }
6091 
6092 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
6093 {
6094 	return cfs_rq->avg.runnable_avg;
6095 }
6096 
6097 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
6098 {
6099 	return cfs_rq->avg.load_avg;
6100 }
6101 
6102 static int sched_balance_newidle(struct rq *this_rq, struct rq_flags *rf)
6103 	__must_hold(__rq_lockp(this_rq));
6104 
6105 static inline unsigned long task_util(struct task_struct *p)
6106 {
6107 	return READ_ONCE(p->se.avg.util_avg);
6108 }
6109 
6110 static inline unsigned long _task_util_est(struct task_struct *p)
6111 {
6112 	return READ_ONCE(p->se.avg.util_est) & ~UTIL_AVG_UNCHANGED;
6113 }
6114 
6115 static inline unsigned long task_util_est(struct task_struct *p)
6116 {
6117 	return max(task_util(p), _task_util_est(p));
6118 }
6119 
6120 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
6121 				    struct task_struct *p)
6122 {
6123 	unsigned int enqueued;
6124 
6125 	if (!sched_feat(UTIL_EST))
6126 		return;
6127 
6128 	/* Update root cfs_rq's estimated utilization */
6129 	enqueued  = cfs_rq->avg.util_est;
6130 	enqueued += _task_util_est(p);
6131 	WRITE_ONCE(cfs_rq->avg.util_est, enqueued);
6132 
6133 	trace_sched_util_est_cfs_tp(cfs_rq);
6134 }
6135 
6136 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
6137 				    struct task_struct *p)
6138 {
6139 	unsigned int enqueued;
6140 
6141 	if (!sched_feat(UTIL_EST))
6142 		return;
6143 
6144 	/* Update root cfs_rq's estimated utilization */
6145 	enqueued  = cfs_rq->avg.util_est;
6146 	enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
6147 	WRITE_ONCE(cfs_rq->avg.util_est, enqueued);
6148 
6149 	trace_sched_util_est_cfs_tp(cfs_rq);
6150 }
6151 
6152 static inline unsigned long get_actual_cpu_capacity(int cpu)
6153 {
6154 	unsigned long capacity = arch_scale_cpu_capacity(cpu);
6155 
6156 	capacity -= max(hw_load_avg(cpu_rq(cpu)), cpufreq_get_pressure(cpu));
6157 
6158 	return capacity;
6159 }
6160 
6161 static inline int util_fits_cpu(unsigned long util,
6162 				unsigned long uclamp_min,
6163 				unsigned long uclamp_max,
6164 				int cpu)
6165 {
6166 	unsigned long capacity = capacity_of(cpu);
6167 	unsigned long capacity_orig;
6168 	bool fits, uclamp_max_fits;
6169 
6170 	/*
6171 	 * Check if the real util fits without any uclamp boost/cap applied.
6172 	 */
6173 	fits = fits_capacity(util, capacity);
6174 
6175 	if (!uclamp_is_used())
6176 		return fits;
6177 
6178 	/*
6179 	 * We must use arch_scale_cpu_capacity() for comparing against uclamp_min and
6180 	 * uclamp_max. We only care about capacity pressure (by using
6181 	 * capacity_of()) for comparing against the real util.
6182 	 *
6183 	 * If a task is boosted to 1024 for example, we don't want a tiny
6184 	 * pressure to skew the check whether it fits a CPU or not.
6185 	 *
6186 	 * Similarly if a task is capped to arch_scale_cpu_capacity(little_cpu), it
6187 	 * should fit a little cpu even if there's some pressure.
6188 	 *
6189 	 * Only exception is for HW or cpufreq pressure since it has a direct impact
6190 	 * on available OPP of the system.
6191 	 *
6192 	 * We honour it for uclamp_min only as a drop in performance level
6193 	 * could result in not getting the requested minimum performance level.
6194 	 *
6195 	 * For uclamp_max, we can tolerate a drop in performance level as the
6196 	 * goal is to cap the task. So it's okay if it's getting less.
6197 	 */
6198 	capacity_orig = arch_scale_cpu_capacity(cpu);
6199 
6200 	/*
6201 	 * We want to force a task to fit a cpu as implied by uclamp_max.
6202 	 * But we do have some corner cases to cater for..
6203 	 *
6204 	 *
6205 	 *                                 C=z
6206 	 *   |                             ___
6207 	 *   |                  C=y       |   |
6208 	 *   |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _  uclamp_max
6209 	 *   |      C=x        |   |      |   |
6210 	 *   |      ___        |   |      |   |
6211 	 *   |     |   |       |   |      |   |    (util somewhere in this region)
6212 	 *   |     |   |       |   |      |   |
6213 	 *   |     |   |       |   |      |   |
6214 	 *   +----------------------------------------
6215 	 *         CPU0        CPU1       CPU2
6216 	 *
6217 	 *   In the above example if a task is capped to a specific performance
6218 	 *   point, y, then when:
6219 	 *
6220 	 *   * util = 80% of x then it does not fit on CPU0 and should migrate
6221 	 *     to CPU1
6222 	 *   * util = 80% of y then it is forced to fit on CPU1 to honour
6223 	 *     uclamp_max request.
6224 	 *
6225 	 *   which is what we're enforcing here. A task always fits if
6226 	 *   uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
6227 	 *   the normal upmigration rules should withhold still.
6228 	 *
6229 	 *   Only exception is when we are on max capacity, then we need to be
6230 	 *   careful not to block overutilized state. This is so because:
6231 	 *
6232 	 *     1. There's no concept of capping at max_capacity! We can't go
6233 	 *        beyond this performance level anyway.
6234 	 *     2. The system is being saturated when we're operating near
6235 	 *        max capacity, it doesn't make sense to block overutilized.
6236 	 */
6237 	uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
6238 	uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
6239 	fits = fits || uclamp_max_fits;
6240 
6241 	/*
6242 	 *
6243 	 *                                 C=z
6244 	 *   |                             ___       (region a, capped, util >= uclamp_max)
6245 	 *   |                  C=y       |   |
6246 	 *   |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
6247 	 *   |      C=x        |   |      |   |
6248 	 *   |      ___        |   |      |   |      (region b, uclamp_min <= util <= uclamp_max)
6249 	 *   |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
6250 	 *   |     |   |       |   |      |   |
6251 	 *   |     |   |       |   |      |   |      (region c, boosted, util < uclamp_min)
6252 	 *   +----------------------------------------
6253 	 *         CPU0        CPU1       CPU2
6254 	 *
6255 	 * a) If util > uclamp_max, then we're capped, we don't care about
6256 	 *    actual fitness value here. We only care if uclamp_max fits
6257 	 *    capacity without taking margin/pressure into account.
6258 	 *    See comment above.
6259 	 *
6260 	 * b) If uclamp_min <= util <= uclamp_max, then the normal
6261 	 *    fits_capacity() rules apply. Except we need to ensure that we
6262 	 *    enforce we remain within uclamp_max, see comment above.
6263 	 *
6264 	 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
6265 	 *    need to take into account the boosted value fits the CPU without
6266 	 *    taking margin/pressure into account.
6267 	 *
6268 	 * Cases (a) and (b) are handled in the 'fits' variable already. We
6269 	 * just need to consider an extra check for case (c) after ensuring we
6270 	 * handle the case uclamp_min > uclamp_max.
6271 	 */
6272 	uclamp_min = min(uclamp_min, uclamp_max);
6273 	if (fits && (util < uclamp_min) &&
6274 	    (uclamp_min > get_actual_cpu_capacity(cpu)))
6275 		return -1;
6276 
6277 	return fits;
6278 }
6279 
6280 static inline int task_fits_cpu(struct task_struct *p, int cpu)
6281 {
6282 	unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
6283 	unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
6284 	unsigned long util = task_util_est(p);
6285 	/*
6286 	 * Return true only if the cpu fully fits the task requirements, which
6287 	 * include the utilization but also the performance hints.
6288 	 */
6289 	return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
6290 }
6291 
6292 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
6293 {
6294 	int cpu = cpu_of(rq);
6295 
6296 	if (!sched_asym_cpucap_active())
6297 		return;
6298 
6299 	/*
6300 	 * Affinity allows us to go somewhere higher?  Or are we on biggest
6301 	 * available CPU already? Or do we fit into this CPU ?
6302 	 */
6303 	if (!p || (p->nr_cpus_allowed == 1) ||
6304 	    (arch_scale_cpu_capacity(cpu) == p->max_allowed_capacity) ||
6305 	    task_fits_cpu(p, cpu)) {
6306 
6307 		rq->misfit_task_load = 0;
6308 		return;
6309 	}
6310 
6311 	/*
6312 	 * Make sure that misfit_task_load will not be null even if
6313 	 * task_h_load() returns 0.
6314 	 */
6315 	rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
6316 }
6317 
6318 void __setparam_fair(struct task_struct *p, const struct sched_attr *attr)
6319 {
6320 	struct sched_entity *se = &p->se;
6321 
6322 	p->static_prio = NICE_TO_PRIO(attr->sched_nice);
6323 	if (attr->sched_runtime) {
6324 		se->custom_slice = 1;
6325 		se->slice = clamp_t(u64, attr->sched_runtime,
6326 				      NSEC_PER_MSEC/10,   /* HZ=1000 * 10 */
6327 				      NSEC_PER_MSEC*100); /* HZ=100  / 10 */
6328 	} else {
6329 		se->custom_slice = 0;
6330 		se->slice = sysctl_sched_base_slice;
6331 	}
6332 }
6333 
6334 static void
6335 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
6336 {
6337 	u64 vslice, vruntime = avg_vruntime(cfs_rq);
6338 	unsigned int nr_queued = cfs_rq->h_nr_queued;
6339 	bool update_zero = false;
6340 	s64 lag = 0;
6341 
6342 	if (!se->custom_slice)
6343 		se->slice = sysctl_sched_base_slice;
6344 	vslice = calc_delta_fair(se->slice, se);
6345 
6346 	if (flags & ENQUEUE_QUEUED)
6347 		nr_queued -= 1;
6348 
6349 	/*
6350 	 * Due to how V is constructed as the weighted average of entities,
6351 	 * adding tasks with positive lag, or removing tasks with negative lag
6352 	 * will move 'time' backwards, this can screw around with the lag of
6353 	 * other tasks.
6354 	 *
6355 	 * EEVDF: placement strategy #1 / #2
6356 	 */
6357 	if (sched_feat(PLACE_LAG) && nr_queued && se->vlag) {
6358 		struct sched_entity *curr = cfs_rq->curr;
6359 		long load, weight;
6360 
6361 		lag = se->vlag;
6362 
6363 		/*
6364 		 * If we want to place a task and preserve lag, we have to
6365 		 * consider the effect of the new entity on the weighted
6366 		 * average and compensate for this, otherwise lag can quickly
6367 		 * evaporate.
6368 		 *
6369 		 * Lag is defined as:
6370 		 *
6371 		 *   lag_i = S - s_i = w_i * (V - v_i)
6372 		 *
6373 		 * To avoid the 'w_i' term all over the place, we only track
6374 		 * the virtual lag:
6375 		 *
6376 		 *   vl_i = V - v_i <=> v_i = V - vl_i
6377 		 *
6378 		 * And we take V to be the weighted average of all v:
6379 		 *
6380 		 *   V = (\Sum w_j*v_j) / W
6381 		 *
6382 		 * Where W is: \Sum w_j
6383 		 *
6384 		 * Then, the weighted average after adding an entity with lag
6385 		 * vl_i is given by:
6386 		 *
6387 		 *   V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i)
6388 		 *      = (W*V + w_i*(V - vl_i)) / (W + w_i)
6389 		 *      = (W*V + w_i*V - w_i*vl_i) / (W + w_i)
6390 		 *      = (V*(W + w_i) - w_i*vl_i) / (W + w_i)
6391 		 *      = V - w_i*vl_i / (W + w_i)
6392 		 *
6393 		 * And the actual lag after adding an entity with vl_i is:
6394 		 *
6395 		 *   vl'_i = V' - v_i
6396 		 *         = V - w_i*vl_i / (W + w_i) - (V - vl_i)
6397 		 *         = vl_i - w_i*vl_i / (W + w_i)
6398 		 *
6399 		 * Which is strictly less than vl_i. So in order to preserve lag
6400 		 * we should inflate the lag before placement such that the
6401 		 * effective lag after placement comes out right.
6402 		 *
6403 		 * As such, invert the above relation for vl'_i to get the vl_i
6404 		 * we need to use such that the lag after placement is the lag
6405 		 * we computed before dequeue.
6406 		 *
6407 		 *   vl'_i = vl_i - w_i*vl_i / (W + w_i)
6408 		 *         = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)
6409 		 *
6410 		 *   (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i
6411 		 *                   = W*vl_i
6412 		 *
6413 		 *   vl_i = (W + w_i)*vl'_i / W
6414 		 */
6415 		load = cfs_rq->sum_weight;
6416 		if (curr && curr->on_rq)
6417 			load += avg_vruntime_weight(cfs_rq, curr->h_load.weight);
6418 		if (WARN_ON_ONCE(!load))
6419 			load = 1;
6420 
6421 		weight = avg_vruntime_weight(cfs_rq, se->h_load.weight);
6422 
6423 		/*
6424 		 * Rewrite, because the product (W + w_i)*vl'_i might
6425 		 * overflow s64:
6426 		 *
6427 		 *   vl_i = (W + w_i)*vl'_i / W
6428 		 *        = (W*vl'_i + w_i*vl'_i) / W
6429 		 *        = vl'_i + w_i*vl'_i/W
6430 		 */
6431 		lag += div64_long(lag * weight, load);
6432 
6433 		/*
6434 		 * A heavy entity (relative to the tree) will pull the
6435 		 * avg_vruntime close to its vruntime position on enqueue. But
6436 		 * the zero_vruntime point is only updated at the next
6437 		 * update_deadline()/place_entity()/update_entity_lag().
6438 		 *
6439 		 * Specifically (see the comment near avg_vruntime_weight()):
6440 		 *
6441 		 *   sum_w_vruntime = \Sum (v_i - v0) * w_i
6442 		 *
6443 		 * Note that if v0 is near a light entity, both terms will be
6444 		 * small for the light entity, while in that case both terms
6445 		 * are large for the heavy entity, leading to risk of
6446 		 * overflow.
6447 		 *
6448 		 * OTOH if v0 is near the heavy entity, then the difference is
6449 		 * larger for the light entity, but the factor is small, while
6450 		 * for the heavy entity the difference is small but the factor
6451 		 * is large. Avoiding the multiplication overflow.
6452 		 */
6453 		if (weight > load)
6454 			update_zero = true;
6455 	}
6456 
6457 	se->vruntime = vruntime - lag;
6458 
6459 	if (update_zero)
6460 		update_zero_vruntime(cfs_rq, -lag);
6461 
6462 	if (sched_feat(PLACE_REL_DEADLINE) && se->rel_deadline) {
6463 		se->deadline += se->vruntime;
6464 		se->rel_deadline = 0;
6465 		return;
6466 	}
6467 
6468 	/*
6469 	 * When joining the competition; the existing tasks will be,
6470 	 * on average, halfway through their slice, as such start tasks
6471 	 * off with half a slice to ease into the competition.
6472 	 */
6473 	if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL))
6474 		vslice /= 2;
6475 
6476 	/*
6477 	 * EEVDF: vd_i = ve_i + r_i/w_i
6478 	 */
6479 	se->deadline = se->vruntime + vslice;
6480 }
6481 
6482 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
6483 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
6484 
6485 static void
6486 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
6487 {
6488 	/*
6489 	 * When enqueuing a sched_entity, we must:
6490 	 *   - Update loads to have both entity and cfs_rq synced with now.
6491 	 *   - For group_entity, update its runnable_weight to reflect the new
6492 	 *     h_nr_runnable of its group cfs_rq.
6493 	 *   - For group_entity, update its weight to reflect the new share of
6494 	 *     its group cfs_rq
6495 	 *   - Add its new weight to cfs_rq->load.weight
6496 	 */
6497 	update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
6498 	se_update_runnable(se);
6499 	/*
6500 	 * XXX update_load_avg() above will have attached us to the pelt sum;
6501 	 * but update_cfs_group() here will re-adjust the weight and have to
6502 	 * undo/redo all that. Seems wasteful.
6503 	 */
6504 	update_cfs_group(se);
6505 
6506 	account_entity_enqueue(cfs_rq, se);
6507 
6508 	/* Entity has migrated, no longer consider this task hot */
6509 	if (flags & ENQUEUE_MIGRATED)
6510 		se->exec_start = 0;
6511 
6512 	check_schedstat_required();
6513 	update_stats_enqueue_fair(cfs_rq, se, flags);
6514 	se->on_rq = 1;
6515 
6516 	if (cfs_rq->nr_queued == 1) {
6517 		check_enqueue_throttle(cfs_rq);
6518 		list_add_leaf_cfs_rq(cfs_rq);
6519 #ifdef CONFIG_CFS_BANDWIDTH
6520 		if (cfs_rq->pelt_clock_throttled) {
6521 			struct rq *rq = rq_of(cfs_rq);
6522 
6523 			cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
6524 				cfs_rq->throttled_clock_pelt;
6525 			cfs_rq->pelt_clock_throttled = 0;
6526 		}
6527 #endif
6528 	}
6529 }
6530 
6531 static void set_next_buddy(struct cfs_rq *cfs_rq, struct sched_entity *se)
6532 {
6533 	if (WARN_ON_ONCE(!se->on_rq || se->sched_delayed))
6534 		return;
6535 	if (se_is_idle(se))
6536 		return;
6537 	cfs_rq->next = se;
6538 }
6539 
6540 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
6541 {
6542 	if (cfs_rq->next == se)
6543 		cfs_rq->next = NULL;
6544 }
6545 
6546 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
6547 
6548 static void set_delayed(struct sched_entity *se)
6549 {
6550 	/*
6551 	 * Delayed se of cfs_rq have no tasks queued on them.
6552 	 * Do not adjust h_nr_runnable since __dequeue_task()
6553 	 * will account it for blocked tasks.
6554 	 *
6555 	 * This check can be removed because when flat pick
6556 	 * patches get merged as only task can get delayed,
6557 	 * same for clear_delayed().
6558 	 */
6559 	if (!entity_is_task(se)) {
6560 		se->sched_delayed = 1;
6561 		return;
6562 	}
6563 
6564 	/*
6565 	 * Drop a task leaving the runnable set.
6566 	 * Needs to be called before sched_delayed is set.
6567 	 * clear_delayed() mirrors this after clearing the flag.
6568 	 */
6569 	pref_llc_running_dec(rq_of(cfs_rq_of(se)), task_of(se));
6570 	se->sched_delayed = 1;
6571 
6572 	for_each_sched_entity(se) {
6573 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
6574 
6575 		cfs_rq->h_nr_runnable--;
6576 	}
6577 }
6578 
6579 static void clear_delayed(struct sched_entity *se)
6580 {
6581 	se->sched_delayed = 0;
6582 
6583 	/*
6584 	 * Delayed se of cfs_rq have no tasks queued on them.
6585 	 * Do not adjust h_nr_runnable since a dequeue has
6586 	 * already accounted for it or an enqueue of a task
6587 	 * below it will account for it in enqueue_task_fair().
6588 	 */
6589 	if (!entity_is_task(se))
6590 		return;
6591 
6592 	/*
6593 	 * Re-add on wake, after sched_delayed is cleared. On a final delayed
6594 	 * dequeue account_llc_dequeue() already cleared pref_llc_queued, so
6595 	 * this does nothing.
6596 	 */
6597 	pref_llc_running_inc(rq_of(cfs_rq_of(se)), task_of(se));
6598 
6599 	for_each_sched_entity(se) {
6600 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
6601 
6602 		cfs_rq->h_nr_runnable++;
6603 	}
6604 }
6605 
6606 static void
6607 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
6608 {
6609 	int action = UPDATE_TG;
6610 
6611 	if (entity_is_task(se)) {
6612 		if (task_on_rq_migrating(task_of(se)))
6613 			action |= DO_DETACH;
6614 
6615 		if ((flags & DEQUEUE_SLEEP) && !(flags & DEQUEUE_DELAYED))
6616 			action |= UPDATE_UTIL_EST;
6617 	}
6618 
6619 	/*
6620 	 * When dequeuing a sched_entity, we must:
6621 	 *   - Update loads to have both entity and cfs_rq synced with now.
6622 	 *   - For group_entity, update its runnable_weight to reflect the new
6623 	 *     h_nr_runnable of its group cfs_rq.
6624 	 *   - Subtract its previous weight from cfs_rq->load.weight.
6625 	 *   - For group entity, update its weight to reflect the new share
6626 	 *     of its group cfs_rq.
6627 	 */
6628 	update_load_avg(cfs_rq, se, action);
6629 	se_update_runnable(se);
6630 
6631 	update_stats_dequeue_fair(cfs_rq, se, flags);
6632 
6633 	se->on_rq = 0;
6634 	account_entity_dequeue(cfs_rq, se);
6635 
6636 	/* return excess runtime on last dequeue */
6637 	return_cfs_rq_runtime(cfs_rq);
6638 
6639 	update_cfs_group(se);
6640 
6641 	if (cfs_rq->nr_queued == 0) {
6642 		update_idle_cfs_rq_clock_pelt(cfs_rq);
6643 #ifdef CONFIG_CFS_BANDWIDTH
6644 		if (throttled_hierarchy(cfs_rq)) {
6645 			struct rq *rq = rq_of(cfs_rq);
6646 
6647 			list_del_leaf_cfs_rq(cfs_rq);
6648 			cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
6649 			cfs_rq->pelt_clock_throttled = 1;
6650 		}
6651 #endif
6652 	}
6653 }
6654 
6655 static void
6656 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
6657 {
6658 	/* 'current' is not kept within the tree. */
6659 	if (se->on_rq) {
6660 		/*
6661 		 * Any task has to be enqueued before it get to execute on
6662 		 * a CPU. So account for the time it spent waiting on the
6663 		 * runqueue.
6664 		 */
6665 		update_stats_wait_end_fair(cfs_rq, se);
6666 		update_load_avg(cfs_rq, se, UPDATE_TG);
6667 	}
6668 
6669 	update_stats_curr_start(cfs_rq, se);
6670 	WARN_ON_ONCE(cfs_rq->h_curr);
6671 	cfs_rq->h_curr = se;
6672 
6673 	/*
6674 	 * Track our maximum slice length, if the CPU's load is at
6675 	 * least twice that of our own weight (i.e. don't track it
6676 	 * when there are only lesser-weight tasks around):
6677 	 */
6678 	if (schedstat_enabled() &&
6679 	    rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
6680 		struct sched_statistics *stats;
6681 
6682 		stats = __schedstats_from_se(se);
6683 		__schedstat_set(stats->slice_max,
6684 				max((u64)stats->slice_max,
6685 				    se->sum_exec_runtime - se->prev_sum_exec_runtime));
6686 	}
6687 
6688 	se->prev_sum_exec_runtime = se->sum_exec_runtime;
6689 }
6690 
6691 static bool __dequeue_task(struct rq *rq, struct task_struct *p, int flags);
6692 
6693 static struct sched_entity *
6694 pick_next_entity(struct rq *rq, bool protect)
6695 {
6696 	struct cfs_rq *cfs_rq = &rq->cfs;
6697 	struct sched_entity *se;
6698 
6699 	se = pick_eevdf(cfs_rq, protect);
6700 	if (se->sched_delayed) {
6701 		__dequeue_task(rq, task_of(se), DEQUEUE_SLEEP | DEQUEUE_DELAYED);
6702 		/*
6703 		 * Must not reference @se again, see __block_task().
6704 		 */
6705 		return NULL;
6706 	}
6707 	return se;
6708 }
6709 
6710 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
6711 {
6712 	/*
6713 	 * If still on the runqueue then deactivate_task()
6714 	 * was not called and update_curr() has to be done:
6715 	 */
6716 	if (prev->on_rq)
6717 		update_curr(cfs_rq);
6718 
6719 	if (prev->on_rq) {
6720 		update_stats_wait_start_fair(cfs_rq, prev);
6721 		/* in !on_rq case, update occurred at dequeue */
6722 		update_load_avg(cfs_rq, prev, 0);
6723 	}
6724 	WARN_ON_ONCE(cfs_rq->h_curr != prev);
6725 	cfs_rq->h_curr = NULL;
6726 }
6727 
6728 static void
6729 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
6730 {
6731 	/*
6732 	 * Update run-time statistics of the 'current'.
6733 	 */
6734 	update_curr(cfs_rq);
6735 
6736 	/*
6737 	 * Ensure that runnable average is periodically updated.
6738 	 */
6739 	update_load_avg(cfs_rq, curr, UPDATE_TG);
6740 	update_cfs_group(curr);
6741 
6742 #ifdef CONFIG_SCHED_HRTICK
6743 	/*
6744 	 * queued ticks are scheduled to match the slice, so don't bother
6745 	 * validating it and just reschedule.
6746 	 */
6747 	if (queued) {
6748 		resched_curr(rq_of(cfs_rq));
6749 		return;
6750 	}
6751 #endif
6752 }
6753 
6754 
6755 /**************************************************
6756  * CFS bandwidth control machinery
6757  */
6758 
6759 #ifdef CONFIG_CFS_BANDWIDTH
6760 
6761 #ifdef CONFIG_JUMP_LABEL
6762 static struct static_key __cfs_bandwidth_used;
6763 
6764 static inline bool cfs_bandwidth_used(void)
6765 {
6766 	return static_key_false(&__cfs_bandwidth_used);
6767 }
6768 
6769 void cfs_bandwidth_usage_inc(void)
6770 {
6771 	static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
6772 }
6773 
6774 void cfs_bandwidth_usage_dec(void)
6775 {
6776 	static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
6777 }
6778 #else /* !CONFIG_JUMP_LABEL: */
6779 static bool cfs_bandwidth_used(void)
6780 {
6781 	return true;
6782 }
6783 
6784 void cfs_bandwidth_usage_inc(void) {}
6785 void cfs_bandwidth_usage_dec(void) {}
6786 #endif /* !CONFIG_JUMP_LABEL */
6787 
6788 static inline u64 sched_cfs_bandwidth_slice(void)
6789 {
6790 	return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
6791 }
6792 
6793 /*
6794  * Replenish runtime according to assigned quota. We use sched_clock_cpu
6795  * directly instead of rq->clock to avoid adding additional synchronization
6796  * around rq->lock.
6797  *
6798  * requires cfs_b->lock
6799  */
6800 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
6801 {
6802 	s64 runtime;
6803 
6804 	if (unlikely(cfs_b->quota == RUNTIME_INF))
6805 		return;
6806 
6807 	cfs_b->runtime += cfs_b->quota;
6808 	runtime = cfs_b->runtime_snap - cfs_b->runtime;
6809 	if (runtime > 0) {
6810 		cfs_b->burst_time += runtime;
6811 		cfs_b->nr_burst++;
6812 	}
6813 
6814 	cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
6815 	cfs_b->runtime_snap = cfs_b->runtime;
6816 }
6817 
6818 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6819 {
6820 	return &tg->cfs_bandwidth;
6821 }
6822 
6823 /* returns 0 on failure to allocate runtime */
6824 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
6825 				   struct cfs_rq *cfs_rq, u64 target_runtime)
6826 {
6827 	u64 min_amount, amount = 0;
6828 
6829 	lockdep_assert_held(&cfs_b->lock);
6830 
6831 	/* note: this is a positive sum as runtime_remaining <= 0 */
6832 	min_amount = target_runtime - cfs_rq->runtime_remaining;
6833 
6834 	if (cfs_b->quota == RUNTIME_INF)
6835 		amount = min_amount;
6836 	else {
6837 		start_cfs_bandwidth(cfs_b);
6838 
6839 		if (cfs_b->runtime > 0) {
6840 			amount = min(cfs_b->runtime, min_amount);
6841 			cfs_b->runtime -= amount;
6842 			cfs_b->idle = 0;
6843 		}
6844 	}
6845 
6846 	cfs_rq->runtime_remaining += amount;
6847 
6848 	return cfs_rq->runtime_remaining > 0;
6849 }
6850 
6851 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq);
6852 
6853 static bool __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
6854 {
6855 	/* dock delta_exec before expiring quota (as it could span periods) */
6856 	cfs_rq->runtime_remaining -= delta_exec;
6857 
6858 	if (likely(cfs_rq->runtime_remaining > 0))
6859 		return false;
6860 
6861 	if (cfs_rq->throttled)
6862 		return true;
6863 	/*
6864 	 * throttle_cfs_rq() will try to extend the runtime first
6865 	 * before throttling the hierarchy.
6866 	 */
6867 	return throttle_cfs_rq(cfs_rq);
6868 }
6869 
6870 static __always_inline
6871 bool account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
6872 {
6873 	if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
6874 		return false;
6875 
6876 	return __account_cfs_rq_runtime(cfs_rq, delta_exec);
6877 }
6878 
6879 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6880 {
6881 	return cfs_bandwidth_used() && cfs_rq->throttled;
6882 }
6883 
6884 static inline bool cfs_rq_pelt_clock_throttled(struct cfs_rq *cfs_rq)
6885 {
6886 	return cfs_bandwidth_used() && cfs_rq->pelt_clock_throttled;
6887 }
6888 
6889 /* check whether cfs_rq, or any parent, is throttled */
6890 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6891 {
6892 	return cfs_bandwidth_used() && cfs_rq->throttle_count;
6893 }
6894 
6895 static inline int lb_throttled_hierarchy(struct task_struct *p, int dst_cpu)
6896 {
6897 	return throttled_hierarchy(tg_cfs_rq(task_group(p), dst_cpu));
6898 }
6899 
6900 static inline bool task_is_throttled(struct task_struct *p)
6901 {
6902 	return cfs_bandwidth_used() && p->throttled;
6903 }
6904 
6905 static bool dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags);
6906 static void throttle_cfs_rq_work(struct callback_head *work)
6907 {
6908 	struct task_struct *p = container_of(work, struct task_struct, sched_throttle_work);
6909 	struct sched_entity *se;
6910 	struct cfs_rq *cfs_rq;
6911 	struct rq *rq;
6912 
6913 	WARN_ON_ONCE(p != current);
6914 	p->sched_throttle_work.next = &p->sched_throttle_work;
6915 
6916 	/*
6917 	 * If task is exiting, then there won't be a return to userspace, so we
6918 	 * don't have to bother with any of this.
6919 	 */
6920 	if ((p->flags & PF_EXITING))
6921 		return;
6922 
6923 	scoped_guard(task_rq_lock, p) {
6924 		se = &p->se;
6925 		cfs_rq = cfs_rq_of(se);
6926 
6927 		/* Raced, forget */
6928 		if (p->sched_class != &fair_sched_class)
6929 			return;
6930 
6931 		/*
6932 		 * If not in limbo, then either replenish has happened or this
6933 		 * task got migrated out of the throttled cfs_rq, move along.
6934 		 */
6935 		if (!cfs_rq->throttle_count)
6936 			return;
6937 		rq = scope.rq;
6938 		update_rq_clock(rq);
6939 		WARN_ON_ONCE(p->throttled || !list_empty(&p->throttle_node));
6940 		dequeue_task_fair(rq, p, DEQUEUE_SLEEP | DEQUEUE_THROTTLE);
6941 		list_add(&p->throttle_node, &cfs_rq->throttled_limbo_list);
6942 		/*
6943 		 * Must not set throttled before dequeue or dequeue will
6944 		 * mistakenly regard this task as an already throttled one.
6945 		 */
6946 		p->throttled = true;
6947 		resched_curr(rq);
6948 	}
6949 }
6950 
6951 void init_cfs_throttle_work(struct task_struct *p)
6952 {
6953 	init_task_work(&p->sched_throttle_work, throttle_cfs_rq_work);
6954 	/* Protect against double add, see throttle_cfs_rq() and throttle_cfs_rq_work() */
6955 	p->sched_throttle_work.next = &p->sched_throttle_work;
6956 	INIT_LIST_HEAD(&p->throttle_node);
6957 }
6958 
6959 /*
6960  * Task is throttled and someone wants to dequeue it again:
6961  * it could be sched/core when core needs to do things like
6962  * task affinity change, task group change, task sched class
6963  * change etc. and in these cases, DEQUEUE_SLEEP is not set;
6964  * or the task is blocked after throttled due to freezer etc.
6965  * and in these cases, DEQUEUE_SLEEP is set.
6966  */
6967 static void detach_task_cfs_rq(struct task_struct *p);
6968 static void dequeue_throttled_task(struct task_struct *p, int flags)
6969 {
6970 	WARN_ON_ONCE(p->se.on_rq);
6971 	list_del_init(&p->throttle_node);
6972 
6973 	/* Task blocked after throttling or is leaving the fair class: */
6974 	if (flags & (DEQUEUE_SLEEP | DEQUEUE_CLASS)) {
6975 		p->throttled = false;
6976 		return;
6977 	}
6978 
6979 	/*
6980 	 * task is migrating off its old cfs_rq, detach
6981 	 * the task's load from its old cfs_rq.
6982 	 */
6983 	if (task_on_rq_migrating(p))
6984 		detach_task_cfs_rq(p);
6985 }
6986 
6987 static bool enqueue_throttled_task(struct task_struct *p)
6988 {
6989 	struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
6990 
6991 	/* @p should have gone through dequeue_throttled_task() first */
6992 	WARN_ON_ONCE(!list_empty(&p->throttle_node));
6993 
6994 	/*
6995 	 * If the throttled task @p is enqueued to a throttled cfs_rq,
6996 	 * take the fast path by directly putting the task on the
6997 	 * target cfs_rq's limbo list.
6998 	 *
6999 	 * Do not do that when @p is current because the following race can
7000 	 * cause @p's group_node to be incorectly re-insterted in its rq's
7001 	 * cfs_tasks list, despite being throttled:
7002 	 *
7003 	 *     cpuX                       cpuY
7004 	 *   p ret2user
7005 	 *  throttle_cfs_rq_work()  sched_move_task(p)
7006 	 *  LOCK task_rq_lock
7007 	 *  dequeue_task_fair(p)
7008 	 *  UNLOCK task_rq_lock
7009 	 *                          LOCK task_rq_lock
7010 	 *                          task_current_donor(p) == true
7011 	 *                          task_on_rq_queued(p) == true
7012 	 *                          dequeue_task(p)
7013 	 *                          put_prev_task(p)
7014 	 *                          sched_change_group()
7015 	 *                          enqueue_task(p) -> p's new cfs_rq
7016 	 *                                             is throttled, go
7017 	 *                                             fast path and skip
7018 	 *                                             actual enqueue
7019 	 *                          set_next_task(p)
7020 	 *                    list_move(&se->group_node, &rq->cfs_tasks); // bug
7021 	 *  schedule()
7022 	 *
7023 	 * In the above race case, @p current cfs_rq is in the same rq as
7024 	 * its previous cfs_rq because sched_move_task() only moves a task
7025 	 * to a different group from the same rq, so we can use its current
7026 	 * cfs_rq to derive rq and test if the task is current.
7027 	 */
7028 	if (throttled_hierarchy(cfs_rq) &&
7029 	    !task_current_donor(rq_of(cfs_rq), p)) {
7030 		list_add(&p->throttle_node, &cfs_rq->throttled_limbo_list);
7031 		return true;
7032 	}
7033 
7034 	/* we can't take the fast path, do an actual enqueue*/
7035 	p->throttled = false;
7036 	return false;
7037 }
7038 
7039 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags);
7040 static int tg_unthrottle_up(struct task_group *tg, void *data)
7041 {
7042 	struct rq *rq = data;
7043 	struct cfs_rq *cfs_rq = tg_cfs_rq(tg, cpu_of(rq));
7044 	struct task_struct *p, *tmp;
7045 	LIST_HEAD(throttled_tasks);
7046 
7047 	/*
7048 	 * If cfs_rq->curr is set, the cfs_rq might not have caught up
7049 	 * since the last clock update. Do it now before we begin
7050 	 * queueing task onto it to save the need for unnecessarily
7051 	 * unthrottle the hierarchy for this cfs_rq to be throttled
7052 	 * right back again.
7053 	 */
7054 	update_curr(cfs_rq);
7055 
7056 	if (--cfs_rq->throttle_count)
7057 		return 0;
7058 
7059 	if (cfs_rq->pelt_clock_throttled) {
7060 		cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
7061 					     cfs_rq->throttled_clock_pelt;
7062 		cfs_rq->pelt_clock_throttled = 0;
7063 	}
7064 
7065 	if (cfs_rq->throttled_clock_self) {
7066 		u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;
7067 
7068 		cfs_rq->throttled_clock_self = 0;
7069 
7070 		if (WARN_ON_ONCE((s64)delta < 0))
7071 			delta = 0;
7072 
7073 		cfs_rq->throttled_clock_self_time += delta;
7074 	}
7075 
7076 	/*
7077 	 * Move the tasks to a local list since an update_curr() during
7078 	 * enqueue_task_fair() can throttle a higher cfs_rq, and it can
7079 	 * see the "throttled_limbo_list" being non-empty in
7080 	 * tg_throttle_down() if throttle_count turned 0 above.
7081 	 */
7082 	list_splice_init(&cfs_rq->throttled_limbo_list, &throttled_tasks);
7083 
7084 	/* Re-enqueue the tasks that have been throttled at this level. */
7085 	list_for_each_entry_safe(p, tmp, &throttled_tasks, throttle_node) {
7086 		/*
7087 		 * Back to being throttled! Break out and put the remaining
7088 		 * tasks back onto the limbo_list to prevent running them
7089 		 * unnecessarily.
7090 		 */
7091 		if (cfs_rq->throttle_count)
7092 			break;
7093 
7094 		list_del_init(&p->throttle_node);
7095 		p->throttled = false;
7096 		enqueue_task_fair(rq, p, ENQUEUE_WAKEUP);
7097 	}
7098 
7099 	list_splice(&throttled_tasks, &cfs_rq->throttled_limbo_list);
7100 
7101 	/* Add cfs_rq with load or one or more already running entities to the list */
7102 	if (!cfs_rq_is_decayed(cfs_rq))
7103 		list_add_leaf_cfs_rq(cfs_rq);
7104 
7105 	return 0;
7106 }
7107 
7108 static inline bool task_has_throttle_work(struct task_struct *p)
7109 {
7110 	return p->sched_throttle_work.next != &p->sched_throttle_work;
7111 }
7112 
7113 static inline void task_throttle_setup_work(struct task_struct *p)
7114 {
7115 	if (task_has_throttle_work(p))
7116 		return;
7117 
7118 	/*
7119 	 * Kthreads and exiting tasks don't return to userspace, so adding the
7120 	 * work is pointless
7121 	 */
7122 	if ((p->flags & (PF_EXITING | PF_KTHREAD)))
7123 		return;
7124 
7125 	task_work_add(p, &p->sched_throttle_work, TWA_RESUME);
7126 }
7127 
7128 static void record_throttle_clock(struct cfs_rq *cfs_rq)
7129 {
7130 	struct rq *rq = rq_of(cfs_rq);
7131 
7132 	if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)
7133 		cfs_rq->throttled_clock = rq_clock(rq);
7134 
7135 	if (!cfs_rq->throttled_clock_self)
7136 		cfs_rq->throttled_clock_self = rq_clock(rq);
7137 }
7138 
7139 static int tg_throttle_down(struct task_group *tg, void *data)
7140 {
7141 	struct rq *rq = data;
7142 	struct cfs_rq *cfs_rq = tg_cfs_rq(tg, cpu_of(rq));
7143 
7144 	if (cfs_rq->throttle_count++)
7145 		return 0;
7146 
7147 	/*
7148 	 * For cfs_rqs that still have entities enqueued, PELT clock
7149 	 * stop happens at dequeue time when all entities are dequeued.
7150 	 */
7151 	if (!cfs_rq->nr_queued) {
7152 		list_del_leaf_cfs_rq(cfs_rq);
7153 		cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
7154 		cfs_rq->pelt_clock_throttled = 1;
7155 	}
7156 
7157 	WARN_ON_ONCE(cfs_rq->throttled_clock_self);
7158 	WARN_ON_ONCE(!list_empty(&cfs_rq->throttled_limbo_list));
7159 	return 0;
7160 }
7161 
7162 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
7163 {
7164 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
7165 	struct sched_entity *curr = cfs_rq->h_curr;
7166 	struct rq *rq = rq_of(cfs_rq);
7167 
7168 	scoped_guard(raw_spinlock, &cfs_b->lock) {
7169 		u64 target_runtime = 1;
7170 
7171 		/*
7172 		 * If cfs_rq->h_curr is still runnable, we are here from an
7173 		 * update_curr(). Request sysctl_sched_cfs_bandwidth_slice
7174 		 * worth of bandwidth to continue running.
7175 		 *
7176 		 * If the curr is not runnable, just request enough bandwidth
7177 		 * to be runnable next time the pick selects this cfs_rq.
7178 		 */
7179 		if (curr && curr->on_rq)
7180 			target_runtime = sched_cfs_bandwidth_slice();
7181 
7182 		/*
7183 		 * Check if We have raced with bandwidth becoming available. If
7184 		 * we actually throttled the timer might not unthrottle us for
7185 		 * an entire period. We additionally needed to make sure that
7186 		 * any subsequent check_cfs_rq_runtime calls agree not to
7187 		 * throttle us, as we may commit to do cfs put_prev+pick_next,
7188 		 * so we ask for 1ns of runtime rather than just check cfs_b.
7189 		 *
7190 		 * This will start the period timer if necessary.
7191 		 */
7192 		if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, target_runtime))
7193 			return false;
7194 
7195 		/*
7196 		 * No bandwidth available; Add ourselves on the list to be
7197 		 * unthrottled later.
7198 		 */
7199 		list_add_tail_rcu(&cfs_rq->throttled_list,
7200 				  &cfs_b->throttled_cfs_rq);
7201 	}
7202 
7203 	/* freeze hierarchy runnable averages while throttled */
7204 	scoped_guard(rcu)
7205 		walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
7206 
7207 	/*
7208 	 * Note: distribution will already see us throttled via the
7209 	 * throttled-list.  rq->lock protects completion.
7210 	 */
7211 	cfs_rq->throttled = 1;
7212 	WARN_ON_ONCE(cfs_rq->throttled_clock);
7213 
7214 	/*
7215 	 * If current hierarchy was throttled, add throttle work to the
7216 	 * current donor. In case of proxy-execution, the execution
7217 	 * context cannot exit to the userspace while holding a mutex
7218 	 * and the rule of throttle deferral to only throttle the
7219 	 * throttled context at exit to userspace is still preserved.
7220 	 */
7221 	if (curr && curr->on_rq)
7222 		task_throttle_setup_work(rq->donor);
7223 
7224 	return true;
7225 }
7226 
7227 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
7228 {
7229 	struct rq *rq = rq_of(cfs_rq);
7230 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
7231 	struct sched_entity *se = cfs_rq_se(cfs_rq);
7232 
7233 	/*
7234 	 * It's possible we are called with runtime_remaining < 0 due to things
7235 	 * like async unthrottled us with a positive runtime_remaining but other
7236 	 * still running entities consumed those runtime before we reached here.
7237 	 *
7238 	 * We can't unthrottle this cfs_rq without any runtime remaining because
7239 	 * any enqueue in tg_unthrottle_up() will immediately trigger a throttle,
7240 	 * which is not supposed to happen on unthrottle path.
7241 	 *
7242 	 * Catch up on the remaining runtime since last clock update before
7243 	 * checking runtime remaining.
7244 	 */
7245 	update_curr(cfs_rq);
7246 	if (cfs_rq->runtime_enabled && cfs_rq->runtime_remaining <= 0)
7247 		return;
7248 
7249 	cfs_rq->throttled = 0;
7250 
7251 	scoped_guard(raw_spinlock, &cfs_b->lock) {
7252 		list_del_rcu(&cfs_rq->throttled_list);
7253 
7254 		if (!cfs_rq->throttled_clock)
7255 			break;
7256 
7257 		cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
7258 		cfs_rq->throttled_clock = 0;
7259 	}
7260 
7261 	/* update hierarchical throttle state */
7262 	walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
7263 
7264 	if (!cfs_rq->load.weight) {
7265 		if (!cfs_rq->on_list)
7266 			return;
7267 		/*
7268 		 * Nothing to run but something to decay (on_list)?
7269 		 * Complete the branch.
7270 		 */
7271 		for_each_sched_entity(se) {
7272 			if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
7273 				break;
7274 		}
7275 	}
7276 
7277 	assert_list_leaf_cfs_rq(rq);
7278 
7279 	/* Determine whether we need to wake up potentially idle CPU: */
7280 	if (rq->curr == rq->idle && rq->cfs.h_nr_queued)
7281 		resched_curr(rq);
7282 }
7283 
7284 static void __cfsb_csd_unthrottle(void *arg)
7285 {
7286 	struct cfs_rq *cursor, *tmp;
7287 	struct rq *rq = arg;
7288 
7289 	guard(rq_lock)(rq);
7290 
7291 	/*
7292 	 * Iterating over the list can trigger several call to
7293 	 * update_rq_clock() in unthrottle_cfs_rq().
7294 	 * Do it once and skip the potential next ones.
7295 	 */
7296 	update_rq_clock(rq);
7297 	rq_clock_start_loop_update(rq);
7298 
7299 	/*
7300 	 * Since we hold rq lock we're safe from concurrent manipulation of
7301 	 * the CSD list. However, this RCU critical section annotates the
7302 	 * fact that we pair with sched_free_group_rcu(), so that we cannot
7303 	 * race with group being freed in the window between removing it
7304 	 * from the list and advancing to the next entry in the list.
7305 	 */
7306 	guard(rcu)();
7307 
7308 	list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
7309 				 throttled_csd_list) {
7310 		list_del_init(&cursor->throttled_csd_list);
7311 
7312 		if (cfs_rq_throttled(cursor))
7313 			unthrottle_cfs_rq(cursor);
7314 	}
7315 
7316 	rq_clock_stop_loop_update(rq);
7317 }
7318 
7319 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
7320 {
7321 	struct rq *rq = rq_of(cfs_rq);
7322 	bool first;
7323 
7324 	if (rq == this_rq()) {
7325 		update_rq_clock(rq);
7326 		unthrottle_cfs_rq(cfs_rq);
7327 		return;
7328 	}
7329 
7330 	/* Already enqueued */
7331 	if (WARN_ON_ONCE(!list_empty(&cfs_rq->throttled_csd_list)))
7332 		return;
7333 
7334 	first = list_empty(&rq->cfsb_csd_list);
7335 	list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
7336 	if (first)
7337 		smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
7338 }
7339 
7340 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
7341 {
7342 	lockdep_assert_rq_held(rq_of(cfs_rq));
7343 
7344 	if (WARN_ON_ONCE(!cfs_rq_throttled(cfs_rq) ||
7345 	    cfs_rq->runtime_remaining <= 0))
7346 		return;
7347 
7348 	__unthrottle_cfs_rq_async(cfs_rq);
7349 }
7350 
7351 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
7352 {
7353 	bool throttled = false, unthrottle_local = false;
7354 	int this_cpu = smp_processor_id();
7355 	u64 runtime, remaining = 1;
7356 	struct cfs_rq *cfs_rq;
7357 	struct rq *rq;
7358 
7359 	guard(rcu)();
7360 
7361 	list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
7362 				throttled_list) {
7363 		rq = rq_of(cfs_rq);
7364 
7365 		if (!remaining) {
7366 			throttled = true;
7367 			break;
7368 		}
7369 
7370 		guard(rq_lock_irqsave)(rq);
7371 
7372 		if (!cfs_rq_throttled(cfs_rq))
7373 			continue;
7374 
7375 		/* Already queued for async unthrottle */
7376 		if (!list_empty(&cfs_rq->throttled_csd_list))
7377 			continue;
7378 
7379 		if (cfs_rq->h_curr) {
7380 			update_rq_clock(rq);
7381 			update_curr(cfs_rq);
7382 		}
7383 
7384 		/* By the above checks, this should never be true */
7385 		WARN_ON_ONCE(cfs_rq->runtime_remaining > 0);
7386 
7387 		scoped_guard(raw_spinlock, &cfs_b->lock) {
7388 			runtime = -cfs_rq->runtime_remaining + 1;
7389 			if (runtime > cfs_b->runtime)
7390 				runtime = cfs_b->runtime;
7391 			cfs_b->runtime -= runtime;
7392 			remaining = cfs_b->runtime;
7393 		}
7394 
7395 		cfs_rq->runtime_remaining += runtime;
7396 
7397 		/*
7398 		 * Ran out of bandwidth during distribution!
7399 		 * Indicate throttled entities and break early.
7400 		 */
7401 		if (cfs_rq->runtime_remaining <= 0) {
7402 			throttled = true;
7403 			break;
7404 		}
7405 
7406 		/* we check whether we're throttled above */
7407 		if (cpu_of(rq) != this_cpu) {
7408 			unthrottle_cfs_rq_async(cfs_rq);
7409 			continue;
7410 		}
7411 
7412 		/*
7413 		 * Allow a parallel async unthrottle to unthrottle
7414 		 * this cfs_rq too via __cfsb_csd_unthrottle().
7415 		 * If we are first, do it ourselves at the end and
7416 		 * save on an IPI from remote CPUs.
7417 		 */
7418 		unthrottle_local = list_empty(&rq->cfsb_csd_list);
7419 		list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
7420 	}
7421 
7422 	if (unthrottle_local) {
7423 		/*
7424 		 * Protect against an IPI that is also trying to flush
7425 		 * the unthrottled cfs_rq(s) from this CPU's csd_list.
7426 		 */
7427 		scoped_guard(irqsave)
7428 			__cfsb_csd_unthrottle(cpu_rq(this_cpu));
7429 	}
7430 
7431 	return throttled;
7432 }
7433 
7434 /*
7435  * Responsible for refilling a task_group's bandwidth and unthrottling its
7436  * cfs_rqs as appropriate. If there has been no activity within the last
7437  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
7438  * used to track this state.
7439  */
7440 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
7441 	__must_hold(&cfs_b->lock)
7442 {
7443 	int throttled;
7444 
7445 	/* no need to continue the timer with no bandwidth constraint */
7446 	if (cfs_b->quota == RUNTIME_INF)
7447 		goto out_deactivate;
7448 
7449 	throttled = !list_empty(&cfs_b->throttled_cfs_rq);
7450 	cfs_b->nr_periods += overrun;
7451 
7452 	/* Refill extra burst quota even if cfs_b->idle */
7453 	__refill_cfs_bandwidth_runtime(cfs_b);
7454 
7455 	/*
7456 	 * idle depends on !throttled (for the case of a large deficit), and if
7457 	 * we're going inactive then everything else can be deferred
7458 	 */
7459 	if (cfs_b->idle && !throttled)
7460 		goto out_deactivate;
7461 
7462 	if (!throttled) {
7463 		/* mark as potentially idle for the upcoming period */
7464 		cfs_b->idle = 1;
7465 		return 0;
7466 	}
7467 
7468 	/* account preceding periods in which throttling occurred */
7469 	cfs_b->nr_throttled += overrun;
7470 
7471 	/*
7472 	 * This check is repeated as we release cfs_b->lock while we unthrottle.
7473 	 */
7474 	while (throttled && cfs_b->runtime > 0) {
7475 		raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
7476 		/* we can't nest cfs_b->lock while distributing bandwidth */
7477 		throttled = distribute_cfs_runtime(cfs_b);
7478 		raw_spin_lock_irqsave(&cfs_b->lock, flags);
7479 	}
7480 
7481 	/*
7482 	 * While we are ensured activity in the period following an
7483 	 * unthrottle, this also covers the case in which the new bandwidth is
7484 	 * insufficient to cover the existing bandwidth deficit.  (Forcing the
7485 	 * timer to remain active while there are any throttled entities.)
7486 	 */
7487 	cfs_b->idle = 0;
7488 
7489 	return 0;
7490 
7491 out_deactivate:
7492 	return 1;
7493 }
7494 
7495 /* a cfs_rq won't donate quota below this amount */
7496 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
7497 /* minimum remaining period time to redistribute slack quota */
7498 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
7499 /* how long we wait to gather additional slack before distributing */
7500 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
7501 
7502 /*
7503  * Are we near the end of the current quota period?
7504  *
7505  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
7506  * hrtimer base being cleared by hrtimer_start. In the case of
7507  * migrate_hrtimers, base is never cleared, so we are fine.
7508  */
7509 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
7510 {
7511 	struct hrtimer *refresh_timer = &cfs_b->period_timer;
7512 	s64 remaining;
7513 
7514 	/* if the call-back is running a quota refresh is already occurring */
7515 	if (hrtimer_callback_running(refresh_timer))
7516 		return 1;
7517 
7518 	/* is a quota refresh about to occur? */
7519 	remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
7520 	if (remaining < (s64)min_expire)
7521 		return 1;
7522 
7523 	return 0;
7524 }
7525 
7526 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
7527 {
7528 	u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
7529 
7530 	/* if there's a quota refresh soon don't bother with slack */
7531 	if (runtime_refresh_within(cfs_b, min_left))
7532 		return;
7533 
7534 	/* don't push forwards an existing deferred unthrottle */
7535 	if (cfs_b->slack_started)
7536 		return;
7537 	cfs_b->slack_started = true;
7538 
7539 	hrtimer_start(&cfs_b->slack_timer,
7540 			ns_to_ktime(cfs_bandwidth_slack_period),
7541 			HRTIMER_MODE_REL);
7542 }
7543 
7544 /* we know any runtime found here is valid as update_curr() precedes return */
7545 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
7546 {
7547 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
7548 	s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
7549 
7550 	if (slack_runtime <= 0)
7551 		return;
7552 
7553 	guard(raw_spinlock)(&cfs_b->lock);
7554 
7555 	if (cfs_b->quota != RUNTIME_INF) {
7556 		cfs_b->runtime += slack_runtime;
7557 
7558 		/* we are under rq->lock, defer unthrottling using a timer */
7559 		if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
7560 		    !list_empty(&cfs_b->throttled_cfs_rq))
7561 			start_cfs_slack_bandwidth(cfs_b);
7562 	}
7563 
7564 	/* even if it's not valid for return we don't want to try again */
7565 	cfs_rq->runtime_remaining -= slack_runtime;
7566 }
7567 
7568 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
7569 {
7570 	if (!cfs_bandwidth_used())
7571 		return;
7572 
7573 	if (!cfs_rq->runtime_enabled || cfs_rq->nr_queued)
7574 		return;
7575 
7576 	__return_cfs_rq_runtime(cfs_rq);
7577 }
7578 
7579 /*
7580  * This is done with a timer (instead of inline with bandwidth return) since
7581  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
7582  */
7583 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
7584 {
7585 	/* confirm we're still not at a refresh boundary */
7586 	scoped_guard(raw_spinlock_irqsave, &cfs_b->lock) {
7587 		u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
7588 
7589 		cfs_b->slack_started = false;
7590 
7591 		if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
7592 			return;
7593 
7594 		if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
7595 			runtime = cfs_b->runtime;
7596 
7597 		if (!runtime)
7598 			return;
7599 	}
7600 
7601 	distribute_cfs_runtime(cfs_b);
7602 }
7603 
7604 /*
7605  * When a group wakes up we want to make sure that its quota is not already
7606  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
7607  * runtime as update_curr() throttling can not trigger until it's on-rq.
7608  */
7609 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
7610 {
7611 	if (!cfs_bandwidth_used())
7612 		return;
7613 
7614 	/* an active group must be handled by the update_curr() path */
7615 	if (!cfs_rq->runtime_enabled || cfs_rq->h_curr)
7616 		return;
7617 
7618 	/* ensure the group is not already throttled */
7619 	if (cfs_rq_throttled(cfs_rq))
7620 		return;
7621 
7622 	/* update runtime allocation */
7623 	account_cfs_rq_runtime(cfs_rq, 0);
7624 }
7625 
7626 static void sync_throttle(struct task_group *tg, int cpu)
7627 {
7628 	struct cfs_rq *pcfs_rq, *cfs_rq;
7629 
7630 	if (!cfs_bandwidth_used())
7631 		return;
7632 
7633 	if (!tg->parent)
7634 		return;
7635 
7636 	cfs_rq = tg_cfs_rq(tg, cpu);
7637 	pcfs_rq = tg_cfs_rq(tg->parent, cpu);
7638 
7639 	cfs_rq->throttle_count = pcfs_rq->throttle_count;
7640 	cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
7641 
7642 	/*
7643 	 * It is not enough to sync the "pelt_clock_throttled" indicator
7644 	 * with the parent cfs_rq when the hierarchy is not queued.
7645 	 * Always join a throttled hierarchy with PELT clock throttled
7646 	 * and leaf it to the first enqueue, or distribution to
7647 	 * unthrottle the PELT clock.
7648 	 */
7649 	if (cfs_rq->throttle_count)
7650 		cfs_rq->pelt_clock_throttled = 1;
7651 }
7652 
7653 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
7654 {
7655 	struct cfs_bandwidth *cfs_b =
7656 		container_of(timer, struct cfs_bandwidth, slack_timer);
7657 
7658 	do_sched_cfs_slack_timer(cfs_b);
7659 
7660 	return HRTIMER_NORESTART;
7661 }
7662 
7663 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
7664 {
7665 	struct cfs_bandwidth *cfs_b =
7666 		container_of(timer, struct cfs_bandwidth, period_timer);
7667 	int overrun;
7668 	int idle = 0;
7669 	int count = 0;
7670 
7671 	CLASS(raw_spinlock_irqsave, cfsb_guard)(&cfs_b->lock);
7672 
7673 	for (;;) {
7674 		overrun = hrtimer_forward_now(timer, cfs_b->period);
7675 		if (!overrun)
7676 			break;
7677 
7678 		idle = do_sched_cfs_period_timer(cfs_b, overrun, cfsb_guard.flags);
7679 
7680 		if (++count > 3) {
7681 			u64 new, old = ktime_to_ns(cfs_b->period);
7682 
7683 			/*
7684 			 * Grow period by a factor of 2 to avoid losing precision.
7685 			 * Precision loss in the quota/period ratio can cause __cfs_schedulable
7686 			 * to fail.
7687 			 */
7688 			new = old * 2;
7689 			if (new < max_bw_quota_period_us * NSEC_PER_USEC) {
7690 				cfs_b->period = ns_to_ktime(new);
7691 				cfs_b->quota *= 2;
7692 				cfs_b->burst *= 2;
7693 
7694 				pr_warn_ratelimited(
7695 	"cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
7696 					smp_processor_id(),
7697 					div_u64(new, NSEC_PER_USEC),
7698 					div_u64(cfs_b->quota, NSEC_PER_USEC));
7699 			} else {
7700 				pr_warn_ratelimited(
7701 	"cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
7702 					smp_processor_id(),
7703 					div_u64(old, NSEC_PER_USEC),
7704 					div_u64(cfs_b->quota, NSEC_PER_USEC));
7705 			}
7706 
7707 			/* reset count so we don't come right back in here */
7708 			count = 0;
7709 		}
7710 	}
7711 
7712 	if (idle) {
7713 		cfs_b->period_active = 0;
7714 		return HRTIMER_NORESTART;
7715 	}
7716 
7717 	return HRTIMER_RESTART;
7718 }
7719 
7720 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent)
7721 {
7722 	raw_spin_lock_init(&cfs_b->lock);
7723 	cfs_b->runtime = 0;
7724 	cfs_b->quota = RUNTIME_INF;
7725 	cfs_b->period = us_to_ktime(default_bw_period_us());
7726 	cfs_b->burst = 0;
7727 	cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;
7728 
7729 	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
7730 	hrtimer_setup(&cfs_b->period_timer, sched_cfs_period_timer, CLOCK_MONOTONIC,
7731 		      HRTIMER_MODE_ABS_PINNED);
7732 
7733 	/* Add a random offset so that timers interleave */
7734 	hrtimer_set_expires(&cfs_b->period_timer,
7735 			    get_random_u32_below(cfs_b->period));
7736 	hrtimer_setup(&cfs_b->slack_timer, sched_cfs_slack_timer, CLOCK_MONOTONIC,
7737 		      HRTIMER_MODE_REL);
7738 	cfs_b->slack_started = false;
7739 }
7740 
7741 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
7742 {
7743 	cfs_rq->runtime_enabled = 0;
7744 	INIT_LIST_HEAD(&cfs_rq->throttled_list);
7745 	INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
7746 	INIT_LIST_HEAD(&cfs_rq->throttled_limbo_list);
7747 }
7748 
7749 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
7750 {
7751 	lockdep_assert_held(&cfs_b->lock);
7752 
7753 	if (cfs_b->period_active)
7754 		return;
7755 
7756 	cfs_b->period_active = 1;
7757 	hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
7758 	hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
7759 }
7760 
7761 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
7762 {
7763 	int __maybe_unused i;
7764 
7765 	/* init_cfs_bandwidth() was not called */
7766 	if (!cfs_b->throttled_cfs_rq.next)
7767 		return;
7768 
7769 	hrtimer_cancel(&cfs_b->period_timer);
7770 	hrtimer_cancel(&cfs_b->slack_timer);
7771 
7772 	/*
7773 	 * It is possible that we still have some cfs_rq's pending on a CSD
7774 	 * list, though this race is very rare. In order for this to occur, we
7775 	 * must have raced with the last task leaving the group while there
7776 	 * exist throttled cfs_rq(s), and the period_timer must have queued the
7777 	 * CSD item but the remote cpu has not yet processed it. To handle this,
7778 	 * we can simply flush all pending CSD work inline here. We're
7779 	 * guaranteed at this point that no additional cfs_rq of this group can
7780 	 * join a CSD list.
7781 	 */
7782 	for_each_possible_cpu(i) {
7783 		struct rq *rq = cpu_rq(i);
7784 
7785 		if (list_empty(&rq->cfsb_csd_list))
7786 			continue;
7787 
7788 		scoped_guard(irqsave)
7789 			__cfsb_csd_unthrottle(rq);
7790 	}
7791 }
7792 
7793 /*
7794  * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
7795  *
7796  * The race is harmless, since modifying bandwidth settings of unhooked group
7797  * bits doesn't do much.
7798  */
7799 
7800 /* cpu online callback */
7801 static void __maybe_unused update_runtime_enabled(struct rq *rq)
7802 {
7803 	struct task_group *tg;
7804 
7805 	lockdep_assert_rq_held(rq);
7806 
7807 	guard(rcu)();
7808 
7809 	list_for_each_entry_rcu(tg, &task_groups, list) {
7810 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7811 		struct cfs_rq *cfs_rq = tg_cfs_rq(tg, cpu_of(rq));
7812 
7813 		scoped_guard(raw_spinlock, &cfs_b->lock)
7814 			cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
7815 	}
7816 }
7817 
7818 /* cpu offline callback */
7819 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
7820 {
7821 	struct task_group *tg;
7822 
7823 	lockdep_assert_rq_held(rq);
7824 
7825 	// Do not unthrottle for an active CPU
7826 	if (cpumask_test_cpu(cpu_of(rq), cpu_active_mask))
7827 		return;
7828 
7829 	/*
7830 	 * The rq clock has already been updated in the
7831 	 * set_rq_offline(), so we should skip updating
7832 	 * the rq clock again in unthrottle_cfs_rq().
7833 	 */
7834 	rq_clock_start_loop_update(rq);
7835 
7836 	guard(rcu)();
7837 
7838 	list_for_each_entry_rcu(tg, &task_groups, list) {
7839 		struct cfs_rq *cfs_rq = tg_cfs_rq(tg, cpu_of(rq));
7840 
7841 		if (!cfs_rq->runtime_enabled)
7842 			continue;
7843 
7844 		/*
7845 		 * Offline rq is schedulable till CPU is completely disabled
7846 		 * in take_cpu_down(), so we prevent new cfs throttling here.
7847 		 */
7848 		cfs_rq->runtime_enabled = 0;
7849 
7850 		if (!cfs_rq_throttled(cfs_rq))
7851 			continue;
7852 
7853 		/*
7854 		 * clock_task is not advancing so we just need to make sure
7855 		 * there's some valid quota amount
7856 		 */
7857 		cfs_rq->runtime_remaining = 1;
7858 		unthrottle_cfs_rq(cfs_rq);
7859 	}
7860 
7861 	rq_clock_stop_loop_update(rq);
7862 }
7863 
7864 bool cfs_task_bw_constrained(struct task_struct *p)
7865 {
7866 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
7867 
7868 	if (!cfs_bandwidth_used())
7869 		return false;
7870 
7871 	if (cfs_rq->runtime_enabled ||
7872 	    tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)
7873 		return true;
7874 
7875 	return false;
7876 }
7877 
7878 #ifdef CONFIG_NO_HZ_FULL
7879 /* called from pick_next_task_fair() */
7880 static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
7881 {
7882 	int cpu = cpu_of(rq);
7883 
7884 	if (!cfs_bandwidth_used())
7885 		return;
7886 
7887 	if (!tick_nohz_full_cpu(cpu))
7888 		return;
7889 
7890 	if (rq->nr_running != 1)
7891 		return;
7892 
7893 	/*
7894 	 *  We know there is only one task runnable and we've just picked it. The
7895 	 *  normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
7896 	 *  be otherwise able to stop the tick. Just need to check if we are using
7897 	 *  bandwidth control.
7898 	 */
7899 	if (cfs_task_bw_constrained(p))
7900 		tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
7901 }
7902 #endif /* CONFIG_NO_HZ_FULL */
7903 
7904 #else /* !CONFIG_CFS_BANDWIDTH: */
7905 
7906 static bool account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) { return false; }
7907 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
7908 static inline void sync_throttle(struct task_group *tg, int cpu) {}
7909 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
7910 static void task_throttle_setup_work(struct task_struct *p) {}
7911 static bool task_is_throttled(struct task_struct *p) { return false; }
7912 static void dequeue_throttled_task(struct task_struct *p, int flags) {}
7913 static bool enqueue_throttled_task(struct task_struct *p) { return false; }
7914 static void record_throttle_clock(struct cfs_rq *cfs_rq) {}
7915 
7916 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
7917 {
7918 	return 0;
7919 }
7920 
7921 static inline bool cfs_rq_pelt_clock_throttled(struct cfs_rq *cfs_rq)
7922 {
7923 	return false;
7924 }
7925 
7926 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
7927 {
7928 	return 0;
7929 }
7930 
7931 static inline int lb_throttled_hierarchy(struct task_struct *p, int dst_cpu)
7932 {
7933 	return 0;
7934 }
7935 
7936 #ifdef CONFIG_FAIR_GROUP_SCHED
7937 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {}
7938 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
7939 #endif
7940 
7941 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
7942 {
7943 	return NULL;
7944 }
7945 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
7946 static inline void update_runtime_enabled(struct rq *rq) {}
7947 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
7948 #ifdef CONFIG_CGROUP_SCHED
7949 bool cfs_task_bw_constrained(struct task_struct *p)
7950 {
7951 	return false;
7952 }
7953 #endif
7954 #endif /* !CONFIG_CFS_BANDWIDTH */
7955 
7956 #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL)
7957 static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {}
7958 #endif
7959 
7960 /**************************************************
7961  * CFS operations on tasks:
7962  */
7963 
7964 #ifdef CONFIG_SCHED_HRTICK
7965 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
7966 {
7967 	struct sched_entity *se = &p->se;
7968 	unsigned long scale = 1024;
7969 	unsigned long util = 0;
7970 	u64 vdelta;
7971 	u64 delta;
7972 
7973 	WARN_ON_ONCE(task_rq(p) != rq);
7974 
7975 	if (rq->cfs.h_nr_queued <= 1)
7976 		return;
7977 
7978 	/*
7979 	 * Compute time until virtual deadline
7980 	 */
7981 	vdelta = se->deadline - se->vruntime;
7982 	if ((s64)vdelta < 0) {
7983 		if (task_current_donor(rq, p))
7984 			resched_curr(rq);
7985 		return;
7986 	}
7987 	delta = (se->h_load.weight * vdelta) / NICE_0_LOAD;
7988 
7989 	/*
7990 	 * Correct for instantaneous load of other classes.
7991 	 */
7992 	util += cpu_util_irq(rq);
7993 	if (util && util < 1024) {
7994 		scale *= 1024;
7995 		scale /= (1024 - util);
7996 	}
7997 
7998 	hrtick_start(rq, (scale * delta) / 1024);
7999 }
8000 
8001 /*
8002  * Called on enqueue to start the hrtick when h_nr_queued becomes more than 1.
8003  */
8004 static void hrtick_update(struct rq *rq)
8005 {
8006 	struct task_struct *donor = rq->donor;
8007 
8008 	if (!hrtick_enabled_fair(rq) || donor->sched_class != &fair_sched_class)
8009 		return;
8010 
8011 	if (hrtick_active(rq))
8012 		return;
8013 
8014 	hrtick_start_fair(rq, donor);
8015 }
8016 #else /* !CONFIG_SCHED_HRTICK: */
8017 static inline void
8018 hrtick_start_fair(struct rq *rq, struct task_struct *p)
8019 {
8020 }
8021 
8022 static inline void hrtick_update(struct rq *rq)
8023 {
8024 }
8025 #endif /* !CONFIG_SCHED_HRTICK */
8026 
8027 static inline bool cpu_overutilized(int cpu)
8028 {
8029 	unsigned long rq_util_max;
8030 
8031 	if (!sched_energy_enabled())
8032 		return false;
8033 
8034 	rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
8035 
8036 	/* Return true only if the utilization doesn't fit CPU's capacity */
8037 	return !util_fits_cpu(cpu_util_cfs(cpu), 0, rq_util_max, cpu);
8038 }
8039 
8040 /*
8041  * overutilized value make sense only if EAS is enabled
8042  */
8043 static inline bool is_rd_overutilized(struct root_domain *rd)
8044 {
8045 	return !sched_energy_enabled() || READ_ONCE(rd->overutilized);
8046 }
8047 
8048 static inline void set_rd_overutilized(struct root_domain *rd, bool flag)
8049 {
8050 	if (!sched_energy_enabled())
8051 		return;
8052 
8053 	WRITE_ONCE(rd->overutilized, flag);
8054 	trace_sched_overutilized_tp(rd, flag);
8055 }
8056 
8057 static inline void check_update_overutilized_status(struct rq *rq)
8058 {
8059 	/*
8060 	 * overutilized field is used for load balancing decisions only
8061 	 * if energy aware scheduler is being used
8062 	 */
8063 
8064 	if (!is_rd_overutilized(rq->rd) && cpu_overutilized(rq->cpu))
8065 		set_rd_overutilized(rq->rd, 1);
8066 }
8067 
8068 /* Runqueue only has SCHED_IDLE tasks enqueued */
8069 static int sched_idle_rq(struct rq *rq)
8070 {
8071 	return unlikely(rq->nr_running == rq->cfs.h_nr_idle &&
8072 			rq->nr_running);
8073 }
8074 
8075 static int choose_sched_idle_rq(struct rq *rq, struct task_struct *p)
8076 {
8077 	return sched_idle_rq(rq) && !task_has_idle_policy(p);
8078 }
8079 
8080 static int choose_idle_cpu(int cpu, struct task_struct *p)
8081 {
8082 	return available_idle_cpu(cpu) ||
8083 	       choose_sched_idle_rq(cpu_rq(cpu), p);
8084 }
8085 
8086 static void
8087 requeue_delayed_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
8088 {
8089 	/*
8090 	 * se->sched_delayed should imply: se->on_rq == 1.
8091 	 * Because a delayed entity is one that is still on
8092 	 * the runqueue competing until elegibility.
8093 	 */
8094 	WARN_ON_ONCE(!se->sched_delayed);
8095 	WARN_ON_ONCE(!se->on_rq);
8096 
8097 	if (update_entity_lag(cfs_rq, se)) {
8098 		cfs_rq->h_nr_queued--;
8099 		if (se != cfs_rq->curr)
8100 			__dequeue_entity(cfs_rq, se);
8101 		place_entity(cfs_rq, se, 0);
8102 		if (se != cfs_rq->curr)
8103 			__enqueue_entity(cfs_rq, se);
8104 		cfs_rq->h_nr_queued++;
8105 	}
8106 
8107 	update_load_avg(cfs_rq, se, 0);
8108 	clear_delayed(se);
8109 }
8110 
8111 static unsigned long enqueue_hierarchy(struct task_struct *p, int flags)
8112 {
8113 	unsigned long weight = NICE_0_LOAD;
8114 	int task_new = !(flags & ENQUEUE_WAKEUP);
8115 	struct sched_entity *se = &p->se;
8116 	int h_nr_idle = task_has_idle_policy(p);
8117 	int h_nr_runnable = 1;
8118 
8119 	if (task_new && se->sched_delayed)
8120 		h_nr_runnable = 0;
8121 
8122 	for_each_sched_entity(se) {
8123 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
8124 
8125 		update_curr(cfs_rq);
8126 
8127 		if (!se->on_rq) {
8128 			enqueue_entity(cfs_rq, se, flags);
8129 		} else {
8130 			update_load_avg(cfs_rq, se, UPDATE_TG);
8131 			se_update_runnable(se);
8132 			update_cfs_group(se);
8133 		}
8134 
8135 		cfs_rq->h_nr_runnable += h_nr_runnable;
8136 		cfs_rq->h_nr_queued++;
8137 		cfs_rq->h_nr_idle += h_nr_idle;
8138 
8139 		if (cfs_rq_is_idle(cfs_rq))
8140 			h_nr_idle = 1;
8141 
8142 		weight = __calc_prop_weight(cfs_rq, se, weight);
8143 
8144 		flags = ENQUEUE_WAKEUP;
8145 	}
8146 
8147 	return weight;
8148 }
8149 
8150 /* Update curr's vruntime before placing entity or updating lag */
8151 static inline void update_curr_eevdf(struct cfs_rq *cfs_rq)
8152 {
8153 	if (!cfs_rq->curr)
8154 		return;
8155 
8156 	update_curr(cfs_rq_of(cfs_rq->curr));
8157 }
8158 
8159 /*
8160  * The enqueue_task method is called before nr_running is
8161  * increased. Here we update the fair scheduling stats and
8162  * then put the task into the rbtree:
8163  */
8164 static void
8165 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
8166 {
8167 	int rq_h_nr_queued = rq->cfs.h_nr_queued;
8168 	int task_new = !(flags & ENQUEUE_WAKEUP);
8169 	struct sched_entity *se = &p->se;
8170 	struct cfs_rq *cfs_rq = &rq->cfs;
8171 	unsigned long weight;
8172 	bool curr;
8173 
8174 	if (task_is_throttled(p) && enqueue_throttled_task(p))
8175 		return;
8176 
8177 	/*
8178 	 * The code below (indirectly) updates schedutil which looks at
8179 	 * the cfs_rq utilization to select a frequency.
8180 	 * Let's add the task's estimated utilization to the cfs_rq's
8181 	 * estimated utilization, before we update schedutil.
8182 	 */
8183 	if (!p->se.sched_delayed || (flags & ENQUEUE_DELAYED))
8184 		util_est_enqueue(cfs_rq, p);
8185 
8186 	update_curr_eevdf(cfs_rq);
8187 
8188 	if (flags & ENQUEUE_DELAYED) {
8189 		requeue_delayed_entity(cfs_rq, se);
8190 		return;
8191 	}
8192 
8193 	/*
8194 	 * If in_iowait is set, the code below may not trigger any cpufreq
8195 	 * utilization updates, so do it here explicitly with the IOWAIT flag
8196 	 * passed.
8197 	 */
8198 	if (p->in_iowait)
8199 		cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
8200 
8201 	curr = (cfs_rq->curr == se);
8202 	/*
8203 	 * This can actually happen with CFS_BANDWIDTH, a task can be throttled
8204 	 * and dequeued, unthrottled from another CPU and then enqueued again
8205 	 * before schedule().
8206 	 */
8207 	if (curr)
8208 		place_entity(cfs_rq, se, flags);
8209 
8210 	if (se->on_rq && se->sched_delayed)
8211 		requeue_delayed_entity(cfs_rq, se);
8212 
8213 	weight = enqueue_hierarchy(p, flags);
8214 
8215 	if (!curr) {
8216 		reweight_eevdf(cfs_rq, se, weight, false);
8217 		place_entity(cfs_rq, se, flags | ENQUEUE_QUEUED);
8218 		__enqueue_entity(cfs_rq, se);
8219 	}
8220 
8221 	if (!rq_h_nr_queued && rq->cfs.h_nr_queued)
8222 		dl_server_start(&rq->fair_server);
8223 
8224 	/* At this point se is NULL and we are at root level*/
8225 	add_nr_running(rq, 1);
8226 
8227 	/*
8228 	 * Since new tasks are assigned an initial util_avg equal to
8229 	 * half of the spare capacity of their CPU, tiny tasks have the
8230 	 * ability to cross the overutilized threshold, which will
8231 	 * result in the load balancer ruining all the task placement
8232 	 * done by EAS. As a way to mitigate that effect, do not account
8233 	 * for the first enqueue operation of new tasks during the
8234 	 * overutilized flag detection.
8235 	 *
8236 	 * A better way of solving this problem would be to wait for
8237 	 * the PELT signals of tasks to converge before taking them
8238 	 * into account, but that is not straightforward to implement,
8239 	 * and the following generally works well enough in practice.
8240 	 */
8241 	if (!task_new)
8242 		check_update_overutilized_status(rq);
8243 
8244 	assert_list_leaf_cfs_rq(rq);
8245 
8246 	hrtick_update(rq);
8247 }
8248 
8249 static void dequeue_hierarchy(struct task_struct *p, int flags)
8250 {
8251 	struct sched_entity *se = &p->se;
8252 	bool task_sleep = flags & DEQUEUE_SLEEP;
8253 	bool task_delayed = flags & DEQUEUE_DELAYED;
8254 	bool task_throttled = flags & DEQUEUE_THROTTLE;
8255 	int h_nr_runnable = 0;
8256 	int h_nr_idle = task_has_idle_policy(p);
8257 	bool dequeue = true;
8258 
8259 	if (task_sleep || task_delayed || !se->sched_delayed)
8260 		h_nr_runnable = 1;
8261 
8262 	for_each_sched_entity(se) {
8263 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
8264 
8265 		update_curr(cfs_rq);
8266 
8267 		if (dequeue) {
8268 			dequeue_entity(cfs_rq, se, flags);
8269 			/* Don't dequeue parent if it has other entities besides us */
8270 			if (cfs_rq->load.weight)
8271 				dequeue = false;
8272 		} else {
8273 			update_load_avg(cfs_rq, se, UPDATE_TG);
8274 			se_update_runnable(se);
8275 			update_cfs_group(se);
8276 		}
8277 
8278 		cfs_rq->h_nr_runnable -= h_nr_runnable;
8279 		cfs_rq->h_nr_queued--;
8280 		cfs_rq->h_nr_idle -= h_nr_idle;
8281 
8282 		if (cfs_rq_is_idle(cfs_rq))
8283 			h_nr_idle = 1;
8284 
8285 		if (throttled_hierarchy(cfs_rq) && task_throttled)
8286 			record_throttle_clock(cfs_rq);
8287 
8288 		flags |= DEQUEUE_SLEEP;
8289 		flags &= ~(DEQUEUE_DELAYED | DEQUEUE_SPECIAL);
8290 	}
8291 }
8292 
8293 /*
8294  * The part of dequeue_task_fair() that is needed to dequeue delayed tasks.
8295  *
8296  * Returns:
8297  *   true  - dequeued
8298  *   false - delayed
8299  */
8300 static bool __dequeue_task(struct rq *rq, struct task_struct *p, int flags)
8301 {
8302 	struct sched_entity *se = &p->se;
8303 	struct cfs_rq *cfs_rq = &rq->cfs;
8304 	bool was_sched_idle = sched_idle_rq(rq);
8305 	bool task_sleep = flags & DEQUEUE_SLEEP;
8306 	bool task_delayed = flags & DEQUEUE_DELAYED;
8307 
8308 	clear_buddies(cfs_rq, se);
8309 
8310 	update_curr_eevdf(cfs_rq);
8311 	update_entity_lag(cfs_rq, se);
8312 
8313 	if (flags & DEQUEUE_DELAYED) {
8314 		WARN_ON_ONCE(!se->sched_delayed);
8315 	} else {
8316 		bool delay = task_sleep;
8317 		/*
8318 		 * DELAY_DEQUEUE relies on spurious wakeups, special task
8319 		 * states must not suffer spurious wakeups, excempt them.
8320 		 */
8321 		if (flags & (DEQUEUE_SPECIAL | DEQUEUE_THROTTLE))
8322 			delay = false;
8323 
8324 		WARN_ON_ONCE(delay && se->sched_delayed);
8325 
8326 		if (sched_feat(DELAY_DEQUEUE) && delay &&
8327 		    !entity_eligible(cfs_rq, se)) {
8328 			update_load_avg(cfs_rq_of(se), se, UPDATE_UTIL_EST);
8329 			set_delayed(se);
8330 			return false;
8331 		}
8332 	}
8333 
8334 	dequeue_hierarchy(p, flags);
8335 
8336 	if (sched_feat(PLACE_REL_DEADLINE) && !task_sleep) {
8337 		se->deadline -= se->vruntime;
8338 		se->rel_deadline = 1;
8339 	}
8340 	if (se != cfs_rq->curr)
8341 		__dequeue_entity(cfs_rq, se);
8342 
8343 	sub_nr_running(rq, 1);
8344 
8345 	/* balance early to pull high priority tasks */
8346 	if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
8347 		rq->next_balance = jiffies;
8348 
8349 	if (task_delayed) {
8350 		clear_delayed(se);
8351 
8352 		WARN_ON_ONCE(!task_sleep);
8353 		WARN_ON_ONCE(p->on_rq != 1);
8354 
8355 		/*
8356 		 * Fix-up what block_task() skipped.
8357 		 *
8358 		 * Must be last, @p might not be valid after this.
8359 		 */
8360 		__block_task(rq, p);
8361 	}
8362 
8363 	return true;
8364 }
8365 
8366 /*
8367  * The dequeue_task method is called before nr_running is
8368  * decreased. We remove the task from the rbtree and
8369  * update the fair scheduling stats:
8370  */
8371 static bool dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
8372 {
8373 	if (task_is_throttled(p)) {
8374 		dequeue_throttled_task(p, flags);
8375 		return true;
8376 	}
8377 
8378 	if (!p->se.sched_delayed)
8379 		util_est_dequeue(&rq->cfs, p);
8380 
8381 	if (!__dequeue_task(rq, p, flags))
8382 		return false;
8383 
8384 	/*
8385 	 * Must not reference @p after __dequeue_task(DEQUEUE_DELAYED).
8386 	 */
8387 	return true;
8388 }
8389 
8390 static inline unsigned int cfs_h_nr_delayed(struct rq *rq)
8391 {
8392 	return (rq->cfs.h_nr_queued - rq->cfs.h_nr_runnable);
8393 }
8394 
8395 /* Working cpumask for: sched_balance_rq(), sched_balance_newidle(). */
8396 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8397 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
8398 static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask);
8399 
8400 #ifdef CONFIG_NO_HZ_COMMON
8401 
8402 static struct {
8403 	cpumask_var_t idle_cpus_mask;
8404 	int has_blocked_load;		/* Idle CPUS has blocked load */
8405 	int needs_update;		/* Newly idle CPUs need their next_balance collated */
8406 	unsigned long next_balance;     /* in jiffy units */
8407 	unsigned long next_blocked;	/* Next update of blocked load in jiffies */
8408 } nohz ____cacheline_aligned;
8409 
8410 #endif /* CONFIG_NO_HZ_COMMON */
8411 
8412 static unsigned long cpu_load(struct rq *rq)
8413 {
8414 	return cfs_rq_load_avg(&rq->cfs);
8415 }
8416 
8417 /*
8418  * cpu_load_without - compute CPU load without any contributions from *p
8419  * @cpu: the CPU which load is requested
8420  * @p: the task which load should be discounted
8421  *
8422  * The load of a CPU is defined by the load of tasks currently enqueued on that
8423  * CPU as well as tasks which are currently sleeping after an execution on that
8424  * CPU.
8425  *
8426  * This method returns the load of the specified CPU by discounting the load of
8427  * the specified task, whenever the task is currently contributing to the CPU
8428  * load.
8429  */
8430 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
8431 {
8432 	struct cfs_rq *cfs_rq;
8433 	unsigned int load;
8434 
8435 	/* Task has no contribution or is new */
8436 	if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8437 		return cpu_load(rq);
8438 
8439 	cfs_rq = &rq->cfs;
8440 	load = READ_ONCE(cfs_rq->avg.load_avg);
8441 
8442 	/* Discount task's util from CPU's util */
8443 	lsub_positive(&load, task_h_load(p));
8444 
8445 	return load;
8446 }
8447 
8448 static unsigned long cpu_runnable(struct rq *rq)
8449 {
8450 	return cfs_rq_runnable_avg(&rq->cfs);
8451 }
8452 
8453 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
8454 {
8455 	struct cfs_rq *cfs_rq;
8456 	unsigned int runnable;
8457 
8458 	/* Task has no contribution or is new */
8459 	if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8460 		return cpu_runnable(rq);
8461 
8462 	cfs_rq = &rq->cfs;
8463 	runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
8464 
8465 	/* Discount task's runnable from CPU's runnable */
8466 	lsub_positive(&runnable, p->se.avg.runnable_avg);
8467 
8468 	return runnable;
8469 }
8470 
8471 static unsigned long capacity_of(int cpu)
8472 {
8473 	return cpu_rq(cpu)->cpu_capacity;
8474 }
8475 
8476 static void record_wakee(struct task_struct *p)
8477 {
8478 	/*
8479 	 * Only decay a single time; tasks that have less then 1 wakeup per
8480 	 * jiffy will not have built up many flips.
8481 	 */
8482 	if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
8483 		current->wakee_flips >>= 1;
8484 		current->wakee_flip_decay_ts = jiffies;
8485 	}
8486 
8487 	if (current->last_wakee != p) {
8488 		current->last_wakee = p;
8489 		current->wakee_flips++;
8490 	}
8491 }
8492 
8493 /*
8494  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
8495  *
8496  * A waker of many should wake a different task than the one last awakened
8497  * at a frequency roughly N times higher than one of its wakees.
8498  *
8499  * In order to determine whether we should let the load spread vs consolidating
8500  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
8501  * partner, and a factor of lls_size higher frequency in the other.
8502  *
8503  * With both conditions met, we can be relatively sure that the relationship is
8504  * non-monogamous, with partner count exceeding socket size.
8505  *
8506  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
8507  * whatever is irrelevant, spread criteria is apparent partner count exceeds
8508  * socket size.
8509  */
8510 static int wake_wide(struct task_struct *p)
8511 {
8512 	unsigned int master = current->wakee_flips;
8513 	unsigned int slave = p->wakee_flips;
8514 	int factor = __this_cpu_read(sd_llc_size);
8515 
8516 	if (master < slave)
8517 		swap(master, slave);
8518 	if (slave < factor || master < slave * factor)
8519 		return 0;
8520 	return 1;
8521 }
8522 
8523 /*
8524  * The purpose of wake_affine() is to quickly determine on which CPU we can run
8525  * soonest. For the purpose of speed we only consider the waking and previous
8526  * CPU.
8527  *
8528  * wake_affine_idle() - only considers 'now', it check if the waking CPU is
8529  *			cache-affine and is (or	will be) idle.
8530  *
8531  * wake_affine_weight() - considers the weight to reflect the average
8532  *			  scheduling latency of the CPUs. This seems to work
8533  *			  for the overloaded case.
8534  */
8535 static int
8536 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
8537 {
8538 	/*
8539 	 * If this_cpu is idle, it implies the wakeup is from interrupt
8540 	 * context. Only allow the move if cache is shared. Otherwise an
8541 	 * interrupt intensive workload could force all tasks onto one
8542 	 * node depending on the IO topology or IRQ affinity settings.
8543 	 *
8544 	 * If the prev_cpu is idle and cache affine then avoid a migration.
8545 	 * There is no guarantee that the cache hot data from an interrupt
8546 	 * is more important than cache hot data on the prev_cpu and from
8547 	 * a cpufreq perspective, it's better to have higher utilisation
8548 	 * on one CPU.
8549 	 */
8550 	if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
8551 		return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
8552 
8553 	if (sync) {
8554 		struct rq *rq = cpu_rq(this_cpu);
8555 
8556 		if ((rq->nr_running - cfs_h_nr_delayed(rq)) == 1)
8557 			return this_cpu;
8558 	}
8559 
8560 	if (available_idle_cpu(prev_cpu))
8561 		return prev_cpu;
8562 
8563 	return nr_cpumask_bits;
8564 }
8565 
8566 static int
8567 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
8568 		   int this_cpu, int prev_cpu, int sync)
8569 {
8570 	s64 this_eff_load, prev_eff_load;
8571 	unsigned long task_load;
8572 
8573 	this_eff_load = cpu_load(cpu_rq(this_cpu));
8574 
8575 	if (sync) {
8576 		unsigned long current_load = task_h_load(current);
8577 
8578 		if (current_load > this_eff_load)
8579 			return this_cpu;
8580 
8581 		this_eff_load -= current_load;
8582 	}
8583 
8584 	task_load = task_h_load(p);
8585 
8586 	this_eff_load += task_load;
8587 	if (sched_feat(WA_BIAS))
8588 		this_eff_load *= 100;
8589 	this_eff_load *= capacity_of(prev_cpu);
8590 
8591 	prev_eff_load = cpu_load(cpu_rq(prev_cpu));
8592 	prev_eff_load -= task_load;
8593 	if (sched_feat(WA_BIAS))
8594 		prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
8595 	prev_eff_load *= capacity_of(this_cpu);
8596 
8597 	/*
8598 	 * If sync, adjust the weight of prev_eff_load such that if
8599 	 * prev_eff == this_eff that select_idle_sibling() will consider
8600 	 * stacking the wakee on top of the waker if no other CPU is
8601 	 * idle.
8602 	 */
8603 	if (sync)
8604 		prev_eff_load += 1;
8605 
8606 	return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
8607 }
8608 
8609 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
8610 		       int this_cpu, int prev_cpu, int sync)
8611 {
8612 	int target = nr_cpumask_bits;
8613 
8614 	if (sched_feat(WA_IDLE))
8615 		target = wake_affine_idle(this_cpu, prev_cpu, sync);
8616 
8617 	if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
8618 		target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
8619 
8620 	schedstat_inc(p->stats.nr_wakeups_affine_attempts);
8621 	if (target != this_cpu)
8622 		return prev_cpu;
8623 
8624 	schedstat_inc(sd->ttwu_move_affine);
8625 	schedstat_inc(p->stats.nr_wakeups_affine);
8626 	return target;
8627 }
8628 
8629 static struct sched_group *
8630 sched_balance_find_dst_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
8631 
8632 /*
8633  * sched_balance_find_dst_group_cpu - find the idlest CPU among the CPUs in the group.
8634  */
8635 static int
8636 sched_balance_find_dst_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
8637 {
8638 	unsigned long load, min_load = ULONG_MAX;
8639 	u64 min_exit_latency = U64_MAX;
8640 	unsigned int nr_candidates = 0;
8641 	int least_loaded_cpu = this_cpu;
8642 	int shallowest_idle_cpu = -1;
8643 	int i;
8644 
8645 	/* Check if we have any choice: */
8646 	if (group->group_weight == 1)
8647 		return cpumask_first(sched_group_span(group));
8648 
8649 	/* Traverse only the allowed CPUs */
8650 	for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
8651 		struct rq *rq = cpu_rq(i);
8652 
8653 		if (!sched_core_cookie_match(rq, p))
8654 			continue;
8655 
8656 		if (choose_sched_idle_rq(rq, p))
8657 			return i;
8658 
8659 		if (available_idle_cpu(i)) {
8660 			struct cpuidle_state *idle = idle_get_state(rq);
8661 			u64 exit_latency = idle ? idle->exit_latency : U64_MAX;
8662 
8663 			if (shallowest_idle_cpu == -1 || exit_latency < min_exit_latency) {
8664 				min_exit_latency = exit_latency;
8665 				shallowest_idle_cpu = i;
8666 				nr_candidates = 1;
8667 			} else if (exit_latency == min_exit_latency) {
8668 				nr_candidates++;
8669 				if (!reciprocal_scale(sched_rng(), nr_candidates))
8670 					shallowest_idle_cpu = i;
8671 			}
8672 		} else if (shallowest_idle_cpu == -1) {
8673 			load = cpu_load(cpu_rq(i));
8674 			if (load < min_load) {
8675 				min_load = load;
8676 				least_loaded_cpu = i;
8677 			}
8678 		}
8679 	}
8680 
8681 	return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
8682 }
8683 
8684 static inline int sched_balance_find_dst_cpu(struct sched_domain *sd, struct task_struct *p,
8685 				  int cpu, int prev_cpu, int sd_flag)
8686 {
8687 	int new_cpu = cpu;
8688 
8689 	if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
8690 		return prev_cpu;
8691 
8692 	/*
8693 	 * We need task's util for cpu_util_without, sync it up to
8694 	 * prev_cpu's last_update_time.
8695 	 */
8696 	if (!(sd_flag & SD_BALANCE_FORK))
8697 		sync_entity_load_avg(&p->se);
8698 
8699 	while (sd) {
8700 		struct sched_group *group;
8701 		struct sched_domain *tmp;
8702 		int weight;
8703 
8704 		if (!(sd->flags & sd_flag)) {
8705 			sd = sd->child;
8706 			continue;
8707 		}
8708 
8709 		group = sched_balance_find_dst_group(sd, p, cpu);
8710 		if (!group) {
8711 			sd = sd->child;
8712 			continue;
8713 		}
8714 
8715 		new_cpu = sched_balance_find_dst_group_cpu(group, p, cpu);
8716 		if (new_cpu == cpu) {
8717 			/* Now try balancing at a lower domain level of 'cpu': */
8718 			sd = sd->child;
8719 			continue;
8720 		}
8721 
8722 		/* Now try balancing at a lower domain level of 'new_cpu': */
8723 		cpu = new_cpu;
8724 		weight = sd->span_weight;
8725 		sd = NULL;
8726 		for_each_domain(cpu, tmp) {
8727 			if (weight <= tmp->span_weight)
8728 				break;
8729 			if (tmp->flags & sd_flag)
8730 				sd = tmp;
8731 		}
8732 	}
8733 
8734 	return new_cpu;
8735 }
8736 
8737 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
8738 {
8739 	if (choose_idle_cpu(cpu, p) && sched_cpu_cookie_match(cpu_rq(cpu), p))
8740 		return cpu;
8741 
8742 	return -1;
8743 }
8744 
8745 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
8746 EXPORT_SYMBOL_GPL(sched_smt_present);
8747 
8748 static inline void set_idle_cores(int cpu, int val)
8749 {
8750 	struct sched_domain_shared *sds;
8751 
8752 	sds = rcu_dereference_all(per_cpu(sd_balance_shared, cpu));
8753 	if (sds)
8754 		WRITE_ONCE(sds->has_idle_cores, val);
8755 }
8756 
8757 static inline bool test_idle_cores(int cpu)
8758 {
8759 	struct sched_domain_shared *sds;
8760 
8761 	sds = rcu_dereference_all(per_cpu(sd_balance_shared, cpu));
8762 	if (sds)
8763 		return READ_ONCE(sds->has_idle_cores);
8764 
8765 	return false;
8766 }
8767 
8768 /*
8769  * Redirect a CPU to a higher-priority available sibling in its SMT domain,
8770  * subject to task affinity.
8771  */
8772 static inline int select_idle_smt_cpu(struct task_struct *p, int cpu)
8773 {
8774 	struct sched_domain *sd;
8775 	int best = cpu;
8776 	int sibling;
8777 
8778 	if (!sched_smt_active())
8779 		return cpu;
8780 
8781 	sd = rcu_dereference_all(cpu_rq(cpu)->sd);
8782 	if (!sd || !(sd->flags & SD_SHARE_CPUCAPACITY) ||
8783 	    !(sd->flags & SD_ASYM_PACKING))
8784 		return cpu;
8785 
8786 	for_each_cpu_and(sibling, sched_domain_span(sd), p->cpus_ptr) {
8787 		if (sibling == best || !choose_idle_cpu(sibling, p))
8788 			continue;
8789 
8790 		if (sched_asym_prefer(sibling, best))
8791 			best = sibling;
8792 	}
8793 
8794 	return best;
8795 }
8796 
8797 /*
8798  * Scans the local SMT mask to see if the entire core is idle, and records this
8799  * information in sd_balance_shared->has_idle_cores.
8800  *
8801  * Since SMT siblings share all cache levels, inspecting this limited remote
8802  * state should be fairly cheap.
8803  */
8804 void __update_idle_core(struct rq *rq)
8805 {
8806 	int core = cpu_of(rq);
8807 	int cpu;
8808 
8809 	rcu_read_lock();
8810 	if (test_idle_cores(core))
8811 		goto unlock;
8812 
8813 	for_each_cpu(cpu, cpu_smt_mask(core)) {
8814 		if (cpu == core)
8815 			continue;
8816 
8817 		if (!available_idle_cpu(cpu))
8818 			goto unlock;
8819 	}
8820 
8821 	set_idle_cores(core, 1);
8822 unlock:
8823 	rcu_read_unlock();
8824 }
8825 
8826 /*
8827  * Scan the entire LLC domain for idle cores; this dynamically switches off if
8828  * there are no idle cores left in the system; tracked through
8829  * sd_balance_shared->has_idle_cores and enabled through update_idle_core()
8830  * above.
8831  */
8832 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
8833 {
8834 	bool idle = true;
8835 	int cpu;
8836 
8837 	for_each_cpu(cpu, cpu_smt_mask(core)) {
8838 		if (!available_idle_cpu(cpu)) {
8839 			idle = false;
8840 			if (*idle_cpu == -1) {
8841 				if (choose_sched_idle_rq(cpu_rq(cpu), p) &&
8842 				    cpumask_test_cpu(cpu, cpus)) {
8843 					*idle_cpu = cpu;
8844 					break;
8845 				}
8846 				continue;
8847 			}
8848 			break;
8849 		}
8850 		if (*idle_cpu == -1 && cpumask_test_cpu(cpu, cpus))
8851 			*idle_cpu = cpu;
8852 	}
8853 
8854 	if (idle)
8855 		return core;
8856 
8857 	cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
8858 	return -1;
8859 }
8860 
8861 /*
8862  * Scan the local SMT mask for idle CPUs.
8863  */
8864 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
8865 {
8866 	int cpu;
8867 
8868 	for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
8869 		if (cpu == target)
8870 			continue;
8871 		/*
8872 		 * Check if the CPU is in the LLC scheduling domain of @target.
8873 		 * Due to isolcpus, there is no guarantee that all the siblings are in the domain.
8874 		 */
8875 		if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
8876 			continue;
8877 		if (choose_idle_cpu(cpu, p))
8878 			return cpu;
8879 	}
8880 
8881 	return -1;
8882 }
8883 
8884 /*
8885  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
8886  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
8887  * average idle time for this rq (as found in rq->avg_idle).
8888  */
8889 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
8890 {
8891 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
8892 	int i, cpu, idle_cpu = -1, nr = INT_MAX;
8893 
8894 	if (sched_feat(SIS_UTIL) && sd->shared) {
8895 		/*
8896 		 * Increment because !--nr is the condition to stop scan.
8897 		 *
8898 		 * Since "sd" is "sd_llc" for target CPU dereferenced in the
8899 		 * caller, it is safe to directly dereference "sd->shared".
8900 		 * Topology bits always ensure it assigned for "sd_llc" abd it
8901 		 * cannot disappear as long as we have a RCU protected
8902 		 * reference to one the associated "sd" here.
8903 		 */
8904 		nr = READ_ONCE(sd->shared->nr_idle_scan) + 1;
8905 		/* overloaded LLC is unlikely to have idle cpu/core */
8906 		if (nr == 1)
8907 			return -1;
8908 	}
8909 
8910 	if (!cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr))
8911 		return -1;
8912 
8913 	if (static_branch_unlikely(&sched_cluster_active)) {
8914 		struct sched_group *sg = sd->groups;
8915 
8916 		if (sg->flags & SD_CLUSTER) {
8917 			for_each_cpu_wrap(cpu, sched_group_span(sg), target + 1) {
8918 				if (!cpumask_test_cpu(cpu, cpus))
8919 					continue;
8920 
8921 				if (has_idle_core) {
8922 					i = select_idle_core(p, cpu, cpus, &idle_cpu);
8923 					if ((unsigned int)i < nr_cpumask_bits)
8924 						return i;
8925 				} else {
8926 					if (--nr <= 0)
8927 						return -1;
8928 					idle_cpu = __select_idle_cpu(cpu, p);
8929 					if ((unsigned int)idle_cpu < nr_cpumask_bits)
8930 						return idle_cpu;
8931 				}
8932 			}
8933 			cpumask_andnot(cpus, cpus, sched_group_span(sg));
8934 		}
8935 	}
8936 
8937 	for_each_cpu_wrap(cpu, cpus, target + 1) {
8938 		if (has_idle_core) {
8939 			i = select_idle_core(p, cpu, cpus, &idle_cpu);
8940 			if ((unsigned int)i < nr_cpumask_bits)
8941 				return i;
8942 
8943 		} else {
8944 			if (--nr <= 0)
8945 				return -1;
8946 			idle_cpu = __select_idle_cpu(cpu, p);
8947 			if ((unsigned int)idle_cpu < nr_cpumask_bits)
8948 				break;
8949 		}
8950 	}
8951 
8952 	if (has_idle_core)
8953 		set_idle_cores(target, false);
8954 
8955 	return idle_cpu;
8956 }
8957 
8958 /*
8959  * Idle-capacity scan converts util_fits_cpu() outcomes into preference ranks,
8960  * where lower values indicate a better fit - see select_idle_capacity().
8961  *
8962  * A CPU that both fits the task and sits on a fully-idle SMT core is returned
8963  * immediately and is never assigned one of these ranks. On !SMT every CPU is
8964  * its own "core", so the early return covers all fits-and-idle cases and the
8965  * core-tier ranks below become unreachable.
8966  *
8967  *   Rank                            Val  Tier    Meaning
8968  *   ------------------------------  ---  ------  ---------------------------
8969  *   ASYM_IDLE_UCLAMP_MISFIT         -4   core    Idle core; capacity fits
8970  *                                                util but uclamp_min misses.
8971  *   ASYM_IDLE_COMPLETE_MISFIT       -3   core    Idle core; capacity does
8972  *                                                not fit. Still beats every
8973  *                                                thread-tier rank: a busy
8974  *                                                sibling cuts effective
8975  *                                                capacity more than a
8976  *                                                misfit hurts a quiet core.
8977  *   ASYM_IDLE_THREAD_FITS           -2   thread  Busy SMT sibling; capacity
8978  *                                                fits util + uclamp.
8979  *   ASYM_IDLE_THREAD_UCLAMP_MISFIT  -1   thread  Busy SMT sibling; capacity
8980  *                                                fits but uclamp_min misses
8981  *                                                (native util_fits_cpu()
8982  *                                                return value).
8983  *   ASYM_IDLE_THREAD_MISFIT          0   thread  Busy SMT sibling; capacity
8984  *                                                does not fit.
8985  *
8986  * ASYM_IDLE_CORE_BIAS (-3) is an offset, not a state. On an idle core,
8987  * fits += ASYM_IDLE_CORE_BIAS rebases thread-tier ranks into the core tier:
8988  *
8989  *   ASYM_IDLE_THREAD_UCLAMP_MISFIT (-1) + BIAS -> ASYM_IDLE_UCLAMP_MISFIT   (-4)
8990  *   ASYM_IDLE_THREAD_MISFIT         (0) + BIAS -> ASYM_IDLE_COMPLETE_MISFIT (-3)
8991  *
8992  * ASYM_IDLE_THREAD_FITS (-2) is never rebased because a fully-fitting idle-core
8993  * candidate early-returns from select_idle_capacity().
8994  */
8995 enum asym_fits_state {
8996 	ASYM_IDLE_UCLAMP_MISFIT = -4,
8997 	ASYM_IDLE_COMPLETE_MISFIT,
8998 	ASYM_IDLE_THREAD_FITS,
8999 	ASYM_IDLE_THREAD_UCLAMP_MISFIT,
9000 	ASYM_IDLE_THREAD_MISFIT,
9001 
9002 	/* util_fits_cpu() bias for idle core */
9003 	ASYM_IDLE_CORE_BIAS = -3,
9004 };
9005 
9006 /*
9007  * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
9008  * the task fits. If no CPU is big enough, but there are idle ones, try to
9009  * maximize capacity.
9010  */
9011 static int
9012 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
9013 {
9014 	/*
9015 	 * On !SMT systems, has_idle_core is always false and preferred_core
9016 	 * is always true (CPU == core), so the SMT preference logic below
9017 	 * collapses to the plain capacity scan.
9018 	 */
9019 	bool has_idle_core = sched_smt_active() && test_idle_cores(target);
9020 	unsigned long task_util, util_min, util_max, best_cap = 0;
9021 	int fits, best_fits = ASYM_IDLE_THREAD_MISFIT;
9022 	int cpu, best_cpu = -1;
9023 	struct cpumask *cpus;
9024 	int nr = INT_MAX;
9025 
9026 	cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
9027 	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
9028 
9029 	task_util = task_util_est(p);
9030 	util_min = uclamp_eff_value(p, UCLAMP_MIN);
9031 	util_max = uclamp_eff_value(p, UCLAMP_MAX);
9032 
9033 	if (sched_feat(SIS_UTIL) && sd->shared) {
9034 		/*
9035 		 * Same nr_idle_scan hint as select_idle_cpu(), nr only limits
9036 		 * the scan when not preferring an idle core.
9037 		 */
9038 		nr = READ_ONCE(sd->shared->nr_idle_scan) + 1;
9039 		/* overloaded domain is unlikely to have idle cpu/core */
9040 		if (nr == 1)
9041 			return -1;
9042 	}
9043 
9044 	for_each_cpu_wrap(cpu, cpus, target) {
9045 		bool preferred_core = !has_idle_core || is_core_idle(cpu);
9046 		unsigned long cpu_cap = capacity_of(cpu);
9047 
9048 		/*
9049 		 * Stop when the nr_idle_scan is exhausted (mirrors
9050 		 * select_idle_cpu() logic).
9051 		 */
9052 		if (!has_idle_core && --nr <= 0)
9053 			return best_cpu;
9054 
9055 		if (!choose_idle_cpu(cpu, p))
9056 			continue;
9057 
9058 		fits = util_fits_cpu(task_util, util_min, util_max, cpu);
9059 
9060 		/*
9061 		 * Perfect fit: capacity satisfies util + uclamp and the CPU
9062 		 * sits on a fully-idle SMT core, this is a !SMT system, or
9063 		 * there is no idle core to find.
9064 		 * Short-circuit the rank-based selection and return
9065 		 * immediately.
9066 		 */
9067 		if (fits > 0 && preferred_core)
9068 			return cpu;
9069 		/*
9070 		 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
9071 		 * Look for the CPU with best capacity.
9072 		 */
9073 		else if (fits < 0)
9074 			cpu_cap = get_actual_cpu_capacity(cpu);
9075 		/*
9076 		 * fits > 0 implies we are not on a preferred core, but the util
9077 		 * fits CPU capacity. Set fits to ASYM_IDLE_THREAD_FITS
9078 		 * so the effective range becomes
9079 		 * [ASYM_IDLE_THREAD_FITS, ASYM_IDLE_THREAD_MISFIT], where:
9080 		 *    ASYM_IDLE_THREAD_MISFIT - does not fit
9081 		 *    ASYM_IDLE_THREAD_UCLAMP_MISFIT - fits with the exception of UCLAMP_MIN
9082 		 *    ASYM_IDLE_THREAD_FITS - fits with the exception of preferred_core
9083 		 */
9084 		else if (fits > 0)
9085 			fits = ASYM_IDLE_THREAD_FITS;
9086 
9087 		/*
9088 		 * If we are on a preferred core, translate the range of fits
9089 		 * of [ASYM_IDLE_THREAD_UCLAMP_MISFIT, ASYM_IDLE_THREAD_MISFIT] to
9090 		 * [ASYM_IDLE_UCLAMP_MISFIT, ASYM_IDLE_COMPLETE_MISFIT].
9091 		 * This ensures that an idle core is always given priority over
9092 		 * (partially) busy core.
9093 		 *
9094 		 * A fully fitting idle core would have returned early and hence
9095 		 * fits > 0 for preferred_core need not be dealt with.
9096 		 */
9097 		if (preferred_core)
9098 			fits += ASYM_IDLE_CORE_BIAS;
9099 
9100 		/*
9101 		 * First, select CPU which fits better (lower is more preferred).
9102 		 * Then, select the one with best capacity at same level.
9103 		 */
9104 		if ((fits < best_fits) ||
9105 		    ((fits == best_fits) && (cpu_cap > best_cap))) {
9106 			best_cap = cpu_cap;
9107 			best_cpu = cpu;
9108 			best_fits = fits;
9109 		}
9110 	}
9111 
9112 	/*
9113 	 * A value in the [ASYM_IDLE_UCLAMP_MISFIT, ASYM_IDLE_COMPLETE_MISFIT]
9114 	 * range means the chosen CPU is in a fully idle SMT core. Values above
9115 	 * ASYM_IDLE_COMPLETE_MISFIT mean we never ranked such a CPU best.
9116 	 *
9117 	 * The asym-capacity wakeup path returns from select_idle_sibling()
9118 	 * after this function and never runs select_idle_cpu(), so the usual
9119 	 * select_idle_cpu() tail that clears idle cores must live here when the
9120 	 * idle-core preference did not win.
9121 	 */
9122 	if (has_idle_core && best_fits > ASYM_IDLE_COMPLETE_MISFIT)
9123 		set_idle_cores(target, false);
9124 
9125 	return best_cpu;
9126 }
9127 
9128 static inline bool asym_fits_cpu(unsigned long util,
9129 				 unsigned long util_min,
9130 				 unsigned long util_max,
9131 				 int cpu)
9132 {
9133 	if (sched_asym_cpucap_active()) {
9134 		/*
9135 		 * Return true only if the cpu fully fits the task requirements
9136 		 * which include the utilization and the performance hints.
9137 		 *
9138 		 * When SMT is active, also require that the core has no busy
9139 		 * siblings.
9140 		 *
9141 		 * Note: gating on is_core_idle() also makes the early-bailout
9142 		 * candidates in select_idle_sibling() (target, prev,
9143 		 * recent_used_cpu) idle-core-aware on ASYM+SMT, which the
9144 		 * NO_ASYM path does not do.
9145 		 */
9146 		return (!sched_smt_active() || is_core_idle(cpu)) &&
9147 		       (util_fits_cpu(util, util_min, util_max, cpu) > 0);
9148 	}
9149 
9150 	return true;
9151 }
9152 
9153 /*
9154  * Try and locate an idle core/thread in the LLC cache domain.
9155  */
9156 static int select_idle_sibling(struct task_struct *p, int prev, int target)
9157 {
9158 	bool has_idle_core = false;
9159 	struct sched_domain *sd;
9160 	unsigned long task_util, util_min, util_max;
9161 	int i, recent_used_cpu, prev_aff = -1;
9162 
9163 	/*
9164 	 * On asymmetric system, update task utilization because we will check
9165 	 * that the task fits with CPU's capacity.
9166 	 */
9167 	if (sched_asym_cpucap_active()) {
9168 		sync_entity_load_avg(&p->se);
9169 		task_util = task_util_est(p);
9170 		util_min = uclamp_eff_value(p, UCLAMP_MIN);
9171 		util_max = uclamp_eff_value(p, UCLAMP_MAX);
9172 	}
9173 
9174 	/*
9175 	 * per-cpu select_rq_mask usage
9176 	 */
9177 	lockdep_assert_irqs_disabled();
9178 
9179 	if (choose_idle_cpu(target, p) &&
9180 	    asym_fits_cpu(task_util, util_min, util_max, target))
9181 		goto select_smt_priority;
9182 
9183 	/*
9184 	 * If the previous CPU is cache affine and idle, don't be stupid:
9185 	 */
9186 	if (prev != target && cpus_share_cache(prev, target) &&
9187 	    choose_idle_cpu(prev, p) &&
9188 	    asym_fits_cpu(task_util, util_min, util_max, prev)) {
9189 
9190 		if (!static_branch_unlikely(&sched_cluster_active) ||
9191 		    cpus_share_resources(prev, target)) {
9192 			target = prev;
9193 			goto select_smt_priority;
9194 		}
9195 
9196 		prev_aff = prev;
9197 	}
9198 
9199 	/*
9200 	 * Allow a per-cpu kthread to stack with the wakee if the
9201 	 * kworker thread and the tasks previous CPUs are the same.
9202 	 * The assumption is that the wakee queued work for the
9203 	 * per-cpu kthread that is now complete and the wakeup is
9204 	 * essentially a sync wakeup. An obvious example of this
9205 	 * pattern is IO completions.
9206 	 */
9207 	if (is_per_cpu_kthread(current) &&
9208 	    in_task() &&
9209 	    prev == smp_processor_id() &&
9210 	    this_rq()->nr_running <= 1 &&
9211 	    asym_fits_cpu(task_util, util_min, util_max, prev)) {
9212 		target = prev;
9213 		goto select_smt_priority;
9214 	}
9215 
9216 	/* Check a recently used CPU as a potential idle candidate: */
9217 	recent_used_cpu = p->recent_used_cpu;
9218 	p->recent_used_cpu = prev;
9219 	if (recent_used_cpu != prev &&
9220 	    recent_used_cpu != target &&
9221 	    cpus_share_cache(recent_used_cpu, target) &&
9222 	    choose_idle_cpu(recent_used_cpu, p) &&
9223 	    cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&
9224 	    asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
9225 
9226 		if (!static_branch_unlikely(&sched_cluster_active) ||
9227 		    cpus_share_resources(recent_used_cpu, target)) {
9228 			target = recent_used_cpu;
9229 			goto select_smt_priority;
9230 		}
9231 
9232 	} else {
9233 		recent_used_cpu = -1;
9234 	}
9235 
9236 	/*
9237 	 * For asymmetric CPU capacity systems, our domain of interest is
9238 	 * sd_asym_cpucapacity rather than sd_llc.
9239 	 */
9240 	if (sched_asym_cpucap_active()) {
9241 		sd = rcu_dereference_all(per_cpu(sd_asym_cpucapacity, target));
9242 		/*
9243 		 * On an asymmetric CPU capacity system where an exclusive
9244 		 * cpuset defines a symmetric island (i.e. one unique
9245 		 * capacity_orig value through the cpuset), the key will be set
9246 		 * but the CPUs within that cpuset will not have a domain with
9247 		 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
9248 		 * capacity path.
9249 		 */
9250 		if (sd) {
9251 			i = select_idle_capacity(p, sd, target);
9252 			if ((unsigned int)i < nr_cpumask_bits) {
9253 				target = i;
9254 				goto select_smt_priority;
9255 			}
9256 			return target;
9257 		}
9258 	}
9259 
9260 	sd = rcu_dereference_all(per_cpu(sd_llc, target));
9261 	if (!sd)
9262 		return target;
9263 
9264 	if (sched_smt_active()) {
9265 		has_idle_core = test_idle_cores(target);
9266 
9267 		if (!has_idle_core && cpus_share_cache(prev, target)) {
9268 			i = select_idle_smt(p, sd, prev);
9269 			if ((unsigned int)i < nr_cpumask_bits) {
9270 				target = i;
9271 				goto select_smt_priority;
9272 			}
9273 		}
9274 	}
9275 
9276 	i = select_idle_cpu(p, sd, has_idle_core, target);
9277 	if ((unsigned int)i < nr_cpumask_bits) {
9278 		target = i;
9279 		goto select_smt_priority;
9280 	}
9281 
9282 	/*
9283 	 * For cluster machines which have lower sharing cache like L2 or
9284 	 * LLC Tag, we tend to find an idle CPU in the target's cluster
9285 	 * first. But prev_cpu or recent_used_cpu may also be a good candidate,
9286 	 * use them if possible when no idle CPU found in select_idle_cpu().
9287 	 */
9288 	if ((unsigned int)prev_aff < nr_cpumask_bits) {
9289 		target = prev_aff;
9290 		goto select_smt_priority;
9291 	}
9292 	if ((unsigned int)recent_used_cpu < nr_cpumask_bits) {
9293 		target = recent_used_cpu;
9294 		goto select_smt_priority;
9295 	}
9296 
9297 	return target;
9298 
9299 select_smt_priority:
9300 	return select_idle_smt_cpu(p, target);
9301 }
9302 
9303 /**
9304  * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
9305  * @cpu: the CPU to get the utilization for
9306  * @p: task for which the CPU utilization should be predicted or NULL
9307  * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
9308  * @boost: 1 to enable boosting, otherwise 0
9309  *
9310  * The unit of the return value must be the same as the one of CPU capacity
9311  * so that CPU utilization can be compared with CPU capacity.
9312  *
9313  * CPU utilization is the sum of running time of runnable tasks plus the
9314  * recent utilization of currently non-runnable tasks on that CPU.
9315  * It represents the amount of CPU capacity currently used by CFS tasks in
9316  * the range [0..max CPU capacity] with max CPU capacity being the CPU
9317  * capacity at f_max.
9318  *
9319  * The estimated CPU utilization is defined as the maximum between CPU
9320  * utilization and sum of the estimated utilization of the currently
9321  * runnable tasks on that CPU. It preserves a utilization "snapshot" of
9322  * previously-executed tasks, which helps better deduce how busy a CPU will
9323  * be when a long-sleeping task wakes up. The contribution to CPU utilization
9324  * of such a task would be significantly decayed at this point of time.
9325  *
9326  * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization).
9327  * CPU contention for CFS tasks can be detected by CPU runnable > CPU
9328  * utilization. Boosting is implemented in cpu_util() so that internal
9329  * users (e.g. EAS) can use it next to external users (e.g. schedutil),
9330  * latter via cpu_util_cfs_boost().
9331  *
9332  * CPU utilization can be higher than the current CPU capacity
9333  * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because
9334  * of rounding errors as well as task migrations or wakeups of new tasks.
9335  * CPU utilization has to be capped to fit into the [0..max CPU capacity]
9336  * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%)
9337  * could be seen as over-utilized even though CPU1 has 20% of spare CPU
9338  * capacity. CPU utilization is allowed to overshoot current CPU capacity
9339  * though since this is useful for predicting the CPU capacity required
9340  * after task migrations (scheduler-driven DVFS).
9341  *
9342  * Return: (Boosted) (estimated) utilization for the specified CPU.
9343  */
9344 static unsigned long
9345 cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
9346 {
9347 	bool add_task = p && task_cpu(p) != cpu && dst_cpu == cpu;
9348 	bool sub_task = p && task_cpu(p) == cpu && dst_cpu != cpu;
9349 	struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
9350 	unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
9351 	unsigned long runnable;
9352 
9353 	/*
9354 	 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
9355 	 * contribution. If @p migrates from another CPU to @cpu add its
9356 	 * contribution. In all the other cases @cpu is not impacted by the
9357 	 * migration so its util_avg is already correct.
9358 	 */
9359 	if (add_task)
9360 		util += task_util(p);
9361 	else if (sub_task)
9362 		lsub_positive(&util, task_util(p));
9363 
9364 	if (boost) {
9365 		runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
9366 		if (add_task)
9367 			runnable += READ_ONCE(p->se.avg.runnable_avg);
9368 		else if (sub_task)
9369 			lsub_positive(&runnable,
9370 				      READ_ONCE(p->se.avg.runnable_avg));
9371 		util = max(util, runnable);
9372 	}
9373 
9374 	if (sched_feat(UTIL_EST)) {
9375 		unsigned long util_est;
9376 
9377 		util_est = READ_ONCE(cfs_rq->avg.util_est);
9378 
9379 		/*
9380 		 * During wake-up @p isn't enqueued yet and doesn't contribute
9381 		 * to any cpu_rq(cpu)->cfs.avg.util_est.
9382 		 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
9383 		 * has been enqueued.
9384 		 *
9385 		 * During exec (@dst_cpu = -1) @p is enqueued and does
9386 		 * contribute to cpu_rq(cpu)->cfs.util_est.
9387 		 * Remove it to "simulate" cpu_util without @p's contribution.
9388 		 *
9389 		 * Despite the task_on_rq_queued(@p) check there is still a
9390 		 * small window for a possible race when an exec
9391 		 * select_task_rq_fair() races with LB's detach_task().
9392 		 *
9393 		 *   detach_task()
9394 		 *     deactivate_task()
9395 		 *       p->on_rq = TASK_ON_RQ_MIGRATING;
9396 		 *       -------------------------------- A
9397 		 *       dequeue_task()                    \
9398 		 *         dequeue_task_fair()              + Race Time
9399 		 *           util_est_dequeue()            /
9400 		 *       -------------------------------- B
9401 		 *
9402 		 * The additional check "current == p" is required to further
9403 		 * reduce the race window.
9404 		 */
9405 		if (dst_cpu == cpu)
9406 			util_est += _task_util_est(p);
9407 		else if (p && unlikely(task_on_rq_queued(p) || current == p))
9408 			lsub_positive(&util_est, _task_util_est(p));
9409 
9410 		util = max(util, util_est);
9411 	}
9412 
9413 	return min(util, arch_scale_cpu_capacity(cpu));
9414 }
9415 
9416 unsigned long cpu_util_cfs(int cpu)
9417 {
9418 	return cpu_util(cpu, NULL, -1, 0);
9419 }
9420 
9421 unsigned long cpu_util_cfs_boost(int cpu)
9422 {
9423 	return cpu_util(cpu, NULL, -1, 1);
9424 }
9425 
9426 /*
9427  * cpu_util_without: compute cpu utilization without any contributions from *p
9428  * @cpu: the CPU which utilization is requested
9429  * @p: the task which utilization should be discounted
9430  *
9431  * The utilization of a CPU is defined by the utilization of tasks currently
9432  * enqueued on that CPU as well as tasks which are currently sleeping after an
9433  * execution on that CPU.
9434  *
9435  * This method returns the utilization of the specified CPU by discounting the
9436  * utilization of the specified task, whenever the task is currently
9437  * contributing to the CPU utilization.
9438  */
9439 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
9440 {
9441 	/* Task has no contribution or is new */
9442 	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
9443 		p = NULL;
9444 
9445 	return cpu_util(cpu, p, -1, 0);
9446 }
9447 
9448 /*
9449  * This function computes an effective utilization for the given CPU, to be
9450  * used for frequency selection given the linear relation: f = u * f_max.
9451  *
9452  * The scheduler tracks the following metrics:
9453  *
9454  *   cpu_util_{cfs,rt,dl,irq}()
9455  *   cpu_bw_dl()
9456  *
9457  * Where the cfs,rt and dl util numbers are tracked with the same metric and
9458  * synchronized windows and are thus directly comparable.
9459  *
9460  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
9461  * which excludes things like IRQ and steal-time. These latter are then accrued
9462  * in the IRQ utilization.
9463  *
9464  * The DL bandwidth number OTOH is not a measured metric but a value computed
9465  * based on the task model parameters and gives the minimal utilization
9466  * required to meet deadlines.
9467  */
9468 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
9469 				 unsigned long *min,
9470 				 unsigned long *max)
9471 {
9472 	unsigned long util, irq, scale;
9473 	struct rq *rq = cpu_rq(cpu);
9474 
9475 	scale = arch_scale_cpu_capacity(cpu);
9476 
9477 	/*
9478 	 * Early check to see if IRQ/steal time saturates the CPU, can be
9479 	 * because of inaccuracies in how we track these -- see
9480 	 * update_irq_load_avg().
9481 	 */
9482 	irq = cpu_util_irq(rq);
9483 	if (unlikely(irq >= scale)) {
9484 		if (min)
9485 			*min = scale;
9486 		if (max)
9487 			*max = scale;
9488 		return scale;
9489 	}
9490 
9491 	if (min) {
9492 		/*
9493 		 * The minimum utilization returns the highest level between:
9494 		 * - the computed DL bandwidth needed with the IRQ pressure which
9495 		 *   steals time to the deadline task.
9496 		 * - The minimum performance requirement for CFS and/or RT.
9497 		 */
9498 		*min = max(irq + cpu_bw_dl(rq), uclamp_rq_get(rq, UCLAMP_MIN));
9499 
9500 		/*
9501 		 * When an RT task is runnable and uclamp is not used, we must
9502 		 * ensure that the task will run at maximum compute capacity.
9503 		 */
9504 		if (!uclamp_is_used() && rt_rq_is_runnable(&rq->rt))
9505 			*min = max(*min, scale);
9506 	}
9507 
9508 	/*
9509 	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
9510 	 * CFS tasks and we use the same metric to track the effective
9511 	 * utilization (PELT windows are synchronized) we can directly add them
9512 	 * to obtain the CPU's actual utilization.
9513 	 */
9514 	util = util_cfs + cpu_util_rt(rq);
9515 	util += cpu_util_dl(rq);
9516 
9517 	/*
9518 	 * The maximum hint is a soft bandwidth requirement, which can be lower
9519 	 * than the actual utilization because of uclamp_max requirements.
9520 	 */
9521 	if (max)
9522 		*max = min(scale, uclamp_rq_get(rq, UCLAMP_MAX));
9523 
9524 	if (util >= scale)
9525 		return scale;
9526 
9527 	/*
9528 	 * There is still idle time; further improve the number by using the
9529 	 * IRQ metric. Because IRQ/steal time is hidden from the task clock we
9530 	 * need to scale the task numbers:
9531 	 *
9532 	 *              max - irq
9533 	 *   U' = irq + --------- * U
9534 	 *                 max
9535 	 */
9536 	util = scale_irq_capacity(util, irq, scale);
9537 	util += irq;
9538 
9539 	return min(scale, util);
9540 }
9541 
9542 unsigned long sched_cpu_util(int cpu)
9543 {
9544 	return effective_cpu_util(cpu, cpu_util_cfs(cpu), NULL, NULL);
9545 }
9546 
9547 /*
9548  * energy_env - Utilization landscape for energy estimation.
9549  * @task_busy_time: Utilization contribution by the task for which we test the
9550  *                  placement. Given by eenv_task_busy_time().
9551  * @pd_busy_time:   Utilization of the whole perf domain without the task
9552  *                  contribution. Given by eenv_pd_busy_time().
9553  * @cpu_cap:        Maximum CPU capacity for the perf domain.
9554  * @pd_cap:         Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
9555  */
9556 struct energy_env {
9557 	unsigned long task_busy_time;
9558 	unsigned long pd_busy_time;
9559 	unsigned long cpu_cap;
9560 	unsigned long pd_cap;
9561 };
9562 
9563 /*
9564  * Compute the task busy time for compute_energy(). This time cannot be
9565  * injected directly into effective_cpu_util() because of the IRQ scaling.
9566  * The latter only makes sense with the most recent CPUs where the task has
9567  * run.
9568  */
9569 static inline void eenv_task_busy_time(struct energy_env *eenv,
9570 				       struct task_struct *p, int prev_cpu)
9571 {
9572 	unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
9573 	unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
9574 
9575 	if (unlikely(irq >= max_cap))
9576 		busy_time = max_cap;
9577 	else
9578 		busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
9579 
9580 	eenv->task_busy_time = busy_time;
9581 }
9582 
9583 /*
9584  * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
9585  * utilization for each @pd_cpus, it however doesn't take into account
9586  * clamping since the ratio (utilization / cpu_capacity) is already enough to
9587  * scale the EM reported power consumption at the (eventually clamped)
9588  * cpu_capacity.
9589  *
9590  * The contribution of the task @p for which we want to estimate the
9591  * energy cost is removed (by cpu_util()) and must be calculated
9592  * separately (see eenv_task_busy_time). This ensures:
9593  *
9594  *   - A stable PD utilization, no matter which CPU of that PD we want to place
9595  *     the task on.
9596  *
9597  *   - A fair comparison between CPUs as the task contribution (task_util())
9598  *     will always be the same no matter which CPU utilization we rely on
9599  *     (util_avg or util_est).
9600  *
9601  * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
9602  * exceed @eenv->pd_cap.
9603  */
9604 static inline void eenv_pd_busy_time(struct energy_env *eenv,
9605 				     struct cpumask *pd_cpus,
9606 				     struct task_struct *p)
9607 {
9608 	unsigned long busy_time = 0;
9609 	int cpu;
9610 
9611 	for_each_cpu(cpu, pd_cpus) {
9612 		unsigned long util = cpu_util(cpu, p, -1, 0);
9613 
9614 		busy_time += effective_cpu_util(cpu, util, NULL, NULL);
9615 	}
9616 
9617 	eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
9618 }
9619 
9620 /*
9621  * Compute the maximum utilization for compute_energy() when the task @p
9622  * is placed on the cpu @dst_cpu.
9623  *
9624  * Returns the maximum utilization among @eenv->cpus. This utilization can't
9625  * exceed @eenv->cpu_cap.
9626  */
9627 static inline unsigned long
9628 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
9629 		 struct task_struct *p, int dst_cpu)
9630 {
9631 	unsigned long max_util = 0;
9632 	int cpu;
9633 
9634 	for_each_cpu(cpu, pd_cpus) {
9635 		struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
9636 		unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
9637 		unsigned long eff_util, min, max;
9638 
9639 		/*
9640 		 * Performance domain frequency: utilization clamping
9641 		 * must be considered since it affects the selection
9642 		 * of the performance domain frequency.
9643 		 * NOTE: in case RT tasks are running, by default the min
9644 		 * utilization can be max OPP.
9645 		 */
9646 		eff_util = effective_cpu_util(cpu, util, &min, &max);
9647 
9648 		/* Task's uclamp can modify min and max value */
9649 		if (tsk && uclamp_is_used()) {
9650 			min = max(min, uclamp_eff_value(p, UCLAMP_MIN));
9651 
9652 			/*
9653 			 * If there is no active max uclamp constraint,
9654 			 * directly use task's one, otherwise keep max.
9655 			 */
9656 			if (uclamp_rq_is_idle(cpu_rq(cpu)))
9657 				max = uclamp_eff_value(p, UCLAMP_MAX);
9658 			else
9659 				max = max(max, uclamp_eff_value(p, UCLAMP_MAX));
9660 		}
9661 
9662 		eff_util = sugov_effective_cpu_perf(cpu, eff_util, min, max);
9663 		max_util = max(max_util, eff_util);
9664 	}
9665 
9666 	return min(max_util, eenv->cpu_cap);
9667 }
9668 
9669 /*
9670  * compute_energy(): Use the Energy Model to estimate the energy that @pd would
9671  * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
9672  * contribution is ignored.
9673  */
9674 static inline unsigned long
9675 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
9676 	       struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
9677 {
9678 	unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
9679 	unsigned long busy_time = eenv->pd_busy_time;
9680 	unsigned long energy;
9681 
9682 	if (dst_cpu >= 0)
9683 		busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
9684 
9685 	energy = em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
9686 
9687 	trace_sched_compute_energy_tp(p, dst_cpu, energy, max_util, busy_time);
9688 
9689 	return energy;
9690 }
9691 
9692 /*
9693  * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
9694  * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
9695  * spare capacity in each performance domain and uses it as a potential
9696  * candidate to execute the task. Then, it uses the Energy Model to figure
9697  * out which of the CPU candidates is the most energy-efficient.
9698  *
9699  * The rationale for this heuristic is as follows. In a performance domain,
9700  * all the most energy efficient CPU candidates (according to the Energy
9701  * Model) are those for which we'll request a low frequency. When there are
9702  * several CPUs for which the frequency request will be the same, we don't
9703  * have enough data to break the tie between them, because the Energy Model
9704  * only includes active power costs. With this model, if we assume that
9705  * frequency requests follow utilization (e.g. using schedutil), the CPU with
9706  * the maximum spare capacity in a performance domain is guaranteed to be among
9707  * the best candidates of the performance domain.
9708  *
9709  * In practice, it could be preferable from an energy standpoint to pack
9710  * small tasks on a CPU in order to let other CPUs go in deeper idle states,
9711  * but that could also hurt our chances to go cluster idle, and we have no
9712  * ways to tell with the current Energy Model if this is actually a good
9713  * idea or not. So, find_energy_efficient_cpu() basically favors
9714  * cluster-packing, and spreading inside a cluster. That should at least be
9715  * a good thing for latency, and this is consistent with the idea that most
9716  * of the energy savings of EAS come from the asymmetry of the system, and
9717  * not so much from breaking the tie between identical CPUs. That's also the
9718  * reason why EAS is enabled in the topology code only for systems where
9719  * SD_ASYM_CPUCAPACITY is set.
9720  *
9721  * NOTE: Forkees are not accepted in the energy-aware wake-up path because
9722  * they don't have any useful utilization data yet and it's not possible to
9723  * forecast their impact on energy consumption. Consequently, they will be
9724  * placed by sched_balance_find_dst_cpu() on the least loaded CPU, which might turn out
9725  * to be energy-inefficient in some use-cases. The alternative would be to
9726  * bias new tasks towards specific types of CPUs first, or to try to infer
9727  * their util_avg from the parent task, but those heuristics could hurt
9728  * other use-cases too. So, until someone finds a better way to solve this,
9729  * let's keep things simple by re-using the existing slow path.
9730  */
9731 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
9732 {
9733 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
9734 	unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
9735 	unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
9736 	unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
9737 	struct root_domain *rd = this_rq()->rd;
9738 	int cpu, best_energy_cpu, target = -1;
9739 	int prev_fits = -1, best_fits = -1;
9740 	unsigned long best_actual_cap = 0;
9741 	unsigned long prev_actual_cap = 0;
9742 	struct sched_domain *sd;
9743 	struct perf_domain *pd;
9744 	struct energy_env eenv;
9745 
9746 	pd = rcu_dereference_all(rd->pd);
9747 	if (!pd)
9748 		return target;
9749 
9750 	/*
9751 	 * Energy-aware wake-up happens on the lowest sched_domain starting
9752 	 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
9753 	 */
9754 	sd = rcu_dereference_all(*this_cpu_ptr(&sd_asym_cpucapacity));
9755 	while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
9756 		sd = sd->parent;
9757 	if (!sd)
9758 		return target;
9759 
9760 	target = prev_cpu;
9761 
9762 	sync_entity_load_avg(&p->se);
9763 	if (!task_util_est(p) && p_util_min == 0)
9764 		return target;
9765 
9766 	eenv_task_busy_time(&eenv, p, prev_cpu);
9767 
9768 	for (; pd; pd = pd->next) {
9769 		unsigned long util_min = p_util_min, util_max = p_util_max;
9770 		unsigned long cpu_cap, cpu_actual_cap, util;
9771 		long prev_spare_cap = -1, max_spare_cap = -1;
9772 		unsigned long rq_util_min, rq_util_max;
9773 		unsigned long cur_delta, base_energy;
9774 		int max_spare_cap_cpu = -1;
9775 		int fits, max_fits = -1;
9776 
9777 		if (!cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask))
9778 			continue;
9779 
9780 		/* Account external pressure for the energy estimation */
9781 		cpu = cpumask_first(cpus);
9782 		cpu_actual_cap = get_actual_cpu_capacity(cpu);
9783 
9784 		eenv.cpu_cap = cpu_actual_cap;
9785 		eenv.pd_cap = 0;
9786 
9787 		for_each_cpu(cpu, cpus) {
9788 			struct rq *rq = cpu_rq(cpu);
9789 
9790 			eenv.pd_cap += cpu_actual_cap;
9791 
9792 			if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
9793 				continue;
9794 
9795 			if (!cpumask_test_cpu(cpu, p->cpus_ptr))
9796 				continue;
9797 
9798 			util = cpu_util(cpu, p, cpu, 0);
9799 			cpu_cap = capacity_of(cpu);
9800 
9801 			/*
9802 			 * Skip CPUs that cannot satisfy the capacity request.
9803 			 * IOW, placing the task there would make the CPU
9804 			 * overutilized. Take uclamp into account to see how
9805 			 * much capacity we can get out of the CPU; this is
9806 			 * aligned with sched_cpu_util().
9807 			 */
9808 			if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
9809 				/*
9810 				 * Open code uclamp_rq_util_with() except for
9811 				 * the clamp() part. I.e.: apply max aggregation
9812 				 * only. util_fits_cpu() logic requires to
9813 				 * operate on non clamped util but must use the
9814 				 * max-aggregated uclamp_{min, max}.
9815 				 */
9816 				rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
9817 				rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
9818 
9819 				util_min = max(rq_util_min, p_util_min);
9820 				util_max = max(rq_util_max, p_util_max);
9821 			}
9822 
9823 			fits = util_fits_cpu(util, util_min, util_max, cpu);
9824 			if (!fits)
9825 				continue;
9826 
9827 			lsub_positive(&cpu_cap, util);
9828 
9829 			if (cpu == prev_cpu) {
9830 				/* Always use prev_cpu as a candidate. */
9831 				prev_spare_cap = cpu_cap;
9832 				prev_fits = fits;
9833 			} else if ((fits > max_fits) ||
9834 				   ((fits == max_fits) && ((long)cpu_cap > max_spare_cap))) {
9835 				/*
9836 				 * Find the CPU with the maximum spare capacity
9837 				 * among the remaining CPUs in the performance
9838 				 * domain.
9839 				 */
9840 				max_spare_cap = cpu_cap;
9841 				max_spare_cap_cpu = cpu;
9842 				max_fits = fits;
9843 			}
9844 		}
9845 
9846 		if (max_spare_cap_cpu < 0 && prev_spare_cap < 0)
9847 			continue;
9848 
9849 		eenv_pd_busy_time(&eenv, cpus, p);
9850 		/* Compute the 'base' energy of the pd, without @p */
9851 		base_energy = compute_energy(&eenv, pd, cpus, p, -1);
9852 
9853 		/* Evaluate the energy impact of using prev_cpu. */
9854 		if (prev_spare_cap > -1) {
9855 			prev_delta = compute_energy(&eenv, pd, cpus, p,
9856 						    prev_cpu);
9857 			/* CPU utilization has changed */
9858 			if (prev_delta < base_energy)
9859 				return target;
9860 			prev_delta -= base_energy;
9861 			prev_actual_cap = cpu_actual_cap;
9862 			best_delta = min(best_delta, prev_delta);
9863 		}
9864 
9865 		/* Evaluate the energy impact of using max_spare_cap_cpu. */
9866 		if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
9867 			/* Current best energy cpu fits better */
9868 			if (max_fits < best_fits)
9869 				continue;
9870 
9871 			/*
9872 			 * Both don't fit performance hint (i.e. uclamp_min)
9873 			 * but best energy cpu has better capacity.
9874 			 */
9875 			if ((max_fits < 0) &&
9876 			    (cpu_actual_cap <= best_actual_cap))
9877 				continue;
9878 
9879 			cur_delta = compute_energy(&eenv, pd, cpus, p,
9880 						   max_spare_cap_cpu);
9881 			/* CPU utilization has changed */
9882 			if (cur_delta < base_energy)
9883 				return target;
9884 			cur_delta -= base_energy;
9885 
9886 			/*
9887 			 * Both fit for the task but best energy cpu has lower
9888 			 * energy impact.
9889 			 */
9890 			if ((max_fits > 0) && (best_fits > 0) &&
9891 			    (cur_delta >= best_delta))
9892 				continue;
9893 
9894 			best_delta = cur_delta;
9895 			best_energy_cpu = max_spare_cap_cpu;
9896 			best_fits = max_fits;
9897 			best_actual_cap = cpu_actual_cap;
9898 		}
9899 	}
9900 
9901 	if ((best_fits > prev_fits) ||
9902 	    ((best_fits > 0) && (best_delta < prev_delta)) ||
9903 	    ((best_fits < 0) && (best_actual_cap > prev_actual_cap)))
9904 		target = best_energy_cpu;
9905 
9906 	return target;
9907 }
9908 
9909 /*
9910  * select_task_rq_fair: Select target runqueue for the waking task in domains
9911  * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
9912  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
9913  *
9914  * Balances load by selecting the idlest CPU in the idlest group, or under
9915  * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
9916  *
9917  * Returns the target CPU number.
9918  */
9919 static int
9920 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
9921 {
9922 	int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
9923 	struct sched_domain *tmp, *sd = NULL;
9924 	int cpu = smp_processor_id();
9925 	int new_cpu = prev_cpu;
9926 	int want_affine = 0;
9927 	/* SD_flags and WF_flags share the first nibble */
9928 	int sd_flag = wake_flags & 0xF;
9929 
9930 	/*
9931 	 * required for stable ->cpus_allowed
9932 	 */
9933 	lockdep_assert_held(&p->pi_lock);
9934 	if (wake_flags & WF_TTWU) {
9935 		record_wakee(p);
9936 
9937 		if ((wake_flags & WF_CURRENT_CPU) &&
9938 		    cpumask_test_cpu(cpu, p->cpus_ptr))
9939 			return cpu;
9940 
9941 		if (!is_rd_overutilized(this_rq()->rd)) {
9942 			new_cpu = find_energy_efficient_cpu(p, prev_cpu);
9943 			if (new_cpu >= 0)
9944 				return new_cpu;
9945 			new_cpu = prev_cpu;
9946 		}
9947 
9948 		want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
9949 	}
9950 
9951 	for_each_domain(cpu, tmp) {
9952 		/*
9953 		 * If both 'cpu' and 'prev_cpu' are part of this domain,
9954 		 * cpu is a valid SD_WAKE_AFFINE target.
9955 		 */
9956 		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
9957 		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
9958 			if (cpu != prev_cpu)
9959 				new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
9960 
9961 			sd = NULL; /* Prefer wake_affine over balance flags */
9962 			break;
9963 		}
9964 
9965 		/*
9966 		 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
9967 		 * usually do not have SD_BALANCE_WAKE set. That means wakeup
9968 		 * will usually go to the fast path.
9969 		 */
9970 		if (tmp->flags & sd_flag)
9971 			sd = tmp;
9972 		else if (!want_affine)
9973 			break;
9974 	}
9975 
9976 	/* Slow path */
9977 	if (unlikely(sd)) {
9978 		new_cpu = sched_balance_find_dst_cpu(sd, p, cpu, prev_cpu, sd_flag);
9979 		return select_idle_smt_cpu(p, new_cpu);
9980 	}
9981 
9982 	/* Fast path */
9983 	if (wake_flags & WF_TTWU)
9984 		return select_idle_sibling(p, prev_cpu, new_cpu);
9985 
9986 	return new_cpu;
9987 }
9988 
9989 /*
9990  * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
9991  * cfs_rq_of(p) references at time of call are still valid and identify the
9992  * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
9993  */
9994 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
9995 {
9996 	struct sched_entity *se = &p->se;
9997 
9998 	if (!task_on_rq_migrating(p)) {
9999 		remove_entity_load_avg(se);
10000 
10001 		/*
10002 		 * Here, the task's PELT values have been updated according to
10003 		 * the current rq's clock. But if that clock hasn't been
10004 		 * updated in a while, a substantial idle time will be missed,
10005 		 * leading to an inflation after wake-up on the new rq.
10006 		 *
10007 		 * Estimate the missing time from the cfs_rq last_update_time
10008 		 * and update sched_avg to improve the PELT continuity after
10009 		 * migration.
10010 		 */
10011 		migrate_se_pelt_lag(se);
10012 	}
10013 
10014 	/* Tell new CPU we are migrated */
10015 	se->avg.last_update_time = 0;
10016 
10017 	update_scan_period(p, new_cpu);
10018 }
10019 
10020 static void task_dead_fair(struct task_struct *p)
10021 {
10022 	struct sched_entity *se = &p->se;
10023 	remove_entity_load_avg(se);
10024 }
10025 
10026 /*
10027  * Set the max capacity the task is allowed to run at for misfit detection.
10028  */
10029 static void set_task_max_allowed_capacity(struct task_struct *p)
10030 {
10031 	struct asym_cap_data *entry;
10032 
10033 	if (!sched_asym_cpucap_active())
10034 		return;
10035 
10036 	rcu_read_lock();
10037 	list_for_each_entry_rcu(entry, &asym_cap_list, link) {
10038 		cpumask_t *cpumask;
10039 
10040 		cpumask = cpu_capacity_span(entry);
10041 		if (!cpumask_intersects(p->cpus_ptr, cpumask))
10042 			continue;
10043 
10044 		p->max_allowed_capacity = entry->capacity;
10045 		break;
10046 	}
10047 	rcu_read_unlock();
10048 }
10049 
10050 static void set_cpus_allowed_fair(struct task_struct *p, struct affinity_context *ctx)
10051 {
10052 	set_cpus_allowed_common(p, ctx);
10053 	set_task_max_allowed_capacity(p);
10054 }
10055 
10056 enum preempt_wakeup_action {
10057 	PREEMPT_WAKEUP_NONE,	/* No preemption. */
10058 	PREEMPT_WAKEUP_SHORT,	/* Ignore slice protection. */
10059 	PREEMPT_WAKEUP_PICK,	/* Let pick_eevdf() decide. */
10060 	PREEMPT_WAKEUP_RESCHED,	/* Force reschedule. */
10061 };
10062 
10063 static inline bool set_preempt_buddy(struct cfs_rq *cfs_rq, struct sched_entity *pse)
10064 {
10065 	/*
10066 	 * Keep existing buddy if the deadline is sooner than pse.
10067 	 * The older buddy may be cache cold and completely unrelated
10068 	 * to the current wakeup but that is unpredictable where as
10069 	 * obeying the deadline is more in line with EEVDF objectives.
10070 	 */
10071 	if (cfs_rq->next && entity_before(cfs_rq->next, pse))
10072 		return false;
10073 
10074 	set_next_buddy(cfs_rq, pse);
10075 	return true;
10076 }
10077 
10078 static inline bool set_short_buddy(struct cfs_rq *cfs_rq, struct sched_entity *pse)
10079 {
10080 	if (cfs_rq->next) {
10081 		if (cfs_rq->next->slice < pse->slice)
10082 			return false;
10083 
10084 		if (cfs_rq->next->slice == pse->slice &&
10085 		    entity_before(cfs_rq->next, pse))
10086 			return false;
10087 	}
10088 
10089 	set_next_buddy(cfs_rq, pse);
10090 	return true;
10091 }
10092 
10093 /*
10094  * WF_SYNC|WF_TTWU indicates the waker expects to sleep but it is not
10095  * strictly enforced because the hint is either misunderstood or
10096  * multiple tasks must be woken up.
10097  */
10098 static inline enum preempt_wakeup_action
10099 preempt_sync(struct rq *rq, int wake_flags,
10100 	     struct sched_entity *pse, struct sched_entity *se)
10101 {
10102 	u64 threshold, delta;
10103 
10104 	/*
10105 	 * WF_SYNC without WF_TTWU is not expected so warn if it happens even
10106 	 * though it is likely harmless.
10107 	 */
10108 	WARN_ON_ONCE(!(wake_flags & WF_TTWU));
10109 
10110 	threshold = sysctl_sched_migration_cost;
10111 	delta = rq_clock_task(rq) - se->exec_start;
10112 	if ((s64)delta < 0)
10113 		delta = 0;
10114 
10115 	/*
10116 	 * WF_RQ_SELECTED implies the tasks are stacking on a CPU when they
10117 	 * could run on other CPUs. Reduce the threshold before preemption is
10118 	 * allowed to an arbitrary lower value as it is more likely (but not
10119 	 * guaranteed) the waker requires the wakee to finish.
10120 	 */
10121 	if (wake_flags & WF_RQ_SELECTED)
10122 		threshold >>= 2;
10123 
10124 	/*
10125 	 * As WF_SYNC is not strictly obeyed, allow some runtime for batch
10126 	 * wakeups to be issued.
10127 	 */
10128 	if (entity_before(pse, se) && delta >= threshold)
10129 		return PREEMPT_WAKEUP_RESCHED;
10130 
10131 	return PREEMPT_WAKEUP_NONE;
10132 }
10133 
10134 /*
10135  * Preempt the current task with a newly woken task if needed:
10136  */
10137 static void wakeup_preempt_fair(struct rq *rq, struct task_struct *p, int wake_flags)
10138 {
10139 	enum preempt_wakeup_action preempt_action = PREEMPT_WAKEUP_PICK;
10140 	struct task_struct *donor = rq->donor;
10141 	struct sched_entity *nse, *se = &donor->se, *pse = &p->se;
10142 	struct cfs_rq *cfs_rq = &rq->cfs;
10143 	int cse_is_idle, pse_is_idle;
10144 
10145 	/*
10146 	 * XXX Getting preempted by higher class, try and find idle CPU?
10147 	 */
10148 	if (p->sched_class != &fair_sched_class ||
10149 	    donor->sched_class != &fair_sched_class)
10150 		return;
10151 
10152 	if (unlikely(se == pse))
10153 		return;
10154 
10155 	/*
10156 	 * This is possible from callers such as attach_tasks(), in which we
10157 	 * unconditionally wakeup_preempt() after an enqueue (which may have
10158 	 * lead to a throttle).  This both saves work and prevents false
10159 	 * next-buddy nomination below.
10160 	 */
10161 	if (task_is_throttled(p))
10162 		return;
10163 
10164 	/*
10165 	 * We can come here with TIF_NEED_RESCHED already set from new task
10166 	 * wake up path.
10167 	 *
10168 	 * Note: this also catches the edge-case of curr being in a throttled
10169 	 * group (e.g. via set_curr_task), since update_curr() (in the
10170 	 * enqueue of curr) will have resulted in resched being set.  This
10171 	 * prevents us from potentially nominating it as a false LAST_BUDDY
10172 	 * below.
10173 	 */
10174 	if (!sched_feat(PREEMPT_SHORT) && test_tsk_need_resched(rq->curr))
10175 		return;
10176 
10177 	if (!sched_feat(WAKEUP_PREEMPTION))
10178 		return;
10179 
10180 	WARN_ON_ONCE(!pse);
10181 
10182 	cse_is_idle = se_is_idle(se);
10183 	pse_is_idle = se_is_idle(pse);
10184 
10185 	nse = se;
10186 	/*
10187 	 * Preempt an idle entity in favor of a non-idle entity (and don't preempt
10188 	 * in the inverse case).
10189 	 */
10190 	if (cse_is_idle && !pse_is_idle)
10191 		goto preempt;
10192 
10193 	update_curr_fair(rq);
10194 
10195 	if (cse_is_idle != pse_is_idle)
10196 		goto update;
10197 
10198 	/*
10199 	 * BATCH and IDLE tasks do not preempt others.
10200 	 */
10201 	if (unlikely(!normal_policy(p->policy)))
10202 		goto update;
10203 
10204 	/*
10205 	 * Do not preempt for tasks that are sched_delayed as it would violate
10206 	 * EEVDF to forcibly queue an ineligible task.
10207 	 */
10208 	if (pse->sched_delayed)
10209 		goto update;
10210 
10211 	/*
10212 	 * If @p has a shorter slice than current and @p is eligible, override
10213 	 * current's slice protection in order to allow preemption.
10214 	 */
10215 	if (sched_feat(PREEMPT_SHORT) && (pse->slice < se->slice)) {
10216 		preempt_action = PREEMPT_WAKEUP_SHORT;
10217 		goto pick;
10218 	}
10219 
10220 	/*
10221 	 * Ignore wakee preemption on WF_FORK as it is less likely that
10222 	 * there is shared data as exec often follow fork.
10223 	 */
10224 	if (wake_flags & WF_FORK)
10225 		goto update;
10226 
10227 	/* Prefer picking wakee soon if appropriate. */
10228 	if (sched_feat(NEXT_BUDDY) && set_preempt_buddy(cfs_rq, pse)) {
10229 		/*
10230 		 * Decide whether to obey WF_SYNC hint for a new buddy. Old
10231 		 * buddies are ignored as they may not be relevant to the
10232 		 * waker and less likely to be cache hot.
10233 		 */
10234 		if (wake_flags & WF_SYNC)
10235 			preempt_action = preempt_sync(rq, wake_flags, pse, se);
10236 	}
10237 
10238 	switch (preempt_action) {
10239 	case PREEMPT_WAKEUP_NONE:
10240 		return;
10241 	case PREEMPT_WAKEUP_RESCHED:
10242 		goto preempt;
10243 	case PREEMPT_WAKEUP_SHORT:
10244 		fallthrough;
10245 	case PREEMPT_WAKEUP_PICK:
10246 		break;
10247 	}
10248 
10249 pick:
10250 	if (cfs_rq->h_nr_queued) {
10251 		nse = pick_next_entity(rq, preempt_action != PREEMPT_WAKEUP_SHORT);
10252 		if (unlikely(!nse))
10253 			goto pick;
10254 
10255 		/* If @p has become the most eligible task, force preemption */
10256 		if (nse == pse)
10257 			goto preempt;
10258 	}
10259 
10260 	/*
10261 	 * If @p is eligible but not the next task to run then cancel protection
10262 	 * to prevent large scheduling latency
10263 	 */
10264 	if (preempt_action == PREEMPT_WAKEUP_SHORT && entity_eligible(cfs_rq, pse))
10265 		goto preempt;
10266 update:
10267 	if (sched_feat(RUN_TO_PARITY))
10268 		update_protect_slice(cfs_rq, se);
10269 
10270 	return;
10271 
10272 preempt:
10273 	cancel_protect_slice(se);
10274 
10275 	if (preempt_action == PREEMPT_WAKEUP_SHORT)
10276 		set_short_buddy(cfs_rq, pse);
10277 
10278 	resched_curr_lazy(rq);
10279 }
10280 
10281 struct task_struct *pick_task_fair(struct rq *rq, struct rq_flags *rf)
10282 	__must_hold(__rq_lockp(rq))
10283 {
10284 	struct cfs_rq *cfs_rq = &rq->cfs;
10285 	struct sched_entity *se;
10286 	struct task_struct *p;
10287 	int new_tasks;
10288 
10289 again:
10290 	if (!cfs_rq->h_nr_queued)
10291 		goto idle;
10292 
10293 	/* Might not have done put_prev_entity() */
10294 	if (cfs_rq->curr && cfs_rq->curr->on_rq)
10295 		update_curr_eevdf(cfs_rq);
10296 
10297 	se = pick_next_entity(rq, true);
10298 	if (!se)
10299 		goto again;
10300 
10301 	p = task_of(se);
10302 	return p;
10303 
10304 idle:
10305 	if (sched_core_enabled(rq))
10306 		return NULL;
10307 
10308 	new_tasks = sched_balance_newidle(rq, rf);
10309 	if (new_tasks < 0)
10310 		return RETRY_TASK;
10311 	if (new_tasks > 0)
10312 		goto again;
10313 	return NULL;
10314 }
10315 
10316 static struct task_struct *
10317 fair_server_pick_task(struct sched_dl_entity *dl_se, struct rq_flags *rf)
10318 	__must_hold(__rq_lockp(dl_se->rq))
10319 {
10320 	return pick_task_fair(dl_se->rq, rf);
10321 }
10322 
10323 void fair_server_init(struct rq *rq)
10324 {
10325 	struct sched_dl_entity *dl_se = &rq->fair_server;
10326 
10327 	init_dl_entity(dl_se);
10328 
10329 	dl_server_init(dl_se, rq, fair_server_pick_task);
10330 }
10331 
10332 /*
10333  * Account for a descheduled task:
10334  */
10335 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev, struct task_struct *next)
10336 {
10337 	struct sched_entity *se = &prev->se;
10338 	struct cfs_rq *cfs_rq = &rq->cfs;
10339 	struct sched_entity *nse = NULL;
10340 
10341 #ifdef CONFIG_FAIR_GROUP_SCHED
10342 	if (next && next->sched_class == &fair_sched_class)
10343 		nse = &next->se;
10344 #endif
10345 
10346 	while (se) {
10347 		cfs_rq = cfs_rq_of(se);
10348 		if (!nse || cfs_rq->h_curr)
10349 			put_prev_entity(cfs_rq, se);
10350 #ifdef CONFIG_FAIR_GROUP_SCHED
10351 		if (nse) {
10352 			if (is_same_group(se, nse))
10353 				break;
10354 
10355 			int d = nse->depth - se->depth;
10356 			if (d >= 0) {
10357 				/* nse has equal or greater depth, ascend */
10358 				nse = parent_entity(nse);
10359 				/* if nse is the deeper, do not ascend se */
10360 				if (d > 0)
10361 					continue;
10362 			}
10363 		}
10364 #endif
10365 		se = parent_entity(se);
10366 	}
10367 
10368 	/* Put 'current' back into the tree. */
10369 	cfs_rq = &rq->cfs;
10370 	se = &prev->se;
10371 	WARN_ON_ONCE(cfs_rq->curr != se);
10372 	cfs_rq->curr = NULL;
10373 	if (se->on_rq)
10374 		__enqueue_entity(cfs_rq, se);
10375 }
10376 
10377 /*
10378  * sched_yield() is very simple
10379  */
10380 static void yield_task_fair(struct rq *rq)
10381 {
10382 	struct task_struct *curr = rq->donor;
10383 	struct sched_entity *se = &curr->se;
10384 	struct cfs_rq *cfs_rq = &rq->cfs;
10385 
10386 	/*
10387 	 * Are we the only task in the tree?
10388 	 */
10389 	if (unlikely(rq->nr_running == 1))
10390 		return;
10391 
10392 	clear_buddies(cfs_rq, se);
10393 
10394 	update_rq_clock(rq);
10395 	/*
10396 	 * Update run-time statistics of the 'current'.
10397 	 */
10398 	update_curr_eevdf(cfs_rq);
10399 	/*
10400 	 * Tell update_rq_clock() that we've just updated,
10401 	 * so we don't do microscopic update in schedule()
10402 	 * and double the fastpath cost.
10403 	 */
10404 	rq_clock_skip_update(rq);
10405 
10406 	/*
10407 	 * Forfeit the remaining vruntime, only if the entity is eligible. This
10408 	 * condition is necessary because in core scheduling we prefer to run
10409 	 * ineligible tasks rather than force idling. If this happens we may
10410 	 * end up in a loop where the core scheduler picks the yielding task,
10411 	 * which yields immediately again; without the condition the vruntime
10412 	 * ends up quickly running away.
10413 	 */
10414 	if (entity_eligible(cfs_rq, se)) {
10415 		se->vruntime = se->deadline;
10416 		update_deadline(cfs_rq, se);
10417 	}
10418 }
10419 
10420 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
10421 {
10422 	struct sched_entity *se = &p->se;
10423 
10424 	/* !se->on_rq also covers throttled task */
10425 	if (!se->on_rq || se->sched_delayed)
10426 		return false;
10427 
10428 	/* Tell the scheduler that we'd really like se to run next. */
10429 	set_next_buddy(&task_rq(p)->cfs, se);
10430 
10431 	yield_task_fair(rq);
10432 
10433 	return true;
10434 }
10435 
10436 /**************************************************
10437  * Fair scheduling class load-balancing methods.
10438  *
10439  * BASICS
10440  *
10441  * The purpose of load-balancing is to achieve the same basic fairness the
10442  * per-CPU scheduler provides, namely provide a proportional amount of compute
10443  * time to each task. This is expressed in the following equation:
10444  *
10445  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
10446  *
10447  * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
10448  * W_i,0 is defined as:
10449  *
10450  *   W_i,0 = \Sum_j w_i,j                                             (2)
10451  *
10452  * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
10453  * is derived from the nice value as per sched_prio_to_weight[].
10454  *
10455  * The weight average is an exponential decay average of the instantaneous
10456  * weight:
10457  *
10458  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
10459  *
10460  * C_i is the compute capacity of CPU i, typically it is the
10461  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
10462  * can also include other factors [XXX].
10463  *
10464  * To achieve this balance we define a measure of imbalance which follows
10465  * directly from (1):
10466  *
10467  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
10468  *
10469  * We them move tasks around to minimize the imbalance. In the continuous
10470  * function space it is obvious this converges, in the discrete case we get
10471  * a few fun cases generally called infeasible weight scenarios.
10472  *
10473  * [XXX expand on:
10474  *     - infeasible weights;
10475  *     - local vs global optima in the discrete case. ]
10476  *
10477  *
10478  * SCHED DOMAINS
10479  *
10480  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
10481  * for all i,j solution, we create a tree of CPUs that follows the hardware
10482  * topology where each level pairs two lower groups (or better). This results
10483  * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
10484  * tree to only the first of the previous level and we decrease the frequency
10485  * of load-balance at each level inversely proportional to the number of CPUs in
10486  * the groups.
10487  *
10488  * This yields:
10489  *
10490  *     log_2 n     1     n
10491  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
10492  *     i = 0      2^i   2^i
10493  *                               `- size of each group
10494  *         |         |     `- number of CPUs doing load-balance
10495  *         |         `- freq
10496  *         `- sum over all levels
10497  *
10498  * Coupled with a limit on how many tasks we can migrate every balance pass,
10499  * this makes (5) the runtime complexity of the balancer.
10500  *
10501  * An important property here is that each CPU is still (indirectly) connected
10502  * to every other CPU in at most O(log n) steps:
10503  *
10504  * The adjacency matrix of the resulting graph is given by:
10505  *
10506  *             log_2 n
10507  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
10508  *             k = 0
10509  *
10510  * And you'll find that:
10511  *
10512  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
10513  *
10514  * Showing there's indeed a path between every CPU in at most O(log n) steps.
10515  * The task movement gives a factor of O(m), giving a convergence complexity
10516  * of:
10517  *
10518  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
10519  *
10520  *
10521  * WORK CONSERVING
10522  *
10523  * In order to avoid CPUs going idle while there's still work to do, new idle
10524  * balancing is more aggressive and has the newly idle CPU iterate up the domain
10525  * tree itself instead of relying on other CPUs to bring it work.
10526  *
10527  * This adds some complexity to both (5) and (8) but it reduces the total idle
10528  * time.
10529  *
10530  * [XXX more?]
10531  *
10532  *
10533  * CGROUPS
10534  *
10535  * Cgroups make a horror show out of (2), instead of a simple sum we get:
10536  *
10537  *                                s_k,i
10538  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
10539  *                                 S_k
10540  *
10541  * Where
10542  *
10543  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
10544  *
10545  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
10546  *
10547  * The big problem is S_k, its a global sum needed to compute a local (W_i)
10548  * property.
10549  *
10550  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
10551  *      rewrite all of this once again.]
10552  */
10553 
10554 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
10555 
10556 enum fbq_type { regular, remote, all };
10557 
10558 /*
10559  * 'group_type' describes the group of CPUs at the moment of load balancing.
10560  *
10561  * The enum is ordered by pulling priority, with the group with lowest priority
10562  * first so the group_type can simply be compared when selecting the busiest
10563  * group. See update_sd_pick_busiest().
10564  */
10565 enum group_type {
10566 	/* The group has spare capacity that can be used to run more tasks.  */
10567 	group_has_spare = 0,
10568 	/*
10569 	 * The group is fully used and the tasks don't compete for more CPU
10570 	 * cycles. Nevertheless, some tasks might wait before running.
10571 	 */
10572 	group_fully_busy,
10573 	/*
10574 	 * One task doesn't fit with CPU's capacity and must be migrated to a
10575 	 * more powerful CPU.
10576 	 */
10577 	group_misfit_task,
10578 	/*
10579 	 * Balance SMT group that's fully busy. Can benefit from migration
10580 	 * a task on SMT with busy sibling to another CPU on idle core.
10581 	 */
10582 	group_smt_balance,
10583 	/*
10584 	 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
10585 	 * and the task should be migrated to it instead of running on the
10586 	 * current CPU.
10587 	 */
10588 	group_asym_packing,
10589 	/*
10590 	 * The tasks' affinity constraints previously prevented the scheduler
10591 	 * from balancing the load across the system.
10592 	 */
10593 	group_imbalanced,
10594 	/*
10595 	 * There are tasks running on non-preferred LLC, possible to move
10596 	 * them to their preferred LLC without creating too much imbalance.
10597 	 * The priority of group_llc_balance is lower than that of
10598 	 * group_overloaded and higher than that of all other group types.
10599 	 * This is because group_llc_balance may exacerbate load imbalance.
10600 	 * If the LLC balancing attempt fails, the nr_balance_failed
10601 	 * mechanism will trigger other group types to rebalance the load.
10602 	 */
10603 	group_llc_balance,
10604 	/*
10605 	 * The CPU is overloaded and can't provide expected CPU cycles to all
10606 	 * tasks.
10607 	 */
10608 	group_overloaded
10609 };
10610 
10611 enum migration_type {
10612 	migrate_load = 0,
10613 	migrate_util,
10614 	migrate_task,
10615 	migrate_misfit,
10616 	migrate_llc_task
10617 };
10618 
10619 #define LBF_ALL_PINNED	0x01
10620 #define LBF_NEED_BREAK	0x02
10621 #define LBF_DST_PINNED  0x04
10622 #define LBF_SOME_PINNED	0x08
10623 #define LBF_ACTIVE_LB	0x10
10624 #define LBF_LLC_PINNED	0x20
10625 #define LBF_ACTIVE_LB_LLC	0x40
10626 
10627 struct lb_env {
10628 	struct sched_domain	*sd;
10629 
10630 	struct rq		*src_rq;
10631 	int			src_cpu;
10632 
10633 	int			dst_cpu;
10634 	struct rq		*dst_rq;
10635 	bool			dst_core_idle;
10636 
10637 	struct cpumask		*dst_grpmask;
10638 	int			new_dst_cpu;
10639 	enum cpu_idle_type	idle;
10640 	long			imbalance;
10641 	/* The set of CPUs under consideration for load-balancing */
10642 	struct cpumask		*cpus;
10643 
10644 	unsigned int		flags;
10645 
10646 	unsigned int		loop;
10647 	unsigned int		loop_break;
10648 	unsigned int		loop_max;
10649 
10650 	enum fbq_type		fbq_type;
10651 	enum migration_type	migration_type;
10652 	struct list_head	tasks;
10653 };
10654 
10655 /*
10656  * Is this task likely cache-hot:
10657  */
10658 static int task_hot(struct task_struct *p, struct lb_env *env)
10659 {
10660 	s64 delta;
10661 
10662 	lockdep_assert_rq_held(env->src_rq);
10663 
10664 	if (p->sched_class != &fair_sched_class)
10665 		return 0;
10666 
10667 	if (unlikely(task_has_idle_policy(p)))
10668 		return 0;
10669 
10670 	/* SMT siblings share cache */
10671 	if (env->sd->flags & SD_SHARE_CPUCAPACITY)
10672 		return 0;
10673 
10674 	/*
10675 	 * Buddy candidates are cache hot:
10676 	 */
10677 	if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
10678 	    (&p->se == cfs_rq_of(&p->se)->next))
10679 		return 1;
10680 
10681 	if (sysctl_sched_migration_cost == -1)
10682 		return 1;
10683 
10684 	/*
10685 	 * Don't migrate task if the task's cookie does not match
10686 	 * with the destination CPU's core cookie.
10687 	 */
10688 	if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
10689 		return 1;
10690 
10691 	if (sysctl_sched_migration_cost == 0)
10692 		return 0;
10693 
10694 	delta = rq_clock_task(env->src_rq) - p->se.exec_start;
10695 
10696 	return delta < (s64)sysctl_sched_migration_cost;
10697 }
10698 
10699 #ifdef CONFIG_NUMA_BALANCING
10700 /*
10701  * Returns a positive value, if task migration degrades locality.
10702  * Returns 0, if task migration is not affected by locality.
10703  * Returns a negative value, if task migration improves locality i.e migration preferred.
10704  */
10705 static long migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
10706 {
10707 	struct numa_group *numa_group = rcu_dereference_all(p->numa_group);
10708 	unsigned long src_weight, dst_weight;
10709 	int src_nid, dst_nid, dist;
10710 
10711 	if (!static_branch_likely(&sched_numa_balancing))
10712 		return 0;
10713 
10714 	if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
10715 		return 0;
10716 
10717 	src_nid = cpu_to_node(env->src_cpu);
10718 	dst_nid = cpu_to_node(env->dst_cpu);
10719 
10720 	if (src_nid == dst_nid)
10721 		return 0;
10722 
10723 	/* Migrating away from the preferred node is always bad. */
10724 	if (src_nid == p->numa_preferred_nid) {
10725 		if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
10726 			return 1;
10727 		else
10728 			return 0;
10729 	}
10730 
10731 	/* Encourage migration to the preferred node. */
10732 	if (dst_nid == p->numa_preferred_nid)
10733 		return -1;
10734 
10735 	/* Leaving a core idle is often worse than degrading locality. */
10736 	if (env->idle == CPU_IDLE)
10737 		return 0;
10738 
10739 	dist = node_distance(src_nid, dst_nid);
10740 	if (numa_group) {
10741 		src_weight = group_weight(p, src_nid, dist);
10742 		dst_weight = group_weight(p, dst_nid, dist);
10743 	} else {
10744 		src_weight = task_weight(p, src_nid, dist);
10745 		dst_weight = task_weight(p, dst_nid, dist);
10746 	}
10747 
10748 	return src_weight - dst_weight;
10749 }
10750 
10751 #else /* !CONFIG_NUMA_BALANCING: */
10752 static inline long migrate_degrades_locality(struct task_struct *p,
10753 					     struct lb_env *env)
10754 {
10755 	return 0;
10756 }
10757 #endif /* !CONFIG_NUMA_BALANCING */
10758 
10759 /*
10760  * Check whether the task is ineligible on the destination cpu
10761  *
10762  * When the PLACE_LAG scheduling feature is enabled and
10763  * dst_cfs_rq->nr_queued is greater than 1, if the task
10764  * is ineligible, it will also be ineligible when
10765  * it is migrated to the destination cpu.
10766  */
10767 static inline int task_is_ineligible_on_dst_cpu(struct task_struct *p, int dest_cpu)
10768 {
10769 	struct cfs_rq *dst_cfs_rq = &cpu_rq(dest_cpu)->cfs;
10770 
10771 	if (sched_feat(PLACE_LAG) && dst_cfs_rq->h_nr_queued &&
10772 	    !entity_eligible(&task_rq(p)->cfs, &p->se))
10773 		return 1;
10774 
10775 	return 0;
10776 }
10777 
10778 #ifdef CONFIG_SCHED_CACHE
10779 /*
10780  * The margin used when comparing LLC utilization with CPU capacity.
10781  * It determines the LLC load level where active LLC aggregation is
10782  * done.
10783  * Derived from fits_capacity().
10784  *
10785  * (default: ~50%, tunable via debugfs)
10786  */
10787 static bool fits_llc_capacity(unsigned long util, unsigned long max)
10788 {
10789 	u32 aggr_pct = llc_overaggr_pct;
10790 
10791 	/*
10792 	 * For single core systems, raise the aggregation
10793 	 * threshold to accommodate more tasks.
10794 	 */
10795 	if (cpu_smt_num_threads == 1)
10796 		aggr_pct = (aggr_pct * 3 / 2);
10797 
10798 	return util * 100 < max * aggr_pct;
10799 }
10800 
10801 /*
10802  * The margin used when comparing utilization.
10803  * is 'util1' noticeably greater than 'util2'
10804  * Derived from capacity_greater().
10805  * Bias is in perentage.
10806  */
10807 /* Allows dst util to be bigger than src util by up to bias percent */
10808 #define util_greater(util1, util2) \
10809 	((util1) * 100 > (util2) * (100 + llc_imb_pct))
10810 
10811 static __maybe_unused bool get_llc_stats(int cpu, unsigned long *util,
10812 					 unsigned long *cap)
10813 {
10814 	struct sched_domain_shared *sd_share;
10815 
10816 	sd_share = rcu_dereference_all(per_cpu(sd_llc_shared, cpu));
10817 	if (!sd_share)
10818 		return false;
10819 
10820 	*util = READ_ONCE(sd_share->util_avg);
10821 	*cap = READ_ONCE(sd_share->capacity);
10822 
10823 	return true;
10824 }
10825 
10826 /*
10827  * Decision matrix according to the LLC utilization. To
10828  * decide whether we can do task aggregation across LLC.
10829  *
10830  * By default, 50% is the threshold for treating the LLC
10831  * as busy. The reason for choosing 50% is to avoid saturation
10832  * of SMT-2, and it is also a safe cutoff for other SMT-n
10833  * platforms. SMT-1 has higher threshold because it is
10834  * supposed to accommodate more tasks, see fits_llc_capacity().
10835  *
10836  * 20% is the utilization imbalance percentage to decide
10837  * if the preferred LLC is busier than the non-preferred LLC.
10838  * 20 is a little higher than the LLC domain's imbalance_pct
10839  * 17. The hysteresis is used to avoid task bouncing between the
10840  * preferred LLC and the non-preferred LLC, and it will
10841  * be turned into tunable debugfs.
10842  *
10843  * 1. moving towards the preferred LLC, dst is the preferred
10844  *    LLC, src is not.
10845  *
10846  * src \ dst      30%  40%  50%  60%
10847  * 30%            Y    Y    Y    N
10848  * 40%            Y    Y    Y    Y
10849  * 50%            Y    Y    G    G
10850  * 60%            Y    Y    G    G
10851  *
10852  * 2. moving out of the preferred LLC, src is the preferred
10853  *    LLC, dst is not:
10854  *
10855  * src \ dst      30%  40%  50%  60%
10856  * 30%            N    N    N    N
10857  * 40%            N    N    N    N
10858  * 50%            N    N    G    G
10859  * 60%            Y    N    G    G
10860  *
10861  * src :      src_util
10862  * dst :      dst_util
10863  * Y :        Yes, migrate
10864  * N :        No, do not migrate
10865  * G :        let the Generic load balance to even the load.
10866  *
10867  * The intention is that if both LLCs are quite busy, cache aware
10868  * load balance should not be performed, and generic load balance
10869  * should take effect. However, if one is busy and the other is not,
10870  * the preferred LLC capacity(50%) and imbalance criteria(20%) should
10871  * be considered to determine whether LLC aggregation should be
10872  * performed to bias the load towards the preferred LLC.
10873  */
10874 
10875 /* migration decision, 3 states are orthogonal. */
10876 enum llc_mig {
10877 	mig_forbid = 0,		/* N: Don't migrate task, respect LLC preference */
10878 	mig_llc,		/* Y: Do LLC preference based migration */
10879 	mig_unrestricted	/* G: Don't restrict generic load balance migration */
10880 };
10881 
10882 /*
10883  * Check if task can be moved from the source LLC to the
10884  * destination LLC without breaking cache aware preferrence.
10885  * src_cpu and dst_cpu are arbitrary CPUs within the source
10886  * and destination LLCs, respectively.
10887  */
10888 static enum llc_mig can_migrate_llc(int src_cpu, int dst_cpu,
10889 				    unsigned long tsk_util,
10890 				    bool to_pref)
10891 {
10892 	unsigned long src_util, dst_util, src_cap, dst_cap;
10893 
10894 	if (!get_llc_stats(src_cpu, &src_util, &src_cap) ||
10895 	    !get_llc_stats(dst_cpu, &dst_util, &dst_cap))
10896 		return mig_unrestricted;
10897 
10898 	src_util = src_util < tsk_util ? 0 : src_util - tsk_util;
10899 	dst_util = dst_util + tsk_util;
10900 
10901 	if (!fits_llc_capacity(dst_util, dst_cap) &&
10902 	    !fits_llc_capacity(src_util, src_cap))
10903 		return mig_unrestricted;
10904 
10905 	if (to_pref) {
10906 		/*
10907 		 * Don't migrate if we will get preferred LLC too
10908 		 * heavily loaded and if the dest is much busier
10909 		 * than the src, in which case migration will
10910 		 * increase the imbalance too much.
10911 		 */
10912 		if (!fits_llc_capacity(dst_util, dst_cap) &&
10913 		    util_greater(dst_util, src_util))
10914 			return mig_forbid;
10915 	} else {
10916 		/*
10917 		 * Don't migrate if we will leave preferred LLC
10918 		 * too idle, or if this migration leads to the
10919 		 * non-preferred LLC falls within sysctl_aggr_imb percent
10920 		 * of preferred LLC, leading to migration again
10921 		 * back to preferred LLC.
10922 		 */
10923 		if (fits_llc_capacity(src_util, src_cap) ||
10924 		    !util_greater(src_util, dst_util))
10925 			return mig_forbid;
10926 	}
10927 	return mig_llc;
10928 }
10929 
10930 static inline bool task_misfits_asym_cpu(struct lb_env *env, struct task_struct *p)
10931 {
10932 	/*
10933 	 * On asymmetric CPU capacity domains, do not let cache-aware
10934 	 * balancing pull the task onto a destination CPU that cannot
10935 	 * accommodate it. Doing so would turn the task into a misfit on
10936 	 * the destination, trading a cache-locality gain for a capacity
10937 	 * loss. If the task already does not fit its source CPU, the move
10938 	 * cannot make things worse, so let the LLC preference decide.
10939 	 */
10940 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && p &&
10941 	    !task_fits_cpu(p, env->dst_cpu) &&
10942 	    task_fits_cpu(p, env->src_cpu))
10943 		return true;
10944 
10945 	return false;
10946 }
10947 
10948 static inline bool sched_asym(struct sched_domain *sd, int dst_cpu, int src_cpu);
10949 
10950 /*
10951  * Check if task p can migrate from source LLC to
10952  * destination LLC in terms of cache aware load balance.
10953  */
10954 static enum llc_mig can_migrate_llc_task(struct lb_env *env,
10955 					 struct task_struct *p)
10956 {
10957 	struct sched_cache_group *grp;
10958 	bool to_pref;
10959 	int cpu, src_cpu, dst_cpu;
10960 
10961 	if (task_misfits_asym_cpu(env, p))
10962 		return mig_forbid;
10963 
10964 	src_cpu = env->src_cpu;
10965 	dst_cpu = env->dst_cpu;
10966 	grp = rcu_dereference_all(p->sched_cache_grp);
10967 	if (!grp)
10968 		return mig_unrestricted;
10969 
10970 	cpu = READ_ONCE(grp->cpu);
10971 	if (cpu < 0 || cpus_share_cache(src_cpu, dst_cpu))
10972 		return mig_unrestricted;
10973 
10974 	/* Prioritize asym packing to idle core over cache awareness */
10975 	if (env->idle && sched_asym(env->sd, dst_cpu, src_cpu))
10976 		return mig_unrestricted;
10977 
10978 	/* skip cache aware load balance for too many threads */
10979 	if (invalid_llc_nr(grp, p, dst_cpu) ||
10980 	    exceed_llc_capacity(grp, dst_cpu)) {
10981 		if (READ_ONCE(grp->cpu) != -1)
10982 			WRITE_ONCE(grp->cpu, -1);
10983 		return mig_unrestricted;
10984 	}
10985 
10986 	if (cpus_share_cache(dst_cpu, cpu))
10987 		to_pref = true;
10988 	else if (cpus_share_cache(src_cpu, cpu))
10989 		to_pref = false;
10990 	else
10991 		return mig_unrestricted;
10992 
10993 	return can_migrate_llc(src_cpu, dst_cpu,
10994 			       task_util(p), to_pref);
10995 }
10996 
10997 /*
10998  * Check if active load balance breaks LLC locality in
10999  * terms of cache aware load balance. The load level and
11000  * imbalance do not warrant breaking LLC preference per
11001  * the can_migrate_llc() policy. Here, the benefit of
11002  * LLC locality outweighs the power efficiency gained from
11003  * migrating the only runnable task away.
11004  */
11005 static inline bool
11006 alb_break_llc(struct lb_env *env)
11007 {
11008 	if (!sched_cache_enabled())
11009 		return false;
11010 
11011 	if (cpus_share_cache(env->src_cpu, env->dst_cpu))
11012 		return false;
11013 	/*
11014 	 * All tasks prefer to stay on their current CPU.
11015 	 * Do not pull a task from its preferred CPU if:
11016 	 * 1. It is the only task running and does not exceed
11017 	 *    imbalance allowance; OR
11018 	 * 2. Migrating it away from its preferred LLC would violate
11019 	 *    the cache-aware scheduling policy.
11020 	 */
11021 	if (env->src_rq->nr_pref_llc_running &&
11022 	    env->src_rq->nr_pref_llc_running == env->src_rq->cfs.h_nr_runnable) {
11023 		unsigned long util = 0;
11024 		struct task_struct *cur;
11025 
11026 		/*
11027 		 * Migrating misfit tasks from current CPU
11028 		 * to CPU with a better fit.
11029 		 * Prioritize that over LLC preference.
11030 		 */
11031 		if (env->migration_type == migrate_misfit)
11032 			return false;
11033 
11034 		if (env->src_rq->nr_running <= 1)
11035 			return true;
11036 
11037 		cur = rcu_dereference_all(env->src_rq->curr);
11038 		/*
11039 		 * Active balance only detaches CFS tasks, so a non-fair
11040 		 * curr is never migrated. Ignore it and avoid using its
11041 		 * state util in the misfit and LLC checks.
11042 		 */
11043 		if (cur && cur->sched_class != &fair_sched_class)
11044 			cur = NULL;
11045 		if (cur)
11046 			util = task_util(cur);
11047 
11048 		if (task_misfits_asym_cpu(env, cur) ||
11049 		    can_migrate_llc(env->src_cpu, env->dst_cpu,
11050 				    util, false) == mig_forbid)
11051 			return true;
11052 	}
11053 
11054 	return false;
11055 }
11056 
11057 /*
11058  * Returns true if p's preferred LLC does not match the destination CPU
11059  * under migrate_llc_task semantics. Passive LB passes migrate_llc_task
11060  * in env->migration_type, while active LB carries LBF_ACTIVE_LB_LLC in
11061  * env->flags to avoid overwriting env->migration_type.
11062  */
11063 static inline bool
11064 migrate_llc_task_wrong_dst(struct task_struct *p, struct lb_env *env)
11065 {
11066 	return sched_cache_enabled() &&
11067 	       (env->migration_type == migrate_llc_task ||
11068 		env->flags & LBF_ACTIVE_LB_LLC) &&
11069 	       READ_ONCE(p->preferred_llc) != llc_id(env->dst_cpu);
11070 }
11071 
11072 /*
11073  * Check if migrating task p from env->src_cpu to
11074  * env->dst_cpu breaks LLC localiy.
11075  */
11076 static bool migrate_degrades_llc(struct task_struct *p, struct lb_env *env)
11077 {
11078 	if (!sched_cache_enabled())
11079 		return false;
11080 
11081 	if (task_has_sched_core(p))
11082 		return false;
11083 	/*
11084 	 * Skip over tasks that would degrade LLC locality;
11085 	 * only when nr_balanced_failed is sufficiently high do we
11086 	 * ignore this constraint.
11087 	 *
11088 	 * Threshold of cache_nice_tries is set to 1 higher
11089 	 * than nr_balance_failed to avoid excessive task
11090 	 * migration at the same time.
11091 	 */
11092 	if (env->sd->nr_balance_failed >= env->sd->cache_nice_tries + 1)
11093 		return false;
11094 
11095 	/*
11096 	 * We know the env->src_cpu has some tasks prefer to
11097 	 * run on env->dst_cpu, skip the tasks do not prefer
11098 	 * env->dst_cpu, and find the one that prefers.
11099 	 */
11100 	if (migrate_llc_task_wrong_dst(p, env))
11101 		return true;
11102 
11103 	if (can_migrate_llc_task(env, p) != mig_forbid)
11104 		return false;
11105 
11106 	return true;
11107 }
11108 
11109 #else
11110 static inline bool get_llc_stats(int cpu, unsigned long *util,
11111 				 unsigned long *cap)
11112 {
11113 	return false;
11114 }
11115 
11116 static inline bool
11117 alb_break_llc(struct lb_env *env)
11118 {
11119 	return false;
11120 }
11121 
11122 static inline bool
11123 migrate_llc_task_wrong_dst(struct task_struct *p, struct lb_env *env)
11124 {
11125 	return false;
11126 }
11127 
11128 static inline bool
11129 migrate_degrades_llc(struct task_struct *p, struct lb_env *env)
11130 {
11131 	return false;
11132 }
11133 #endif
11134 /*
11135  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
11136  */
11137 static
11138 int can_migrate_task(struct task_struct *p, struct lb_env *env)
11139 {
11140 	long degrades, hot;
11141 
11142 	lockdep_assert_rq_held(env->src_rq);
11143 	if (p->sched_task_hot)
11144 		p->sched_task_hot = 0;
11145 
11146 	/*
11147 	 * We do not migrate tasks that are:
11148 	 * 1) delayed dequeued unless we migrate load, or
11149 	 * 2) target cfs_rq is in throttled hierarchy, or
11150 	 * 3) cannot be migrated to this CPU due to cpus_ptr, or
11151 	 * 4) running (obviously), or
11152 	 * 5) are cache-hot on their current CPU, or
11153 	 * 6) are blocked on mutexes (if SCHED_PROXY_EXEC is enabled)
11154 	 */
11155 	if ((p->se.sched_delayed) && (env->migration_type != migrate_load))
11156 		return 0;
11157 
11158 	if (lb_throttled_hierarchy(p, env->dst_cpu))
11159 		return 0;
11160 
11161 	/*
11162 	 * We want to prioritize the migration of eligible tasks.
11163 	 * For ineligible tasks we soft-limit them and only allow
11164 	 * them to migrate when nr_balance_failed is non-zero to
11165 	 * avoid load-balancing trying very hard to balance the load.
11166 	 */
11167 	if (!env->sd->nr_balance_failed &&
11168 	    task_is_ineligible_on_dst_cpu(p, env->dst_cpu))
11169 		return 0;
11170 
11171 	/* Disregard percpu kthreads; they are where they need to be. */
11172 	if (kthread_is_per_cpu(p))
11173 		return 0;
11174 
11175 	if (task_is_blocked(p))
11176 		return 0;
11177 
11178 	if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
11179 		int cpu;
11180 
11181 		schedstat_inc(p->stats.nr_failed_migrations_affine);
11182 
11183 		env->flags |= LBF_SOME_PINNED;
11184 
11185 		/*
11186 		 * Remember if this task can be migrated to any other CPU in
11187 		 * our sched_group. We may want to revisit it if we couldn't
11188 		 * meet load balance goals by pulling other tasks on src_cpu.
11189 		 *
11190 		 * Avoid computing new_dst_cpu
11191 		 * - for NEWLY_IDLE
11192 		 * - if we have already computed one in current iteration
11193 		 * - if it's an active balance
11194 		 */
11195 		if (env->idle == CPU_NEWLY_IDLE ||
11196 		    env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
11197 			return 0;
11198 
11199 		/* Prevent to re-select dst_cpu via env's CPUs: */
11200 		cpu = cpumask_first_and_and(env->dst_grpmask, env->cpus, p->cpus_ptr);
11201 
11202 		if (cpu < nr_cpu_ids) {
11203 			env->flags |= LBF_DST_PINNED;
11204 			env->new_dst_cpu = cpu;
11205 		}
11206 
11207 		return 0;
11208 	}
11209 
11210 	/* Record that we found at least one task that could run on dst_cpu */
11211 	env->flags &= ~LBF_ALL_PINNED;
11212 
11213 	if (task_on_cpu(env->src_rq, p) ||
11214 	    task_current_donor(env->src_rq, p)) {
11215 		schedstat_inc(p->stats.nr_failed_migrations_running);
11216 		return 0;
11217 	}
11218 
11219 	/*
11220 	 * Aggressive migration if:
11221 	 * 1) active balance
11222 	 * 2) destination numa is preferred
11223 	 * 3) task is cache cold, or
11224 	 * 4) too many balance attempts have failed.
11225 	 */
11226 	if (env->flags & LBF_ACTIVE_LB)
11227 		return !migrate_llc_task_wrong_dst(p, env);
11228 
11229 	degrades = migrate_degrades_locality(p, env);
11230 	if (!degrades) {
11231 		/*
11232 		 * If the NUMA locality is not broken,
11233 		 * further check if migration would hurt
11234 		 * LLC locality.
11235 		 */
11236 		if (migrate_degrades_llc(p, env)) {
11237 			/*
11238 			 * If regular load balancing fails to pull a task
11239 			 * due to LLC locality, this is expected behavior
11240 			 * and we set LBF_LLC_PINNED so we don't increase
11241 			 * nr_balance_failed unecessarily.
11242 			 */
11243 			if (env->migration_type != migrate_llc_task)
11244 				env->flags |= LBF_LLC_PINNED;
11245 
11246 			return 0;
11247 		}
11248 
11249 		hot = task_hot(p, env);
11250 	} else {
11251 		hot = degrades > 0;
11252 	}
11253 
11254 	if (!hot || env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
11255 		if (hot)
11256 			p->sched_task_hot = 1;
11257 		return 1;
11258 	}
11259 
11260 	schedstat_inc(p->stats.nr_failed_migrations_hot);
11261 	return 0;
11262 }
11263 
11264 /*
11265  * detach_task() -- detach the task for the migration specified in env
11266  */
11267 static void detach_task(struct task_struct *p, struct lb_env *env)
11268 {
11269 	lockdep_assert_rq_held(env->src_rq);
11270 
11271 	if (p->sched_task_hot) {
11272 		p->sched_task_hot = 0;
11273 		schedstat_inc(env->sd->lb_hot_gained[env->idle]);
11274 		schedstat_inc(p->stats.nr_forced_migrations);
11275 	}
11276 
11277 	WARN_ON(task_current(env->src_rq, p));
11278 	WARN_ON(task_current_donor(env->src_rq, p));
11279 
11280 	deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
11281 	set_task_cpu(p, env->dst_cpu);
11282 }
11283 
11284 /*
11285  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
11286  * part of active balancing operations within "domain".
11287  *
11288  * Returns a task if successful and NULL otherwise.
11289  */
11290 static struct task_struct *detach_one_task(struct lb_env *env)
11291 {
11292 	struct task_struct *p;
11293 
11294 	lockdep_assert_rq_held(env->src_rq);
11295 
11296 	list_for_each_entry_reverse(p,
11297 			&env->src_rq->cfs_tasks, se.group_node) {
11298 		if (!can_migrate_task(p, env))
11299 			continue;
11300 
11301 		detach_task(p, env);
11302 
11303 		/*
11304 		 * Right now, this is only the second place where
11305 		 * lb_gained[env->idle] is updated (other is detach_tasks)
11306 		 * so we can safely collect stats here rather than
11307 		 * inside detach_tasks().
11308 		 */
11309 		schedstat_inc(env->sd->lb_gained[env->idle]);
11310 		return p;
11311 	}
11312 	return NULL;
11313 }
11314 
11315 /*
11316  * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
11317  * busiest_rq, as part of a balancing operation within domain "sd".
11318  *
11319  * Returns number of detached tasks if successful and 0 otherwise.
11320  */
11321 static int detach_tasks(struct lb_env *env)
11322 {
11323 	struct list_head *tasks = &env->src_rq->cfs_tasks;
11324 	unsigned long util, load;
11325 	struct task_struct *p;
11326 	int detached = 0;
11327 
11328 	lockdep_assert_rq_held(env->src_rq);
11329 
11330 	/*
11331 	 * Source run queue has been emptied by another CPU, clear
11332 	 * LBF_ALL_PINNED flag as we will not test any task.
11333 	 */
11334 	if (env->src_rq->nr_running <= 1) {
11335 		env->flags &= ~LBF_ALL_PINNED;
11336 		return 0;
11337 	}
11338 
11339 	if (env->imbalance <= 0)
11340 		return 0;
11341 
11342 	while (!list_empty(tasks)) {
11343 		/*
11344 		 * We don't want to steal all, otherwise we may be treated likewise,
11345 		 * which could at worst lead to a livelock crash.
11346 		 */
11347 		if (env->idle && env->src_rq->nr_running <= 1)
11348 			break;
11349 
11350 		env->loop++;
11351 		/* We've more or less seen every task there is, call it quits */
11352 		if (env->loop > env->loop_max)
11353 			break;
11354 
11355 		/* take a breather every nr_migrate tasks */
11356 		if (env->loop > env->loop_break) {
11357 			env->loop_break += SCHED_NR_MIGRATE_BREAK;
11358 			env->flags |= LBF_NEED_BREAK;
11359 			break;
11360 		}
11361 
11362 		p = list_last_entry(tasks, struct task_struct, se.group_node);
11363 
11364 		if (!can_migrate_task(p, env))
11365 			goto next;
11366 
11367 		switch (env->migration_type) {
11368 		case migrate_load:
11369 			/*
11370 			 * Depending of the number of CPUs and tasks and the
11371 			 * cgroup hierarchy, task_h_load() can return a null
11372 			 * value. Make sure that env->imbalance decreases
11373 			 * otherwise detach_tasks() will stop only after
11374 			 * detaching up to loop_max tasks.
11375 			 */
11376 			load = max_t(unsigned long, task_h_load(p), 1);
11377 
11378 			if (sched_feat(LB_MIN) &&
11379 			    load < 16 && !env->sd->nr_balance_failed)
11380 				goto next;
11381 
11382 			/*
11383 			 * Make sure that we don't migrate too much load.
11384 			 * Nevertheless, let relax the constraint if
11385 			 * scheduler fails to find a good waiting task to
11386 			 * migrate.
11387 			 */
11388 			if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
11389 				goto next;
11390 
11391 			env->imbalance -= load;
11392 			break;
11393 
11394 		case migrate_util:
11395 			util = task_util_est(p);
11396 
11397 			if (shr_bound(util, env->sd->nr_balance_failed) > env->imbalance)
11398 				goto next;
11399 
11400 			env->imbalance -= util;
11401 			break;
11402 
11403 		case migrate_task:
11404 			env->imbalance--;
11405 			break;
11406 
11407 		case migrate_misfit:
11408 			/* This is not a misfit task */
11409 			if (task_fits_cpu(p, env->src_cpu))
11410 				goto next;
11411 
11412 			env->imbalance = 0;
11413 			break;
11414 
11415 		case migrate_llc_task:
11416 			env->imbalance--;
11417 			break;
11418 		}
11419 
11420 		detach_task(p, env);
11421 		list_add(&p->se.group_node, &env->tasks);
11422 
11423 		detached++;
11424 
11425 #ifdef CONFIG_PREEMPTION
11426 		/*
11427 		 * NEWIDLE balancing is a source of latency, so preemptible
11428 		 * kernels will stop after the first task is detached to minimize
11429 		 * the critical section.
11430 		 */
11431 		if (env->idle == CPU_NEWLY_IDLE)
11432 			break;
11433 #endif
11434 
11435 		/*
11436 		 * We only want to steal up to the prescribed amount of
11437 		 * load/util/tasks.
11438 		 */
11439 		if (env->imbalance <= 0)
11440 			break;
11441 
11442 		continue;
11443 next:
11444 		if (p->sched_task_hot)
11445 			schedstat_inc(p->stats.nr_failed_migrations_hot);
11446 
11447 		list_move(&p->se.group_node, tasks);
11448 	}
11449 
11450 	/*
11451 	 * Right now, this is one of only two places we collect this stat
11452 	 * so we can safely collect detach_one_task() stats here rather
11453 	 * than inside detach_one_task().
11454 	 */
11455 	schedstat_add(env->sd->lb_gained[env->idle], detached);
11456 
11457 	return detached;
11458 }
11459 
11460 /*
11461  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
11462  * new rq.
11463  */
11464 static void attach_tasks(struct lb_env *env)
11465 {
11466 	__attach_tasks(env->dst_rq, &env->tasks);
11467 }
11468 
11469 #ifdef CONFIG_NO_HZ_COMMON
11470 static inline bool cfs_rq_has_blocked_load(struct cfs_rq *cfs_rq)
11471 {
11472 	if (cfs_rq->avg.load_avg)
11473 		return true;
11474 
11475 	if (cfs_rq->avg.util_avg)
11476 		return true;
11477 
11478 	return false;
11479 }
11480 
11481 static inline bool others_have_blocked(struct rq *rq)
11482 {
11483 	if (cpu_util_rt(rq))
11484 		return true;
11485 
11486 	if (cpu_util_dl(rq))
11487 		return true;
11488 
11489 	if (hw_load_avg(rq))
11490 		return true;
11491 
11492 	if (cpu_util_irq(rq))
11493 		return true;
11494 
11495 	return false;
11496 }
11497 
11498 static inline void update_blocked_load_tick(struct rq *rq)
11499 {
11500 	WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
11501 }
11502 
11503 static inline void update_has_blocked_load_status(struct rq *rq, bool has_blocked_load)
11504 {
11505 	if (!has_blocked_load)
11506 		rq->has_blocked_load = 0;
11507 }
11508 #else /* !CONFIG_NO_HZ_COMMON: */
11509 static inline bool cfs_rq_has_blocked_load(struct cfs_rq *cfs_rq) { return false; }
11510 static inline bool others_have_blocked(struct rq *rq) { return false; }
11511 static inline void update_blocked_load_tick(struct rq *rq) {}
11512 static inline void update_has_blocked_load_status(struct rq *rq, bool has_blocked_load) {}
11513 #endif /* !CONFIG_NO_HZ_COMMON */
11514 
11515 static bool __update_blocked_others(struct rq *rq, bool *done)
11516 {
11517 	bool updated;
11518 
11519 	/*
11520 	 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
11521 	 * DL and IRQ signals have been updated before updating CFS.
11522 	 */
11523 	updated = update_other_load_avgs(rq);
11524 
11525 	if (others_have_blocked(rq))
11526 		*done = false;
11527 
11528 	return updated;
11529 }
11530 
11531 #ifdef CONFIG_FAIR_GROUP_SCHED
11532 
11533 static bool __update_blocked_fair(struct rq *rq, bool *done)
11534 {
11535 	struct cfs_rq *cfs_rq, *pos;
11536 	bool decayed = false;
11537 
11538 	/*
11539 	 * Iterates the task_group tree in a bottom up fashion, see
11540 	 * list_add_leaf_cfs_rq() for details.
11541 	 */
11542 	for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
11543 		struct sched_entity *se;
11544 
11545 		if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
11546 			update_tg_load_avg(cfs_rq);
11547 
11548 			if (cfs_rq->nr_queued == 0)
11549 				update_idle_cfs_rq_clock_pelt(cfs_rq);
11550 
11551 			if (cfs_rq == &rq->cfs)
11552 				decayed = true;
11553 		}
11554 
11555 		/* Propagate pending load changes to the parent, if any: */
11556 		se = cfs_rq_se(cfs_rq);
11557 		if (se && !skip_blocked_update(se))
11558 			update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11559 
11560 		/*
11561 		 * There can be a lot of idle CPU cgroups.  Don't let fully
11562 		 * decayed cfs_rqs linger on the list.
11563 		 */
11564 		if (cfs_rq_is_decayed(cfs_rq))
11565 			list_del_leaf_cfs_rq(cfs_rq);
11566 
11567 		/* Don't need periodic decay once load/util_avg are null */
11568 		if (cfs_rq_has_blocked_load(cfs_rq))
11569 			*done = false;
11570 	}
11571 
11572 	return decayed;
11573 }
11574 
11575 /*
11576  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
11577  * This needs to be done in a top-down fashion because the load of a child
11578  * group is a fraction of its parents load.
11579  */
11580 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
11581 {
11582 	struct sched_entity *se = cfs_rq_se(cfs_rq);
11583 	unsigned long now = jiffies;
11584 	unsigned long load;
11585 
11586 	if (cfs_rq->last_h_load_update == now)
11587 		return;
11588 
11589 	WRITE_ONCE(cfs_rq->h_load_next, NULL);
11590 	for_each_sched_entity(se) {
11591 		cfs_rq = cfs_rq_of(se);
11592 		WRITE_ONCE(cfs_rq->h_load_next, se);
11593 		if (cfs_rq->last_h_load_update == now)
11594 			break;
11595 	}
11596 
11597 	if (!se) {
11598 		cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
11599 		cfs_rq->last_h_load_update = now;
11600 	}
11601 
11602 	while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
11603 		load = cfs_rq->h_load;
11604 		load = div64_ul(load * se->avg.load_avg,
11605 				cfs_rq_load_avg(cfs_rq) + 1);
11606 		cfs_rq = group_cfs_rq(se);
11607 		cfs_rq->h_load = load;
11608 		cfs_rq->last_h_load_update = now;
11609 	}
11610 }
11611 
11612 static unsigned long task_h_load(struct task_struct *p)
11613 {
11614 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
11615 
11616 	update_cfs_rq_h_load(cfs_rq);
11617 	return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
11618 			cfs_rq_load_avg(cfs_rq) + 1);
11619 }
11620 #else /* !CONFIG_FAIR_GROUP_SCHED: */
11621 static bool __update_blocked_fair(struct rq *rq, bool *done)
11622 {
11623 	struct cfs_rq *cfs_rq = &rq->cfs;
11624 	bool decayed;
11625 
11626 	decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
11627 	if (cfs_rq_has_blocked_load(cfs_rq))
11628 		*done = false;
11629 
11630 	return decayed;
11631 }
11632 
11633 static unsigned long task_h_load(struct task_struct *p)
11634 {
11635 	return p->se.avg.load_avg;
11636 }
11637 #endif /* !CONFIG_FAIR_GROUP_SCHED */
11638 
11639 static void __sched_balance_update_blocked_averages(struct rq *rq)
11640 {
11641 	bool decayed = false, done = true;
11642 
11643 	update_blocked_load_tick(rq);
11644 
11645 	decayed |= __update_blocked_others(rq, &done);
11646 	decayed |= __update_blocked_fair(rq, &done);
11647 
11648 	update_has_blocked_load_status(rq, !done);
11649 	if (decayed)
11650 		cpufreq_update_util(rq, 0);
11651 }
11652 
11653 static void sched_balance_update_blocked_averages(int cpu)
11654 {
11655 	struct rq *rq = cpu_rq(cpu);
11656 
11657 	guard(rq_lock_irqsave)(rq);
11658 	update_rq_clock(rq);
11659 	__sched_balance_update_blocked_averages(rq);
11660 }
11661 
11662 /********** Helpers for sched_balance_find_src_group ************************/
11663 
11664 /*
11665  * sg_lb_stats - stats of a sched_group required for load-balancing:
11666  */
11667 struct sg_lb_stats {
11668 	unsigned long avg_load;			/* Avg load            over the CPUs of the group */
11669 	unsigned long group_load;		/* Total load          over the CPUs of the group */
11670 	unsigned long group_capacity;		/* Capacity            over the CPUs of the group */
11671 	unsigned long group_util;		/* Total utilization   over the CPUs of the group */
11672 	unsigned long group_runnable;		/* Total runnable time over the CPUs of the group */
11673 	unsigned int sum_nr_running;		/* Nr of all tasks running in the group */
11674 	unsigned int sum_h_nr_running;		/* Nr of CFS tasks running in the group */
11675 	unsigned int idle_cpus;                 /* Nr of idle CPUs         in the group */
11676 	unsigned int group_weight;
11677 	enum group_type group_type;
11678 	unsigned int group_asym_packing;	/* Tasks should be moved to preferred CPU */
11679 	unsigned int group_smt_balance;		/* Task on busy SMT be moved */
11680 	unsigned int group_llc_balance;		/* Tasks should be moved to preferred LLC */
11681 	unsigned long group_misfit_task_load;	/* A CPU has a task too big for its capacity */
11682 	unsigned int group_overutilized;	/* At least one CPU is overutilized in the group */
11683 #ifdef CONFIG_NUMA_BALANCING
11684 	unsigned int nr_numa_running;
11685 	unsigned int nr_preferred_running;
11686 #endif
11687 #ifdef CONFIG_SCHED_CACHE
11688 	unsigned int nr_pref_dst_llc;
11689 #endif
11690 };
11691 
11692 /*
11693  * sd_lb_stats - stats of a sched_domain required for load-balancing:
11694  */
11695 struct sd_lb_stats {
11696 	struct sched_group *busiest;		/* Busiest group in this sd */
11697 	struct sched_group *local;		/* Local group in this sd */
11698 	unsigned long total_load;		/* Total load of all groups in sd */
11699 	unsigned long total_capacity;		/* Total capacity of all groups in sd */
11700 	unsigned long avg_load;			/* Average load across all groups in sd */
11701 	unsigned int prefer_sibling;		/* Tasks should go to sibling first */
11702 
11703 	struct sg_lb_stats busiest_stat;	/* Statistics of the busiest group */
11704 	struct sg_lb_stats local_stat;		/* Statistics of the local group */
11705 };
11706 
11707 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
11708 {
11709 	/*
11710 	 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
11711 	 * local_stat because update_sg_lb_stats() does a full clear/assignment.
11712 	 * We must however set busiest_stat::group_type and
11713 	 * busiest_stat::idle_cpus to the worst busiest group because
11714 	 * update_sd_pick_busiest() reads these before assignment.
11715 	 */
11716 	*sds = (struct sd_lb_stats){
11717 		.busiest = NULL,
11718 		.local = NULL,
11719 		.total_load = 0UL,
11720 		.total_capacity = 0UL,
11721 		.busiest_stat = {
11722 			.idle_cpus = UINT_MAX,
11723 			.group_type = group_has_spare,
11724 		},
11725 	};
11726 }
11727 
11728 static unsigned long scale_rt_capacity(int cpu)
11729 {
11730 	unsigned long max = get_actual_cpu_capacity(cpu);
11731 	struct rq *rq = cpu_rq(cpu);
11732 	unsigned long used, free;
11733 	unsigned long irq;
11734 
11735 	irq = cpu_util_irq(rq);
11736 
11737 	if (unlikely(irq >= max))
11738 		return 1;
11739 
11740 	/*
11741 	 * avg_rt.util_avg and avg_dl.util_avg track binary signals
11742 	 * (running and not running) with weights 0 and 1024 respectively.
11743 	 */
11744 	used = cpu_util_rt(rq);
11745 	used += cpu_util_dl(rq);
11746 
11747 	if (unlikely(used >= max))
11748 		return 1;
11749 
11750 	free = max - used;
11751 
11752 	return scale_irq_capacity(free, irq, max);
11753 }
11754 
11755 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
11756 {
11757 	unsigned long capacity = scale_rt_capacity(cpu);
11758 	struct sched_group *sdg = sd->groups;
11759 
11760 	if (!capacity)
11761 		capacity = 1;
11762 
11763 	cpu_rq(cpu)->cpu_capacity = capacity;
11764 	trace_sched_cpu_capacity_tp(cpu_rq(cpu));
11765 
11766 	sdg->sgc->capacity = capacity;
11767 	sdg->sgc->min_capacity = capacity;
11768 	sdg->sgc->max_capacity = capacity;
11769 }
11770 
11771 void update_group_capacity(struct sched_domain *sd, int cpu)
11772 {
11773 	struct sched_domain *child = sd->child;
11774 	struct sched_group *group, *sdg = sd->groups;
11775 	unsigned long capacity, min_capacity, max_capacity;
11776 	unsigned long interval;
11777 
11778 	interval = msecs_to_jiffies(sd->balance_interval);
11779 	interval = clamp(interval, 1UL, max_load_balance_interval);
11780 	sdg->sgc->next_update = jiffies + interval;
11781 
11782 	if (!child) {
11783 		update_cpu_capacity(sd, cpu);
11784 		return;
11785 	}
11786 
11787 	capacity = 0;
11788 	min_capacity = ULONG_MAX;
11789 	max_capacity = 0;
11790 
11791 	if (child->flags & SD_NUMA) {
11792 		/*
11793 		 * SD_NUMA domains cannot assume that child groups
11794 		 * span the current group.
11795 		 */
11796 
11797 		for_each_cpu(cpu, sched_group_span(sdg)) {
11798 			unsigned long cpu_cap = capacity_of(cpu);
11799 
11800 			capacity += cpu_cap;
11801 			min_capacity = min(cpu_cap, min_capacity);
11802 			max_capacity = max(cpu_cap, max_capacity);
11803 		}
11804 	} else  {
11805 		/*
11806 		 * !SD_NUMA domains can assume that child groups
11807 		 * span the current group.
11808 		 */
11809 
11810 		group = child->groups;
11811 		do {
11812 			struct sched_group_capacity *sgc = group->sgc;
11813 
11814 			capacity += sgc->capacity;
11815 			min_capacity = min(sgc->min_capacity, min_capacity);
11816 			max_capacity = max(sgc->max_capacity, max_capacity);
11817 			group = group->next;
11818 		} while (group != child->groups);
11819 	}
11820 
11821 	sdg->sgc->capacity = capacity;
11822 	sdg->sgc->min_capacity = min_capacity;
11823 	sdg->sgc->max_capacity = max_capacity;
11824 }
11825 
11826 /*
11827  * Check whether the capacity of the rq has been noticeably reduced by side
11828  * activity. The imbalance_pct is used for the threshold.
11829  * Return true is the capacity is reduced
11830  */
11831 static inline int
11832 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
11833 {
11834 	return ((rq->cpu_capacity * sd->imbalance_pct) <
11835 				(arch_scale_cpu_capacity(cpu_of(rq)) * 100));
11836 }
11837 
11838 /* Check if the rq has a misfit task */
11839 static inline bool check_misfit_status(struct rq *rq)
11840 {
11841 	return rq->misfit_task_load;
11842 }
11843 
11844 /*
11845  * Group imbalance indicates (and tries to solve) the problem where balancing
11846  * groups is inadequate due to ->cpus_ptr constraints.
11847  *
11848  * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
11849  * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
11850  * Something like:
11851  *
11852  *	{ 0 1 2 3 } { 4 5 6 7 }
11853  *	        *     * * *
11854  *
11855  * If we were to balance group-wise we'd place two tasks in the first group and
11856  * two tasks in the second group. Clearly this is undesired as it will overload
11857  * cpu 3 and leave one of the CPUs in the second group unused.
11858  *
11859  * The current solution to this issue is detecting the skew in the first group
11860  * by noticing the lower domain failed to reach balance and had difficulty
11861  * moving tasks due to affinity constraints.
11862  *
11863  * When this is so detected; this group becomes a candidate for busiest; see
11864  * update_sd_pick_busiest(). And calculate_imbalance() and
11865  * sched_balance_find_src_group() avoid some of the usual balance conditions to allow it
11866  * to create an effective group imbalance.
11867  *
11868  * This is a somewhat tricky proposition since the next run might not find the
11869  * group imbalance and decide the groups need to be balanced again. A most
11870  * subtle and fragile situation.
11871  */
11872 
11873 static inline int sg_imbalanced(struct sched_group *group)
11874 {
11875 	return group->sgc->imbalance;
11876 }
11877 
11878 /*
11879  * group_has_capacity returns true if the group has spare capacity that could
11880  * be used by some tasks.
11881  * We consider that a group has spare capacity if the number of task is
11882  * smaller than the number of CPUs or if the utilization is lower than the
11883  * available capacity for CFS tasks.
11884  * For the latter, we use a threshold to stabilize the state, to take into
11885  * account the variance of the tasks' load and to return true if the available
11886  * capacity in meaningful for the load balancer.
11887  * As an example, an available capacity of 1% can appear but it doesn't make
11888  * any benefit for the load balance.
11889  */
11890 static inline bool
11891 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
11892 {
11893 	if (sgs->sum_nr_running < sgs->group_weight)
11894 		return true;
11895 
11896 	if ((sgs->group_capacity * imbalance_pct) <
11897 			(sgs->group_runnable * 100))
11898 		return false;
11899 
11900 	if ((sgs->group_capacity * 100) >
11901 			(sgs->group_util * imbalance_pct))
11902 		return true;
11903 
11904 	return false;
11905 }
11906 
11907 /*
11908  *  group_is_overloaded returns true if the group has more tasks than it can
11909  *  handle.
11910  *  group_is_overloaded is not equals to !group_has_capacity because a group
11911  *  with the exact right number of tasks, has no more spare capacity but is not
11912  *  overloaded so both group_has_capacity and group_is_overloaded return
11913  *  false.
11914  */
11915 static inline bool
11916 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
11917 {
11918 	/*
11919 	 * With EAS and uclamp, 1 CPU in the group must be overutilized to
11920 	 * consider the group overloaded.
11921 	 */
11922 	if (sched_energy_enabled() && !sgs->group_overutilized)
11923 		return false;
11924 
11925 	if (sgs->sum_nr_running <= sgs->group_weight)
11926 		return false;
11927 
11928 	if ((sgs->group_capacity * 100) <
11929 			(sgs->group_util * imbalance_pct))
11930 		return true;
11931 
11932 	if ((sgs->group_capacity * imbalance_pct) <
11933 			(sgs->group_runnable * 100))
11934 		return true;
11935 
11936 	return false;
11937 }
11938 
11939 static inline enum
11940 group_type group_classify(unsigned int imbalance_pct,
11941 			  struct sched_group *group,
11942 			  struct sg_lb_stats *sgs)
11943 {
11944 	if (group_is_overloaded(imbalance_pct, sgs))
11945 		return group_overloaded;
11946 
11947 	if (sgs->group_llc_balance)
11948 		return group_llc_balance;
11949 
11950 	if (sg_imbalanced(group))
11951 		return group_imbalanced;
11952 
11953 	if (sgs->group_asym_packing)
11954 		return group_asym_packing;
11955 
11956 	if (sgs->group_smt_balance)
11957 		return group_smt_balance;
11958 
11959 	if (sgs->group_misfit_task_load)
11960 		return group_misfit_task;
11961 
11962 	if (!group_has_capacity(imbalance_pct, sgs))
11963 		return group_fully_busy;
11964 
11965 	return group_has_spare;
11966 }
11967 
11968 /**
11969  * sched_use_asym_prio - Check whether asym_packing priority must be used
11970  * @sd:		The scheduling domain of the load balancing
11971  * @cpu:	A CPU
11972  *
11973  * Always use CPU priority when balancing load between SMT siblings. When
11974  * balancing load between cores, it is not sufficient that @cpu is idle. Only
11975  * use CPU priority if the whole core is idle.
11976  *
11977  * Returns: True if the priority of @cpu must be followed. False otherwise.
11978  */
11979 static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
11980 {
11981 	if (!(sd->flags & SD_ASYM_PACKING))
11982 		return false;
11983 
11984 	if (!sched_smt_active())
11985 		return true;
11986 
11987 	return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);
11988 }
11989 
11990 static inline bool sched_asym(struct sched_domain *sd, int dst_cpu, int src_cpu)
11991 {
11992 	/*
11993 	 * First check if @dst_cpu can do asym_packing load balance. Only do it
11994 	 * if it has higher priority than @src_cpu.
11995 	 */
11996 	return sched_use_asym_prio(sd, dst_cpu) &&
11997 		sched_asym_prefer(dst_cpu, src_cpu);
11998 }
11999 
12000 /**
12001  * sched_group_asym - Check if the destination CPU can do asym_packing balance
12002  * @env:	The load balancing environment
12003  * @sgs:	Load-balancing statistics of the candidate busiest group
12004  * @group:	The candidate busiest group
12005  *
12006  * @env::dst_cpu can do asym_packing if it has higher priority than the
12007  * preferred CPU of @group.
12008  *
12009  * Return: true if @env::dst_cpu can do with asym_packing load balance. False
12010  * otherwise.
12011  */
12012 static inline bool
12013 sched_group_asym(struct lb_env *env, struct sg_lb_stats *sgs, struct sched_group *group)
12014 {
12015 	/*
12016 	 * CPU priorities do not make sense for SMT cores with more than one
12017 	 * busy sibling.
12018 	 */
12019 	if ((group->flags & SD_SHARE_CPUCAPACITY) &&
12020 	    (sgs->group_weight - sgs->idle_cpus != 1))
12021 		return false;
12022 
12023 	return sched_asym(env->sd, env->dst_cpu, READ_ONCE(group->asym_prefer_cpu));
12024 }
12025 
12026 /* One group has more than one SMT CPU while the other group does not */
12027 static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1,
12028 				    struct sched_group *sg2)
12029 {
12030 	if (!sg1 || !sg2)
12031 		return false;
12032 
12033 	return (sg1->flags & SD_SHARE_CPUCAPACITY) !=
12034 		(sg2->flags & SD_SHARE_CPUCAPACITY);
12035 }
12036 
12037 static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs,
12038 			       struct sched_group *group)
12039 {
12040 	if (!env->idle)
12041 		return false;
12042 
12043 	/*
12044 	 * For SMT source group, it is better to move a task
12045 	 * to a CPU that doesn't have multiple tasks sharing its CPU capacity.
12046 	 * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY
12047 	 * will not be on.
12048 	 */
12049 	if (group->flags & SD_SHARE_CPUCAPACITY &&
12050 	    sgs->sum_h_nr_running > 1)
12051 		return true;
12052 
12053 	return false;
12054 }
12055 
12056 static inline long sibling_imbalance(struct lb_env *env,
12057 				    struct sd_lb_stats *sds,
12058 				    struct sg_lb_stats *busiest,
12059 				    struct sg_lb_stats *local)
12060 {
12061 	int ncores_busiest, ncores_local;
12062 	long imbalance;
12063 
12064 	if (!env->idle || !busiest->sum_nr_running)
12065 		return 0;
12066 
12067 	ncores_busiest = sds->busiest->cores;
12068 	ncores_local = sds->local->cores;
12069 
12070 	if (ncores_busiest == ncores_local) {
12071 		imbalance = busiest->sum_nr_running;
12072 		lsub_positive(&imbalance, local->sum_nr_running);
12073 		return imbalance;
12074 	}
12075 
12076 	/* Balance such that nr_running/ncores ratio are same on both groups */
12077 	imbalance = ncores_local * busiest->sum_nr_running;
12078 	lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);
12079 	/* Normalize imbalance and do rounding on normalization */
12080 	imbalance = 2 * imbalance + ncores_local + ncores_busiest;
12081 	imbalance /= ncores_local + ncores_busiest;
12082 
12083 	/* Take advantage of resource in an empty sched group */
12084 	if (imbalance <= 1 && local->sum_nr_running == 0 &&
12085 	    busiest->sum_nr_running > 1)
12086 		imbalance = 2;
12087 
12088 	return imbalance;
12089 }
12090 
12091 static inline bool
12092 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
12093 {
12094 	/*
12095 	 * When there is more than 1 task, the group_overloaded case already
12096 	 * takes care of cpu with reduced capacity
12097 	 */
12098 	if (rq->cfs.h_nr_runnable != 1)
12099 		return false;
12100 
12101 	return check_cpu_capacity(rq, sd);
12102 }
12103 
12104 #ifdef CONFIG_SCHED_CACHE
12105 /*
12106  * Record the statistics for this scheduler group for later
12107  * use. These values guide load balancing on aggregating tasks
12108  * to a LLC.
12109  */
12110 static void record_sg_llc_stats(struct lb_env *env,
12111 				struct sg_lb_stats *sgs,
12112 				struct sched_group *group)
12113 {
12114 	struct sched_domain_shared *sd_share;
12115 	int cpu;
12116 
12117 	if (!sched_cache_enabled() || env->idle == CPU_NEWLY_IDLE)
12118 		return;
12119 
12120 	/* Only care about sched domain spanning multiple LLCs */
12121 	if (env->sd->child != rcu_dereference_all(per_cpu(sd_llc, env->dst_cpu)))
12122 		return;
12123 
12124 	/*
12125 	 * A redo repeats stats published on the first pass. Hotplug can
12126 	 * still leave env->cpus partial, so util would not match capacity.
12127 	 */
12128 	if ((env->flags & LBF_ALL_PINNED) ||
12129 	    !cpumask_subset(sched_group_span(group), env->cpus))
12130 		return;
12131 
12132 	/*
12133 	 * At this point we know this group spans a LLC domain.
12134 	 * Record the statistic of this group in its corresponding
12135 	 * shared LLC domain.
12136 	 * Note: sd_share cannot be obtained via sd->child->shared,
12137 	 * because the latter refers to the domain that covers the
12138 	 * local group. Instead, sd_share should be located using
12139 	 * the first CPU of the LLC group.
12140 	 */
12141 	cpu = cpumask_first(sched_group_span(group));
12142 	sd_share = rcu_dereference_all(per_cpu(sd_llc_shared, cpu));
12143 	if (!sd_share)
12144 		return;
12145 
12146 	if (READ_ONCE(sd_share->util_avg) != sgs->group_util)
12147 		WRITE_ONCE(sd_share->util_avg, sgs->group_util);
12148 
12149 	if (unlikely(READ_ONCE(sd_share->capacity) != sgs->group_capacity))
12150 		WRITE_ONCE(sd_share->capacity, sgs->group_capacity);
12151 }
12152 
12153 /*
12154  * Do LLC balance on sched group that contains LLC, and have tasks preferring
12155  * to run on LLC in idle dst_cpu.
12156  */
12157 static inline bool llc_balance(struct lb_env *env, struct sg_lb_stats *sgs,
12158 			       struct sched_group *group)
12159 {
12160 	if (!sched_cache_enabled())
12161 		return false;
12162 
12163 	if (env->sd->flags & SD_SHARE_LLC)
12164 		return false;
12165 
12166 	/*
12167 	 * On asymmetric domains, group_misfit_task_load
12168 	 * should be prioritized to move tasks to CPU that fit them
12169 	 * over aggregating tasks to their preferred LLC.
12170 	 */
12171 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
12172 	    sgs->group_misfit_task_load)
12173 		return false;
12174 
12175 	/*
12176 	 * On asym packing domains, if the destination CPU
12177 	 * has higher priority than all CPUs in the source group,
12178 	 * prioritize asym packing.
12179 	 */
12180 	if (sgs->group_asym_packing)
12181 		return false;
12182 
12183 	/*
12184 	 * Skip cache aware tagging if nr_balanced_failed is sufficiently high.
12185 	 * Threshold of cache_nice_tries is set to 1 higher than nr_balance_failed
12186 	 * to avoid excessive task migration at the same time.
12187 	 */
12188 	if (env->sd->nr_balance_failed >= env->sd->cache_nice_tries + 1)
12189 		return false;
12190 
12191 	if (sgs->nr_pref_dst_llc &&
12192 	    can_migrate_llc(cpumask_first(sched_group_span(group)),
12193 			    env->dst_cpu, 0, true) == mig_llc)
12194 		return true;
12195 
12196 	return false;
12197 }
12198 
12199 static bool update_llc_busiest(struct lb_env *env,
12200 			       struct sg_lb_stats *busiest,
12201 			       struct sg_lb_stats *sgs)
12202 {
12203 	/*
12204 	 * There are more tasks that want to run on dst_cpu's LLC.
12205 	 */
12206 	return sgs->nr_pref_dst_llc > busiest->nr_pref_dst_llc;
12207 }
12208 #else
12209 static inline void record_sg_llc_stats(struct lb_env *env, struct sg_lb_stats *sgs,
12210 				       struct sched_group *group)
12211 {
12212 }
12213 
12214 static inline bool llc_balance(struct lb_env *env, struct sg_lb_stats *sgs,
12215 			       struct sched_group *group)
12216 {
12217 	return false;
12218 }
12219 
12220 static bool update_llc_busiest(struct lb_env *env,
12221 			       struct sg_lb_stats *busiest,
12222 			       struct sg_lb_stats *sgs)
12223 {
12224 	return false;
12225 }
12226 #endif
12227 
12228 /**
12229  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
12230  * @env: The load balancing environment.
12231  * @sds: Load-balancing data with statistics of the local group.
12232  * @group: sched_group whose statistics are to be updated.
12233  * @sgs: variable to hold the statistics for this group.
12234  * @sg_overloaded: sched_group is overloaded
12235  */
12236 static inline void update_sg_lb_stats(struct lb_env *env,
12237 				      struct sd_lb_stats *sds,
12238 				      struct sched_group *group,
12239 				      struct sg_lb_stats *sgs,
12240 				      bool *sg_overloaded)
12241 {
12242 	int i, nr_running, local_group, sd_flags = env->sd->flags;
12243 	bool balancing_at_rd = !env->sd->parent;
12244 
12245 	memset(sgs, 0, sizeof(*sgs));
12246 
12247 	local_group = group == sds->local;
12248 
12249 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
12250 		struct rq *rq = cpu_rq(i);
12251 		unsigned long load = cpu_load(rq);
12252 
12253 		sgs->group_load += load;
12254 		sgs->group_util += cpu_util_cfs(i);
12255 		sgs->group_runnable += cpu_runnable(rq);
12256 		sgs->sum_h_nr_running += rq->cfs.h_nr_runnable;
12257 
12258 		nr_running = rq->nr_running;
12259 		sgs->sum_nr_running += nr_running;
12260 
12261 		if (cpu_overutilized(i))
12262 			sgs->group_overutilized = 1;
12263 
12264 #ifdef CONFIG_SCHED_CACHE
12265 		if (sched_cache_enabled()) {
12266 			struct sched_domain *sd_tmp;
12267 			int dst_llc;
12268 
12269 			dst_llc = llc_id(env->dst_cpu);
12270 			if (llc_id(i) != dst_llc) {
12271 				sd_tmp = rcu_dereference_all(rq->sd);
12272 				if (sd_tmp && (unsigned int)dst_llc < sd_tmp->llc_max)
12273 					sgs->nr_pref_dst_llc += sd_tmp->llc_counts[dst_llc];
12274 			}
12275 		}
12276 #endif
12277 
12278 		/*
12279 		 * No need to call idle_cpu() if nr_running is not 0
12280 		 */
12281 		if (!nr_running && idle_cpu(i)) {
12282 			sgs->idle_cpus++;
12283 			/* Idle cpu can't have misfit task */
12284 			continue;
12285 		}
12286 
12287 		/* Overload indicator is only updated at root domain */
12288 		if (balancing_at_rd && nr_running > 1)
12289 			*sg_overloaded = 1;
12290 
12291 #ifdef CONFIG_NUMA_BALANCING
12292 		/* Only fbq_classify_group() uses this to classify NUMA groups */
12293 		if (sd_flags & SD_NUMA) {
12294 			sgs->nr_numa_running += rq->nr_numa_running;
12295 			sgs->nr_preferred_running += rq->nr_preferred_running;
12296 		}
12297 #endif
12298 		if (local_group)
12299 			continue;
12300 
12301 		if (sd_flags & SD_ASYM_CPUCAPACITY) {
12302 			if (rq->misfit_task_load) {
12303 				/*
12304 				 * Always mark the root domain overloaded so big
12305 				 * CPUs can pick up misfit tasks via newly idle
12306 				 * balance.
12307 				 */
12308 				if (balancing_at_rd)
12309 					*sg_overloaded = 1;
12310 
12311 				/*
12312 				 * Only account misfit load if @dst_cpu can
12313 				 * help; otherwise, the group may be classified
12314 				 * as misfit_task and update_sd_pick_busiest()
12315 				 * will skip it.
12316 				 */
12317 				if (capacity_greater(capacity_of(env->dst_cpu),
12318 						     group->sgc->max_capacity) &&
12319 				    (sgs->group_misfit_task_load < rq->misfit_task_load))
12320 					sgs->group_misfit_task_load = rq->misfit_task_load;
12321 			}
12322 		} else if (env->idle && sched_reduced_capacity(rq, env->sd)) {
12323 			/* Check for a task running on a CPU with reduced capacity */
12324 			if (sgs->group_misfit_task_load < load)
12325 				sgs->group_misfit_task_load = load;
12326 		}
12327 	}
12328 
12329 	sgs->group_capacity = group->sgc->capacity;
12330 
12331 	sgs->group_weight = group->group_weight;
12332 
12333 	if (!local_group) {
12334 		/* Check if dst CPU is idle and preferred to this group */
12335 		if (env->idle && sgs->sum_h_nr_running &&
12336 		    sched_group_asym(env, sgs, group))
12337 			sgs->group_asym_packing = 1;
12338 
12339 		/* Check for loaded SMT group to be balanced to dst CPU */
12340 		if (smt_balance(env, sgs, group))
12341 			sgs->group_smt_balance = 1;
12342 
12343 		/* Check for tasks in this group can be moved to their preferred LLC */
12344 		if (llc_balance(env, sgs, group))
12345 			sgs->group_llc_balance = 1;
12346 	}
12347 
12348 	sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
12349 
12350 	record_sg_llc_stats(env, sgs, group);
12351 	/* Computing avg_load makes sense only when group is overloaded */
12352 	if (sgs->group_type == group_overloaded)
12353 		sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
12354 				sgs->group_capacity;
12355 }
12356 
12357 /**
12358  * update_sd_pick_busiest - return 1 on busiest group
12359  * @env: The load balancing environment.
12360  * @sds: sched_domain statistics
12361  * @sg: sched_group candidate to be checked for being the busiest
12362  * @sgs: sched_group statistics
12363  *
12364  * Determine if @sg is a busier group than the previously selected
12365  * busiest group.
12366  *
12367  * Return: %true if @sg is a busier group than the previously selected
12368  * busiest group. %false otherwise.
12369  */
12370 static bool update_sd_pick_busiest(struct lb_env *env,
12371 				   struct sd_lb_stats *sds,
12372 				   struct sched_group *sg,
12373 				   struct sg_lb_stats *sgs)
12374 {
12375 	struct sg_lb_stats *busiest = &sds->busiest_stat;
12376 
12377 	/* Make sure that there is at least one task to pull */
12378 	if (!sgs->sum_h_nr_running)
12379 		return false;
12380 
12381 	/*
12382 	 * Don't try to pull misfit tasks we can't help.
12383 	 * We can use max_capacity here as reduction in capacity on some
12384 	 * CPUs in the group should either be possible to resolve
12385 	 * internally or be covered by avg_load imbalance (eventually).
12386 	 *
12387 	 * When SMT is active, only pull a misfit to dst_cpu if it is on a
12388 	 * fully idle core; otherwise the effective capacity of the core is
12389 	 * reduced and we may not actually provide more capacity than the
12390 	 * source.
12391 	 */
12392 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
12393 	    (sgs->group_type == group_misfit_task) &&
12394 	    (!env->dst_core_idle ||
12395 	     !capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
12396 	     sds->local_stat.group_type != group_has_spare))
12397 		return false;
12398 
12399 	/*
12400 	 * Candidate sg has no more than one task per CPU and has higher
12401 	 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
12402 	 * throughput. Maximize throughput, power/energy consequences are not
12403 	 * considered.
12404 	 */
12405 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
12406 	    (sgs->group_type <= group_fully_busy) &&
12407 	    (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
12408 		return false;
12409 
12410 	if (sgs->group_type > busiest->group_type)
12411 		return true;
12412 
12413 	if (sgs->group_type < busiest->group_type)
12414 		return false;
12415 
12416 	/*
12417 	 * The candidate and the current busiest group are the same type of
12418 	 * group. Let check which one is the busiest according to the type.
12419 	 */
12420 
12421 	switch (sgs->group_type) {
12422 	case group_overloaded:
12423 		/* Select the overloaded group with highest avg_load. */
12424 		return sgs->avg_load > busiest->avg_load;
12425 
12426 	case group_llc_balance:
12427 		/* Select the group with most tasks preferring dst LLC */
12428 		return update_llc_busiest(env, busiest, sgs);
12429 
12430 	case group_imbalanced:
12431 		/*
12432 		 * Select the 1st imbalanced group as we don't have any way to
12433 		 * choose one more than another.
12434 		 */
12435 		return false;
12436 
12437 	case group_asym_packing:
12438 		/* Prefer to move from lowest priority CPU's work */
12439 		return sched_asym_prefer(READ_ONCE(sds->busiest->asym_prefer_cpu),
12440 					 READ_ONCE(sg->asym_prefer_cpu));
12441 
12442 	case group_misfit_task:
12443 		/*
12444 		 * If we have more than one misfit sg go with the biggest
12445 		 * misfit.
12446 		 */
12447 		return sgs->group_misfit_task_load > busiest->group_misfit_task_load;
12448 
12449 	case group_smt_balance:
12450 		/*
12451 		 * Check if we have spare CPUs on either SMT group to
12452 		 * choose has spare or fully busy handling.
12453 		 */
12454 		if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)
12455 			goto has_spare;
12456 
12457 		fallthrough;
12458 
12459 	case group_fully_busy:
12460 		/*
12461 		 * Select the fully busy group with highest avg_load. In
12462 		 * theory, there is no need to pull task from such kind of
12463 		 * group because tasks have all compute capacity that they need
12464 		 * but we can still improve the overall throughput by reducing
12465 		 * contention when accessing shared HW resources.
12466 		 *
12467 		 * XXX for now avg_load is not computed and always 0 so we
12468 		 * select the 1st one, except if @sg is composed of SMT
12469 		 * siblings.
12470 		 */
12471 
12472 		if (sgs->avg_load < busiest->avg_load)
12473 			return false;
12474 
12475 		if (sgs->avg_load == busiest->avg_load) {
12476 			/*
12477 			 * SMT sched groups need more help than non-SMT groups.
12478 			 * If @sg happens to also be SMT, either choice is good.
12479 			 */
12480 			if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)
12481 				return false;
12482 		}
12483 
12484 		break;
12485 
12486 	case group_has_spare:
12487 		/*
12488 		 * Do not pick sg with SMT CPUs over sg with pure CPUs,
12489 		 * as we do not want to pull task off SMT core with one task
12490 		 * and make the core idle.
12491 		 */
12492 		if (smt_vs_nonsmt_groups(sds->busiest, sg)) {
12493 			if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)
12494 				return false;
12495 			else
12496 				return true;
12497 		}
12498 has_spare:
12499 
12500 		/*
12501 		 * Select not overloaded group with lowest number of idle CPUs
12502 		 * and highest number of running tasks. We could also compare
12503 		 * the spare capacity which is more stable but it can end up
12504 		 * that the group has less spare capacity but finally more idle
12505 		 * CPUs which means less opportunity to pull tasks.
12506 		 */
12507 		if (sgs->idle_cpus > busiest->idle_cpus)
12508 			return false;
12509 		else if ((sgs->idle_cpus == busiest->idle_cpus) &&
12510 			 (sgs->sum_nr_running <= busiest->sum_nr_running))
12511 			return false;
12512 
12513 		break;
12514 	}
12515 
12516 	return true;
12517 }
12518 
12519 #ifdef CONFIG_NUMA_BALANCING
12520 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
12521 {
12522 	if (sgs->sum_h_nr_running > sgs->nr_numa_running)
12523 		return regular;
12524 	if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
12525 		return remote;
12526 	return all;
12527 }
12528 
12529 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
12530 {
12531 	if (rq->nr_running > rq->nr_numa_running)
12532 		return regular;
12533 	if (rq->nr_running > rq->nr_preferred_running)
12534 		return remote;
12535 	return all;
12536 }
12537 #else /* !CONFIG_NUMA_BALANCING: */
12538 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
12539 {
12540 	return all;
12541 }
12542 
12543 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
12544 {
12545 	return regular;
12546 }
12547 #endif /* !CONFIG_NUMA_BALANCING */
12548 
12549 
12550 struct sg_lb_stats;
12551 
12552 /*
12553  * task_running_on_cpu - return 1 if @p is running on @cpu.
12554  */
12555 
12556 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
12557 {
12558 	/* Task has no contribution or is new */
12559 	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
12560 		return 0;
12561 
12562 	if (task_on_rq_queued(p))
12563 		return 1;
12564 
12565 	return 0;
12566 }
12567 
12568 /**
12569  * idle_cpu_without - would a given CPU be idle without p ?
12570  * @cpu: the processor on which idleness is tested.
12571  * @p: task which should be ignored.
12572  *
12573  * Return: 1 if the CPU would be idle. 0 otherwise.
12574  */
12575 static int idle_cpu_without(int cpu, struct task_struct *p)
12576 {
12577 	struct rq *rq = cpu_rq(cpu);
12578 
12579 	if (rq->curr != rq->idle && rq->curr != p)
12580 		return 0;
12581 
12582 	/*
12583 	 * rq->nr_running can't be used but an updated version without the
12584 	 * impact of p on cpu must be used instead. The updated nr_running
12585 	 * be computed and tested before calling idle_cpu_without().
12586 	 */
12587 
12588 	if (rq->ttwu_pending)
12589 		return 0;
12590 
12591 	return 1;
12592 }
12593 
12594 /*
12595  * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
12596  * @sd: The sched_domain level to look for idlest group.
12597  * @group: sched_group whose statistics are to be updated.
12598  * @sgs: variable to hold the statistics for this group.
12599  * @p: The task for which we look for the idlest group/CPU.
12600  */
12601 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
12602 					  struct sched_group *group,
12603 					  struct sg_lb_stats *sgs,
12604 					  struct task_struct *p)
12605 {
12606 	int i, nr_running;
12607 
12608 	memset(sgs, 0, sizeof(*sgs));
12609 
12610 	/* Assume that task can't fit any CPU of the group */
12611 	if (sd->flags & SD_ASYM_CPUCAPACITY)
12612 		sgs->group_misfit_task_load = 1;
12613 
12614 	for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
12615 		struct rq *rq = cpu_rq(i);
12616 		unsigned int local;
12617 
12618 		sgs->group_load += cpu_load_without(rq, p);
12619 		sgs->group_util += cpu_util_without(i, p);
12620 		sgs->group_runnable += cpu_runnable_without(rq, p);
12621 		local = task_running_on_cpu(i, p);
12622 		sgs->sum_h_nr_running += rq->cfs.h_nr_runnable - local;
12623 
12624 		nr_running = rq->nr_running - local;
12625 		sgs->sum_nr_running += nr_running;
12626 
12627 		/*
12628 		 * No need to call idle_cpu_without() if nr_running is not 0
12629 		 */
12630 		if (!nr_running && idle_cpu_without(i, p))
12631 			sgs->idle_cpus++;
12632 
12633 		/* Check if task fits in the CPU */
12634 		if (sd->flags & SD_ASYM_CPUCAPACITY &&
12635 		    sgs->group_misfit_task_load &&
12636 		    task_fits_cpu(p, i))
12637 			sgs->group_misfit_task_load = 0;
12638 
12639 	}
12640 
12641 	sgs->group_capacity = group->sgc->capacity;
12642 
12643 	sgs->group_weight = group->group_weight;
12644 
12645 	sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
12646 
12647 	/*
12648 	 * Computing avg_load makes sense only when group is fully busy or
12649 	 * overloaded
12650 	 */
12651 	if (sgs->group_type == group_fully_busy ||
12652 		sgs->group_type == group_overloaded)
12653 		sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
12654 				sgs->group_capacity;
12655 }
12656 
12657 static bool update_pick_idlest(struct sched_group *idlest,
12658 			       struct sg_lb_stats *idlest_sgs,
12659 			       struct sched_group *group,
12660 			       struct sg_lb_stats *sgs)
12661 {
12662 	if (sgs->group_type < idlest_sgs->group_type)
12663 		return true;
12664 
12665 	if (sgs->group_type > idlest_sgs->group_type)
12666 		return false;
12667 
12668 	/*
12669 	 * The candidate and the current idlest group are the same type of
12670 	 * group. Let check which one is the idlest according to the type.
12671 	 */
12672 
12673 	switch (sgs->group_type) {
12674 	case group_overloaded:
12675 	case group_fully_busy:
12676 		/* Select the group with lowest avg_load. */
12677 		if (idlest_sgs->avg_load <= sgs->avg_load)
12678 			return false;
12679 		break;
12680 
12681 	case group_llc_balance:
12682 	case group_imbalanced:
12683 	case group_asym_packing:
12684 	case group_smt_balance:
12685 		/* Those types are not used in the slow wakeup path */
12686 		return false;
12687 
12688 	case group_misfit_task:
12689 		/* Select group with the highest max capacity */
12690 		if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
12691 			return false;
12692 		break;
12693 
12694 	case group_has_spare:
12695 		/* Select group with most idle CPUs */
12696 		if (idlest_sgs->idle_cpus > sgs->idle_cpus)
12697 			return false;
12698 
12699 		/* Select group with lowest group_util */
12700 		if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
12701 			idlest_sgs->group_util <= sgs->group_util)
12702 			return false;
12703 
12704 		break;
12705 	}
12706 
12707 	return true;
12708 }
12709 
12710 /*
12711  * sched_balance_find_dst_group() finds and returns the least busy CPU group within the
12712  * domain.
12713  *
12714  * Assumes p is allowed on at least one CPU in sd.
12715  */
12716 static struct sched_group *
12717 sched_balance_find_dst_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
12718 {
12719 	struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
12720 	struct sg_lb_stats local_sgs, tmp_sgs;
12721 	struct sg_lb_stats *sgs;
12722 	unsigned long imbalance;
12723 	struct sg_lb_stats idlest_sgs = {
12724 			.avg_load = UINT_MAX,
12725 			.group_type = group_overloaded,
12726 	};
12727 
12728 	do {
12729 		int local_group;
12730 
12731 		/* Skip over this group if it has no CPUs allowed */
12732 		if (!cpumask_intersects(sched_group_span(group),
12733 					p->cpus_ptr))
12734 			continue;
12735 
12736 		/* Skip over this group if no cookie matched */
12737 		if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
12738 			continue;
12739 
12740 		local_group = cpumask_test_cpu(this_cpu,
12741 					       sched_group_span(group));
12742 
12743 		if (local_group) {
12744 			sgs = &local_sgs;
12745 			local = group;
12746 		} else {
12747 			sgs = &tmp_sgs;
12748 		}
12749 
12750 		update_sg_wakeup_stats(sd, group, sgs, p);
12751 
12752 		if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
12753 			idlest = group;
12754 			idlest_sgs = *sgs;
12755 		}
12756 
12757 	} while (group = group->next, group != sd->groups);
12758 
12759 
12760 	/* There is no idlest group to push tasks to */
12761 	if (!idlest)
12762 		return NULL;
12763 
12764 	/* The local group has been skipped because of CPU affinity */
12765 	if (!local)
12766 		return idlest;
12767 
12768 	/*
12769 	 * If the local group is idler than the selected idlest group
12770 	 * don't try and push the task.
12771 	 */
12772 	if (local_sgs.group_type < idlest_sgs.group_type)
12773 		return NULL;
12774 
12775 	/*
12776 	 * If the local group is busier than the selected idlest group
12777 	 * try and push the task.
12778 	 */
12779 	if (local_sgs.group_type > idlest_sgs.group_type)
12780 		return idlest;
12781 
12782 	switch (local_sgs.group_type) {
12783 	case group_overloaded:
12784 	case group_fully_busy:
12785 
12786 		/* Calculate allowed imbalance based on load */
12787 		imbalance = scale_load_down(NICE_0_LOAD) *
12788 				(sd->imbalance_pct-100) / 100;
12789 
12790 		/*
12791 		 * When comparing groups across NUMA domains, it's possible for
12792 		 * the local domain to be very lightly loaded relative to the
12793 		 * remote domains but "imbalance" skews the comparison making
12794 		 * remote CPUs look much more favourable. When considering
12795 		 * cross-domain, add imbalance to the load on the remote node
12796 		 * and consider staying local.
12797 		 */
12798 
12799 		if ((sd->flags & SD_NUMA) &&
12800 		    ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
12801 			return NULL;
12802 
12803 		/*
12804 		 * If the local group is less loaded than the selected
12805 		 * idlest group don't try and push any tasks.
12806 		 */
12807 		if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
12808 			return NULL;
12809 
12810 		if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
12811 			return NULL;
12812 		break;
12813 
12814 	case group_llc_balance:
12815 	case group_imbalanced:
12816 	case group_asym_packing:
12817 	case group_smt_balance:
12818 		/* Those type are not used in the slow wakeup path */
12819 		return NULL;
12820 
12821 	case group_misfit_task:
12822 		/* Select group with the highest max capacity */
12823 		if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
12824 			return NULL;
12825 		break;
12826 
12827 	case group_has_spare:
12828 #ifdef CONFIG_NUMA
12829 		if (sd->flags & SD_NUMA) {
12830 			int imb_numa_nr = sd->imb_numa_nr;
12831 #ifdef CONFIG_NUMA_BALANCING
12832 			int idlest_cpu;
12833 			/*
12834 			 * If there is spare capacity at NUMA, try to select
12835 			 * the preferred node
12836 			 */
12837 			if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
12838 				return NULL;
12839 
12840 			idlest_cpu = cpumask_first(sched_group_span(idlest));
12841 			if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
12842 				return idlest;
12843 #endif /* CONFIG_NUMA_BALANCING */
12844 			/*
12845 			 * Otherwise, keep the task close to the wakeup source
12846 			 * and improve locality if the number of running tasks
12847 			 * would remain below threshold where an imbalance is
12848 			 * allowed while accounting for the possibility the
12849 			 * task is pinned to a subset of CPUs. If there is a
12850 			 * real need of migration, periodic load balance will
12851 			 * take care of it.
12852 			 */
12853 			if (p->nr_cpus_allowed != NR_CPUS) {
12854 				unsigned int w = cpumask_weight_and(p->cpus_ptr,
12855 								sched_group_span(local));
12856 				imb_numa_nr = min(w, sd->imb_numa_nr);
12857 			}
12858 
12859 			imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
12860 			if (!adjust_numa_imbalance(imbalance,
12861 						   local_sgs.sum_nr_running + 1,
12862 						   imb_numa_nr)) {
12863 				return NULL;
12864 			}
12865 		}
12866 #endif /* CONFIG_NUMA */
12867 
12868 		/*
12869 		 * Select group with highest number of idle CPUs. We could also
12870 		 * compare the utilization which is more stable but it can end
12871 		 * up that the group has less spare capacity but finally more
12872 		 * idle CPUs which means more opportunity to run task.
12873 		 */
12874 		if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
12875 			return NULL;
12876 		break;
12877 	}
12878 
12879 	return idlest;
12880 }
12881 
12882 static void update_idle_cpu_scan(struct lb_env *env,
12883 				 unsigned long sum_util)
12884 {
12885 	struct sched_domain_shared *sd_share;
12886 	struct sched_domain *sd = env->sd;
12887 	int llc_weight, pct;
12888 	u64 x, y, tmp;
12889 	/*
12890 	 * Update the number of CPUs to scan in LLC domain, which could
12891 	 * be used as a hint in select_idle_cpu(). The update of sd_share
12892 	 * could be expensive because it is within a shared cache line.
12893 	 * So the write of this hint only occurs during periodic load
12894 	 * balancing, rather than CPU_NEWLY_IDLE, because the latter
12895 	 * can fire way more frequently than the former.
12896 	 */
12897 	if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
12898 		return;
12899 
12900 	sd_share = sd->shared;
12901 	if (!sd_share)
12902 		return;
12903 
12904 	/*
12905 	 * The number of CPUs to search drops as sum_util increases, when
12906 	 * sum_util hits 85% or above, the scan stops.
12907 	 * The reason to choose 85% as the threshold is because this is the
12908 	 * imbalance_pct(117) when a LLC sched group is overloaded.
12909 	 *
12910 	 * let y = SCHED_CAPACITY_SCALE - p * x^2                       [1]
12911 	 * and y'= y / SCHED_CAPACITY_SCALE
12912 	 *
12913 	 * x is the ratio of sum_util compared to the CPU capacity:
12914 	 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
12915 	 * y' is the ratio of CPUs to be scanned in the LLC domain,
12916 	 * and the number of CPUs to scan is calculated by:
12917 	 *
12918 	 * nr_scan = llc_weight * y'                                    [2]
12919 	 *
12920 	 * When x hits the threshold of overloaded, AKA, when
12921 	 * x = 100 / pct, y drops to 0. According to [1],
12922 	 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
12923 	 *
12924 	 * Scale x by SCHED_CAPACITY_SCALE:
12925 	 * x' = sum_util / llc_weight;                                  [3]
12926 	 *
12927 	 * and finally [1] becomes:
12928 	 * y = SCHED_CAPACITY_SCALE -
12929 	 *     x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE)            [4]
12930 	 *
12931 	 */
12932 	/* equation [3] */
12933 	x = sum_util;
12934 	llc_weight = sd->span_weight;
12935 	do_div(x, llc_weight);
12936 
12937 	/* equation [4] */
12938 	pct = sd->imbalance_pct;
12939 	tmp = x * x * pct * pct;
12940 	do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
12941 	tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
12942 	y = SCHED_CAPACITY_SCALE - tmp;
12943 
12944 	/* equation [2] */
12945 	y *= llc_weight;
12946 	do_div(y, SCHED_CAPACITY_SCALE);
12947 	if ((int)y != sd_share->nr_idle_scan)
12948 		WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
12949 }
12950 
12951 /**
12952  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
12953  * @env: The load balancing environment.
12954  * @sds: variable to hold the statistics for this sched_domain.
12955  */
12956 
12957 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
12958 {
12959 	struct sched_group *sg = env->sd->groups;
12960 	struct sg_lb_stats *local = &sds->local_stat;
12961 	struct sg_lb_stats tmp_sgs;
12962 	unsigned long sum_util = 0;
12963 	bool sg_overloaded = 0, sg_overutilized = 0;
12964 
12965 	env->dst_core_idle = !sched_smt_active() || is_core_idle(env->dst_cpu);
12966 
12967 	do {
12968 		struct sg_lb_stats *sgs = &tmp_sgs;
12969 		int local_group;
12970 
12971 		local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
12972 		if (local_group) {
12973 			sds->local = sg;
12974 			sgs = local;
12975 
12976 			if (env->idle != CPU_NEWLY_IDLE ||
12977 			    time_after_eq(jiffies, sg->sgc->next_update))
12978 				update_group_capacity(env->sd, env->dst_cpu);
12979 		}
12980 
12981 		update_sg_lb_stats(env, sds, sg, sgs, &sg_overloaded);
12982 
12983 		if (!local_group && update_sd_pick_busiest(env, sds, sg, sgs)) {
12984 			sds->busiest = sg;
12985 			sds->busiest_stat = *sgs;
12986 		}
12987 
12988 		sg_overutilized |= sgs->group_overutilized;
12989 
12990 		/* Now, start updating sd_lb_stats */
12991 		sds->total_load += sgs->group_load;
12992 		sds->total_capacity += sgs->group_capacity;
12993 
12994 		sum_util += sgs->group_util;
12995 		sg = sg->next;
12996 	} while (sg != env->sd->groups);
12997 
12998 	/*
12999 	 * Indicate that the child domain of the busiest group prefers tasks
13000 	 * go to a child's sibling domains first. NB the flags of a sched group
13001 	 * are those of the child domain.
13002 	 */
13003 	if (sds->busiest)
13004 		sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);
13005 
13006 
13007 	if (env->sd->flags & SD_NUMA)
13008 		env->fbq_type = fbq_classify_group(&sds->busiest_stat);
13009 
13010 	if (!env->sd->parent) {
13011 		/* update overload indicator if we are at root domain */
13012 		set_rd_overloaded(env->dst_rq->rd, sg_overloaded);
13013 
13014 		/* Update over-utilization (tipping point, U >= 0) indicator */
13015 		set_rd_overutilized(env->dst_rq->rd, sg_overutilized);
13016 	} else if (sg_overutilized) {
13017 		set_rd_overutilized(env->dst_rq->rd, sg_overutilized);
13018 	}
13019 
13020 	update_idle_cpu_scan(env, sum_util);
13021 }
13022 
13023 /**
13024  * calculate_imbalance - Calculate the amount of imbalance present within the
13025  *			 groups of a given sched_domain during load balance.
13026  * @env: load balance environment
13027  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
13028  */
13029 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
13030 {
13031 	struct sg_lb_stats *local, *busiest;
13032 
13033 	local = &sds->local_stat;
13034 	busiest = &sds->busiest_stat;
13035 
13036 	if (busiest->group_type == group_misfit_task) {
13037 		if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
13038 			/* Set imbalance to allow misfit tasks to be balanced. */
13039 			env->migration_type = migrate_misfit;
13040 			env->imbalance = 1;
13041 		} else {
13042 			/*
13043 			 * Set load imbalance to allow moving task from cpu
13044 			 * with reduced capacity.
13045 			 */
13046 			env->migration_type = migrate_load;
13047 			env->imbalance = busiest->group_misfit_task_load;
13048 		}
13049 		return;
13050 	}
13051 
13052 	if (busiest->group_type == group_asym_packing) {
13053 		/*
13054 		 * In case of asym capacity, we will try to migrate all load to
13055 		 * the preferred CPU.
13056 		 */
13057 		env->migration_type = migrate_task;
13058 		env->imbalance = busiest->sum_h_nr_running;
13059 		return;
13060 	}
13061 
13062 	if (busiest->group_type == group_smt_balance) {
13063 		/* Reduce number of tasks sharing CPU capacity */
13064 		env->migration_type = migrate_task;
13065 		env->imbalance = 1;
13066 		return;
13067 	}
13068 
13069 #ifdef CONFIG_SCHED_CACHE
13070 	if (busiest->group_type == group_llc_balance) {
13071 		/* Move a task that prefer local LLC */
13072 		env->migration_type = migrate_llc_task;
13073 		env->imbalance = 1;
13074 		return;
13075 	}
13076 #endif
13077 
13078 	if (busiest->group_type == group_imbalanced) {
13079 		/*
13080 		 * In the group_imb case we cannot rely on group-wide averages
13081 		 * to ensure CPU-load equilibrium, try to move any task to fix
13082 		 * the imbalance. The next load balance will take care of
13083 		 * balancing back the system.
13084 		 */
13085 		env->migration_type = migrate_task;
13086 		env->imbalance = 1;
13087 		return;
13088 	}
13089 
13090 	/*
13091 	 * Try to use spare capacity of local group without overloading it or
13092 	 * emptying busiest.
13093 	 */
13094 	if (local->group_type == group_has_spare) {
13095 		if ((busiest->group_type > group_fully_busy) &&
13096 		    !(env->sd->flags & SD_SHARE_LLC)) {
13097 			/*
13098 			 * If busiest is overloaded, try to fill spare
13099 			 * capacity. This might end up creating spare capacity
13100 			 * in busiest or busiest still being overloaded but
13101 			 * there is no simple way to directly compute the
13102 			 * amount of load to migrate in order to balance the
13103 			 * system.
13104 			 */
13105 			env->migration_type = migrate_util;
13106 			env->imbalance = max(local->group_capacity, local->group_util) -
13107 					 local->group_util;
13108 
13109 			/*
13110 			 * In some cases, the group's utilization is max or even
13111 			 * higher than capacity because of migrations but the
13112 			 * local CPU is (newly) idle. There is at least one
13113 			 * waiting task in this overloaded busiest group. Let's
13114 			 * try to pull it.
13115 			 */
13116 			if (env->idle && env->imbalance == 0) {
13117 				env->migration_type = migrate_task;
13118 				env->imbalance = 1;
13119 			}
13120 
13121 			return;
13122 		}
13123 
13124 		if (busiest->group_weight == 1 || sds->prefer_sibling) {
13125 			/*
13126 			 * When prefer sibling, evenly spread running tasks on
13127 			 * groups.
13128 			 */
13129 			env->migration_type = migrate_task;
13130 			env->imbalance = sibling_imbalance(env, sds, busiest, local);
13131 		} else {
13132 
13133 			/*
13134 			 * If there is no overload, we just want to even the number of
13135 			 * idle CPUs.
13136 			 */
13137 			env->migration_type = migrate_task;
13138 			env->imbalance = max_t(long, 0,
13139 					       (local->idle_cpus - busiest->idle_cpus));
13140 		}
13141 
13142 #ifdef CONFIG_NUMA
13143 		/* Consider allowing a small imbalance between NUMA groups */
13144 		if (env->sd->flags & SD_NUMA) {
13145 			env->imbalance = adjust_numa_imbalance(env->imbalance,
13146 							       local->sum_nr_running + 1,
13147 							       env->sd->imb_numa_nr);
13148 		}
13149 #endif
13150 
13151 		/* Number of tasks to move to restore balance */
13152 		env->imbalance >>= 1;
13153 
13154 		return;
13155 	}
13156 
13157 	/*
13158 	 * Local is fully busy but has to take more load to relieve the
13159 	 * busiest group
13160 	 */
13161 	if (local->group_type < group_overloaded) {
13162 		/*
13163 		 * Local will become overloaded so the avg_load metrics are
13164 		 * finally needed.
13165 		 */
13166 
13167 		local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
13168 				  local->group_capacity;
13169 
13170 		/*
13171 		 * If the local group is more loaded than the selected
13172 		 * busiest group don't try to pull any tasks.
13173 		 */
13174 		if (local->avg_load >= busiest->avg_load) {
13175 			env->imbalance = 0;
13176 			return;
13177 		}
13178 
13179 		sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
13180 				sds->total_capacity;
13181 
13182 		/*
13183 		 * If the local group is more loaded than the average system
13184 		 * load, don't try to pull any tasks.
13185 		 */
13186 		if (local->avg_load >= sds->avg_load) {
13187 			env->imbalance = 0;
13188 			return;
13189 		}
13190 
13191 	}
13192 
13193 	/*
13194 	 * Both group are or will become overloaded and we're trying to get all
13195 	 * the CPUs to the average_load, so we don't want to push ourselves
13196 	 * above the average load, nor do we wish to reduce the max loaded CPU
13197 	 * below the average load. At the same time, we also don't want to
13198 	 * reduce the group load below the group capacity. Thus we look for
13199 	 * the minimum possible imbalance.
13200 	 */
13201 	env->migration_type = migrate_load;
13202 	env->imbalance = min(
13203 		(busiest->avg_load - sds->avg_load) * busiest->group_capacity,
13204 		(sds->avg_load - local->avg_load) * local->group_capacity
13205 	) / SCHED_CAPACITY_SCALE;
13206 }
13207 
13208 /******* sched_balance_find_src_group() helpers end here *********************/
13209 
13210 /*
13211  * Decision matrix according to the local and busiest group type:
13212  *
13213  * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
13214  * has_spare        nr_idle   balanced   N/A    N/A  balanced   balanced
13215  * fully_busy       nr_idle   nr_idle    N/A    N/A  balanced   balanced
13216  * misfit_task      force     N/A        N/A    N/A  N/A        N/A
13217  * asym_packing     force     force      N/A    N/A  force      force
13218  * imbalanced       force     force      N/A    N/A  force      force
13219  * overloaded       force     force      N/A    N/A  force      avg_load
13220  *
13221  * N/A :      Not Applicable because already filtered while updating
13222  *            statistics.
13223  * balanced : The system is balanced for these 2 groups.
13224  * force :    Calculate the imbalance as load migration is probably needed.
13225  * avg_load : Only if imbalance is significant enough.
13226  * nr_idle :  dst_cpu is not busy and the number of idle CPUs is quite
13227  *            different in groups.
13228  */
13229 
13230 /**
13231  * sched_balance_find_src_group - Returns the busiest group within the sched_domain
13232  * if there is an imbalance.
13233  * @env: The load balancing environment.
13234  *
13235  * Also calculates the amount of runnable load which should be moved
13236  * to restore balance.
13237  *
13238  * Return:	- The busiest group if imbalance exists.
13239  */
13240 static struct sched_group *sched_balance_find_src_group(struct lb_env *env)
13241 {
13242 	struct sg_lb_stats *local, *busiest;
13243 	struct sd_lb_stats sds;
13244 
13245 	init_sd_lb_stats(&sds);
13246 
13247 	/*
13248 	 * Compute the various statistics relevant for load balancing at
13249 	 * this level.
13250 	 */
13251 	update_sd_lb_stats(env, &sds);
13252 
13253 	/* There is no busy sibling group to pull tasks from */
13254 	if (!sds.busiest)
13255 		goto out_balanced;
13256 
13257 	busiest = &sds.busiest_stat;
13258 
13259 	/* Misfit tasks should be dealt with regardless of the avg load */
13260 	if (busiest->group_type == group_misfit_task)
13261 		goto force_balance;
13262 
13263 	if (!is_rd_overutilized(env->dst_rq->rd) &&
13264 	    rcu_dereference_all(env->dst_rq->rd->pd))
13265 		goto out_balanced;
13266 
13267 	/* ASYM feature bypasses nice load balance check */
13268 	if (busiest->group_type == group_asym_packing)
13269 		goto force_balance;
13270 
13271 	/*
13272 	 * If the busiest group is imbalanced the below checks don't
13273 	 * work because they assume all things are equal, which typically
13274 	 * isn't true due to cpus_ptr constraints and the like.
13275 	 */
13276 	if (busiest->group_type == group_imbalanced)
13277 		goto force_balance;
13278 
13279 	local = &sds.local_stat;
13280 	/*
13281 	 * If the local group is busier than the selected busiest group
13282 	 * don't try and pull any tasks.
13283 	 */
13284 	if (local->group_type > busiest->group_type)
13285 		goto out_balanced;
13286 
13287 	/*
13288 	 * When groups are overloaded, use the avg_load to ensure fairness
13289 	 * between tasks.
13290 	 */
13291 	if (local->group_type == group_overloaded) {
13292 		/*
13293 		 * If the local group is more loaded than the selected
13294 		 * busiest group don't try to pull any tasks.
13295 		 */
13296 		if (local->avg_load >= busiest->avg_load)
13297 			goto out_balanced;
13298 
13299 		/* XXX broken for overlapping NUMA groups */
13300 		sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
13301 				sds.total_capacity;
13302 
13303 		/*
13304 		 * Don't pull any tasks if this group is already above the
13305 		 * domain average load.
13306 		 */
13307 		if (local->avg_load >= sds.avg_load)
13308 			goto out_balanced;
13309 
13310 		/*
13311 		 * If the busiest group is more loaded, use imbalance_pct to be
13312 		 * conservative.
13313 		 */
13314 		if (100 * busiest->avg_load <=
13315 				env->sd->imbalance_pct * local->avg_load)
13316 			goto out_balanced;
13317 	}
13318 
13319 	/*
13320 	 * Try to move all excess tasks to a sibling domain of the busiest
13321 	 * group's child domain.
13322 	 */
13323 	if (sds.prefer_sibling && local->group_type == group_has_spare &&
13324 	    (busiest->group_type == group_llc_balance ||
13325 	    sibling_imbalance(env, &sds, busiest, local) > 1))
13326 		goto force_balance;
13327 
13328 	if (busiest->group_type != group_overloaded) {
13329 		if (!env->idle) {
13330 			/*
13331 			 * If the busiest group is not overloaded (and as a
13332 			 * result the local one too) but this CPU is already
13333 			 * busy, let another idle CPU try to pull task.
13334 			 */
13335 			goto out_balanced;
13336 		}
13337 
13338 		if (busiest->group_type == group_smt_balance &&
13339 		    smt_vs_nonsmt_groups(sds.local, sds.busiest)) {
13340 			/* Let non SMT CPU pull from SMT CPU sharing with sibling */
13341 			goto force_balance;
13342 		}
13343 
13344 		if (busiest->group_weight > 1 &&
13345 		    local->idle_cpus <= (busiest->idle_cpus + 1)) {
13346 			/*
13347 			 * If the busiest group is not overloaded
13348 			 * and there is no imbalance between this and busiest
13349 			 * group wrt idle CPUs, it is balanced. The imbalance
13350 			 * becomes significant if the diff is greater than 1
13351 			 * otherwise we might end up to just move the imbalance
13352 			 * on another group. Of course this applies only if
13353 			 * there is more than 1 CPU per group.
13354 			 */
13355 			goto out_balanced;
13356 		}
13357 
13358 		if (busiest->sum_h_nr_running == 1) {
13359 			/*
13360 			 * busiest doesn't have any tasks waiting to run
13361 			 */
13362 			goto out_balanced;
13363 		}
13364 	}
13365 
13366 force_balance:
13367 	/* Looks like there is an imbalance. Compute it */
13368 	calculate_imbalance(env, &sds);
13369 	return env->imbalance ? sds.busiest : NULL;
13370 
13371 out_balanced:
13372 	env->imbalance = 0;
13373 	return NULL;
13374 }
13375 
13376 /*
13377  * sched_balance_find_src_rq - find the busiest runqueue among the CPUs in the group.
13378  */
13379 static struct rq *sched_balance_find_src_rq(struct lb_env *env,
13380 				     struct sched_group *group)
13381 {
13382 	struct rq *busiest = NULL, *rq;
13383 	unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
13384 	unsigned int __maybe_unused busiest_pref_llc = 0;
13385 	struct sched_domain __maybe_unused *sd_tmp;
13386 	unsigned int busiest_nr = 0;
13387 	int __maybe_unused dst_llc;
13388 	int i;
13389 
13390 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
13391 		unsigned long capacity, load, util;
13392 		unsigned int nr_running;
13393 		enum fbq_type rt;
13394 
13395 		rq = cpu_rq(i);
13396 		rt = fbq_classify_rq(rq);
13397 
13398 		/*
13399 		 * We classify groups/runqueues into three groups:
13400 		 *  - regular: there are !numa tasks
13401 		 *  - remote:  there are numa tasks that run on the 'wrong' node
13402 		 *  - all:     there is no distinction
13403 		 *
13404 		 * In order to avoid migrating ideally placed numa tasks,
13405 		 * ignore those when there's better options.
13406 		 *
13407 		 * If we ignore the actual busiest queue to migrate another
13408 		 * task, the next balance pass can still reduce the busiest
13409 		 * queue by moving tasks around inside the node.
13410 		 *
13411 		 * If we cannot move enough load due to this classification
13412 		 * the next pass will adjust the group classification and
13413 		 * allow migration of more tasks.
13414 		 *
13415 		 * Both cases only affect the total convergence complexity.
13416 		 */
13417 		if (rt > env->fbq_type)
13418 			continue;
13419 
13420 		nr_running = rq->cfs.h_nr_runnable;
13421 		if (!nr_running)
13422 			continue;
13423 
13424 		capacity = capacity_of(i);
13425 
13426 		/*
13427 		 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
13428 		 * eventually lead to active_balancing high->low capacity.
13429 		 * Higher per-CPU capacity is considered better than balancing
13430 		 * average load.
13431 		 */
13432 		if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
13433 		    nr_running == 1) {
13434 			bool cluster_equal_cap = static_branch_unlikely(&sched_cluster_active) &&
13435 						 (get_actual_cpu_capacity(env->dst_cpu) ==
13436 						  get_actual_cpu_capacity(i));
13437 			bool smt_degraded_cap = sched_smt_active() && !is_core_idle(i);
13438 
13439 			/*
13440 			 * Busy SMT siblings reduce the capacity of CPU @i. Do
13441 			 * not skip it in this case.
13442 			 *
13443 			 * CONFIG_SCHED_CLUSTER requires balancing load across
13444 			 * clusters of identical capacity, accounting for
13445 			 * hardware and cpufreq pressure.
13446 			 */
13447 			if (!smt_degraded_cap && !cluster_equal_cap &&
13448 			    !capacity_greater(capacity_of(env->dst_cpu), capacity))
13449 				continue;
13450 		}
13451 
13452 		/*
13453 		 * Make sure we only pull tasks from a CPU of lower priority
13454 		 * when balancing between SMT siblings.
13455 		 *
13456 		 * If balancing between cores, let lower priority CPUs help
13457 		 * SMT cores with more than one busy sibling.
13458 		 */
13459 		if (sched_asym(env->sd, i, env->dst_cpu) && nr_running == 1)
13460 			continue;
13461 
13462 		switch (env->migration_type) {
13463 		case migrate_load:
13464 			/*
13465 			 * When comparing with load imbalance, use cpu_load()
13466 			 * which is not scaled with the CPU capacity.
13467 			 */
13468 			load = cpu_load(rq);
13469 
13470 			if (nr_running == 1 && load > env->imbalance &&
13471 			    !check_cpu_capacity(rq, env->sd))
13472 				break;
13473 
13474 			/*
13475 			 * For the load comparisons with the other CPUs,
13476 			 * consider the cpu_load() scaled with the CPU
13477 			 * capacity, so that the load can be moved away
13478 			 * from the CPU that is potentially running at a
13479 			 * lower capacity.
13480 			 *
13481 			 * Thus we're looking for max(load_i / capacity_i),
13482 			 * crosswise multiplication to rid ourselves of the
13483 			 * division works out to:
13484 			 * load_i * capacity_j > load_j * capacity_i;
13485 			 * where j is our previous maximum.
13486 			 */
13487 			if (load * busiest_capacity > busiest_load * capacity) {
13488 				busiest_load = load;
13489 				busiest_capacity = capacity;
13490 				busiest = rq;
13491 			}
13492 			break;
13493 
13494 		case migrate_util:
13495 			util = cpu_util_cfs_boost(i);
13496 
13497 			/*
13498 			 * Don't try to pull utilization from a CPU with one
13499 			 * running task. Whatever its utilization, we will fail
13500 			 * detach the task.
13501 			 */
13502 			if (nr_running <= 1)
13503 				continue;
13504 
13505 			if (busiest_util < util) {
13506 				busiest_util = util;
13507 				busiest = rq;
13508 			}
13509 			break;
13510 
13511 		case migrate_task:
13512 			if (busiest_nr < nr_running) {
13513 				busiest_nr = nr_running;
13514 				busiest = rq;
13515 			}
13516 			break;
13517 
13518 		case migrate_misfit:
13519 			/*
13520 			 * For ASYM_CPUCAPACITY domains with misfit tasks we
13521 			 * simply seek the "biggest" misfit task.
13522 			 */
13523 			if (rq->misfit_task_load > busiest_load) {
13524 				busiest_load = rq->misfit_task_load;
13525 				busiest = rq;
13526 			}
13527 
13528 			break;
13529 
13530 		case migrate_llc_task:
13531 #ifdef CONFIG_SCHED_CACHE
13532 			sd_tmp = rcu_dereference_all(rq->sd);
13533 			dst_llc = llc_id(env->dst_cpu);
13534 
13535 			if (sd_tmp && (unsigned)dst_llc < sd_tmp->llc_max) {
13536 				unsigned int this_pref_llc =
13537 					sd_tmp->llc_counts[dst_llc];
13538 
13539 				if (busiest_pref_llc < this_pref_llc) {
13540 					busiest_pref_llc = this_pref_llc;
13541 					busiest = rq;
13542 				}
13543 			}
13544 #endif
13545 			break;
13546 
13547 		}
13548 	}
13549 
13550 	return busiest;
13551 }
13552 
13553 /*
13554  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
13555  * so long as it is large enough.
13556  */
13557 #define MAX_PINNED_INTERVAL	512
13558 
13559 static inline bool
13560 asym_active_balance(struct lb_env *env)
13561 {
13562 	/*
13563 	 * ASYM_PACKING needs to force migrate tasks from busy but lower
13564 	 * priority CPUs in order to pack all tasks in the highest priority
13565 	 * CPUs. When done between cores, do it only if the whole core if the
13566 	 * whole core is idle.
13567 	 *
13568 	 * If @env::src_cpu is an SMT core with busy siblings, let
13569 	 * the lower priority @env::dst_cpu help it. Do not follow
13570 	 * CPU priority.
13571 	 */
13572 	return env->idle && sched_use_asym_prio(env->sd, env->dst_cpu) &&
13573 	       (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
13574 		!sched_use_asym_prio(env->sd, env->src_cpu));
13575 }
13576 
13577 static inline bool
13578 imbalanced_active_balance(struct lb_env *env)
13579 {
13580 	struct sched_domain *sd = env->sd;
13581 
13582 	/*
13583 	 * The imbalanced case includes the case of pinned tasks preventing a fair
13584 	 * distribution of the load on the system but also the even distribution of the
13585 	 * threads on a system with spare capacity
13586 	 */
13587 	if ((env->migration_type == migrate_task) &&
13588 	    (sd->nr_balance_failed > sd->cache_nice_tries+2))
13589 		return 1;
13590 
13591 	return 0;
13592 }
13593 
13594 static int need_active_balance(struct lb_env *env)
13595 {
13596 	struct sched_domain *sd = env->sd;
13597 
13598 	if (asym_active_balance(env))
13599 		return 1;
13600 
13601 	if (alb_break_llc(env))
13602 		return 0;
13603 
13604 	if (imbalanced_active_balance(env))
13605 		return 1;
13606 
13607 	/*
13608 	 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
13609 	 * It's worth migrating the task if the src_cpu's capacity is reduced
13610 	 * because of other sched_class or IRQs if more capacity stays
13611 	 * available on dst_cpu.
13612 	 */
13613 	if (env->idle &&
13614 	    (env->src_rq->cfs.h_nr_runnable == 1)) {
13615 		if ((check_cpu_capacity(env->src_rq, sd)) &&
13616 		    (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
13617 			return 1;
13618 	}
13619 
13620 	if (env->migration_type == migrate_misfit ||
13621 	    env->migration_type == migrate_llc_task)
13622 		return 1;
13623 
13624 	return 0;
13625 }
13626 
13627 static int active_load_balance_cpu_stop(void *data);
13628 static int active_load_balance_llc_cpu_stop(void *data);
13629 
13630 /*
13631  * migration_type is checked elsewhere to decide migration policy, so
13632  * it shouldn't be repurposed just to flag an LLC-directed active
13633  * balance across the stopper. Pick the callback here instead.
13634  */
13635 static inline cpu_stop_fn_t alb_stop_fn(struct lb_env *env)
13636 {
13637 	if (env->migration_type == migrate_llc_task)
13638 		return active_load_balance_llc_cpu_stop;
13639 
13640 	return active_load_balance_cpu_stop;
13641 }
13642 
13643 static int should_we_balance(struct lb_env *env)
13644 {
13645 	struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask);
13646 	struct sched_group *sg = env->sd->groups;
13647 	int cpu, idle_smt = -1;
13648 
13649 	/*
13650 	 * Ensure the balancing environment is consistent; can happen
13651 	 * when the softirq triggers 'during' hotplug.
13652 	 */
13653 	if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
13654 		return 0;
13655 
13656 	/*
13657 	 * In the newly idle case, we will allow all the CPUs
13658 	 * to do the newly idle load balance.
13659 	 *
13660 	 * However, we bail out if we already have tasks or a wakeup pending,
13661 	 * to optimize wakeup latency.
13662 	 */
13663 	if (env->idle == CPU_NEWLY_IDLE) {
13664 		if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
13665 			return 0;
13666 		return 1;
13667 	}
13668 
13669 	cpumask_copy(swb_cpus, group_balance_mask(sg));
13670 	/* Try to find first idle CPU */
13671 	for_each_cpu_and(cpu, swb_cpus, env->cpus) {
13672 		if (!idle_cpu(cpu))
13673 			continue;
13674 
13675 		/*
13676 		 * Don't balance to idle SMT in busy core right away when
13677 		 * balancing cores, but remember the first idle SMT CPU for
13678 		 * later consideration.  Find CPU on an idle core first.
13679 		 */
13680 		if (sched_smt_active() &&
13681 		    !(env->sd->flags & SD_SHARE_CPUCAPACITY) &&
13682 		    !is_core_idle(cpu)) {
13683 			if (idle_smt == -1)
13684 				idle_smt = cpu;
13685 			/*
13686 			 * If the core is not idle, and first SMT sibling which is
13687 			 * idle has been found, then its not needed to check other
13688 			 * SMT siblings for idleness:
13689 			 */
13690 			cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu));
13691 			continue;
13692 		}
13693 
13694 		/*
13695 		 * Are we the first idle core in a non-SMT domain or higher,
13696 		 * or the first idle CPU in a SMT domain?
13697 		 */
13698 		return cpu == env->dst_cpu;
13699 	}
13700 
13701 	/* Are we the first idle CPU with busy siblings? */
13702 	if (idle_smt != -1)
13703 		return idle_smt == env->dst_cpu;
13704 
13705 	/* Are we the first CPU of this group ? */
13706 	return group_balance_cpu(sg) == env->dst_cpu;
13707 }
13708 
13709 static void update_lb_imbalance_stat(struct lb_env *env, struct sched_domain *sd,
13710 				     enum cpu_idle_type idle)
13711 {
13712 	if (!schedstat_enabled())
13713 		return;
13714 
13715 	switch (env->migration_type) {
13716 	case migrate_load:
13717 		__schedstat_add(sd->lb_imbalance_load[idle], env->imbalance);
13718 		break;
13719 	case migrate_util:
13720 		__schedstat_add(sd->lb_imbalance_util[idle], env->imbalance);
13721 		break;
13722 	case migrate_task:
13723 		__schedstat_add(sd->lb_imbalance_task[idle], env->imbalance);
13724 		break;
13725 	case migrate_misfit:
13726 		__schedstat_add(sd->lb_imbalance_misfit[idle], env->imbalance);
13727 		break;
13728 	case migrate_llc_task:
13729 		break;
13730 	}
13731 }
13732 
13733 /*
13734  * This flag serializes load-balancing passes over large domains
13735  * (above the NODE topology level) - only one load-balancing instance
13736  * may run at a time, to reduce overhead on very large systems with
13737  * lots of CPUs and large NUMA distances.
13738  *
13739  * - Note that load-balancing passes triggered while another one
13740  *   is executing are skipped and not re-tried.
13741  *
13742  * - Also note that this does not serialize rebalance_domains()
13743  *   execution, as non-SD_SERIALIZE domains will still be
13744  *   load-balanced in parallel.
13745  */
13746 static atomic_t sched_balance_running = ATOMIC_INIT(0);
13747 
13748 /*
13749  * Check this_cpu to ensure it is balanced within domain. Attempt to move
13750  * tasks if there is an imbalance.
13751  */
13752 static int sched_balance_rq(int this_cpu, struct rq *this_rq,
13753 			struct sched_domain *sd, enum cpu_idle_type idle,
13754 			int *continue_balancing)
13755 {
13756 	int ld_moved, cur_ld_moved, active_balance = 0;
13757 	struct sched_domain *sd_parent = sd->parent;
13758 	struct sched_group *group;
13759 	struct rq *busiest;
13760 	struct rq_flags rf;
13761 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
13762 	struct lb_env env = {
13763 		.sd		= sd,
13764 		.dst_cpu	= this_cpu,
13765 		.dst_rq		= this_rq,
13766 		.dst_grpmask    = group_balance_mask(sd->groups),
13767 		.idle		= idle,
13768 		.loop_break	= SCHED_NR_MIGRATE_BREAK,
13769 		.cpus		= cpus,
13770 		.fbq_type	= all,
13771 		.tasks		= LIST_HEAD_INIT(env.tasks),
13772 	};
13773 	bool need_unlock = false;
13774 
13775 	cpumask_and(cpus, sched_domain_span(sd), cpu_preferred_mask);
13776 
13777 	schedstat_inc(sd->lb_count[idle]);
13778 
13779 redo:
13780 	if (!should_we_balance(&env)) {
13781 		*continue_balancing = 0;
13782 		goto out_balanced;
13783 	}
13784 
13785 	if (!need_unlock && (sd->flags & SD_SERIALIZE)) {
13786 		int zero = 0;
13787 		if (!atomic_try_cmpxchg_acquire(&sched_balance_running, &zero, 1))
13788 			goto out_balanced;
13789 
13790 		need_unlock = true;
13791 	}
13792 
13793 	group = sched_balance_find_src_group(&env);
13794 	if (!group) {
13795 		schedstat_inc(sd->lb_nobusyg[idle]);
13796 		goto out_balanced;
13797 	}
13798 
13799 	busiest = sched_balance_find_src_rq(&env, group);
13800 	if (!busiest) {
13801 		schedstat_inc(sd->lb_nobusyq[idle]);
13802 		goto out_balanced;
13803 	}
13804 
13805 	WARN_ON_ONCE(busiest == env.dst_rq);
13806 
13807 	update_lb_imbalance_stat(&env, sd, idle);
13808 
13809 	env.src_cpu = busiest->cpu;
13810 	env.src_rq = busiest;
13811 
13812 	ld_moved = 0;
13813 	/* Clear this flag as soon as we find a pullable task */
13814 	env.flags |= LBF_ALL_PINNED;
13815 	if (busiest->nr_running > 1) {
13816 		/*
13817 		 * Attempt to move tasks. If sched_balance_find_src_group has found
13818 		 * an imbalance but busiest->nr_running <= 1, the group is
13819 		 * still unbalanced. ld_moved simply stays zero, so it is
13820 		 * correctly treated as an imbalance.
13821 		 */
13822 		env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
13823 
13824 more_balance:
13825 		rq_lock_irqsave(busiest, &rf);
13826 		update_rq_clock(busiest);
13827 
13828 		/*
13829 		 * cur_ld_moved - load moved in current iteration
13830 		 * ld_moved     - cumulative load moved across iterations
13831 		 */
13832 		cur_ld_moved = detach_tasks(&env);
13833 
13834 		/*
13835 		 * We've detached some tasks from busiest_rq. Every
13836 		 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
13837 		 * unlock busiest->lock, and we are able to be sure
13838 		 * that nobody can manipulate the tasks in parallel.
13839 		 * See task_rq_lock() family for the details.
13840 		 */
13841 
13842 		rq_unlock(busiest, &rf);
13843 
13844 		if (cur_ld_moved) {
13845 			attach_tasks(&env);
13846 			ld_moved += cur_ld_moved;
13847 		}
13848 
13849 		local_irq_restore(rf.flags);
13850 
13851 		if (env.flags & LBF_NEED_BREAK) {
13852 			env.flags &= ~LBF_NEED_BREAK;
13853 			goto more_balance;
13854 		}
13855 
13856 		/*
13857 		 * Revisit (affine) tasks on src_cpu that couldn't be moved to
13858 		 * us and move them to an alternate dst_cpu in our sched_group
13859 		 * where they can run. The upper limit on how many times we
13860 		 * iterate on same src_cpu is dependent on number of CPUs in our
13861 		 * sched_group.
13862 		 *
13863 		 * This changes load balance semantics a bit on who can move
13864 		 * load to a given_cpu. In addition to the given_cpu itself
13865 		 * (or a ilb_cpu acting on its behalf where given_cpu is
13866 		 * nohz-idle), we now have balance_cpu in a position to move
13867 		 * load to given_cpu. In rare situations, this may cause
13868 		 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
13869 		 * _independently_ and at _same_ time to move some load to
13870 		 * given_cpu) causing excess load to be moved to given_cpu.
13871 		 * This however should not happen so much in practice and
13872 		 * moreover subsequent load balance cycles should correct the
13873 		 * excess load moved.
13874 		 */
13875 		if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
13876 
13877 			/* Prevent to re-select dst_cpu via env's CPUs */
13878 			__cpumask_clear_cpu(env.dst_cpu, env.cpus);
13879 
13880 			env.dst_rq	 = cpu_rq(env.new_dst_cpu);
13881 			env.dst_cpu	 = env.new_dst_cpu;
13882 			env.flags	&= ~LBF_DST_PINNED;
13883 			env.loop	 = 0;
13884 			env.loop_break	 = SCHED_NR_MIGRATE_BREAK;
13885 
13886 			/*
13887 			 * Go back to "more_balance" rather than "redo" since we
13888 			 * need to continue with same src_cpu.
13889 			 */
13890 			goto more_balance;
13891 		}
13892 
13893 		/*
13894 		 * We failed to reach balance because of affinity.
13895 		 */
13896 		if (sd_parent) {
13897 			int *group_imbalance = &sd_parent->groups->sgc->imbalance;
13898 
13899 			if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
13900 				*group_imbalance = 1;
13901 		}
13902 
13903 		/* All tasks on this runqueue were pinned by CPU affinity */
13904 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
13905 			__cpumask_clear_cpu(cpu_of(busiest), cpus);
13906 			/*
13907 			 * Attempting to continue load balancing at the current
13908 			 * sched_domain level only makes sense if there are
13909 			 * active CPUs remaining as possible busiest CPUs to
13910 			 * pull load from which are not contained within the
13911 			 * destination group that is receiving any migrated
13912 			 * load.
13913 			 */
13914 			if (!cpumask_subset(cpus, env.dst_grpmask)) {
13915 				env.loop = 0;
13916 				env.loop_break = SCHED_NR_MIGRATE_BREAK;
13917 				goto redo;
13918 			}
13919 			goto out_all_pinned;
13920 		}
13921 	}
13922 
13923 	if (ld_moved) {
13924 		sd->nr_balance_failed = 0;
13925 		goto out_unbalanced;
13926 	}
13927 
13928 	schedstat_inc(sd->lb_failed[idle]);
13929 	/*
13930 	 * Increment the failure counter only on periodic balance.
13931 	 * We do not want newidle balance, which can be very
13932 	 * frequent, pollute the failure counter causing
13933 	 * excessive cache_hot migrations and active balances.
13934 	 *
13935 	 * Similarly for migration_misfit which is not related to
13936 	 * load/util migration, don't pollute nr_balance_failed.
13937 	 *
13938 	 * The same for cache aware scheduling's allowance for
13939 	 * load imbalance. If regular load balance does not
13940 	 * migrate task due to LLC locality, it is a expected
13941 	 * behavior and don't pollute nr_balance_failed.
13942 	 * See can_migrate_task().
13943 	 */
13944 	if (idle != CPU_NEWLY_IDLE &&
13945 	    env.migration_type != migrate_misfit &&
13946 	    !(env.flags & LBF_LLC_PINNED))
13947 		sd->nr_balance_failed++;
13948 
13949 	if (!need_active_balance(&env))
13950 		goto out_unbalanced;
13951 
13952 	scoped_guard (raw_spin_rq_lock_irqsave, busiest) {
13953 		/*
13954 		 * Don't kick the active_load_balance_cpu_stop,
13955 		 * if the curr task on busiest CPU can't be
13956 		 * moved to this_cpu:
13957 		 */
13958 		if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr))
13959 			goto out_one_pinned;
13960 
13961 		/* Record that we found at least one task that could run on this_cpu */
13962 		env.flags &= ~LBF_ALL_PINNED;
13963 
13964 		/*
13965 		 * ->active_balance synchronizes accesses to
13966 		 * ->active_balance_work.  Once set, it's cleared
13967 		 * only after active load balance is finished.
13968 		 */
13969 		if (busiest->active_balance)
13970 			goto out_unbalanced;
13971 
13972 		/*
13973 		 * @busiest dropped its rq_lock in the middle of
13974 		 * scheduling out its ->curr task (->on_rq := 0), no
13975 		 * need to forcefully punt it away with active balance.
13976 		 */
13977 		if (!busiest->curr->on_rq)
13978 			goto out_unbalanced;
13979 
13980 		busiest->active_balance = 1;
13981 		busiest->push_cpu = this_cpu;
13982 		active_balance = 1;
13983 		preempt_disable();
13984 	}
13985 	if (active_balance) {
13986 		stop_one_cpu_nowait(cpu_of(busiest),
13987 				    alb_stop_fn(&env), busiest,
13988 				    &busiest->active_balance_work);
13989 	}
13990 	preempt_enable();
13991 
13992 out_unbalanced:
13993 	/* We were unbalanced, so reset the balancing interval */
13994 	sd->balance_interval = sd->min_interval;
13995 	goto out;
13996 
13997 out_balanced:
13998 	/*
13999 	 * We reach balance although we may have faced some affinity
14000 	 * constraints. Clear the imbalance flag only if other tasks got
14001 	 * a chance to move and fix the imbalance.
14002 	 */
14003 	if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
14004 		int *group_imbalance = &sd_parent->groups->sgc->imbalance;
14005 
14006 		if (*group_imbalance)
14007 			*group_imbalance = 0;
14008 	}
14009 
14010 out_all_pinned:
14011 	/*
14012 	 * We reach balance because all tasks are pinned at this level so
14013 	 * we can't migrate them. Let the imbalance flag set so parent level
14014 	 * can try to migrate them.
14015 	 */
14016 	schedstat_inc(sd->lb_balanced[idle]);
14017 
14018 	sd->nr_balance_failed = 0;
14019 
14020 out_one_pinned:
14021 	ld_moved = 0;
14022 
14023 	/*
14024 	 * sched_balance_newidle() disregards balance intervals, so we could
14025 	 * repeatedly reach this code, which would lead to balance_interval
14026 	 * skyrocketing in a short amount of time. Skip the balance_interval
14027 	 * increase logic to avoid that.
14028 	 *
14029 	 * Similarly misfit migration which is not necessarily an indication of
14030 	 * the system being busy and requires lb to backoff to let it settle
14031 	 * down.
14032 	 */
14033 	if (env.idle == CPU_NEWLY_IDLE ||
14034 	    env.migration_type == migrate_misfit)
14035 		goto out;
14036 
14037 	/* tune up the balancing interval */
14038 	if ((env.flags & LBF_ALL_PINNED &&
14039 	     sd->balance_interval < MAX_PINNED_INTERVAL) ||
14040 	    sd->balance_interval < sd->max_interval)
14041 		sd->balance_interval *= 2;
14042 out:
14043 	if (need_unlock)
14044 		atomic_set_release(&sched_balance_running, 0);
14045 
14046 	return ld_moved;
14047 }
14048 
14049 static inline unsigned long
14050 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
14051 {
14052 	unsigned long interval = sd->balance_interval;
14053 
14054 	if (cpu_busy)
14055 		interval *= sd->busy_factor;
14056 
14057 	/* scale ms to jiffies */
14058 	interval = msecs_to_jiffies(interval);
14059 
14060 	/*
14061 	 * Reduce likelihood of busy balancing at higher domains racing with
14062 	 * balancing at lower domains by preventing their balancing periods
14063 	 * from being multiples of each other.
14064 	 */
14065 	if (cpu_busy)
14066 		interval -= 1;
14067 
14068 	interval = clamp(interval, 1UL, max_load_balance_interval);
14069 
14070 	return interval;
14071 }
14072 
14073 static inline void
14074 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
14075 {
14076 	unsigned long interval, next;
14077 
14078 	/* used by idle balance, so cpu_busy = 0 */
14079 	interval = get_sd_balance_interval(sd, 0);
14080 	next = sd->last_balance + interval;
14081 
14082 	if (time_after(*next_balance, next))
14083 		*next_balance = next;
14084 }
14085 
14086 /*
14087  * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
14088  * running tasks off the busiest CPU onto idle CPUs. It requires at
14089  * least 1 task to be running on each physical CPU where possible, and
14090  * avoids physical / logical imbalances.
14091  */
14092 static int __active_load_balance_cpu_stop(void *data, unsigned int lb_flags)
14093 {
14094 	struct rq *busiest_rq = data;
14095 	int busiest_cpu = cpu_of(busiest_rq);
14096 	int target_cpu = busiest_rq->push_cpu;
14097 	struct rq *target_rq = cpu_rq(target_cpu);
14098 	struct sched_domain *sd;
14099 	struct task_struct *p = NULL;
14100 	struct rq_flags rf;
14101 
14102 	rq_lock_irq(busiest_rq, &rf);
14103 	/*
14104 	 * Between queueing the stop-work and running it is a hole in which
14105 	 * CPUs can become inactive. We should not move tasks from or to
14106 	 * inactive CPUs.
14107 	 */
14108 	if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
14109 		goto out_unlock;
14110 
14111 	/* Make sure the requested CPU hasn't gone down in the meantime: */
14112 	if (unlikely(busiest_cpu != smp_processor_id() ||
14113 		     !busiest_rq->active_balance))
14114 		goto out_unlock;
14115 
14116 	/* Is there any task to move? */
14117 	if (busiest_rq->nr_running <= 1)
14118 		goto out_unlock;
14119 
14120 	/*
14121 	 * This condition is "impossible", if it occurs
14122 	 * we need to fix it. Originally reported by
14123 	 * Bjorn Helgaas on a 128-CPU setup.
14124 	 */
14125 	WARN_ON_ONCE(busiest_rq == target_rq);
14126 
14127 	/* Search for an sd spanning us and the target CPU. */
14128 	rcu_read_lock();
14129 	for_each_domain(target_cpu, sd) {
14130 		if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
14131 			break;
14132 	}
14133 
14134 	if (likely(sd)) {
14135 		struct lb_env env = {
14136 			.sd		= sd,
14137 			.dst_cpu	= target_cpu,
14138 			.dst_rq		= target_rq,
14139 			.src_cpu	= busiest_rq->cpu,
14140 			.src_rq		= busiest_rq,
14141 			.idle		= CPU_IDLE,
14142 			.flags		= LBF_ACTIVE_LB | lb_flags,
14143 		};
14144 
14145 		schedstat_inc(sd->alb_count);
14146 		update_rq_clock(busiest_rq);
14147 
14148 		p = detach_one_task(&env);
14149 		if (p) {
14150 			schedstat_inc(sd->alb_pushed);
14151 			/* Active balancing done, reset the failure counter. */
14152 			sd->nr_balance_failed = 0;
14153 		} else {
14154 			schedstat_inc(sd->alb_failed);
14155 		}
14156 	}
14157 	rcu_read_unlock();
14158 out_unlock:
14159 	busiest_rq->active_balance = 0;
14160 	rq_unlock(busiest_rq, &rf);
14161 
14162 	if (p)
14163 		attach_one_task(target_rq, p);
14164 
14165 	local_irq_enable();
14166 
14167 	return 0;
14168 }
14169 
14170 static int active_load_balance_cpu_stop(void *data)
14171 {
14172 	return __active_load_balance_cpu_stop(data, 0);
14173 }
14174 
14175 static int active_load_balance_llc_cpu_stop(void *data)
14176 {
14177 	return __active_load_balance_cpu_stop(data, LBF_ACTIVE_LB_LLC);
14178 }
14179 
14180 /*
14181  * Scale the max sched_balance_rq interval with the number of CPUs in the system.
14182  * This trades load-balance latency on larger machines for less cross talk.
14183  */
14184 void update_max_interval(void)
14185 {
14186 	max_load_balance_interval = HZ*num_online_cpus()/10;
14187 }
14188 
14189 static inline void update_newidle_stats(struct sched_domain *sd, unsigned int success)
14190 {
14191 	sd->newidle_call++;
14192 	sd->newidle_success += success;
14193 
14194 	if (sd->newidle_call >= 1024) {
14195 		u64 now = sched_clock();
14196 		s64 delta = now - sd->newidle_stamp;
14197 		sd->newidle_stamp = now;
14198 		int ratio = 0;
14199 
14200 		if (delta < 0)
14201 			delta = 0;
14202 
14203 		if (sched_feat(NI_RATE)) {
14204 			/*
14205 			 * ratio  delta   freq
14206 			 *
14207 			 * 1024 -  4  s -  128 Hz
14208 			 *  512 -  2  s -  256 Hz
14209 			 *  256 -  1  s -  512 Hz
14210 			 *  128 - .5  s - 1024 Hz
14211 			 *   64 - .25 s - 2048 Hz
14212 			 */
14213 			ratio = delta >> 22;
14214 		}
14215 
14216 		ratio += sd->newidle_success;
14217 
14218 		sd->newidle_ratio = min(1024, ratio);
14219 		sd->newidle_call /= 2;
14220 		sd->newidle_success /= 2;
14221 	}
14222 }
14223 
14224 static inline bool
14225 update_newidle_cost(struct sched_domain *sd, u64 cost, unsigned int success)
14226 {
14227 	unsigned long next_decay = sd->last_decay_max_lb_cost + HZ;
14228 	unsigned long now = jiffies;
14229 
14230 	if (cost)
14231 		update_newidle_stats(sd, success);
14232 
14233 	if (cost > sd->max_newidle_lb_cost) {
14234 		/*
14235 		 * Track max cost of a domain to make sure to not delay the
14236 		 * next wakeup on the CPU.
14237 		 */
14238 		sd->max_newidle_lb_cost = cost;
14239 		sd->last_decay_max_lb_cost = now;
14240 
14241 	} else if (time_after(now, next_decay)) {
14242 		/*
14243 		 * Decay the newidle max times by ~1% per second to ensure that
14244 		 * it is not outdated and the current max cost is actually
14245 		 * shorter.
14246 		 */
14247 		sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
14248 		sd->last_decay_max_lb_cost = now;
14249 		return true;
14250 	}
14251 
14252 	return false;
14253 }
14254 
14255 /*
14256  * It checks each scheduling domain to see if it is due to be balanced,
14257  * and initiates a balancing operation if so.
14258  *
14259  * Balancing parameters are set up in init_sched_domains.
14260  */
14261 static void sched_balance_domains(struct rq *rq, enum cpu_idle_type idle)
14262 {
14263 	int continue_balancing = 1;
14264 	int cpu = rq->cpu;
14265 	int busy = idle != CPU_IDLE && !sched_idle_rq(rq);
14266 	unsigned long interval;
14267 	struct sched_domain *sd;
14268 	/* Earliest time when we have to do rebalance again */
14269 	unsigned long next_balance = jiffies + 60*HZ;
14270 	int update_next_balance = 0;
14271 	int need_decay = 0;
14272 	u64 max_cost = 0;
14273 
14274 	rcu_read_lock();
14275 	for_each_domain(cpu, sd) {
14276 		/*
14277 		 * Decay the newidle max times here because this is a regular
14278 		 * visit to all the domains.
14279 		 */
14280 		need_decay = update_newidle_cost(sd, 0, 0);
14281 		max_cost += sd->max_newidle_lb_cost;
14282 
14283 		/*
14284 		 * Stop the load balance at this level. There is another
14285 		 * CPU in our sched group which is doing load balancing more
14286 		 * actively.
14287 		 */
14288 		if (!continue_balancing) {
14289 			if (need_decay)
14290 				continue;
14291 			break;
14292 		}
14293 
14294 		interval = get_sd_balance_interval(sd, busy);
14295 		if (time_after_eq(jiffies, sd->last_balance + interval)) {
14296 			if (sched_balance_rq(cpu, rq, sd, idle, &continue_balancing)) {
14297 				/*
14298 				 * The LBF_DST_PINNED logic could have changed
14299 				 * env->dst_cpu, so we can't know our idle
14300 				 * state even if we migrated tasks. Update it.
14301 				 */
14302 				idle = idle_cpu(cpu);
14303 				busy = !idle && !sched_idle_rq(rq);
14304 			}
14305 			sd->last_balance = jiffies;
14306 			interval = get_sd_balance_interval(sd, busy);
14307 		}
14308 		if (time_after(next_balance, sd->last_balance + interval)) {
14309 			next_balance = sd->last_balance + interval;
14310 			update_next_balance = 1;
14311 		}
14312 	}
14313 	if (need_decay) {
14314 		/*
14315 		 * Ensure the rq-wide value also decays but keep it at a
14316 		 * reasonable floor to avoid funnies with rq->avg_idle.
14317 		 */
14318 		rq->max_idle_balance_cost =
14319 			max((u64)sysctl_sched_migration_cost, max_cost);
14320 	}
14321 	rcu_read_unlock();
14322 
14323 	/*
14324 	 * next_balance will be updated only when there is a need.
14325 	 * When the cpu is attached to null domain for ex, it will not be
14326 	 * updated.
14327 	 */
14328 	if (likely(update_next_balance))
14329 		rq->next_balance = next_balance;
14330 
14331 }
14332 
14333 static inline int on_null_domain(struct rq *rq)
14334 {
14335 	return unlikely(!rcu_dereference_sched(rq->sd));
14336 }
14337 
14338 #ifdef CONFIG_NO_HZ_COMMON
14339 /*
14340  * NOHZ idle load balancing (ILB) details:
14341  *
14342  * - When one of the busy CPUs notices that there may be an idle rebalancing
14343  *   needed, they will kick the idle load balancer, which then does idle
14344  *   load balancing for all the idle CPUs.
14345  */
14346 static inline int find_new_ilb(void)
14347 {
14348 	struct cpumask *ilb_cpus;
14349 	int ilb_cpu, fallback = -1;
14350 
14351 	lockdep_assert_irqs_disabled();
14352 
14353 	/*
14354 	 * Reuse the per-CPU select_rq_mask, which is protected from concurrent
14355 	 * use on this CPU by having interrupts disabled.
14356 	 */
14357 	ilb_cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
14358 	cpumask_and(ilb_cpus, nohz.idle_cpus_mask,
14359 		    housekeeping_cpumask(HK_TYPE_KERNEL_NOISE));
14360 
14361 	for_each_cpu(ilb_cpu, ilb_cpus) {
14362 		if (!idle_cpu(ilb_cpu)) {
14363 			/*
14364 			 * Once an idle fallback exists, a busy CPU proves that
14365 			 * this core cannot be fully idle. Skip its siblings.
14366 			 */
14367 			if (sched_smt_active() && fallback >= 0)
14368 				cpumask_andnot(ilb_cpus, ilb_cpus, cpu_smt_mask(ilb_cpu));
14369 			continue;
14370 		}
14371 
14372 		/*
14373 		 * Running the idle load balancer on an idle sibling of a busy
14374 		 * SMT core can reduce the capacity available to its sibling. Prefer
14375 		 * a CPU whose entire core is idle, but retain the first idle CPU as
14376 		 * a fallback so idle balancing can still make progress when no fully
14377 		 * idle core exists.
14378 		 */
14379 		if (sched_smt_active() && !is_core_idle(ilb_cpu)) {
14380 			if (fallback < 0)
14381 				fallback = ilb_cpu;
14382 
14383 			/*
14384 			 * The core is not idle, so there is no need to check
14385 			 * any of its other SMT siblings.
14386 			 */
14387 			cpumask_andnot(ilb_cpus, ilb_cpus,
14388 				       cpu_smt_mask(ilb_cpu));
14389 			continue;
14390 		}
14391 
14392 		return ilb_cpu;
14393 	}
14394 
14395 	return fallback;
14396 }
14397 
14398 /*
14399  * Kick a CPU to do the NOHZ balancing, if it is time for it, via a cross-CPU
14400  * SMP function call (IPI).
14401  *
14402  * Prefer a CPU on a fully idle core in the HK_TYPE_KERNEL_NOISE housekeeping
14403  * set. Fall back to the first idle CPU when no fully idle core exists.
14404  */
14405 static void kick_ilb(unsigned int flags)
14406 {
14407 	int ilb_cpu;
14408 
14409 	/*
14410 	 * Increase nohz.next_balance only when if full ilb is triggered but
14411 	 * not if we only update stats.
14412 	 */
14413 	if (flags & NOHZ_BALANCE_KICK)
14414 		nohz.next_balance = jiffies+1;
14415 
14416 	ilb_cpu = find_new_ilb();
14417 	if (ilb_cpu < 0)
14418 		return;
14419 
14420 	/*
14421 	 * Don't bother if no new NOHZ balance work items for ilb_cpu,
14422 	 * i.e. all bits in flags are already set in ilb_cpu.
14423 	 */
14424 	if ((atomic_read(nohz_flags(ilb_cpu)) & flags) == flags)
14425 		return;
14426 
14427 	/*
14428 	 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
14429 	 * the first flag owns it; cleared by nohz_csd_func().
14430 	 */
14431 	flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
14432 	if (flags & NOHZ_KICK_MASK)
14433 		return;
14434 
14435 	/*
14436 	 * This way we generate an IPI on the target CPU which
14437 	 * is idle, and the softirq performing NOHZ idle load balancing
14438 	 * will be run before returning from the IPI.
14439 	 */
14440 	smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
14441 }
14442 
14443 /*
14444  * Current decision point for kicking the idle load balancer in the presence
14445  * of idle CPUs in the system.
14446  */
14447 static void nohz_balancer_kick(struct rq *rq)
14448 {
14449 	unsigned long now = jiffies;
14450 	struct sched_domain_shared *sds;
14451 	struct sched_domain *sd;
14452 	int nr_busy, i, cpu = rq->cpu;
14453 	unsigned int flags = 0;
14454 
14455 	if (unlikely(rq->idle_balance))
14456 		return;
14457 
14458 	/*
14459 	 * We may be recently in ticked or tickless idle mode. At the first
14460 	 * busy tick after returning from idle, we will update the busy stats.
14461 	 */
14462 	nohz_balance_exit_idle(rq);
14463 
14464 	if (READ_ONCE(nohz.has_blocked_load) &&
14465 	    time_after(now, READ_ONCE(nohz.next_blocked)))
14466 		flags = NOHZ_STATS_KICK;
14467 
14468 	/*
14469 	 * Most of the time system is not 100% busy. i.e nohz.nr_cpus > 0
14470 	 * Skip the read if time is not due.
14471 	 *
14472 	 * If none are in tickless mode, there maybe a narrow window
14473 	 * (28 jiffies, HZ=1000) where flags maybe set and kick_ilb called.
14474 	 * But idle load balancing is not done as find_new_ilb fails.
14475 	 * That's very rare. So read nohz.nr_cpus only if time is due.
14476 	 */
14477 	if (time_before(now, nohz.next_balance))
14478 		goto out;
14479 
14480 	/*
14481 	 * None are in tickless mode and hence no need for NOHZ idle load
14482 	 * balancing
14483 	 */
14484 	if (unlikely(cpumask_empty(nohz.idle_cpus_mask)))
14485 		return;
14486 
14487 	if (rq->nr_running >= 2) {
14488 		flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
14489 		goto out;
14490 	}
14491 
14492 	sd = rcu_dereference_all(rq->sd);
14493 	if (sd) {
14494 		/*
14495 		 * If there's a runnable CFS task and the current CPU has reduced
14496 		 * capacity, kick the ILB to see if there's a better CPU to run on:
14497 		 */
14498 		if (rq->cfs.h_nr_runnable >= 1 && check_cpu_capacity(rq, sd)) {
14499 			flags |= NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
14500 			goto out;
14501 		}
14502 	}
14503 
14504 	sd = rcu_dereference_all(per_cpu(sd_asym_packing, cpu));
14505 	if (sd) {
14506 		/*
14507 		 * When ASYM_PACKING; see if there's a more preferred CPU
14508 		 * currently idle; in which case, kick the ILB to move tasks
14509 		 * around.
14510 		 *
14511 		 * When balancing between cores, all the SMT siblings of the
14512 		 * preferred CPU must be idle.
14513 		 */
14514 		for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
14515 			if (sched_asym(sd, i, cpu)) {
14516 				flags |= NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
14517 				goto out;
14518 			}
14519 		}
14520 	}
14521 
14522 	sd = rcu_dereference_all(per_cpu(sd_asym_cpucapacity, cpu));
14523 	if (sd) {
14524 		/*
14525 		 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
14526 		 * to run the misfit task on.
14527 		 */
14528 		if (check_misfit_status(rq))
14529 			flags |= NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
14530 
14531 		/*
14532 		 * For asymmetric systems, we do not want to nicely balance
14533 		 * cache use, instead we want to embrace asymmetry and only
14534 		 * ensure tasks have enough CPU capacity.
14535 		 *
14536 		 * Skip the LLC logic because it's not relevant in that case.
14537 		 */
14538 		goto out;
14539 	}
14540 
14541 	sds = rcu_dereference_all(per_cpu(sd_balance_shared, cpu));
14542 	if (sds) {
14543 		/*
14544 		 * If there is an imbalance between LLC domains (IOW we could
14545 		 * increase the overall cache utilization), we need a less-loaded LLC
14546 		 * domain to pull some load from. Likewise, we may need to spread
14547 		 * load within the current LLC domain (e.g. packed SMT cores but
14548 		 * other CPUs are idle). We can't really know from here how busy
14549 		 * the others are - so just get a NOHZ balance going if it looks
14550 		 * like this LLC domain has tasks we could move.
14551 		 */
14552 		nr_busy = atomic_read(&sds->nr_busy_cpus);
14553 		if (nr_busy > 1)
14554 			flags |= NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
14555 	}
14556 out:
14557 	if (READ_ONCE(nohz.needs_update))
14558 		flags |= NOHZ_NEXT_KICK;
14559 
14560 	if (flags)
14561 		kick_ilb(flags);
14562 }
14563 
14564 static void set_cpu_sd_state_busy(int cpu)
14565 {
14566 	struct sched_domain *sd;
14567 	sd = rcu_dereference_all(per_cpu(sd_llc, cpu));
14568 
14569 	/*
14570 	 * sd->nohz_idle only pairs with nr_busy_cpus on sd->shared; if this
14571 	 * domain has no shared object there is nothing to clear or account.
14572 	 */
14573 	if (!sd || !sd->shared || !sd->nohz_idle)
14574 		return;
14575 	sd->nohz_idle = 0;
14576 
14577 	atomic_inc(&sd->shared->nr_busy_cpus);
14578 }
14579 
14580 void nohz_balance_exit_idle(struct rq *rq)
14581 {
14582 	WARN_ON_ONCE(rq != this_rq());
14583 
14584 	if (likely(!rq->nohz_tick_stopped))
14585 		return;
14586 
14587 	rq->nohz_tick_stopped = 0;
14588 	cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
14589 
14590 	set_cpu_sd_state_busy(rq->cpu);
14591 }
14592 
14593 static void set_cpu_sd_state_idle(int cpu)
14594 {
14595 	struct sched_domain *sd;
14596 	sd = rcu_dereference_all(per_cpu(sd_llc, cpu));
14597 
14598 	/* See set_cpu_sd_state_busy(): nohz_idle is only used with sd->shared. */
14599 	if (!sd || !sd->shared || sd->nohz_idle)
14600 		return;
14601 	sd->nohz_idle = 1;
14602 
14603 	atomic_dec(&sd->shared->nr_busy_cpus);
14604 }
14605 
14606 /*
14607  * This routine will record that the CPU is going idle with tick stopped.
14608  * This info will be used in performing idle load balancing in the future.
14609  */
14610 void nohz_balance_enter_idle(int cpu)
14611 {
14612 	struct rq *rq = cpu_rq(cpu);
14613 
14614 	WARN_ON_ONCE(cpu != smp_processor_id());
14615 
14616 	/* If this CPU is going down, then nothing needs to be done: */
14617 	if (!cpu_active(cpu))
14618 		return;
14619 
14620 	/*
14621 	 * Can be set safely without rq->lock held
14622 	 * If a clear happens, it will have evaluated last additions because
14623 	 * rq->lock is held during the check and the clear
14624 	 */
14625 	rq->has_blocked_load = 1;
14626 
14627 	/*
14628 	 * The tick is still stopped but load could have been added in the
14629 	 * meantime. We set the nohz.has_blocked_load flag to trig a check of the
14630 	 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
14631 	 * of nohz.has_blocked_load can only happen after checking the new load
14632 	 */
14633 	if (rq->nohz_tick_stopped)
14634 		goto out;
14635 
14636 	/* If we're a completely isolated CPU, we don't play: */
14637 	if (on_null_domain(rq))
14638 		return;
14639 
14640 	rq->nohz_tick_stopped = 1;
14641 
14642 	cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
14643 
14644 	/*
14645 	 * Ensures that if nohz_idle_balance() fails to observe our
14646 	 * @idle_cpus_mask store, it must observe the @has_blocked_load
14647 	 * and @needs_update stores.
14648 	 */
14649 	smp_mb__after_atomic();
14650 
14651 	set_cpu_sd_state_idle(cpu);
14652 
14653 	WRITE_ONCE(nohz.needs_update, 1);
14654 out:
14655 	/*
14656 	 * Each time a cpu enter idle, we assume that it has blocked load and
14657 	 * enable the periodic update of the load of idle CPUs
14658 	 */
14659 	WRITE_ONCE(nohz.has_blocked_load, 1);
14660 }
14661 
14662 static bool update_nohz_stats(struct rq *rq)
14663 {
14664 	unsigned int cpu = rq->cpu;
14665 
14666 	if (!rq->has_blocked_load)
14667 		return false;
14668 
14669 	if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
14670 		return false;
14671 
14672 	if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
14673 		return true;
14674 
14675 	sched_balance_update_blocked_averages(cpu);
14676 
14677 	return rq->has_blocked_load;
14678 }
14679 
14680 /*
14681  * Internal function that runs load balance for all idle CPUs. The load balance
14682  * can be a simple update of blocked load or a complete load balance with
14683  * tasks movement depending of flags.
14684  */
14685 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
14686 {
14687 	/* Earliest time when we have to do rebalance again */
14688 	unsigned long now = jiffies;
14689 	unsigned long next_balance = now + 60*HZ;
14690 	bool has_blocked_load = false;
14691 	int update_next_balance = 0;
14692 	int this_cpu = this_rq->cpu;
14693 	int balance_cpu;
14694 	struct rq *rq;
14695 
14696 	WARN_ON_ONCE((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
14697 
14698 	/*
14699 	 * We assume there will be no idle load after this update and clear
14700 	 * the has_blocked_load flag. If a cpu enters idle in the mean time, it will
14701 	 * set the has_blocked_load flag and trigger another update of idle load.
14702 	 * Because a cpu that becomes idle, is added to idle_cpus_mask before
14703 	 * setting the flag, we are sure to not clear the state and not
14704 	 * check the load of an idle cpu.
14705 	 *
14706 	 * Same applies to idle_cpus_mask vs needs_update.
14707 	 */
14708 	if (flags & NOHZ_STATS_KICK)
14709 		WRITE_ONCE(nohz.has_blocked_load, 0);
14710 	if (flags & NOHZ_NEXT_KICK)
14711 		WRITE_ONCE(nohz.needs_update, 0);
14712 
14713 	/*
14714 	 * Ensures that if we miss the CPU, we must see the has_blocked_load
14715 	 * store from nohz_balance_enter_idle().
14716 	 */
14717 	smp_mb();
14718 
14719 	/*
14720 	 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
14721 	 * chance for other idle cpu to pull load.
14722 	 */
14723 	for_each_cpu_wrap(balance_cpu,  nohz.idle_cpus_mask, this_cpu+1) {
14724 		if (!idle_cpu(balance_cpu))
14725 			continue;
14726 
14727 		/*
14728 		 * If this CPU gets work to do, stop the load balancing
14729 		 * work being done for other CPUs. Next load
14730 		 * balancing owner will pick it up.
14731 		 */
14732 		if (!idle_cpu(this_cpu) && need_resched()) {
14733 			if (flags & NOHZ_STATS_KICK)
14734 				has_blocked_load = true;
14735 			if (flags & NOHZ_NEXT_KICK)
14736 				WRITE_ONCE(nohz.needs_update, 1);
14737 			goto abort;
14738 		}
14739 
14740 		rq = cpu_rq(balance_cpu);
14741 
14742 		if (flags & NOHZ_STATS_KICK)
14743 			has_blocked_load |= update_nohz_stats(rq);
14744 
14745 		/*
14746 		 * If time for next balance is due,
14747 		 * do the balance.
14748 		 */
14749 		if (time_after_eq(jiffies, rq->next_balance)) {
14750 			struct rq_flags rf;
14751 
14752 			rq_lock_irqsave(rq, &rf);
14753 			update_rq_clock(rq);
14754 			rq_unlock_irqrestore(rq, &rf);
14755 
14756 			if (flags & NOHZ_BALANCE_KICK)
14757 				sched_balance_domains(rq, CPU_IDLE);
14758 		}
14759 
14760 		if (time_after(next_balance, rq->next_balance)) {
14761 			next_balance = rq->next_balance;
14762 			update_next_balance = 1;
14763 		}
14764 	}
14765 
14766 	/*
14767 	 * next_balance will be updated only when there is a need.
14768 	 * When the CPU is attached to null domain for ex, it will not be
14769 	 * updated.
14770 	 */
14771 	if (likely(update_next_balance))
14772 		nohz.next_balance = next_balance;
14773 
14774 	if (flags & NOHZ_STATS_KICK)
14775 		WRITE_ONCE(nohz.next_blocked,
14776 			   now + msecs_to_jiffies(LOAD_AVG_PERIOD));
14777 
14778 abort:
14779 	/* There is still blocked load, enable periodic update */
14780 	if (has_blocked_load)
14781 		WRITE_ONCE(nohz.has_blocked_load, 1);
14782 }
14783 
14784 /*
14785  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
14786  * rebalancing for all the CPUs for whom scheduler ticks are stopped.
14787  */
14788 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
14789 {
14790 	unsigned int flags = this_rq->nohz_idle_balance;
14791 
14792 	if (!flags)
14793 		return false;
14794 
14795 	this_rq->nohz_idle_balance = 0;
14796 
14797 	if (idle != CPU_IDLE)
14798 		return false;
14799 
14800 	_nohz_idle_balance(this_rq, flags);
14801 
14802 	return true;
14803 }
14804 
14805 /*
14806  * Check if we need to directly run the ILB for updating blocked load before
14807  * entering idle state. Here we run ILB directly without issuing IPIs.
14808  *
14809  * Note that when this function is called, the tick may not yet be stopped on
14810  * this CPU yet. nohz.idle_cpus_mask is updated only when tick is stopped and
14811  * cleared on the next busy tick. In other words, nohz.idle_cpus_mask updates
14812  * don't align with CPUs enter/exit idle to avoid bottlenecks due to high idle
14813  * entry/exit rate (usec). So it is possible that _nohz_idle_balance() is
14814  * called from this function on (this) CPU that's not yet in the mask. That's
14815  * OK because the goal of nohz_run_idle_balance() is to run ILB only for
14816  * updating the blocked load of already idle CPUs without waking up one of
14817  * those idle CPUs and outside the preempt disable / IRQ off phase of the local
14818  * cpu about to enter idle, because it can take a long time.
14819  */
14820 void nohz_run_idle_balance(int cpu)
14821 {
14822 	unsigned int flags;
14823 
14824 	flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
14825 
14826 	/*
14827 	 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
14828 	 * (i.e. NOHZ_STATS_KICK set) and will do the same.
14829 	 */
14830 	if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
14831 		_nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
14832 }
14833 
14834 static void nohz_newidle_balance(struct rq *this_rq)
14835 {
14836 	int this_cpu = this_rq->cpu;
14837 
14838 	/* Will wake up very soon. No time for doing anything else*/
14839 	if (this_rq->avg_idle < sysctl_sched_migration_cost)
14840 		return;
14841 
14842 	/* Don't need to update blocked load of idle CPUs*/
14843 	if (!READ_ONCE(nohz.has_blocked_load) ||
14844 	    time_before(jiffies, READ_ONCE(nohz.next_blocked)))
14845 		return;
14846 
14847 	/*
14848 	 * Set the need to trigger ILB in order to update blocked load
14849 	 * before entering idle state.
14850 	 */
14851 	atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
14852 }
14853 
14854 #else /* !CONFIG_NO_HZ_COMMON: */
14855 static inline void nohz_balancer_kick(struct rq *rq) { }
14856 
14857 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
14858 {
14859 	return false;
14860 }
14861 
14862 static inline void nohz_newidle_balance(struct rq *this_rq) { }
14863 #endif /* !CONFIG_NO_HZ_COMMON */
14864 
14865 /*
14866  * sched_balance_newidle is called by schedule() if this_cpu is about to become
14867  * idle. Attempts to pull tasks from other CPUs.
14868  *
14869  * Returns:
14870  *   < 0 - we released the lock and there are !fair tasks present
14871  *     0 - failed, no new tasks
14872  *   > 0 - success, new (fair) tasks present
14873  */
14874 static int sched_balance_newidle(struct rq *this_rq, struct rq_flags *rf)
14875 	__must_hold(__rq_lockp(this_rq))
14876 {
14877 	unsigned long next_balance = jiffies + HZ;
14878 	int this_cpu = this_rq->cpu;
14879 	int continue_balancing = 1;
14880 	u64 t0, t1, curr_cost = 0;
14881 	struct sched_domain *sd;
14882 	int pulled_task = 0;
14883 
14884 	update_misfit_status(NULL, this_rq);
14885 
14886 	/*
14887 	 * There is a task waiting to run. No need to search for one.
14888 	 * Return 0; the task will be enqueued when switching to idle.
14889 	 */
14890 	if (this_rq->ttwu_pending)
14891 		return 0;
14892 
14893 	/*
14894 	 * We must set idle_stamp _before_ calling sched_balance_rq()
14895 	 * for CPU_NEWLY_IDLE, such that we measure the this duration
14896 	 * as idle time.
14897 	 */
14898 	this_rq->idle_stamp = rq_clock(this_rq);
14899 
14900 	/* Do not pull tasks towards !preferred CPUs */
14901 	if (!cpu_preferred(this_cpu))
14902 		return 0;
14903 
14904 	/*
14905 	 * This is OK, because current is on_cpu, which avoids it being picked
14906 	 * for load-balance and preemption/IRQs are still disabled avoiding
14907 	 * further scheduler activity on it and we're being very careful to
14908 	 * re-start the picking loop.
14909 	 */
14910 	rq_unpin_lock(this_rq, rf);
14911 
14912 	sd = rcu_dereference_sched_domain(this_rq->sd);
14913 	if (!sd)
14914 		goto out;
14915 
14916 	if (!get_rd_overloaded(this_rq->rd) ||
14917 	    this_rq->avg_idle < sd->max_newidle_lb_cost) {
14918 
14919 		update_next_balance(sd, &next_balance);
14920 		goto out;
14921 	}
14922 
14923 	/*
14924 	 * Include sched_balance_update_blocked_averages() in the cost
14925 	 * calculation because it can be quite costly -- this ensures we skip
14926 	 * it when avg_idle gets to be very low.
14927 	 */
14928 	t0 = sched_clock_cpu(this_cpu);
14929 	__sched_balance_update_blocked_averages(this_rq);
14930 
14931 	rq_modified_begin(this_rq, &fair_sched_class);
14932 	raw_spin_rq_unlock(this_rq);
14933 
14934 	for_each_domain(this_cpu, sd) {
14935 		u64 domain_cost;
14936 
14937 		update_next_balance(sd, &next_balance);
14938 
14939 		if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
14940 			break;
14941 
14942 		if (sd->flags & SD_BALANCE_NEWIDLE) {
14943 			unsigned int weight = 1;
14944 
14945 			if (sched_feat(NI_RANDOM) && sd->newidle_ratio < 1024) {
14946 				/*
14947 				 * Throw a 1k sided dice; and only run
14948 				 * newidle_balance according to the success
14949 				 * rate.
14950 				 */
14951 				u32 d1k = sched_rng() % 1024;
14952 				weight = 1 + sd->newidle_ratio;
14953 				if (d1k > weight) {
14954 					update_newidle_stats(sd, 0);
14955 					continue;
14956 				}
14957 				weight = (1024 + weight/2) / weight;
14958 			}
14959 
14960 			pulled_task = sched_balance_rq(this_cpu, this_rq,
14961 						   sd, CPU_NEWLY_IDLE,
14962 						   &continue_balancing);
14963 
14964 			t1 = sched_clock_cpu(this_cpu);
14965 			domain_cost = t1 - t0;
14966 			curr_cost += domain_cost;
14967 			t0 = t1;
14968 
14969 			/*
14970 			 * Track max cost of a domain to make sure to not delay the
14971 			 * next wakeup on the CPU.
14972 			 */
14973 			update_newidle_cost(sd, domain_cost, weight * !!pulled_task);
14974 		}
14975 
14976 		/*
14977 		 * Stop searching for tasks to pull if there are
14978 		 * now runnable tasks on this rq.
14979 		 */
14980 		if (pulled_task || !continue_balancing)
14981 			break;
14982 	}
14983 
14984 	raw_spin_rq_lock(this_rq);
14985 
14986 	if (curr_cost > this_rq->max_idle_balance_cost)
14987 		this_rq->max_idle_balance_cost = curr_cost;
14988 
14989 	/*
14990 	 * While browsing the domains, we released the rq lock, a task could
14991 	 * have been enqueued in the meantime. Since we're not going idle,
14992 	 * pretend we pulled a task.
14993 	 */
14994 	if (this_rq->cfs.h_nr_queued && !pulled_task)
14995 		pulled_task = 1;
14996 
14997 	/* If a higher prio class was modified, restart the pick */
14998 	if (rq_modified_above(this_rq, &fair_sched_class))
14999 		pulled_task = -1;
15000 
15001 out:
15002 	/* Move the next balance forward */
15003 	if (time_after(this_rq->next_balance, next_balance))
15004 		this_rq->next_balance = next_balance;
15005 
15006 	if (pulled_task)
15007 		this_rq->idle_stamp = 0;
15008 	else
15009 		nohz_newidle_balance(this_rq);
15010 
15011 	rq_repin_lock(this_rq, rf);
15012 
15013 	return pulled_task;
15014 }
15015 
15016 /*
15017  * This softirq handler is triggered via SCHED_SOFTIRQ from two places:
15018  *
15019  * - directly from the local sched_tick() for periodic load balancing
15020  *
15021  * - indirectly from a remote sched_tick() for NOHZ idle balancing
15022  *   through the SMP cross-call nohz_csd_func()
15023  */
15024 static __latent_entropy void sched_balance_softirq(void)
15025 {
15026 	struct rq *this_rq = this_rq();
15027 	enum cpu_idle_type idle = this_rq->idle_balance;
15028 	/*
15029 	 * If this CPU has a pending NOHZ_BALANCE_KICK, then do the
15030 	 * balancing on behalf of the other idle CPUs whose ticks are
15031 	 * stopped. Do nohz_idle_balance *before* sched_balance_domains to
15032 	 * give the idle CPUs a chance to load balance. Else we may
15033 	 * load balance only within the local sched_domain hierarchy
15034 	 * and abort nohz_idle_balance altogether if we pull some load.
15035 	 */
15036 	if (nohz_idle_balance(this_rq, idle))
15037 		return;
15038 
15039 	/* normal load balance */
15040 	sched_balance_update_blocked_averages(this_rq->cpu);
15041 	sched_balance_domains(this_rq, idle);
15042 }
15043 
15044 /*
15045  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
15046  */
15047 void sched_balance_trigger(struct rq *rq)
15048 {
15049 	/*
15050 	 * Don't need to rebalance while attached to NULL domain or
15051 	 * runqueue CPU is not active
15052 	 */
15053 	if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
15054 		return;
15055 
15056 	if (time_after_eq(jiffies, rq->next_balance))
15057 		raise_softirq(SCHED_SOFTIRQ);
15058 
15059 	nohz_balancer_kick(rq);
15060 }
15061 
15062 static void rq_online_fair(struct rq *rq)
15063 {
15064 	update_sysctl();
15065 
15066 	update_runtime_enabled(rq);
15067 }
15068 
15069 static void rq_offline_fair(struct rq *rq)
15070 {
15071 	update_sysctl();
15072 
15073 	/* Ensure any throttled groups are reachable by pick_next_task */
15074 	unthrottle_offline_cfs_rqs(rq);
15075 
15076 	/* Ensure that we remove rq contribution to group share: */
15077 	clear_tg_offline_cfs_rqs(rq);
15078 }
15079 
15080 #ifdef CONFIG_SCHED_CORE
15081 static inline bool
15082 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
15083 {
15084 	u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
15085 	u64 slice = se->slice;
15086 
15087 	return (rtime * min_nr_tasks > slice);
15088 }
15089 
15090 #define MIN_NR_TASKS_DURING_FORCEIDLE	2
15091 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
15092 {
15093 	if (!sched_core_enabled(rq))
15094 		return;
15095 
15096 	/*
15097 	 * If runqueue has only one task which used up its slice and
15098 	 * if the sibling is forced idle, then trigger schedule to
15099 	 * give forced idle task a chance.
15100 	 *
15101 	 * __entity_slice_used() considers only this active rq and it gets the
15102 	 * whole slice. But during force idle, we have siblings acting
15103 	 * like a single runqueue and hence we need to consider runnable
15104 	 * tasks on this CPU and the forced idle CPU. Ideally, we should
15105 	 * go through the forced idle rq, but that would be a perf hit.
15106 	 * We can assume that the forced idle CPU has at least
15107 	 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
15108 	 * if we need to give up the CPU.
15109 	 */
15110 	if (rq->core->core_forceidle_count && rq->cfs.h_nr_queued == 1 &&
15111 	    __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
15112 		resched_curr(rq);
15113 }
15114 
15115 /*
15116  * Consider any infeasible weight scenario. Take for instance two tasks,
15117  * each bound to their respective sibling, one with weight 1 and one with
15118  * weight 2. Then the lower weight task will run ahead of the higher weight
15119  * task without bound.
15120  *
15121  * This utterly destroys the concept of a shared time base.
15122  *
15123  * Remember; all this is about a proportionally fair scheduling, where each
15124  * tasks receives:
15125  *
15126  *              w_i
15127  *   dt_i = ---------- dt                                     (1)
15128  *          \Sum_j w_j
15129  *
15130  * which we do by tracking a virtual time, s_i:
15131  *
15132  *          1
15133  *   s_i = --- d[t]_i                                         (2)
15134  *         w_i
15135  *
15136  * Where d[t] is a delta of discrete time, while dt is an infinitesimal.
15137  * The immediate corollary is that the ideal schedule S, where (2) to use
15138  * an infinitesimal delta, is:
15139  *
15140  *           1
15141  *   S = ---------- dt                                        (3)
15142  *       \Sum_i w_i
15143  *
15144  * From which we can define the lag, or deviation from the ideal, as:
15145  *
15146  *   lag(i) = S - s_i                                         (4)
15147  *
15148  * And since the one and only purpose is to approximate S, we get that:
15149  *
15150  *   \Sum_i w_i lag(i) := 0                                   (5)
15151  *
15152  * If this were not so, we no longer converge to S, and we can no longer
15153  * claim our scheduler has any of the properties we derive from S. This is
15154  * exactly what you did above, you broke it!
15155  *
15156  *
15157  * Let's continue for a while though; to see if there is anything useful to
15158  * be learned. We can combine (1)-(3) or (4)-(5) and express S in s_i:
15159  *
15160  *       \Sum_i w_i s_i
15161  *   S = --------------                                       (6)
15162  *         \Sum_i w_i
15163  *
15164  * Which gives us a way to compute S, given our s_i. Now, if you've read
15165  * our code, you know that we do not in fact do this, the reason for this
15166  * is two-fold. Firstly, computing S in that way requires a 64bit division
15167  * for every time we'd use it (see 12), and secondly, this only describes
15168  * the steady-state, it doesn't handle dynamics.
15169  *
15170  * Anyway, in (6):  s_i -> x + (s_i - x), to get:
15171  *
15172  *           \Sum_i w_i (s_i - x)
15173  *   S - x = --------------------                             (7)
15174  *              \Sum_i w_i
15175  *
15176  * Which shows that S and s_i transform alike (which makes perfect sense
15177  * given that S is basically the (weighted) average of s_i).
15178  *
15179  * So the thing to remember is that the above is strictly UP. It is
15180  * possible to generalize to multiple runqueues -- however it gets really
15181  * yuck when you have to add affinity support, as illustrated by our very
15182  * first counter-example.
15183  *
15184  * Luckily I think we can avoid needing a full multi-queue variant for
15185  * core-scheduling (or load-balancing). The crucial observation is that we
15186  * only actually need this comparison in the presence of forced-idle; only
15187  * then do we need to tell if the stalled rq has higher priority over the
15188  * other.
15189  *
15190  * [XXX assumes SMT2; better consider the more general case, I suspect
15191  * it'll work out because our comparison is always between 2 rqs and the
15192  * answer is only interesting if one of them is forced-idle]
15193  *
15194  * And (under assumption of SMT2) when there is forced-idle, there is only
15195  * a single queue, so everything works like normal.
15196  *
15197  * Let, for our runqueue 'k':
15198  *
15199  *   T_k = \Sum_i w_i s_i
15200  *   W_k = \Sum_i w_i      ; for all i of k                  (8)
15201  *
15202  * Then we can write (6) like:
15203  *
15204  *         T_k
15205  *   S_k = ---                                               (9)
15206  *         W_k
15207  *
15208  * From which immediately follows that:
15209  *
15210  *           T_k + T_l
15211  *   S_k+l = ---------                                       (10)
15212  *           W_k + W_l
15213  *
15214  * On which we can define a combined lag:
15215  *
15216  *   lag_k+l(i) := S_k+l - s_i                               (11)
15217  *
15218  * And that gives us the tools to compare tasks across a combined runqueue.
15219  *
15220  *
15221  * Combined this gives the following:
15222  *
15223  *  a) when a runqueue enters force-idle, sync it against it's sibling rq(s)
15224  *     using (7); this only requires storing single 'time'-stamps.
15225  *
15226  *  b) when comparing tasks between 2 runqueues of which one is forced-idle,
15227  *     compare the combined lag, per (11).
15228  *
15229  * Now, of course cgroups (I so hate them) make this more interesting in
15230  * that a) seems to suggest we need to iterate all cgroup on a CPU at such
15231  * boundaries, but I think we can avoid that. The force-idle is for the
15232  * whole CPU, all it's rqs. So we can mark it in the root and lazily
15233  * propagate downward on demand.
15234  */
15235 
15236 /*
15237  * So this sync is basically a relative reset of S to 0.
15238  *
15239  * So with 2 queues, when one goes idle, we drop them both to 0 and one
15240  * then increases due to not being idle, and the idle one builds up lag to
15241  * get re-elected. So far so simple, right?
15242  *
15243  * When there's 3, we can have the situation where 2 run and one is idle,
15244  * we sync to 0 and let the idle one build up lag to get re-election. Now
15245  * suppose another one also drops idle. At this point dropping all to 0
15246  * again would destroy the built-up lag from the queue that was already
15247  * idle, not good.
15248  *
15249  * So instead of syncing everything, we can:
15250  *
15251  *   less := !((s64)(s_a - s_b) <= 0)
15252  *
15253  *   (v_a - S_a) - (v_b - S_b) == v_a - v_b - S_a + S_b
15254  *                             == v_a - (v_b - S_a + S_b)
15255  *
15256  * IOW, we can recast the (lag) comparison to a one-sided difference.
15257  * So if then, instead of syncing the whole queue, sync the idle queue
15258  * against the active queue with S_a + S_b at the point where we sync.
15259  *
15260  * (XXX consider the implication of living in a cyclic group: N / 2^n N)
15261  *
15262  * This gives us means of syncing single queues against the active queue,
15263  * and for already idle queues to preserve their build-up lag.
15264  *
15265  * Of course, then we get the situation where there's 2 active and one
15266  * going idle, who do we pick to sync against? Theory would have us sync
15267  * against the combined S, but as we've already demonstrated, there is no
15268  * such thing in infeasible weight scenarios.
15269  *
15270  * One thing I've considered; and this is where that core_active rudiment
15271  * came from, is having active queues sync up between themselves after
15272  * every tick. This limits the observed divergence due to the work
15273  * conservancy.
15274  *
15275  * On top of that, we can improve upon things by employing (10) here.
15276  */
15277 
15278 /*
15279  * se_fi_update - Update the cfs_rq->zero_vruntime_fi in a CFS hierarchy if needed.
15280  */
15281 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
15282 			 bool forceidle)
15283 {
15284 	for_each_sched_entity(se) {
15285 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
15286 
15287 		if (forceidle) {
15288 			if (cfs_rq->forceidle_seq == fi_seq)
15289 				break;
15290 			cfs_rq->forceidle_seq = fi_seq;
15291 		}
15292 
15293 		cfs_rq->zero_vruntime_fi = cfs_rq->zero_vruntime;
15294 	}
15295 }
15296 
15297 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
15298 {
15299 	struct sched_entity *se = &p->se;
15300 
15301 	if (p->sched_class != &fair_sched_class)
15302 		return;
15303 
15304 	se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
15305 }
15306 
15307 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
15308 			bool in_fi)
15309 {
15310 	struct rq *rq = task_rq(a);
15311 	const struct sched_entity *sea = &a->se;
15312 	const struct sched_entity *seb = &b->se;
15313 	struct cfs_rq *cfs_rqa;
15314 	struct cfs_rq *cfs_rqb;
15315 	s64 delta;
15316 
15317 	WARN_ON_ONCE(task_rq(b)->core != rq->core);
15318 
15319 	cfs_rqa = &task_rq(a)->cfs;
15320 	cfs_rqb = &task_rq(b)->cfs;
15321 
15322 	/*
15323 	 * Find delta after normalizing se's vruntime with its cfs_rq's
15324 	 * zero_vruntime_fi, which would have been updated in prior calls
15325 	 * to se_fi_update().
15326 	 */
15327 	delta = vruntime_op(sea->vruntime, "-", seb->vruntime) +
15328 		vruntime_op(cfs_rqb->zero_vruntime_fi, "-", cfs_rqa->zero_vruntime_fi);
15329 
15330 	return delta > 0;
15331 }
15332 
15333 static int task_is_throttled_fair(struct task_struct *p, int cpu)
15334 {
15335 	struct cfs_rq *cfs_rq;
15336 
15337 #ifdef CONFIG_FAIR_GROUP_SCHED
15338 	cfs_rq = tg_cfs_rq(task_group(p), cpu);
15339 #else
15340 	cfs_rq = &cpu_rq(cpu)->cfs;
15341 #endif
15342 	return throttled_hierarchy(cfs_rq);
15343 }
15344 #else /* !CONFIG_SCHED_CORE: */
15345 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
15346 #endif /* !CONFIG_SCHED_CORE */
15347 
15348 /*
15349  * scheduler tick hitting a task of our scheduling class.
15350  *
15351  * NOTE: This function can be called remotely by the tick offload that
15352  * goes along full dynticks. Therefore no local assumption can be made
15353  * and everything must be accessed through the @rq and @curr passed in
15354  * parameters.
15355  */
15356 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
15357 {
15358 	struct sched_entity *se = &curr->se;
15359 
15360 	if (se->on_rq) {
15361 		unsigned long weight = NICE_0_LOAD;
15362 		struct cfs_rq *cfs_rq;
15363 
15364 		for_each_sched_entity(se) {
15365 			cfs_rq = cfs_rq_of(se);
15366 			entity_tick(cfs_rq, se, queued);
15367 
15368 			weight = __calc_prop_weight(cfs_rq, se, weight);
15369 		}
15370 
15371 		se = &curr->se;
15372 		reweight_eevdf(cfs_rq, se, weight, se->on_rq);
15373 	}
15374 
15375 	if (queued)
15376 		return;
15377 
15378 	if (static_branch_unlikely(&sched_numa_balancing))
15379 		task_tick_numa(rq, curr);
15380 
15381 	task_tick_cache(rq, curr);
15382 
15383 	update_misfit_status(curr, rq);
15384 	check_update_overutilized_status(task_rq(curr));
15385 
15386 	task_tick_core(rq, curr);
15387 }
15388 
15389 /*
15390  * called on fork with the child task as argument from the parent's context
15391  *  - child not yet on the tasklist
15392  *  - preemption disabled
15393  */
15394 static void task_fork_fair(struct task_struct *p)
15395 {
15396 	set_task_max_allowed_capacity(p);
15397 }
15398 
15399 /*
15400  * Priority of the task has changed. Check to see if we preempt
15401  * the current task.
15402  */
15403 static void
15404 prio_changed_fair(struct rq *rq, struct task_struct *p, u64 oldprio)
15405 {
15406 	if (!task_on_rq_queued(p))
15407 		return;
15408 
15409 	if (p->prio == oldprio)
15410 		return;
15411 
15412 	if (rq->cfs.h_nr_queued == 1)
15413 		return;
15414 
15415 	/*
15416 	 * Reschedule if we are currently running on this runqueue and
15417 	 * our priority decreased, or if we are not currently running on
15418 	 * this runqueue and our priority is higher than the current's
15419 	 */
15420 	if (task_current_donor(rq, p)) {
15421 		if (p->prio > oldprio)
15422 			resched_curr(rq);
15423 	} else {
15424 		wakeup_preempt(rq, p, 0);
15425 	}
15426 }
15427 
15428 #ifdef CONFIG_FAIR_GROUP_SCHED
15429 /*
15430  * Propagate the changes of the sched_entity across the tg tree to make it
15431  * visible to the root
15432  */
15433 static void propagate_entity_cfs_rq(struct sched_entity *se)
15434 {
15435 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
15436 
15437 	/*
15438 	 * If a task gets attached to this cfs_rq and before being queued,
15439 	 * it gets migrated to another CPU due to reasons like affinity
15440 	 * change, make sure this cfs_rq stays on leaf cfs_rq list to have
15441 	 * that removed load decayed or it can cause faireness problem.
15442 	 */
15443 	if (!cfs_rq_pelt_clock_throttled(cfs_rq))
15444 		list_add_leaf_cfs_rq(cfs_rq);
15445 
15446 	/* Start to propagate at parent */
15447 	se = se->parent;
15448 
15449 	for_each_sched_entity(se) {
15450 		cfs_rq = cfs_rq_of(se);
15451 
15452 		update_load_avg(cfs_rq, se, UPDATE_TG);
15453 
15454 		if (!cfs_rq_pelt_clock_throttled(cfs_rq))
15455 			list_add_leaf_cfs_rq(cfs_rq);
15456 	}
15457 
15458 	assert_list_leaf_cfs_rq(rq_of(cfs_rq));
15459 }
15460 #else /* !CONFIG_FAIR_GROUP_SCHED: */
15461 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
15462 #endif /* !CONFIG_FAIR_GROUP_SCHED */
15463 
15464 static void detach_entity_cfs_rq(struct sched_entity *se)
15465 {
15466 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
15467 
15468 	/*
15469 	 * In case the task sched_avg hasn't been attached:
15470 	 * - A forked task which hasn't been woken up by wake_up_new_task().
15471 	 * - A task which has been woken up by try_to_wake_up() but is
15472 	 *   waiting for actually being woken up by sched_ttwu_pending().
15473 	 */
15474 	if (!se->avg.last_update_time)
15475 		return;
15476 
15477 	/* Catch up with the cfs_rq and remove our load when we leave */
15478 	update_load_avg(cfs_rq, se, 0);
15479 	detach_entity_load_avg(cfs_rq, se);
15480 	update_tg_load_avg(cfs_rq);
15481 	propagate_entity_cfs_rq(se);
15482 }
15483 
15484 static void attach_entity_cfs_rq(struct sched_entity *se)
15485 {
15486 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
15487 
15488 	/* Synchronize entity with its cfs_rq */
15489 	update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
15490 	attach_entity_load_avg(cfs_rq, se);
15491 	update_tg_load_avg(cfs_rq);
15492 	propagate_entity_cfs_rq(se);
15493 }
15494 
15495 static void detach_task_cfs_rq(struct task_struct *p)
15496 {
15497 	struct sched_entity *se = &p->se;
15498 
15499 	detach_entity_cfs_rq(se);
15500 }
15501 
15502 static void attach_task_cfs_rq(struct task_struct *p)
15503 {
15504 	struct sched_entity *se = &p->se;
15505 
15506 	attach_entity_cfs_rq(se);
15507 }
15508 
15509 static void switching_from_fair(struct rq *rq, struct task_struct *p)
15510 {
15511 	if (p->se.sched_delayed)
15512 		dequeue_task(rq, p, DEQUEUE_SLEEP | DEQUEUE_DELAYED | DEQUEUE_NOCLOCK);
15513 }
15514 
15515 static void switched_from_fair(struct rq *rq, struct task_struct *p)
15516 {
15517 	detach_task_cfs_rq(p);
15518 }
15519 
15520 static void switched_to_fair(struct rq *rq, struct task_struct *p)
15521 {
15522 	WARN_ON_ONCE(p->se.sched_delayed);
15523 
15524 	attach_task_cfs_rq(p);
15525 
15526 	set_task_max_allowed_capacity(p);
15527 
15528 	if (task_on_rq_queued(p)) {
15529 		/*
15530 		 * We were most likely switched from sched_rt, so
15531 		 * kick off the schedule if running, otherwise just see
15532 		 * if we can still preempt the current task.
15533 		 */
15534 		if (task_current_donor(rq, p))
15535 			resched_curr(rq);
15536 		else
15537 			wakeup_preempt(rq, p, 0);
15538 	}
15539 }
15540 
15541 static void set_next_task_fair(struct rq *rq, struct task_struct *p, enum snt_e type)
15542 {
15543 	struct sched_entity *se = &p->se;
15544 	struct cfs_rq *cfs_rq = &rq->cfs;
15545 	unsigned long weight = NICE_0_LOAD;
15546 	bool first = type == SNT_PICK;
15547 	bool throttled = false;
15548 	bool on_rq = se->on_rq;
15549 
15550 	if (type == SNT_REPICK)
15551 		goto repick;
15552 
15553 	clear_buddies(cfs_rq, se);
15554 
15555 	if (on_rq)
15556 		__dequeue_entity(cfs_rq, se);
15557 
15558 	for_each_sched_entity(se) {
15559 		cfs_rq = cfs_rq_of(se);
15560 
15561 		if (!IS_ENABLED(CONFIG_FAIR_GROUP_SCHED) ||
15562 		    !first || !cfs_rq->h_curr)
15563 			set_next_entity(cfs_rq, se);
15564 
15565 		/* ensure bandwidth has been allocated on our new cfs_rq */
15566 		throttled |= account_cfs_rq_runtime(cfs_rq, 0);
15567 
15568 		if (on_rq)
15569 			weight = __calc_prop_weight(cfs_rq, se, weight);
15570 	}
15571 
15572 	if (throttled)
15573 		task_throttle_setup_work(p);
15574 
15575 	se = &p->se;
15576 	cfs_rq->curr = se;
15577 
15578 	if (on_rq) {
15579 		reweight_eevdf(cfs_rq, se, weight, se->on_rq);
15580 		if (first)
15581 			set_protect_slice(cfs_rq, se);
15582 	}
15583 
15584 	if (task_on_rq_queued(p)) {
15585 		/*
15586 		 * Move the next running task to the front of the list, so our
15587 		 * cfs_tasks list becomes MRU one.
15588 		 */
15589 		list_move(&se->group_node, &rq->cfs_tasks);
15590 	}
15591 	if (!first)
15592 		return;
15593 
15594 	WARN_ON_ONCE(se->sched_delayed);
15595 
15596 	update_misfit_status(p, rq);
15597 	sched_fair_update_stop_tick(rq, p);
15598 
15599 repick:
15600 	/*
15601 	 * A same-task repick skips put_prev_task_fair(), but
15602 	 * pick_task_fair() refreshed the entity hrtick_start_fair() reads
15603 	 * before selecting it again. rq->cfs.curr identifies that entity,
15604 	 * including with group scheduling.
15605 	 */
15606 	if (hrtick_enabled_fair(rq))
15607 		hrtick_start_fair(rq, p);
15608 }
15609 
15610 void init_cfs_rq(struct cfs_rq *cfs_rq)
15611 {
15612 	cfs_rq->tasks_timeline = RB_ROOT_CACHED;
15613 	cfs_rq->zero_vruntime = (u64)(-(1LL << 20));
15614 	raw_spin_lock_init(&cfs_rq->removed.lock);
15615 }
15616 
15617 #ifdef CONFIG_FAIR_GROUP_SCHED
15618 static void task_change_group_fair(struct task_struct *p)
15619 {
15620 	/*
15621 	 * We couldn't detach or attach a forked task which
15622 	 * hasn't been woken up by wake_up_new_task().
15623 	 */
15624 	if (READ_ONCE(p->__state) == TASK_NEW)
15625 		return;
15626 
15627 	detach_task_cfs_rq(p);
15628 
15629 	/* Tell se's cfs_rq has been changed -- migrated */
15630 	p->se.avg.last_update_time = 0;
15631 	set_task_rq(p, task_cpu(p));
15632 	attach_task_cfs_rq(p);
15633 }
15634 
15635 void free_fair_sched_group(struct task_group *tg)
15636 {
15637 	free_percpu(tg->cfs_rq);
15638 }
15639 
15640 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
15641 {
15642 	struct cfs_tg_state __percpu *state;
15643 	struct sched_entity *se;
15644 	struct cfs_rq *cfs_rq;
15645 	int i;
15646 
15647 	state = alloc_percpu_gfp(struct cfs_tg_state, GFP_KERNEL);
15648 	if (!state)
15649 		goto err;
15650 
15651 	tg->cfs_rq = &state->cfs_rq;
15652 	tg->shares = NICE_0_LOAD;
15653 
15654 	init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent));
15655 
15656 	for_each_possible_cpu(i) {
15657 		cfs_rq = tg_cfs_rq(tg, i);
15658 		if (!cfs_rq)
15659 			goto err;
15660 
15661 		se = tg_se(tg, i);
15662 		init_cfs_rq(cfs_rq);
15663 		init_tg_cfs_entry(tg, cfs_rq, se, i, tg_se(parent, i));
15664 		init_entity_runnable_average(se);
15665 	}
15666 
15667 	return 1;
15668 
15669 err:
15670 	return 0;
15671 }
15672 
15673 void online_fair_sched_group(struct task_group *tg)
15674 {
15675 	struct sched_entity *se;
15676 	struct rq_flags rf;
15677 	struct rq *rq;
15678 	int i;
15679 
15680 	for_each_possible_cpu(i) {
15681 		rq = cpu_rq(i);
15682 		se = tg_se(tg, i);
15683 		rq_lock_irq(rq, &rf);
15684 		update_rq_clock(rq);
15685 		attach_entity_cfs_rq(se);
15686 		sync_throttle(tg, i);
15687 		rq_unlock_irq(rq, &rf);
15688 	}
15689 }
15690 
15691 void unregister_fair_sched_group(struct task_group *tg)
15692 {
15693 	int cpu;
15694 
15695 	destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
15696 
15697 	for_each_possible_cpu(cpu) {
15698 		struct cfs_rq *cfs_rq = tg_cfs_rq(tg, cpu);
15699 		struct sched_entity *se = tg_se(tg, cpu);
15700 		struct rq *rq = cpu_rq(cpu);
15701 
15702 		if (se)
15703 			remove_entity_load_avg(se);
15704 
15705 		/*
15706 		 * Only empty task groups can be destroyed; so we can speculatively
15707 		 * check on_list without danger of it being re-added.
15708 		 */
15709 		if (cfs_rq->on_list) {
15710 			guard(rq_lock_irqsave)(rq);
15711 			list_del_leaf_cfs_rq(cfs_rq);
15712 		}
15713 	}
15714 }
15715 
15716 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
15717 			struct sched_entity *se, int cpu,
15718 			struct sched_entity *parent)
15719 {
15720 	struct rq *rq = cpu_rq(cpu);
15721 
15722 	cfs_rq->tg = tg;
15723 	cfs_rq->rq = rq;
15724 	init_cfs_rq_runtime(cfs_rq);
15725 
15726 	/* se could be NULL for root_task_group */
15727 	if (!se)
15728 		return;
15729 
15730 	if (!parent) {
15731 		se->cfs_rq = &rq->cfs;
15732 		se->depth = 0;
15733 	} else {
15734 		se->cfs_rq = parent->my_q;
15735 		se->depth = parent->depth + 1;
15736 	}
15737 
15738 	se->my_q = cfs_rq;
15739 	/* guarantee group entities always have weight */
15740 	update_load_set(&se->load, NICE_0_LOAD);
15741 	se->parent = parent;
15742 }
15743 
15744 static DEFINE_MUTEX(shares_mutex);
15745 
15746 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
15747 {
15748 	int i;
15749 
15750 	lockdep_assert_held(&shares_mutex);
15751 
15752 	/*
15753 	 * We can't change the weight of the root cgroup.
15754 	 */
15755 	if (is_root_task_group(tg))
15756 		return -EINVAL;
15757 
15758 	shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
15759 
15760 	if (tg->shares == shares)
15761 		return 0;
15762 
15763 	tg->shares = shares;
15764 	for_each_possible_cpu(i) {
15765 		struct rq *rq = cpu_rq(i);
15766 		struct sched_entity *se = tg_se(tg, i);
15767 		struct rq_flags rf;
15768 
15769 		/* Propagate contribution to hierarchy */
15770 		rq_lock_irqsave(rq, &rf);
15771 		update_rq_clock(rq);
15772 		for_each_sched_entity(se) {
15773 			update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
15774 			update_cfs_group(se);
15775 		}
15776 		rq_unlock_irqrestore(rq, &rf);
15777 	}
15778 
15779 	return 0;
15780 }
15781 
15782 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
15783 {
15784 	int ret;
15785 
15786 	mutex_lock(&shares_mutex);
15787 	if (tg_is_idle(tg))
15788 		ret = -EINVAL;
15789 	else
15790 		ret = __sched_group_set_shares(tg, shares);
15791 	mutex_unlock(&shares_mutex);
15792 
15793 	return ret;
15794 }
15795 
15796 int sched_group_set_idle(struct task_group *tg, long idle)
15797 {
15798 	int i;
15799 
15800 	if (tg == &root_task_group)
15801 		return -EINVAL;
15802 
15803 	if (idle < 0 || idle > 1)
15804 		return -EINVAL;
15805 
15806 	mutex_lock(&shares_mutex);
15807 
15808 	if (tg->idle == idle) {
15809 		mutex_unlock(&shares_mutex);
15810 		return 0;
15811 	}
15812 
15813 	tg->idle = idle;
15814 
15815 	for_each_possible_cpu(i) {
15816 		struct rq *rq = cpu_rq(i);
15817 		struct sched_entity *se = tg_se(tg, i);
15818 		struct cfs_rq *grp_cfs_rq = tg_cfs_rq(tg, i);
15819 		bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
15820 		long idle_task_delta;
15821 		struct rq_flags rf;
15822 
15823 		rq_lock_irqsave(rq, &rf);
15824 
15825 		grp_cfs_rq->idle = idle;
15826 		if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
15827 			goto next_cpu;
15828 
15829 		idle_task_delta = grp_cfs_rq->h_nr_queued -
15830 				  grp_cfs_rq->h_nr_idle;
15831 		if (!cfs_rq_is_idle(grp_cfs_rq))
15832 			idle_task_delta *= -1;
15833 
15834 		for_each_sched_entity(se) {
15835 			struct cfs_rq *cfs_rq = cfs_rq_of(se);
15836 
15837 			if (!se->on_rq)
15838 				break;
15839 
15840 			cfs_rq->h_nr_idle += idle_task_delta;
15841 
15842 			/* Already accounted at parent level and above. */
15843 			if (cfs_rq_is_idle(cfs_rq))
15844 				break;
15845 		}
15846 
15847 next_cpu:
15848 		rq_unlock_irqrestore(rq, &rf);
15849 	}
15850 
15851 	/* Idle groups have minimum weight. */
15852 	if (tg_is_idle(tg))
15853 		__sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
15854 	else
15855 		__sched_group_set_shares(tg, NICE_0_LOAD);
15856 
15857 	mutex_unlock(&shares_mutex);
15858 	return 0;
15859 }
15860 
15861 #endif /* CONFIG_FAIR_GROUP_SCHED */
15862 
15863 
15864 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
15865 {
15866 	struct sched_entity *se = &task->se;
15867 	unsigned int rr_interval = 0;
15868 
15869 	/*
15870 	 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
15871 	 * idle runqueue:
15872 	 */
15873 	if (rq->cfs.load.weight)
15874 		rr_interval = NS_TO_JIFFIES(se->slice);
15875 
15876 	return rr_interval;
15877 }
15878 
15879 /*
15880  * All the scheduling class methods:
15881  */
15882 DEFINE_SCHED_CLASS(fair) = {
15883 	.enqueue_task		= enqueue_task_fair,
15884 	.dequeue_task		= dequeue_task_fair,
15885 	.yield_task		= yield_task_fair,
15886 	.yield_to_task		= yield_to_task_fair,
15887 
15888 	.wakeup_preempt		= wakeup_preempt_fair,
15889 
15890 	.pick_task		= pick_task_fair,
15891 	.put_prev_task		= put_prev_task_fair,
15892 	.set_next_task          = set_next_task_fair,
15893 
15894 	.select_task_rq		= select_task_rq_fair,
15895 	.migrate_task_rq	= migrate_task_rq_fair,
15896 
15897 	.rq_online		= rq_online_fair,
15898 	.rq_offline		= rq_offline_fair,
15899 
15900 	.task_dead		= task_dead_fair,
15901 	.set_cpus_allowed	= set_cpus_allowed_fair,
15902 
15903 	.task_tick		= task_tick_fair,
15904 	.task_fork		= task_fork_fair,
15905 
15906 	.reweight_task		= reweight_task_fair,
15907 	.prio_changed		= prio_changed_fair,
15908 	.switching_from		= switching_from_fair,
15909 	.switched_from		= switched_from_fair,
15910 	.switched_to		= switched_to_fair,
15911 
15912 	.get_rr_interval	= get_rr_interval_fair,
15913 
15914 	.update_curr		= update_curr_fair,
15915 
15916 #ifdef CONFIG_FAIR_GROUP_SCHED
15917 	.task_change_group	= task_change_group_fair,
15918 #endif
15919 
15920 #ifdef CONFIG_SCHED_CORE
15921 	.task_is_throttled	= task_is_throttled_fair,
15922 #endif
15923 
15924 #ifdef CONFIG_UCLAMP_TASK
15925 	.uclamp_enabled		= 1,
15926 #endif
15927 };
15928 
15929 void print_cfs_stats(struct seq_file *m, int cpu)
15930 {
15931 	struct cfs_rq *cfs_rq, *pos;
15932 
15933 	rcu_read_lock();
15934 	for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
15935 		print_cfs_rq(m, cpu, cfs_rq);
15936 	rcu_read_unlock();
15937 }
15938 
15939 #ifdef CONFIG_NUMA_BALANCING
15940 void show_numa_stats(struct task_struct *p, struct seq_file *m)
15941 {
15942 	int node;
15943 	unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
15944 	struct numa_group *ng;
15945 
15946 	rcu_read_lock();
15947 	ng = rcu_dereference_all(p->numa_group);
15948 	for_each_online_node(node) {
15949 		if (p->numa_faults) {
15950 			tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
15951 			tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
15952 		}
15953 		if (ng) {
15954 			gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
15955 			gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
15956 		}
15957 		print_numa_stats(m, node, tsf, tpf, gsf, gpf);
15958 	}
15959 	rcu_read_unlock();
15960 }
15961 #endif /* CONFIG_NUMA_BALANCING */
15962 
15963 __init void init_sched_fair_class(void)
15964 {
15965 	int i;
15966 
15967 	for_each_possible_cpu(i) {
15968 		zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
15969 		zalloc_cpumask_var_node(&per_cpu(select_rq_mask,    i), GFP_KERNEL, cpu_to_node(i));
15970 		zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i),
15971 					GFP_KERNEL, cpu_to_node(i));
15972 
15973 #ifdef CONFIG_CFS_BANDWIDTH
15974 		INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
15975 		INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
15976 #endif
15977 	}
15978 
15979 	open_softirq(SCHED_SOFTIRQ, sched_balance_softirq);
15980 
15981 #ifdef CONFIG_NO_HZ_COMMON
15982 	nohz.next_balance = jiffies;
15983 	nohz.next_blocked = jiffies;
15984 	zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
15985 #endif
15986 }
15987