xref: /linux/kernel/sched/fair.c (revision c924c5e9b8c65b3a479a90e5e37d74cc8cd9fe0a)
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, Thomas Gleixner <tglx@linutronix.de>
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/hugetlb_inline.h>
26 #include <linux/jiffies.h>
27 #include <linux/mm_api.h>
28 #include <linux/highmem.h>
29 #include <linux/spinlock_api.h>
30 #include <linux/cpumask_api.h>
31 #include <linux/lockdep_api.h>
32 #include <linux/softirq.h>
33 #include <linux/refcount_api.h>
34 #include <linux/topology.h>
35 #include <linux/sched/clock.h>
36 #include <linux/sched/cond_resched.h>
37 #include <linux/sched/cputime.h>
38 #include <linux/sched/isolation.h>
39 #include <linux/sched/nohz.h>
40 #include <linux/sched/prio.h>
41 
42 #include <linux/cpuidle.h>
43 #include <linux/interrupt.h>
44 #include <linux/memory-tiers.h>
45 #include <linux/mempolicy.h>
46 #include <linux/mutex_api.h>
47 #include <linux/profile.h>
48 #include <linux/psi.h>
49 #include <linux/ratelimit.h>
50 #include <linux/task_work.h>
51 #include <linux/rbtree_augmented.h>
52 
53 #include <asm/switch_to.h>
54 
55 #include <uapi/linux/sched/types.h>
56 
57 #include "sched.h"
58 #include "stats.h"
59 #include "autogroup.h"
60 
61 /*
62  * The initial- and re-scaling of tunables is configurable
63  *
64  * Options are:
65  *
66  *   SCHED_TUNABLESCALING_NONE - unscaled, always *1
67  *   SCHED_TUNABLESCALING_LOG - scaled logarithmically, *1+ilog(ncpus)
68  *   SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
69  *
70  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
71  */
72 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
73 
74 /*
75  * Minimal preemption granularity for CPU-bound tasks:
76  *
77  * (default: 0.70 msec * (1 + ilog(ncpus)), units: nanoseconds)
78  */
79 unsigned int sysctl_sched_base_slice			= 700000ULL;
80 static unsigned int normalized_sysctl_sched_base_slice	= 700000ULL;
81 
82 __read_mostly unsigned int sysctl_sched_migration_cost	= 500000UL;
83 
setup_sched_thermal_decay_shift(char * str)84 static int __init setup_sched_thermal_decay_shift(char *str)
85 {
86 	pr_warn("Ignoring the deprecated sched_thermal_decay_shift= option\n");
87 	return 1;
88 }
89 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
90 
91 #ifdef CONFIG_SMP
92 /*
93  * For asym packing, by default the lower numbered CPU has higher priority.
94  */
arch_asym_cpu_priority(int cpu)95 int __weak arch_asym_cpu_priority(int cpu)
96 {
97 	return -cpu;
98 }
99 
100 /*
101  * The margin used when comparing utilization with CPU capacity.
102  *
103  * (default: ~20%)
104  */
105 #define fits_capacity(cap, max)	((cap) * 1280 < (max) * 1024)
106 
107 /*
108  * The margin used when comparing CPU capacities.
109  * is 'cap1' noticeably greater than 'cap2'
110  *
111  * (default: ~5%)
112  */
113 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
114 #endif
115 
116 #ifdef CONFIG_CFS_BANDWIDTH
117 /*
118  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
119  * each time a cfs_rq requests quota.
120  *
121  * Note: in the case that the slice exceeds the runtime remaining (either due
122  * to consumption or the quota being specified to be smaller than the slice)
123  * we will always only issue the remaining available time.
124  *
125  * (default: 5 msec, units: microseconds)
126  */
127 static unsigned int sysctl_sched_cfs_bandwidth_slice		= 5000UL;
128 #endif
129 
130 #ifdef CONFIG_NUMA_BALANCING
131 /* Restrict the NUMA promotion throughput (MB/s) for each target node. */
132 static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536;
133 #endif
134 
135 #ifdef CONFIG_SYSCTL
136 static const struct ctl_table sched_fair_sysctls[] = {
137 #ifdef CONFIG_CFS_BANDWIDTH
138 	{
139 		.procname       = "sched_cfs_bandwidth_slice_us",
140 		.data           = &sysctl_sched_cfs_bandwidth_slice,
141 		.maxlen         = sizeof(unsigned int),
142 		.mode           = 0644,
143 		.proc_handler   = proc_dointvec_minmax,
144 		.extra1         = SYSCTL_ONE,
145 	},
146 #endif
147 #ifdef CONFIG_NUMA_BALANCING
148 	{
149 		.procname	= "numa_balancing_promote_rate_limit_MBps",
150 		.data		= &sysctl_numa_balancing_promote_rate_limit,
151 		.maxlen		= sizeof(unsigned int),
152 		.mode		= 0644,
153 		.proc_handler	= proc_dointvec_minmax,
154 		.extra1		= SYSCTL_ZERO,
155 	},
156 #endif /* CONFIG_NUMA_BALANCING */
157 };
158 
sched_fair_sysctl_init(void)159 static int __init sched_fair_sysctl_init(void)
160 {
161 	register_sysctl_init("kernel", sched_fair_sysctls);
162 	return 0;
163 }
164 late_initcall(sched_fair_sysctl_init);
165 #endif
166 
update_load_add(struct load_weight * lw,unsigned long inc)167 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
168 {
169 	lw->weight += inc;
170 	lw->inv_weight = 0;
171 }
172 
update_load_sub(struct load_weight * lw,unsigned long dec)173 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
174 {
175 	lw->weight -= dec;
176 	lw->inv_weight = 0;
177 }
178 
update_load_set(struct load_weight * lw,unsigned long w)179 static inline void update_load_set(struct load_weight *lw, unsigned long w)
180 {
181 	lw->weight = w;
182 	lw->inv_weight = 0;
183 }
184 
185 /*
186  * Increase the granularity value when there are more CPUs,
187  * because with more CPUs the 'effective latency' as visible
188  * to users decreases. But the relationship is not linear,
189  * so pick a second-best guess by going with the log2 of the
190  * number of CPUs.
191  *
192  * This idea comes from the SD scheduler of Con Kolivas:
193  */
get_update_sysctl_factor(void)194 static unsigned int get_update_sysctl_factor(void)
195 {
196 	unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
197 	unsigned int factor;
198 
199 	switch (sysctl_sched_tunable_scaling) {
200 	case SCHED_TUNABLESCALING_NONE:
201 		factor = 1;
202 		break;
203 	case SCHED_TUNABLESCALING_LINEAR:
204 		factor = cpus;
205 		break;
206 	case SCHED_TUNABLESCALING_LOG:
207 	default:
208 		factor = 1 + ilog2(cpus);
209 		break;
210 	}
211 
212 	return factor;
213 }
214 
update_sysctl(void)215 static void update_sysctl(void)
216 {
217 	unsigned int factor = get_update_sysctl_factor();
218 
219 #define SET_SYSCTL(name) \
220 	(sysctl_##name = (factor) * normalized_sysctl_##name)
221 	SET_SYSCTL(sched_base_slice);
222 #undef SET_SYSCTL
223 }
224 
sched_init_granularity(void)225 void __init sched_init_granularity(void)
226 {
227 	update_sysctl();
228 }
229 
230 #define WMULT_CONST	(~0U)
231 #define WMULT_SHIFT	32
232 
__update_inv_weight(struct load_weight * lw)233 static void __update_inv_weight(struct load_weight *lw)
234 {
235 	unsigned long w;
236 
237 	if (likely(lw->inv_weight))
238 		return;
239 
240 	w = scale_load_down(lw->weight);
241 
242 	if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
243 		lw->inv_weight = 1;
244 	else if (unlikely(!w))
245 		lw->inv_weight = WMULT_CONST;
246 	else
247 		lw->inv_weight = WMULT_CONST / w;
248 }
249 
250 /*
251  * delta_exec * weight / lw.weight
252  *   OR
253  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
254  *
255  * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
256  * we're guaranteed shift stays positive because inv_weight is guaranteed to
257  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
258  *
259  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
260  * weight/lw.weight <= 1, and therefore our shift will also be positive.
261  */
__calc_delta(u64 delta_exec,unsigned long weight,struct load_weight * lw)262 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
263 {
264 	u64 fact = scale_load_down(weight);
265 	u32 fact_hi = (u32)(fact >> 32);
266 	int shift = WMULT_SHIFT;
267 	int fs;
268 
269 	__update_inv_weight(lw);
270 
271 	if (unlikely(fact_hi)) {
272 		fs = fls(fact_hi);
273 		shift -= fs;
274 		fact >>= fs;
275 	}
276 
277 	fact = mul_u32_u32(fact, lw->inv_weight);
278 
279 	fact_hi = (u32)(fact >> 32);
280 	if (fact_hi) {
281 		fs = fls(fact_hi);
282 		shift -= fs;
283 		fact >>= fs;
284 	}
285 
286 	return mul_u64_u32_shr(delta_exec, fact, shift);
287 }
288 
289 /*
290  * delta /= w
291  */
calc_delta_fair(u64 delta,struct sched_entity * se)292 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
293 {
294 	if (unlikely(se->load.weight != NICE_0_LOAD))
295 		delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
296 
297 	return delta;
298 }
299 
300 const struct sched_class fair_sched_class;
301 
302 /**************************************************************
303  * CFS operations on generic schedulable entities:
304  */
305 
306 #ifdef CONFIG_FAIR_GROUP_SCHED
307 
308 /* Walk up scheduling entities hierarchy */
309 #define for_each_sched_entity(se) \
310 		for (; se; se = se->parent)
311 
list_add_leaf_cfs_rq(struct cfs_rq * cfs_rq)312 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
313 {
314 	struct rq *rq = rq_of(cfs_rq);
315 	int cpu = cpu_of(rq);
316 
317 	if (cfs_rq->on_list)
318 		return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
319 
320 	cfs_rq->on_list = 1;
321 
322 	/*
323 	 * Ensure we either appear before our parent (if already
324 	 * enqueued) or force our parent to appear after us when it is
325 	 * enqueued. The fact that we always enqueue bottom-up
326 	 * reduces this to two cases and a special case for the root
327 	 * cfs_rq. Furthermore, it also means that we will always reset
328 	 * tmp_alone_branch either when the branch is connected
329 	 * to a tree or when we reach the top of the tree
330 	 */
331 	if (cfs_rq->tg->parent &&
332 	    cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
333 		/*
334 		 * If parent is already on the list, we add the child
335 		 * just before. Thanks to circular linked property of
336 		 * the list, this means to put the child at the tail
337 		 * of the list that starts by parent.
338 		 */
339 		list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
340 			&(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
341 		/*
342 		 * The branch is now connected to its tree so we can
343 		 * reset tmp_alone_branch to the beginning of the
344 		 * list.
345 		 */
346 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
347 		return true;
348 	}
349 
350 	if (!cfs_rq->tg->parent) {
351 		/*
352 		 * cfs rq without parent should be put
353 		 * at the tail of the list.
354 		 */
355 		list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
356 			&rq->leaf_cfs_rq_list);
357 		/*
358 		 * We have reach the top of a tree so we can reset
359 		 * tmp_alone_branch to the beginning of the list.
360 		 */
361 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
362 		return true;
363 	}
364 
365 	/*
366 	 * The parent has not already been added so we want to
367 	 * make sure that it will be put after us.
368 	 * tmp_alone_branch points to the begin of the branch
369 	 * where we will add parent.
370 	 */
371 	list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
372 	/*
373 	 * update tmp_alone_branch to points to the new begin
374 	 * of the branch
375 	 */
376 	rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
377 	return false;
378 }
379 
list_del_leaf_cfs_rq(struct cfs_rq * cfs_rq)380 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
381 {
382 	if (cfs_rq->on_list) {
383 		struct rq *rq = rq_of(cfs_rq);
384 
385 		/*
386 		 * With cfs_rq being unthrottled/throttled during an enqueue,
387 		 * it can happen the tmp_alone_branch points to the leaf that
388 		 * we finally want to delete. In this case, tmp_alone_branch moves
389 		 * to the prev element but it will point to rq->leaf_cfs_rq_list
390 		 * at the end of the enqueue.
391 		 */
392 		if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
393 			rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
394 
395 		list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
396 		cfs_rq->on_list = 0;
397 	}
398 }
399 
assert_list_leaf_cfs_rq(struct rq * rq)400 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
401 {
402 	WARN_ON_ONCE(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
403 }
404 
405 /* Iterate through all leaf cfs_rq's on a runqueue */
406 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)			\
407 	list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list,	\
408 				 leaf_cfs_rq_list)
409 
410 /* Do the two (enqueued) entities belong to the same group ? */
411 static inline struct cfs_rq *
is_same_group(struct sched_entity * se,struct sched_entity * pse)412 is_same_group(struct sched_entity *se, struct sched_entity *pse)
413 {
414 	if (se->cfs_rq == pse->cfs_rq)
415 		return se->cfs_rq;
416 
417 	return NULL;
418 }
419 
parent_entity(const struct sched_entity * se)420 static inline struct sched_entity *parent_entity(const struct sched_entity *se)
421 {
422 	return se->parent;
423 }
424 
425 static void
find_matching_se(struct sched_entity ** se,struct sched_entity ** pse)426 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
427 {
428 	int se_depth, pse_depth;
429 
430 	/*
431 	 * preemption test can be made between sibling entities who are in the
432 	 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
433 	 * both tasks until we find their ancestors who are siblings of common
434 	 * parent.
435 	 */
436 
437 	/* First walk up until both entities are at same depth */
438 	se_depth = (*se)->depth;
439 	pse_depth = (*pse)->depth;
440 
441 	while (se_depth > pse_depth) {
442 		se_depth--;
443 		*se = parent_entity(*se);
444 	}
445 
446 	while (pse_depth > se_depth) {
447 		pse_depth--;
448 		*pse = parent_entity(*pse);
449 	}
450 
451 	while (!is_same_group(*se, *pse)) {
452 		*se = parent_entity(*se);
453 		*pse = parent_entity(*pse);
454 	}
455 }
456 
tg_is_idle(struct task_group * tg)457 static int tg_is_idle(struct task_group *tg)
458 {
459 	return tg->idle > 0;
460 }
461 
cfs_rq_is_idle(struct cfs_rq * cfs_rq)462 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
463 {
464 	return cfs_rq->idle > 0;
465 }
466 
se_is_idle(struct sched_entity * se)467 static int se_is_idle(struct sched_entity *se)
468 {
469 	if (entity_is_task(se))
470 		return task_has_idle_policy(task_of(se));
471 	return cfs_rq_is_idle(group_cfs_rq(se));
472 }
473 
474 #else	/* !CONFIG_FAIR_GROUP_SCHED */
475 
476 #define for_each_sched_entity(se) \
477 		for (; se; se = NULL)
478 
list_add_leaf_cfs_rq(struct cfs_rq * cfs_rq)479 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
480 {
481 	return true;
482 }
483 
list_del_leaf_cfs_rq(struct cfs_rq * cfs_rq)484 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
485 {
486 }
487 
assert_list_leaf_cfs_rq(struct rq * rq)488 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
489 {
490 }
491 
492 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)	\
493 		for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
494 
parent_entity(struct sched_entity * se)495 static inline struct sched_entity *parent_entity(struct sched_entity *se)
496 {
497 	return NULL;
498 }
499 
500 static inline void
find_matching_se(struct sched_entity ** se,struct sched_entity ** pse)501 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
502 {
503 }
504 
tg_is_idle(struct task_group * tg)505 static inline int tg_is_idle(struct task_group *tg)
506 {
507 	return 0;
508 }
509 
cfs_rq_is_idle(struct cfs_rq * cfs_rq)510 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
511 {
512 	return 0;
513 }
514 
se_is_idle(struct sched_entity * se)515 static int se_is_idle(struct sched_entity *se)
516 {
517 	return task_has_idle_policy(task_of(se));
518 }
519 
520 #endif	/* CONFIG_FAIR_GROUP_SCHED */
521 
522 static __always_inline
523 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
524 
525 /**************************************************************
526  * Scheduling class tree data structure manipulation methods:
527  */
528 
max_vruntime(u64 max_vruntime,u64 vruntime)529 static inline __maybe_unused u64 max_vruntime(u64 max_vruntime, u64 vruntime)
530 {
531 	s64 delta = (s64)(vruntime - max_vruntime);
532 	if (delta > 0)
533 		max_vruntime = vruntime;
534 
535 	return max_vruntime;
536 }
537 
min_vruntime(u64 min_vruntime,u64 vruntime)538 static inline __maybe_unused u64 min_vruntime(u64 min_vruntime, u64 vruntime)
539 {
540 	s64 delta = (s64)(vruntime - min_vruntime);
541 	if (delta < 0)
542 		min_vruntime = vruntime;
543 
544 	return min_vruntime;
545 }
546 
entity_before(const struct sched_entity * a,const struct sched_entity * b)547 static inline bool entity_before(const struct sched_entity *a,
548 				 const struct sched_entity *b)
549 {
550 	/*
551 	 * Tiebreak on vruntime seems unnecessary since it can
552 	 * hardly happen.
553 	 */
554 	return (s64)(a->deadline - b->deadline) < 0;
555 }
556 
entity_key(struct cfs_rq * cfs_rq,struct sched_entity * se)557 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
558 {
559 	return (s64)(se->vruntime - cfs_rq->min_vruntime);
560 }
561 
562 #define __node_2_se(node) \
563 	rb_entry((node), struct sched_entity, run_node)
564 
565 /*
566  * Compute virtual time from the per-task service numbers:
567  *
568  * Fair schedulers conserve lag:
569  *
570  *   \Sum lag_i = 0
571  *
572  * Where lag_i is given by:
573  *
574  *   lag_i = S - s_i = w_i * (V - v_i)
575  *
576  * Where S is the ideal service time and V is it's virtual time counterpart.
577  * Therefore:
578  *
579  *   \Sum lag_i = 0
580  *   \Sum w_i * (V - v_i) = 0
581  *   \Sum w_i * V - w_i * v_i = 0
582  *
583  * From which we can solve an expression for V in v_i (which we have in
584  * se->vruntime):
585  *
586  *       \Sum v_i * w_i   \Sum v_i * w_i
587  *   V = -------------- = --------------
588  *          \Sum w_i            W
589  *
590  * Specifically, this is the weighted average of all entity virtual runtimes.
591  *
592  * [[ NOTE: this is only equal to the ideal scheduler under the condition
593  *          that join/leave operations happen at lag_i = 0, otherwise the
594  *          virtual time has non-contiguous motion equivalent to:
595  *
596  *	      V +-= lag_i / W
597  *
598  *	    Also see the comment in place_entity() that deals with this. ]]
599  *
600  * However, since v_i is u64, and the multiplication could easily overflow
601  * transform it into a relative form that uses smaller quantities:
602  *
603  * Substitute: v_i == (v_i - v0) + v0
604  *
605  *     \Sum ((v_i - v0) + v0) * w_i   \Sum (v_i - v0) * w_i
606  * V = ---------------------------- = --------------------- + v0
607  *                  W                            W
608  *
609  * Which we track using:
610  *
611  *                    v0 := cfs_rq->min_vruntime
612  * \Sum (v_i - v0) * w_i := cfs_rq->avg_vruntime
613  *              \Sum w_i := cfs_rq->avg_load
614  *
615  * Since min_vruntime is a monotonic increasing variable that closely tracks
616  * the per-task service, these deltas: (v_i - v), will be in the order of the
617  * maximal (virtual) lag induced in the system due to quantisation.
618  *
619  * Also, we use scale_load_down() to reduce the size.
620  *
621  * As measured, the max (key * weight) value was ~44 bits for a kernel build.
622  */
623 static void
avg_vruntime_add(struct cfs_rq * cfs_rq,struct sched_entity * se)624 avg_vruntime_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
625 {
626 	unsigned long weight = scale_load_down(se->load.weight);
627 	s64 key = entity_key(cfs_rq, se);
628 
629 	cfs_rq->avg_vruntime += key * weight;
630 	cfs_rq->avg_load += weight;
631 }
632 
633 static void
avg_vruntime_sub(struct cfs_rq * cfs_rq,struct sched_entity * se)634 avg_vruntime_sub(struct cfs_rq *cfs_rq, struct sched_entity *se)
635 {
636 	unsigned long weight = scale_load_down(se->load.weight);
637 	s64 key = entity_key(cfs_rq, se);
638 
639 	cfs_rq->avg_vruntime -= key * weight;
640 	cfs_rq->avg_load -= weight;
641 }
642 
643 static inline
avg_vruntime_update(struct cfs_rq * cfs_rq,s64 delta)644 void avg_vruntime_update(struct cfs_rq *cfs_rq, s64 delta)
645 {
646 	/*
647 	 * v' = v + d ==> avg_vruntime' = avg_runtime - d*avg_load
648 	 */
649 	cfs_rq->avg_vruntime -= cfs_rq->avg_load * delta;
650 }
651 
652 /*
653  * Specifically: avg_runtime() + 0 must result in entity_eligible() := true
654  * For this to be so, the result of this function must have a left bias.
655  */
avg_vruntime(struct cfs_rq * cfs_rq)656 u64 avg_vruntime(struct cfs_rq *cfs_rq)
657 {
658 	struct sched_entity *curr = cfs_rq->curr;
659 	s64 avg = cfs_rq->avg_vruntime;
660 	long load = cfs_rq->avg_load;
661 
662 	if (curr && curr->on_rq) {
663 		unsigned long weight = scale_load_down(curr->load.weight);
664 
665 		avg += entity_key(cfs_rq, curr) * weight;
666 		load += weight;
667 	}
668 
669 	if (load) {
670 		/* sign flips effective floor / ceiling */
671 		if (avg < 0)
672 			avg -= (load - 1);
673 		avg = div_s64(avg, load);
674 	}
675 
676 	return cfs_rq->min_vruntime + avg;
677 }
678 
679 /*
680  * lag_i = S - s_i = w_i * (V - v_i)
681  *
682  * However, since V is approximated by the weighted average of all entities it
683  * is possible -- by addition/removal/reweight to the tree -- to move V around
684  * and end up with a larger lag than we started with.
685  *
686  * Limit this to either double the slice length with a minimum of TICK_NSEC
687  * since that is the timing granularity.
688  *
689  * EEVDF gives the following limit for a steady state system:
690  *
691  *   -r_max < lag < max(r_max, q)
692  *
693  * XXX could add max_slice to the augmented data to track this.
694  */
update_entity_lag(struct cfs_rq * cfs_rq,struct sched_entity * se)695 static void update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
696 {
697 	s64 vlag, limit;
698 
699 	WARN_ON_ONCE(!se->on_rq);
700 
701 	vlag = avg_vruntime(cfs_rq) - se->vruntime;
702 	limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se);
703 
704 	se->vlag = clamp(vlag, -limit, limit);
705 }
706 
707 /*
708  * Entity is eligible once it received less service than it ought to have,
709  * eg. lag >= 0.
710  *
711  * lag_i = S - s_i = w_i*(V - v_i)
712  *
713  * lag_i >= 0 -> V >= v_i
714  *
715  *     \Sum (v_i - v)*w_i
716  * V = ------------------ + v
717  *          \Sum w_i
718  *
719  * lag_i >= 0 -> \Sum (v_i - v)*w_i >= (v_i - v)*(\Sum w_i)
720  *
721  * Note: using 'avg_vruntime() > se->vruntime' is inaccurate due
722  *       to the loss in precision caused by the division.
723  */
vruntime_eligible(struct cfs_rq * cfs_rq,u64 vruntime)724 static int vruntime_eligible(struct cfs_rq *cfs_rq, u64 vruntime)
725 {
726 	struct sched_entity *curr = cfs_rq->curr;
727 	s64 avg = cfs_rq->avg_vruntime;
728 	long load = cfs_rq->avg_load;
729 
730 	if (curr && curr->on_rq) {
731 		unsigned long weight = scale_load_down(curr->load.weight);
732 
733 		avg += entity_key(cfs_rq, curr) * weight;
734 		load += weight;
735 	}
736 
737 	return avg >= (s64)(vruntime - cfs_rq->min_vruntime) * load;
738 }
739 
entity_eligible(struct cfs_rq * cfs_rq,struct sched_entity * se)740 int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se)
741 {
742 	return vruntime_eligible(cfs_rq, se->vruntime);
743 }
744 
__update_min_vruntime(struct cfs_rq * cfs_rq,u64 vruntime)745 static u64 __update_min_vruntime(struct cfs_rq *cfs_rq, u64 vruntime)
746 {
747 	u64 min_vruntime = cfs_rq->min_vruntime;
748 	/*
749 	 * open coded max_vruntime() to allow updating avg_vruntime
750 	 */
751 	s64 delta = (s64)(vruntime - min_vruntime);
752 	if (delta > 0) {
753 		avg_vruntime_update(cfs_rq, delta);
754 		min_vruntime = vruntime;
755 	}
756 	return min_vruntime;
757 }
758 
update_min_vruntime(struct cfs_rq * cfs_rq)759 static void update_min_vruntime(struct cfs_rq *cfs_rq)
760 {
761 	struct sched_entity *se = __pick_root_entity(cfs_rq);
762 	struct sched_entity *curr = cfs_rq->curr;
763 	u64 vruntime = cfs_rq->min_vruntime;
764 
765 	if (curr) {
766 		if (curr->on_rq)
767 			vruntime = curr->vruntime;
768 		else
769 			curr = NULL;
770 	}
771 
772 	if (se) {
773 		if (!curr)
774 			vruntime = se->min_vruntime;
775 		else
776 			vruntime = min_vruntime(vruntime, se->min_vruntime);
777 	}
778 
779 	/* ensure we never gain time by being placed backwards. */
780 	cfs_rq->min_vruntime = __update_min_vruntime(cfs_rq, vruntime);
781 }
782 
cfs_rq_min_slice(struct cfs_rq * cfs_rq)783 static inline u64 cfs_rq_min_slice(struct cfs_rq *cfs_rq)
784 {
785 	struct sched_entity *root = __pick_root_entity(cfs_rq);
786 	struct sched_entity *curr = cfs_rq->curr;
787 	u64 min_slice = ~0ULL;
788 
789 	if (curr && curr->on_rq)
790 		min_slice = curr->slice;
791 
792 	if (root)
793 		min_slice = min(min_slice, root->min_slice);
794 
795 	return min_slice;
796 }
797 
__entity_less(struct rb_node * a,const struct rb_node * b)798 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
799 {
800 	return entity_before(__node_2_se(a), __node_2_se(b));
801 }
802 
803 #define vruntime_gt(field, lse, rse) ({ (s64)((lse)->field - (rse)->field) > 0; })
804 
__min_vruntime_update(struct sched_entity * se,struct rb_node * node)805 static inline void __min_vruntime_update(struct sched_entity *se, struct rb_node *node)
806 {
807 	if (node) {
808 		struct sched_entity *rse = __node_2_se(node);
809 		if (vruntime_gt(min_vruntime, se, rse))
810 			se->min_vruntime = rse->min_vruntime;
811 	}
812 }
813 
__min_slice_update(struct sched_entity * se,struct rb_node * node)814 static inline void __min_slice_update(struct sched_entity *se, struct rb_node *node)
815 {
816 	if (node) {
817 		struct sched_entity *rse = __node_2_se(node);
818 		if (rse->min_slice < se->min_slice)
819 			se->min_slice = rse->min_slice;
820 	}
821 }
822 
823 /*
824  * se->min_vruntime = min(se->vruntime, {left,right}->min_vruntime)
825  */
min_vruntime_update(struct sched_entity * se,bool exit)826 static inline bool min_vruntime_update(struct sched_entity *se, bool exit)
827 {
828 	u64 old_min_vruntime = se->min_vruntime;
829 	u64 old_min_slice = se->min_slice;
830 	struct rb_node *node = &se->run_node;
831 
832 	se->min_vruntime = se->vruntime;
833 	__min_vruntime_update(se, node->rb_right);
834 	__min_vruntime_update(se, node->rb_left);
835 
836 	se->min_slice = se->slice;
837 	__min_slice_update(se, node->rb_right);
838 	__min_slice_update(se, node->rb_left);
839 
840 	return se->min_vruntime == old_min_vruntime &&
841 	       se->min_slice == old_min_slice;
842 }
843 
844 RB_DECLARE_CALLBACKS(static, min_vruntime_cb, struct sched_entity,
845 		     run_node, min_vruntime, min_vruntime_update);
846 
847 /*
848  * Enqueue an entity into the rb-tree:
849  */
__enqueue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se)850 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
851 {
852 	avg_vruntime_add(cfs_rq, se);
853 	se->min_vruntime = se->vruntime;
854 	se->min_slice = se->slice;
855 	rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
856 				__entity_less, &min_vruntime_cb);
857 }
858 
__dequeue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se)859 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
860 {
861 	rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
862 				  &min_vruntime_cb);
863 	avg_vruntime_sub(cfs_rq, se);
864 }
865 
__pick_root_entity(struct cfs_rq * cfs_rq)866 struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq)
867 {
868 	struct rb_node *root = cfs_rq->tasks_timeline.rb_root.rb_node;
869 
870 	if (!root)
871 		return NULL;
872 
873 	return __node_2_se(root);
874 }
875 
__pick_first_entity(struct cfs_rq * cfs_rq)876 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
877 {
878 	struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
879 
880 	if (!left)
881 		return NULL;
882 
883 	return __node_2_se(left);
884 }
885 
886 /*
887  * HACK, stash a copy of deadline at the point of pick in vlag,
888  * which isn't used until dequeue.
889  */
set_protect_slice(struct sched_entity * se)890 static inline void set_protect_slice(struct sched_entity *se)
891 {
892 	se->vlag = se->deadline;
893 }
894 
protect_slice(struct sched_entity * se)895 static inline bool protect_slice(struct sched_entity *se)
896 {
897 	return se->vlag == se->deadline;
898 }
899 
cancel_protect_slice(struct sched_entity * se)900 static inline void cancel_protect_slice(struct sched_entity *se)
901 {
902 	if (protect_slice(se))
903 		se->vlag = se->deadline + 1;
904 }
905 
906 /*
907  * Earliest Eligible Virtual Deadline First
908  *
909  * In order to provide latency guarantees for different request sizes
910  * EEVDF selects the best runnable task from two criteria:
911  *
912  *  1) the task must be eligible (must be owed service)
913  *
914  *  2) from those tasks that meet 1), we select the one
915  *     with the earliest virtual deadline.
916  *
917  * We can do this in O(log n) time due to an augmented RB-tree. The
918  * tree keeps the entries sorted on deadline, but also functions as a
919  * heap based on the vruntime by keeping:
920  *
921  *  se->min_vruntime = min(se->vruntime, se->{left,right}->min_vruntime)
922  *
923  * Which allows tree pruning through eligibility.
924  */
pick_eevdf(struct cfs_rq * cfs_rq)925 static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq)
926 {
927 	struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node;
928 	struct sched_entity *se = __pick_first_entity(cfs_rq);
929 	struct sched_entity *curr = cfs_rq->curr;
930 	struct sched_entity *best = NULL;
931 
932 	/*
933 	 * We can safely skip eligibility check if there is only one entity
934 	 * in this cfs_rq, saving some cycles.
935 	 */
936 	if (cfs_rq->nr_queued == 1)
937 		return curr && curr->on_rq ? curr : se;
938 
939 	if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))
940 		curr = NULL;
941 
942 	if (sched_feat(RUN_TO_PARITY) && curr && protect_slice(curr))
943 		return curr;
944 
945 	/* Pick the leftmost entity if it's eligible */
946 	if (se && entity_eligible(cfs_rq, se)) {
947 		best = se;
948 		goto found;
949 	}
950 
951 	/* Heap search for the EEVD entity */
952 	while (node) {
953 		struct rb_node *left = node->rb_left;
954 
955 		/*
956 		 * Eligible entities in left subtree are always better
957 		 * choices, since they have earlier deadlines.
958 		 */
959 		if (left && vruntime_eligible(cfs_rq,
960 					__node_2_se(left)->min_vruntime)) {
961 			node = left;
962 			continue;
963 		}
964 
965 		se = __node_2_se(node);
966 
967 		/*
968 		 * The left subtree either is empty or has no eligible
969 		 * entity, so check the current node since it is the one
970 		 * with earliest deadline that might be eligible.
971 		 */
972 		if (entity_eligible(cfs_rq, se)) {
973 			best = se;
974 			break;
975 		}
976 
977 		node = node->rb_right;
978 	}
979 found:
980 	if (!best || (curr && entity_before(curr, best)))
981 		best = curr;
982 
983 	return best;
984 }
985 
__pick_last_entity(struct cfs_rq * cfs_rq)986 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
987 {
988 	struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
989 
990 	if (!last)
991 		return NULL;
992 
993 	return __node_2_se(last);
994 }
995 
996 /**************************************************************
997  * Scheduling class statistics methods:
998  */
999 #ifdef CONFIG_SMP
sched_update_scaling(void)1000 int sched_update_scaling(void)
1001 {
1002 	unsigned int factor = get_update_sysctl_factor();
1003 
1004 #define WRT_SYSCTL(name) \
1005 	(normalized_sysctl_##name = sysctl_##name / (factor))
1006 	WRT_SYSCTL(sched_base_slice);
1007 #undef WRT_SYSCTL
1008 
1009 	return 0;
1010 }
1011 #endif
1012 
1013 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se);
1014 
1015 /*
1016  * XXX: strictly: vd_i += N*r_i/w_i such that: vd_i > ve_i
1017  * this is probably good enough.
1018  */
update_deadline(struct cfs_rq * cfs_rq,struct sched_entity * se)1019 static bool update_deadline(struct cfs_rq *cfs_rq, struct sched_entity *se)
1020 {
1021 	if ((s64)(se->vruntime - se->deadline) < 0)
1022 		return false;
1023 
1024 	/*
1025 	 * For EEVDF the virtual time slope is determined by w_i (iow.
1026 	 * nice) while the request time r_i is determined by
1027 	 * sysctl_sched_base_slice.
1028 	 */
1029 	if (!se->custom_slice)
1030 		se->slice = sysctl_sched_base_slice;
1031 
1032 	/*
1033 	 * EEVDF: vd_i = ve_i + r_i / w_i
1034 	 */
1035 	se->deadline = se->vruntime + calc_delta_fair(se->slice, se);
1036 
1037 	/*
1038 	 * The task has consumed its request, reschedule.
1039 	 */
1040 	return true;
1041 }
1042 
1043 #include "pelt.h"
1044 #ifdef CONFIG_SMP
1045 
1046 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
1047 static unsigned long task_h_load(struct task_struct *p);
1048 static unsigned long capacity_of(int cpu);
1049 
1050 /* Give new sched_entity start runnable values to heavy its load in infant time */
init_entity_runnable_average(struct sched_entity * se)1051 void init_entity_runnable_average(struct sched_entity *se)
1052 {
1053 	struct sched_avg *sa = &se->avg;
1054 
1055 	memset(sa, 0, sizeof(*sa));
1056 
1057 	/*
1058 	 * Tasks are initialized with full load to be seen as heavy tasks until
1059 	 * they get a chance to stabilize to their real load level.
1060 	 * Group entities are initialized with zero load to reflect the fact that
1061 	 * nothing has been attached to the task group yet.
1062 	 */
1063 	if (entity_is_task(se))
1064 		sa->load_avg = scale_load_down(se->load.weight);
1065 
1066 	/* when this task is enqueued, it will contribute to its cfs_rq's load_avg */
1067 }
1068 
1069 /*
1070  * With new tasks being created, their initial util_avgs are extrapolated
1071  * based on the cfs_rq's current util_avg:
1072  *
1073  *   util_avg = cfs_rq->avg.util_avg / (cfs_rq->avg.load_avg + 1)
1074  *		* se_weight(se)
1075  *
1076  * However, in many cases, the above util_avg does not give a desired
1077  * value. Moreover, the sum of the util_avgs may be divergent, such
1078  * as when the series is a harmonic series.
1079  *
1080  * To solve this problem, we also cap the util_avg of successive tasks to
1081  * only 1/2 of the left utilization budget:
1082  *
1083  *   util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
1084  *
1085  * where n denotes the nth task and cpu_scale the CPU capacity.
1086  *
1087  * For example, for a CPU with 1024 of capacity, a simplest series from
1088  * the beginning would be like:
1089  *
1090  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
1091  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
1092  *
1093  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
1094  * if util_avg > util_avg_cap.
1095  */
post_init_entity_util_avg(struct task_struct * p)1096 void post_init_entity_util_avg(struct task_struct *p)
1097 {
1098 	struct sched_entity *se = &p->se;
1099 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
1100 	struct sched_avg *sa = &se->avg;
1101 	long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
1102 	long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
1103 
1104 	if (p->sched_class != &fair_sched_class) {
1105 		/*
1106 		 * For !fair tasks do:
1107 		 *
1108 		update_cfs_rq_load_avg(now, cfs_rq);
1109 		attach_entity_load_avg(cfs_rq, se);
1110 		switched_from_fair(rq, p);
1111 		 *
1112 		 * such that the next switched_to_fair() has the
1113 		 * expected state.
1114 		 */
1115 		se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
1116 		return;
1117 	}
1118 
1119 	if (cap > 0) {
1120 		if (cfs_rq->avg.util_avg != 0) {
1121 			sa->util_avg  = cfs_rq->avg.util_avg * se_weight(se);
1122 			sa->util_avg /= (cfs_rq->avg.load_avg + 1);
1123 
1124 			if (sa->util_avg > cap)
1125 				sa->util_avg = cap;
1126 		} else {
1127 			sa->util_avg = cap;
1128 		}
1129 	}
1130 
1131 	sa->runnable_avg = sa->util_avg;
1132 }
1133 
1134 #else /* !CONFIG_SMP */
init_entity_runnable_average(struct sched_entity * se)1135 void init_entity_runnable_average(struct sched_entity *se)
1136 {
1137 }
post_init_entity_util_avg(struct task_struct * p)1138 void post_init_entity_util_avg(struct task_struct *p)
1139 {
1140 }
update_tg_load_avg(struct cfs_rq * cfs_rq)1141 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
1142 {
1143 }
1144 #endif /* CONFIG_SMP */
1145 
update_curr_se(struct rq * rq,struct sched_entity * curr)1146 static s64 update_curr_se(struct rq *rq, struct sched_entity *curr)
1147 {
1148 	u64 now = rq_clock_task(rq);
1149 	s64 delta_exec;
1150 
1151 	delta_exec = now - curr->exec_start;
1152 	if (unlikely(delta_exec <= 0))
1153 		return delta_exec;
1154 
1155 	curr->exec_start = now;
1156 	curr->sum_exec_runtime += delta_exec;
1157 
1158 	if (schedstat_enabled()) {
1159 		struct sched_statistics *stats;
1160 
1161 		stats = __schedstats_from_se(curr);
1162 		__schedstat_set(stats->exec_max,
1163 				max(delta_exec, stats->exec_max));
1164 	}
1165 
1166 	return delta_exec;
1167 }
1168 
update_curr_task(struct task_struct * p,s64 delta_exec)1169 static inline void update_curr_task(struct task_struct *p, s64 delta_exec)
1170 {
1171 	trace_sched_stat_runtime(p, delta_exec);
1172 	account_group_exec_runtime(p, delta_exec);
1173 	cgroup_account_cputime(p, delta_exec);
1174 }
1175 
did_preempt_short(struct cfs_rq * cfs_rq,struct sched_entity * curr)1176 static inline bool did_preempt_short(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1177 {
1178 	if (!sched_feat(PREEMPT_SHORT))
1179 		return false;
1180 
1181 	if (curr->vlag == curr->deadline)
1182 		return false;
1183 
1184 	return !entity_eligible(cfs_rq, curr);
1185 }
1186 
do_preempt_short(struct cfs_rq * cfs_rq,struct sched_entity * pse,struct sched_entity * se)1187 static inline bool do_preempt_short(struct cfs_rq *cfs_rq,
1188 				    struct sched_entity *pse, struct sched_entity *se)
1189 {
1190 	if (!sched_feat(PREEMPT_SHORT))
1191 		return false;
1192 
1193 	if (pse->slice >= se->slice)
1194 		return false;
1195 
1196 	if (!entity_eligible(cfs_rq, pse))
1197 		return false;
1198 
1199 	if (entity_before(pse, se))
1200 		return true;
1201 
1202 	if (!entity_eligible(cfs_rq, se))
1203 		return true;
1204 
1205 	return false;
1206 }
1207 
1208 /*
1209  * Used by other classes to account runtime.
1210  */
update_curr_common(struct rq * rq)1211 s64 update_curr_common(struct rq *rq)
1212 {
1213 	struct task_struct *donor = rq->donor;
1214 	s64 delta_exec;
1215 
1216 	delta_exec = update_curr_se(rq, &donor->se);
1217 	if (likely(delta_exec > 0))
1218 		update_curr_task(donor, delta_exec);
1219 
1220 	return delta_exec;
1221 }
1222 
1223 /*
1224  * Update the current task's runtime statistics.
1225  */
update_curr(struct cfs_rq * cfs_rq)1226 static void update_curr(struct cfs_rq *cfs_rq)
1227 {
1228 	struct sched_entity *curr = cfs_rq->curr;
1229 	struct rq *rq = rq_of(cfs_rq);
1230 	s64 delta_exec;
1231 	bool resched;
1232 
1233 	if (unlikely(!curr))
1234 		return;
1235 
1236 	delta_exec = update_curr_se(rq, curr);
1237 	if (unlikely(delta_exec <= 0))
1238 		return;
1239 
1240 	curr->vruntime += calc_delta_fair(delta_exec, curr);
1241 	resched = update_deadline(cfs_rq, curr);
1242 	update_min_vruntime(cfs_rq);
1243 
1244 	if (entity_is_task(curr)) {
1245 		struct task_struct *p = task_of(curr);
1246 
1247 		update_curr_task(p, delta_exec);
1248 
1249 		/*
1250 		 * If the fair_server is active, we need to account for the
1251 		 * fair_server time whether or not the task is running on
1252 		 * behalf of fair_server or not:
1253 		 *  - If the task is running on behalf of fair_server, we need
1254 		 *    to limit its time based on the assigned runtime.
1255 		 *  - Fair task that runs outside of fair_server should account
1256 		 *    against fair_server such that it can account for this time
1257 		 *    and possibly avoid running this period.
1258 		 */
1259 		if (dl_server_active(&rq->fair_server))
1260 			dl_server_update(&rq->fair_server, delta_exec);
1261 	}
1262 
1263 	account_cfs_rq_runtime(cfs_rq, delta_exec);
1264 
1265 	if (cfs_rq->nr_queued == 1)
1266 		return;
1267 
1268 	if (resched || did_preempt_short(cfs_rq, curr)) {
1269 		resched_curr_lazy(rq);
1270 		clear_buddies(cfs_rq, curr);
1271 	}
1272 }
1273 
update_curr_fair(struct rq * rq)1274 static void update_curr_fair(struct rq *rq)
1275 {
1276 	update_curr(cfs_rq_of(&rq->donor->se));
1277 }
1278 
1279 static inline void
update_stats_wait_start_fair(struct cfs_rq * cfs_rq,struct sched_entity * se)1280 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1281 {
1282 	struct sched_statistics *stats;
1283 	struct task_struct *p = NULL;
1284 
1285 	if (!schedstat_enabled())
1286 		return;
1287 
1288 	stats = __schedstats_from_se(se);
1289 
1290 	if (entity_is_task(se))
1291 		p = task_of(se);
1292 
1293 	__update_stats_wait_start(rq_of(cfs_rq), p, stats);
1294 }
1295 
1296 static inline void
update_stats_wait_end_fair(struct cfs_rq * cfs_rq,struct sched_entity * se)1297 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1298 {
1299 	struct sched_statistics *stats;
1300 	struct task_struct *p = NULL;
1301 
1302 	if (!schedstat_enabled())
1303 		return;
1304 
1305 	stats = __schedstats_from_se(se);
1306 
1307 	/*
1308 	 * When the sched_schedstat changes from 0 to 1, some sched se
1309 	 * maybe already in the runqueue, the se->statistics.wait_start
1310 	 * will be 0.So it will let the delta wrong. We need to avoid this
1311 	 * scenario.
1312 	 */
1313 	if (unlikely(!schedstat_val(stats->wait_start)))
1314 		return;
1315 
1316 	if (entity_is_task(se))
1317 		p = task_of(se);
1318 
1319 	__update_stats_wait_end(rq_of(cfs_rq), p, stats);
1320 }
1321 
1322 static inline void
update_stats_enqueue_sleeper_fair(struct cfs_rq * cfs_rq,struct sched_entity * se)1323 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1324 {
1325 	struct sched_statistics *stats;
1326 	struct task_struct *tsk = NULL;
1327 
1328 	if (!schedstat_enabled())
1329 		return;
1330 
1331 	stats = __schedstats_from_se(se);
1332 
1333 	if (entity_is_task(se))
1334 		tsk = task_of(se);
1335 
1336 	__update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
1337 }
1338 
1339 /*
1340  * Task is being enqueued - update stats:
1341  */
1342 static inline void
update_stats_enqueue_fair(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)1343 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1344 {
1345 	if (!schedstat_enabled())
1346 		return;
1347 
1348 	/*
1349 	 * Are we enqueueing a waiting task? (for current tasks
1350 	 * a dequeue/enqueue event is a NOP)
1351 	 */
1352 	if (se != cfs_rq->curr)
1353 		update_stats_wait_start_fair(cfs_rq, se);
1354 
1355 	if (flags & ENQUEUE_WAKEUP)
1356 		update_stats_enqueue_sleeper_fair(cfs_rq, se);
1357 }
1358 
1359 static inline void
update_stats_dequeue_fair(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)1360 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1361 {
1362 
1363 	if (!schedstat_enabled())
1364 		return;
1365 
1366 	/*
1367 	 * Mark the end of the wait period if dequeueing a
1368 	 * waiting task:
1369 	 */
1370 	if (se != cfs_rq->curr)
1371 		update_stats_wait_end_fair(cfs_rq, se);
1372 
1373 	if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1374 		struct task_struct *tsk = task_of(se);
1375 		unsigned int state;
1376 
1377 		/* XXX racy against TTWU */
1378 		state = READ_ONCE(tsk->__state);
1379 		if (state & TASK_INTERRUPTIBLE)
1380 			__schedstat_set(tsk->stats.sleep_start,
1381 				      rq_clock(rq_of(cfs_rq)));
1382 		if (state & TASK_UNINTERRUPTIBLE)
1383 			__schedstat_set(tsk->stats.block_start,
1384 				      rq_clock(rq_of(cfs_rq)));
1385 	}
1386 }
1387 
1388 /*
1389  * We are picking a new current task - update its stats:
1390  */
1391 static inline void
update_stats_curr_start(struct cfs_rq * cfs_rq,struct sched_entity * se)1392 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1393 {
1394 	/*
1395 	 * We are starting a new run period:
1396 	 */
1397 	se->exec_start = rq_clock_task(rq_of(cfs_rq));
1398 }
1399 
1400 /**************************************************
1401  * Scheduling class queueing methods:
1402  */
1403 
is_core_idle(int cpu)1404 static inline bool is_core_idle(int cpu)
1405 {
1406 #ifdef CONFIG_SCHED_SMT
1407 	int sibling;
1408 
1409 	for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1410 		if (cpu == sibling)
1411 			continue;
1412 
1413 		if (!idle_cpu(sibling))
1414 			return false;
1415 	}
1416 #endif
1417 
1418 	return true;
1419 }
1420 
1421 #ifdef CONFIG_NUMA
1422 #define NUMA_IMBALANCE_MIN 2
1423 
1424 static inline long
adjust_numa_imbalance(int imbalance,int dst_running,int imb_numa_nr)1425 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1426 {
1427 	/*
1428 	 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1429 	 * threshold. Above this threshold, individual tasks may be contending
1430 	 * for both memory bandwidth and any shared HT resources.  This is an
1431 	 * approximation as the number of running tasks may not be related to
1432 	 * the number of busy CPUs due to sched_setaffinity.
1433 	 */
1434 	if (dst_running > imb_numa_nr)
1435 		return imbalance;
1436 
1437 	/*
1438 	 * Allow a small imbalance based on a simple pair of communicating
1439 	 * tasks that remain local when the destination is lightly loaded.
1440 	 */
1441 	if (imbalance <= NUMA_IMBALANCE_MIN)
1442 		return 0;
1443 
1444 	return imbalance;
1445 }
1446 #endif /* CONFIG_NUMA */
1447 
1448 #ifdef CONFIG_NUMA_BALANCING
1449 /*
1450  * Approximate time to scan a full NUMA task in ms. The task scan period is
1451  * calculated based on the tasks virtual memory size and
1452  * numa_balancing_scan_size.
1453  */
1454 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1455 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1456 
1457 /* Portion of address space to scan in MB */
1458 unsigned int sysctl_numa_balancing_scan_size = 256;
1459 
1460 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1461 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1462 
1463 /* The page with hint page fault latency < threshold in ms is considered hot */
1464 unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
1465 
1466 struct numa_group {
1467 	refcount_t refcount;
1468 
1469 	spinlock_t lock; /* nr_tasks, tasks */
1470 	int nr_tasks;
1471 	pid_t gid;
1472 	int active_nodes;
1473 
1474 	struct rcu_head rcu;
1475 	unsigned long total_faults;
1476 	unsigned long max_faults_cpu;
1477 	/*
1478 	 * faults[] array is split into two regions: faults_mem and faults_cpu.
1479 	 *
1480 	 * Faults_cpu is used to decide whether memory should move
1481 	 * towards the CPU. As a consequence, these stats are weighted
1482 	 * more by CPU use than by memory faults.
1483 	 */
1484 	unsigned long faults[];
1485 };
1486 
1487 /*
1488  * For functions that can be called in multiple contexts that permit reading
1489  * ->numa_group (see struct task_struct for locking rules).
1490  */
deref_task_numa_group(struct task_struct * p)1491 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1492 {
1493 	return rcu_dereference_check(p->numa_group, p == current ||
1494 		(lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1495 }
1496 
deref_curr_numa_group(struct task_struct * p)1497 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1498 {
1499 	return rcu_dereference_protected(p->numa_group, p == current);
1500 }
1501 
1502 static inline unsigned long group_faults_priv(struct numa_group *ng);
1503 static inline unsigned long group_faults_shared(struct numa_group *ng);
1504 
task_nr_scan_windows(struct task_struct * p)1505 static unsigned int task_nr_scan_windows(struct task_struct *p)
1506 {
1507 	unsigned long rss = 0;
1508 	unsigned long nr_scan_pages;
1509 
1510 	/*
1511 	 * Calculations based on RSS as non-present and empty pages are skipped
1512 	 * by the PTE scanner and NUMA hinting faults should be trapped based
1513 	 * on resident pages
1514 	 */
1515 	nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1516 	rss = get_mm_rss(p->mm);
1517 	if (!rss)
1518 		rss = nr_scan_pages;
1519 
1520 	rss = round_up(rss, nr_scan_pages);
1521 	return rss / nr_scan_pages;
1522 }
1523 
1524 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1525 #define MAX_SCAN_WINDOW 2560
1526 
task_scan_min(struct task_struct * p)1527 static unsigned int task_scan_min(struct task_struct *p)
1528 {
1529 	unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1530 	unsigned int scan, floor;
1531 	unsigned int windows = 1;
1532 
1533 	if (scan_size < MAX_SCAN_WINDOW)
1534 		windows = MAX_SCAN_WINDOW / scan_size;
1535 	floor = 1000 / windows;
1536 
1537 	scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1538 	return max_t(unsigned int, floor, scan);
1539 }
1540 
task_scan_start(struct task_struct * p)1541 static unsigned int task_scan_start(struct task_struct *p)
1542 {
1543 	unsigned long smin = task_scan_min(p);
1544 	unsigned long period = smin;
1545 	struct numa_group *ng;
1546 
1547 	/* Scale the maximum scan period with the amount of shared memory. */
1548 	rcu_read_lock();
1549 	ng = rcu_dereference(p->numa_group);
1550 	if (ng) {
1551 		unsigned long shared = group_faults_shared(ng);
1552 		unsigned long private = group_faults_priv(ng);
1553 
1554 		period *= refcount_read(&ng->refcount);
1555 		period *= shared + 1;
1556 		period /= private + shared + 1;
1557 	}
1558 	rcu_read_unlock();
1559 
1560 	return max(smin, period);
1561 }
1562 
task_scan_max(struct task_struct * p)1563 static unsigned int task_scan_max(struct task_struct *p)
1564 {
1565 	unsigned long smin = task_scan_min(p);
1566 	unsigned long smax;
1567 	struct numa_group *ng;
1568 
1569 	/* Watch for min being lower than max due to floor calculations */
1570 	smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1571 
1572 	/* Scale the maximum scan period with the amount of shared memory. */
1573 	ng = deref_curr_numa_group(p);
1574 	if (ng) {
1575 		unsigned long shared = group_faults_shared(ng);
1576 		unsigned long private = group_faults_priv(ng);
1577 		unsigned long period = smax;
1578 
1579 		period *= refcount_read(&ng->refcount);
1580 		period *= shared + 1;
1581 		period /= private + shared + 1;
1582 
1583 		smax = max(smax, period);
1584 	}
1585 
1586 	return max(smin, smax);
1587 }
1588 
account_numa_enqueue(struct rq * rq,struct task_struct * p)1589 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1590 {
1591 	rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1592 	rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1593 }
1594 
account_numa_dequeue(struct rq * rq,struct task_struct * p)1595 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1596 {
1597 	rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1598 	rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1599 }
1600 
1601 /* Shared or private faults. */
1602 #define NR_NUMA_HINT_FAULT_TYPES 2
1603 
1604 /* Memory and CPU locality */
1605 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1606 
1607 /* Averaged statistics, and temporary buffers. */
1608 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1609 
task_numa_group_id(struct task_struct * p)1610 pid_t task_numa_group_id(struct task_struct *p)
1611 {
1612 	struct numa_group *ng;
1613 	pid_t gid = 0;
1614 
1615 	rcu_read_lock();
1616 	ng = rcu_dereference(p->numa_group);
1617 	if (ng)
1618 		gid = ng->gid;
1619 	rcu_read_unlock();
1620 
1621 	return gid;
1622 }
1623 
1624 /*
1625  * The averaged statistics, shared & private, memory & CPU,
1626  * occupy the first half of the array. The second half of the
1627  * array is for current counters, which are averaged into the
1628  * first set by task_numa_placement.
1629  */
task_faults_idx(enum numa_faults_stats s,int nid,int priv)1630 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1631 {
1632 	return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1633 }
1634 
task_faults(struct task_struct * p,int nid)1635 static inline unsigned long task_faults(struct task_struct *p, int nid)
1636 {
1637 	if (!p->numa_faults)
1638 		return 0;
1639 
1640 	return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1641 		p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1642 }
1643 
group_faults(struct task_struct * p,int nid)1644 static inline unsigned long group_faults(struct task_struct *p, int nid)
1645 {
1646 	struct numa_group *ng = deref_task_numa_group(p);
1647 
1648 	if (!ng)
1649 		return 0;
1650 
1651 	return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1652 		ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1653 }
1654 
group_faults_cpu(struct numa_group * group,int nid)1655 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1656 {
1657 	return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1658 		group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1659 }
1660 
group_faults_priv(struct numa_group * ng)1661 static inline unsigned long group_faults_priv(struct numa_group *ng)
1662 {
1663 	unsigned long faults = 0;
1664 	int node;
1665 
1666 	for_each_online_node(node) {
1667 		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1668 	}
1669 
1670 	return faults;
1671 }
1672 
group_faults_shared(struct numa_group * ng)1673 static inline unsigned long group_faults_shared(struct numa_group *ng)
1674 {
1675 	unsigned long faults = 0;
1676 	int node;
1677 
1678 	for_each_online_node(node) {
1679 		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1680 	}
1681 
1682 	return faults;
1683 }
1684 
1685 /*
1686  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1687  * considered part of a numa group's pseudo-interleaving set. Migrations
1688  * between these nodes are slowed down, to allow things to settle down.
1689  */
1690 #define ACTIVE_NODE_FRACTION 3
1691 
numa_is_active_node(int nid,struct numa_group * ng)1692 static bool numa_is_active_node(int nid, struct numa_group *ng)
1693 {
1694 	return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1695 }
1696 
1697 /* Handle placement on systems where not all nodes are directly connected. */
score_nearby_nodes(struct task_struct * p,int nid,int lim_dist,bool task)1698 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1699 					int lim_dist, bool task)
1700 {
1701 	unsigned long score = 0;
1702 	int node, max_dist;
1703 
1704 	/*
1705 	 * All nodes are directly connected, and the same distance
1706 	 * from each other. No need for fancy placement algorithms.
1707 	 */
1708 	if (sched_numa_topology_type == NUMA_DIRECT)
1709 		return 0;
1710 
1711 	/* sched_max_numa_distance may be changed in parallel. */
1712 	max_dist = READ_ONCE(sched_max_numa_distance);
1713 	/*
1714 	 * This code is called for each node, introducing N^2 complexity,
1715 	 * which should be OK given the number of nodes rarely exceeds 8.
1716 	 */
1717 	for_each_online_node(node) {
1718 		unsigned long faults;
1719 		int dist = node_distance(nid, node);
1720 
1721 		/*
1722 		 * The furthest away nodes in the system are not interesting
1723 		 * for placement; nid was already counted.
1724 		 */
1725 		if (dist >= max_dist || node == nid)
1726 			continue;
1727 
1728 		/*
1729 		 * On systems with a backplane NUMA topology, compare groups
1730 		 * of nodes, and move tasks towards the group with the most
1731 		 * memory accesses. When comparing two nodes at distance
1732 		 * "hoplimit", only nodes closer by than "hoplimit" are part
1733 		 * of each group. Skip other nodes.
1734 		 */
1735 		if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1736 			continue;
1737 
1738 		/* Add up the faults from nearby nodes. */
1739 		if (task)
1740 			faults = task_faults(p, node);
1741 		else
1742 			faults = group_faults(p, node);
1743 
1744 		/*
1745 		 * On systems with a glueless mesh NUMA topology, there are
1746 		 * no fixed "groups of nodes". Instead, nodes that are not
1747 		 * directly connected bounce traffic through intermediate
1748 		 * nodes; a numa_group can occupy any set of nodes.
1749 		 * The further away a node is, the less the faults count.
1750 		 * This seems to result in good task placement.
1751 		 */
1752 		if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1753 			faults *= (max_dist - dist);
1754 			faults /= (max_dist - LOCAL_DISTANCE);
1755 		}
1756 
1757 		score += faults;
1758 	}
1759 
1760 	return score;
1761 }
1762 
1763 /*
1764  * These return the fraction of accesses done by a particular task, or
1765  * task group, on a particular numa node.  The group weight is given a
1766  * larger multiplier, in order to group tasks together that are almost
1767  * evenly spread out between numa nodes.
1768  */
task_weight(struct task_struct * p,int nid,int dist)1769 static inline unsigned long task_weight(struct task_struct *p, int nid,
1770 					int dist)
1771 {
1772 	unsigned long faults, total_faults;
1773 
1774 	if (!p->numa_faults)
1775 		return 0;
1776 
1777 	total_faults = p->total_numa_faults;
1778 
1779 	if (!total_faults)
1780 		return 0;
1781 
1782 	faults = task_faults(p, nid);
1783 	faults += score_nearby_nodes(p, nid, dist, true);
1784 
1785 	return 1000 * faults / total_faults;
1786 }
1787 
group_weight(struct task_struct * p,int nid,int dist)1788 static inline unsigned long group_weight(struct task_struct *p, int nid,
1789 					 int dist)
1790 {
1791 	struct numa_group *ng = deref_task_numa_group(p);
1792 	unsigned long faults, total_faults;
1793 
1794 	if (!ng)
1795 		return 0;
1796 
1797 	total_faults = ng->total_faults;
1798 
1799 	if (!total_faults)
1800 		return 0;
1801 
1802 	faults = group_faults(p, nid);
1803 	faults += score_nearby_nodes(p, nid, dist, false);
1804 
1805 	return 1000 * faults / total_faults;
1806 }
1807 
1808 /*
1809  * If memory tiering mode is enabled, cpupid of slow memory page is
1810  * used to record scan time instead of CPU and PID.  When tiering mode
1811  * is disabled at run time, the scan time (in cpupid) will be
1812  * interpreted as CPU and PID.  So CPU needs to be checked to avoid to
1813  * access out of array bound.
1814  */
cpupid_valid(int cpupid)1815 static inline bool cpupid_valid(int cpupid)
1816 {
1817 	return cpupid_to_cpu(cpupid) < nr_cpu_ids;
1818 }
1819 
1820 /*
1821  * For memory tiering mode, if there are enough free pages (more than
1822  * enough watermark defined here) in fast memory node, to take full
1823  * advantage of fast memory capacity, all recently accessed slow
1824  * memory pages will be migrated to fast memory node without
1825  * considering hot threshold.
1826  */
pgdat_free_space_enough(struct pglist_data * pgdat)1827 static bool pgdat_free_space_enough(struct pglist_data *pgdat)
1828 {
1829 	int z;
1830 	unsigned long enough_wmark;
1831 
1832 	enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT,
1833 			   pgdat->node_present_pages >> 4);
1834 	for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1835 		struct zone *zone = pgdat->node_zones + z;
1836 
1837 		if (!populated_zone(zone))
1838 			continue;
1839 
1840 		if (zone_watermark_ok(zone, 0,
1841 				      promo_wmark_pages(zone) + enough_wmark,
1842 				      ZONE_MOVABLE, 0))
1843 			return true;
1844 	}
1845 	return false;
1846 }
1847 
1848 /*
1849  * For memory tiering mode, when page tables are scanned, the scan
1850  * time will be recorded in struct page in addition to make page
1851  * PROT_NONE for slow memory page.  So when the page is accessed, in
1852  * hint page fault handler, the hint page fault latency is calculated
1853  * via,
1854  *
1855  *	hint page fault latency = hint page fault time - scan time
1856  *
1857  * The smaller the hint page fault latency, the higher the possibility
1858  * for the page to be hot.
1859  */
numa_hint_fault_latency(struct folio * folio)1860 static int numa_hint_fault_latency(struct folio *folio)
1861 {
1862 	int last_time, time;
1863 
1864 	time = jiffies_to_msecs(jiffies);
1865 	last_time = folio_xchg_access_time(folio, time);
1866 
1867 	return (time - last_time) & PAGE_ACCESS_TIME_MASK;
1868 }
1869 
1870 /*
1871  * For memory tiering mode, too high promotion/demotion throughput may
1872  * hurt application latency.  So we provide a mechanism to rate limit
1873  * the number of pages that are tried to be promoted.
1874  */
numa_promotion_rate_limit(struct pglist_data * pgdat,unsigned long rate_limit,int nr)1875 static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
1876 				      unsigned long rate_limit, int nr)
1877 {
1878 	unsigned long nr_cand;
1879 	unsigned int now, start;
1880 
1881 	now = jiffies_to_msecs(jiffies);
1882 	mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr);
1883 	nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1884 	start = pgdat->nbp_rl_start;
1885 	if (now - start > MSEC_PER_SEC &&
1886 	    cmpxchg(&pgdat->nbp_rl_start, start, now) == start)
1887 		pgdat->nbp_rl_nr_cand = nr_cand;
1888 	if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)
1889 		return true;
1890 	return false;
1891 }
1892 
1893 #define NUMA_MIGRATION_ADJUST_STEPS	16
1894 
numa_promotion_adjust_threshold(struct pglist_data * pgdat,unsigned long rate_limit,unsigned int ref_th)1895 static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
1896 					    unsigned long rate_limit,
1897 					    unsigned int ref_th)
1898 {
1899 	unsigned int now, start, th_period, unit_th, th;
1900 	unsigned long nr_cand, ref_cand, diff_cand;
1901 
1902 	now = jiffies_to_msecs(jiffies);
1903 	th_period = sysctl_numa_balancing_scan_period_max;
1904 	start = pgdat->nbp_th_start;
1905 	if (now - start > th_period &&
1906 	    cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
1907 		ref_cand = rate_limit *
1908 			sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
1909 		nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1910 		diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
1911 		unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
1912 		th = pgdat->nbp_threshold ? : ref_th;
1913 		if (diff_cand > ref_cand * 11 / 10)
1914 			th = max(th - unit_th, unit_th);
1915 		else if (diff_cand < ref_cand * 9 / 10)
1916 			th = min(th + unit_th, ref_th * 2);
1917 		pgdat->nbp_th_nr_cand = nr_cand;
1918 		pgdat->nbp_threshold = th;
1919 	}
1920 }
1921 
should_numa_migrate_memory(struct task_struct * p,struct folio * folio,int src_nid,int dst_cpu)1922 bool should_numa_migrate_memory(struct task_struct *p, struct folio *folio,
1923 				int src_nid, int dst_cpu)
1924 {
1925 	struct numa_group *ng = deref_curr_numa_group(p);
1926 	int dst_nid = cpu_to_node(dst_cpu);
1927 	int last_cpupid, this_cpupid;
1928 
1929 	/*
1930 	 * Cannot migrate to memoryless nodes.
1931 	 */
1932 	if (!node_state(dst_nid, N_MEMORY))
1933 		return false;
1934 
1935 	/*
1936 	 * The pages in slow memory node should be migrated according
1937 	 * to hot/cold instead of private/shared.
1938 	 */
1939 	if (folio_use_access_time(folio)) {
1940 		struct pglist_data *pgdat;
1941 		unsigned long rate_limit;
1942 		unsigned int latency, th, def_th;
1943 
1944 		pgdat = NODE_DATA(dst_nid);
1945 		if (pgdat_free_space_enough(pgdat)) {
1946 			/* workload changed, reset hot threshold */
1947 			pgdat->nbp_threshold = 0;
1948 			return true;
1949 		}
1950 
1951 		def_th = sysctl_numa_balancing_hot_threshold;
1952 		rate_limit = sysctl_numa_balancing_promote_rate_limit << \
1953 			(20 - PAGE_SHIFT);
1954 		numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
1955 
1956 		th = pgdat->nbp_threshold ? : def_th;
1957 		latency = numa_hint_fault_latency(folio);
1958 		if (latency >= th)
1959 			return false;
1960 
1961 		return !numa_promotion_rate_limit(pgdat, rate_limit,
1962 						  folio_nr_pages(folio));
1963 	}
1964 
1965 	this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1966 	last_cpupid = folio_xchg_last_cpupid(folio, this_cpupid);
1967 
1968 	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1969 	    !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1970 		return false;
1971 
1972 	/*
1973 	 * Allow first faults or private faults to migrate immediately early in
1974 	 * the lifetime of a task. The magic number 4 is based on waiting for
1975 	 * two full passes of the "multi-stage node selection" test that is
1976 	 * executed below.
1977 	 */
1978 	if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1979 	    (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1980 		return true;
1981 
1982 	/*
1983 	 * Multi-stage node selection is used in conjunction with a periodic
1984 	 * migration fault to build a temporal task<->page relation. By using
1985 	 * a two-stage filter we remove short/unlikely relations.
1986 	 *
1987 	 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1988 	 * a task's usage of a particular page (n_p) per total usage of this
1989 	 * page (n_t) (in a given time-span) to a probability.
1990 	 *
1991 	 * Our periodic faults will sample this probability and getting the
1992 	 * same result twice in a row, given these samples are fully
1993 	 * independent, is then given by P(n)^2, provided our sample period
1994 	 * is sufficiently short compared to the usage pattern.
1995 	 *
1996 	 * This quadric squishes small probabilities, making it less likely we
1997 	 * act on an unlikely task<->page relation.
1998 	 */
1999 	if (!cpupid_pid_unset(last_cpupid) &&
2000 				cpupid_to_nid(last_cpupid) != dst_nid)
2001 		return false;
2002 
2003 	/* Always allow migrate on private faults */
2004 	if (cpupid_match_pid(p, last_cpupid))
2005 		return true;
2006 
2007 	/* A shared fault, but p->numa_group has not been set up yet. */
2008 	if (!ng)
2009 		return true;
2010 
2011 	/*
2012 	 * Destination node is much more heavily used than the source
2013 	 * node? Allow migration.
2014 	 */
2015 	if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
2016 					ACTIVE_NODE_FRACTION)
2017 		return true;
2018 
2019 	/*
2020 	 * Distribute memory according to CPU & memory use on each node,
2021 	 * with 3/4 hysteresis to avoid unnecessary memory migrations:
2022 	 *
2023 	 * faults_cpu(dst)   3   faults_cpu(src)
2024 	 * --------------- * - > ---------------
2025 	 * faults_mem(dst)   4   faults_mem(src)
2026 	 */
2027 	return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
2028 	       group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
2029 }
2030 
2031 /*
2032  * 'numa_type' describes the node at the moment of load balancing.
2033  */
2034 enum numa_type {
2035 	/* The node has spare capacity that can be used to run more tasks.  */
2036 	node_has_spare = 0,
2037 	/*
2038 	 * The node is fully used and the tasks don't compete for more CPU
2039 	 * cycles. Nevertheless, some tasks might wait before running.
2040 	 */
2041 	node_fully_busy,
2042 	/*
2043 	 * The node is overloaded and can't provide expected CPU cycles to all
2044 	 * tasks.
2045 	 */
2046 	node_overloaded
2047 };
2048 
2049 /* Cached statistics for all CPUs within a node */
2050 struct numa_stats {
2051 	unsigned long load;
2052 	unsigned long runnable;
2053 	unsigned long util;
2054 	/* Total compute capacity of CPUs on a node */
2055 	unsigned long compute_capacity;
2056 	unsigned int nr_running;
2057 	unsigned int weight;
2058 	enum numa_type node_type;
2059 	int idle_cpu;
2060 };
2061 
2062 struct task_numa_env {
2063 	struct task_struct *p;
2064 
2065 	int src_cpu, src_nid;
2066 	int dst_cpu, dst_nid;
2067 	int imb_numa_nr;
2068 
2069 	struct numa_stats src_stats, dst_stats;
2070 
2071 	int imbalance_pct;
2072 	int dist;
2073 
2074 	struct task_struct *best_task;
2075 	long best_imp;
2076 	int best_cpu;
2077 };
2078 
2079 static unsigned long cpu_load(struct rq *rq);
2080 static unsigned long cpu_runnable(struct rq *rq);
2081 
2082 static inline enum
numa_classify(unsigned int imbalance_pct,struct numa_stats * ns)2083 numa_type numa_classify(unsigned int imbalance_pct,
2084 			 struct numa_stats *ns)
2085 {
2086 	if ((ns->nr_running > ns->weight) &&
2087 	    (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
2088 	     ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
2089 		return node_overloaded;
2090 
2091 	if ((ns->nr_running < ns->weight) ||
2092 	    (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
2093 	     ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
2094 		return node_has_spare;
2095 
2096 	return node_fully_busy;
2097 }
2098 
2099 #ifdef CONFIG_SCHED_SMT
2100 /* Forward declarations of select_idle_sibling helpers */
2101 static inline bool test_idle_cores(int cpu);
numa_idle_core(int idle_core,int cpu)2102 static inline int numa_idle_core(int idle_core, int cpu)
2103 {
2104 	if (!static_branch_likely(&sched_smt_present) ||
2105 	    idle_core >= 0 || !test_idle_cores(cpu))
2106 		return idle_core;
2107 
2108 	/*
2109 	 * Prefer cores instead of packing HT siblings
2110 	 * and triggering future load balancing.
2111 	 */
2112 	if (is_core_idle(cpu))
2113 		idle_core = cpu;
2114 
2115 	return idle_core;
2116 }
2117 #else
numa_idle_core(int idle_core,int cpu)2118 static inline int numa_idle_core(int idle_core, int cpu)
2119 {
2120 	return idle_core;
2121 }
2122 #endif
2123 
2124 /*
2125  * Gather all necessary information to make NUMA balancing placement
2126  * decisions that are compatible with standard load balancer. This
2127  * borrows code and logic from update_sg_lb_stats but sharing a
2128  * common implementation is impractical.
2129  */
update_numa_stats(struct task_numa_env * env,struct numa_stats * ns,int nid,bool find_idle)2130 static void update_numa_stats(struct task_numa_env *env,
2131 			      struct numa_stats *ns, int nid,
2132 			      bool find_idle)
2133 {
2134 	int cpu, idle_core = -1;
2135 
2136 	memset(ns, 0, sizeof(*ns));
2137 	ns->idle_cpu = -1;
2138 
2139 	rcu_read_lock();
2140 	for_each_cpu(cpu, cpumask_of_node(nid)) {
2141 		struct rq *rq = cpu_rq(cpu);
2142 
2143 		ns->load += cpu_load(rq);
2144 		ns->runnable += cpu_runnable(rq);
2145 		ns->util += cpu_util_cfs(cpu);
2146 		ns->nr_running += rq->cfs.h_nr_runnable;
2147 		ns->compute_capacity += capacity_of(cpu);
2148 
2149 		if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) {
2150 			if (READ_ONCE(rq->numa_migrate_on) ||
2151 			    !cpumask_test_cpu(cpu, env->p->cpus_ptr))
2152 				continue;
2153 
2154 			if (ns->idle_cpu == -1)
2155 				ns->idle_cpu = cpu;
2156 
2157 			idle_core = numa_idle_core(idle_core, cpu);
2158 		}
2159 	}
2160 	rcu_read_unlock();
2161 
2162 	ns->weight = cpumask_weight(cpumask_of_node(nid));
2163 
2164 	ns->node_type = numa_classify(env->imbalance_pct, ns);
2165 
2166 	if (idle_core >= 0)
2167 		ns->idle_cpu = idle_core;
2168 }
2169 
task_numa_assign(struct task_numa_env * env,struct task_struct * p,long imp)2170 static void task_numa_assign(struct task_numa_env *env,
2171 			     struct task_struct *p, long imp)
2172 {
2173 	struct rq *rq = cpu_rq(env->dst_cpu);
2174 
2175 	/* Check if run-queue part of active NUMA balance. */
2176 	if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
2177 		int cpu;
2178 		int start = env->dst_cpu;
2179 
2180 		/* Find alternative idle CPU. */
2181 		for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) {
2182 			if (cpu == env->best_cpu || !idle_cpu(cpu) ||
2183 			    !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
2184 				continue;
2185 			}
2186 
2187 			env->dst_cpu = cpu;
2188 			rq = cpu_rq(env->dst_cpu);
2189 			if (!xchg(&rq->numa_migrate_on, 1))
2190 				goto assign;
2191 		}
2192 
2193 		/* Failed to find an alternative idle CPU */
2194 		return;
2195 	}
2196 
2197 assign:
2198 	/*
2199 	 * Clear previous best_cpu/rq numa-migrate flag, since task now
2200 	 * found a better CPU to move/swap.
2201 	 */
2202 	if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
2203 		rq = cpu_rq(env->best_cpu);
2204 		WRITE_ONCE(rq->numa_migrate_on, 0);
2205 	}
2206 
2207 	if (env->best_task)
2208 		put_task_struct(env->best_task);
2209 	if (p)
2210 		get_task_struct(p);
2211 
2212 	env->best_task = p;
2213 	env->best_imp = imp;
2214 	env->best_cpu = env->dst_cpu;
2215 }
2216 
load_too_imbalanced(long src_load,long dst_load,struct task_numa_env * env)2217 static bool load_too_imbalanced(long src_load, long dst_load,
2218 				struct task_numa_env *env)
2219 {
2220 	long imb, old_imb;
2221 	long orig_src_load, orig_dst_load;
2222 	long src_capacity, dst_capacity;
2223 
2224 	/*
2225 	 * The load is corrected for the CPU capacity available on each node.
2226 	 *
2227 	 * src_load        dst_load
2228 	 * ------------ vs ---------
2229 	 * src_capacity    dst_capacity
2230 	 */
2231 	src_capacity = env->src_stats.compute_capacity;
2232 	dst_capacity = env->dst_stats.compute_capacity;
2233 
2234 	imb = abs(dst_load * src_capacity - src_load * dst_capacity);
2235 
2236 	orig_src_load = env->src_stats.load;
2237 	orig_dst_load = env->dst_stats.load;
2238 
2239 	old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
2240 
2241 	/* Would this change make things worse? */
2242 	return (imb > old_imb);
2243 }
2244 
2245 /*
2246  * Maximum NUMA importance can be 1998 (2*999);
2247  * SMALLIMP @ 30 would be close to 1998/64.
2248  * Used to deter task migration.
2249  */
2250 #define SMALLIMP	30
2251 
2252 /*
2253  * This checks if the overall compute and NUMA accesses of the system would
2254  * be improved if the source tasks was migrated to the target dst_cpu taking
2255  * into account that it might be best if task running on the dst_cpu should
2256  * be exchanged with the source task
2257  */
task_numa_compare(struct task_numa_env * env,long taskimp,long groupimp,bool maymove)2258 static bool task_numa_compare(struct task_numa_env *env,
2259 			      long taskimp, long groupimp, bool maymove)
2260 {
2261 	struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
2262 	struct rq *dst_rq = cpu_rq(env->dst_cpu);
2263 	long imp = p_ng ? groupimp : taskimp;
2264 	struct task_struct *cur;
2265 	long src_load, dst_load;
2266 	int dist = env->dist;
2267 	long moveimp = imp;
2268 	long load;
2269 	bool stopsearch = false;
2270 
2271 	if (READ_ONCE(dst_rq->numa_migrate_on))
2272 		return false;
2273 
2274 	rcu_read_lock();
2275 	cur = rcu_dereference(dst_rq->curr);
2276 	if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
2277 		cur = NULL;
2278 
2279 	/*
2280 	 * Because we have preemption enabled we can get migrated around and
2281 	 * end try selecting ourselves (current == env->p) as a swap candidate.
2282 	 */
2283 	if (cur == env->p) {
2284 		stopsearch = true;
2285 		goto unlock;
2286 	}
2287 
2288 	if (!cur) {
2289 		if (maymove && moveimp >= env->best_imp)
2290 			goto assign;
2291 		else
2292 			goto unlock;
2293 	}
2294 
2295 	/* Skip this swap candidate if cannot move to the source cpu. */
2296 	if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
2297 		goto unlock;
2298 
2299 	/*
2300 	 * Skip this swap candidate if it is not moving to its preferred
2301 	 * node and the best task is.
2302 	 */
2303 	if (env->best_task &&
2304 	    env->best_task->numa_preferred_nid == env->src_nid &&
2305 	    cur->numa_preferred_nid != env->src_nid) {
2306 		goto unlock;
2307 	}
2308 
2309 	/*
2310 	 * "imp" is the fault differential for the source task between the
2311 	 * source and destination node. Calculate the total differential for
2312 	 * the source task and potential destination task. The more negative
2313 	 * the value is, the more remote accesses that would be expected to
2314 	 * be incurred if the tasks were swapped.
2315 	 *
2316 	 * If dst and source tasks are in the same NUMA group, or not
2317 	 * in any group then look only at task weights.
2318 	 */
2319 	cur_ng = rcu_dereference(cur->numa_group);
2320 	if (cur_ng == p_ng) {
2321 		/*
2322 		 * Do not swap within a group or between tasks that have
2323 		 * no group if there is spare capacity. Swapping does
2324 		 * not address the load imbalance and helps one task at
2325 		 * the cost of punishing another.
2326 		 */
2327 		if (env->dst_stats.node_type == node_has_spare)
2328 			goto unlock;
2329 
2330 		imp = taskimp + task_weight(cur, env->src_nid, dist) -
2331 		      task_weight(cur, env->dst_nid, dist);
2332 		/*
2333 		 * Add some hysteresis to prevent swapping the
2334 		 * tasks within a group over tiny differences.
2335 		 */
2336 		if (cur_ng)
2337 			imp -= imp / 16;
2338 	} else {
2339 		/*
2340 		 * Compare the group weights. If a task is all by itself
2341 		 * (not part of a group), use the task weight instead.
2342 		 */
2343 		if (cur_ng && p_ng)
2344 			imp += group_weight(cur, env->src_nid, dist) -
2345 			       group_weight(cur, env->dst_nid, dist);
2346 		else
2347 			imp += task_weight(cur, env->src_nid, dist) -
2348 			       task_weight(cur, env->dst_nid, dist);
2349 	}
2350 
2351 	/* Discourage picking a task already on its preferred node */
2352 	if (cur->numa_preferred_nid == env->dst_nid)
2353 		imp -= imp / 16;
2354 
2355 	/*
2356 	 * Encourage picking a task that moves to its preferred node.
2357 	 * This potentially makes imp larger than it's maximum of
2358 	 * 1998 (see SMALLIMP and task_weight for why) but in this
2359 	 * case, it does not matter.
2360 	 */
2361 	if (cur->numa_preferred_nid == env->src_nid)
2362 		imp += imp / 8;
2363 
2364 	if (maymove && moveimp > imp && moveimp > env->best_imp) {
2365 		imp = moveimp;
2366 		cur = NULL;
2367 		goto assign;
2368 	}
2369 
2370 	/*
2371 	 * Prefer swapping with a task moving to its preferred node over a
2372 	 * task that is not.
2373 	 */
2374 	if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
2375 	    env->best_task->numa_preferred_nid != env->src_nid) {
2376 		goto assign;
2377 	}
2378 
2379 	/*
2380 	 * If the NUMA importance is less than SMALLIMP,
2381 	 * task migration might only result in ping pong
2382 	 * of tasks and also hurt performance due to cache
2383 	 * misses.
2384 	 */
2385 	if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
2386 		goto unlock;
2387 
2388 	/*
2389 	 * In the overloaded case, try and keep the load balanced.
2390 	 */
2391 	load = task_h_load(env->p) - task_h_load(cur);
2392 	if (!load)
2393 		goto assign;
2394 
2395 	dst_load = env->dst_stats.load + load;
2396 	src_load = env->src_stats.load - load;
2397 
2398 	if (load_too_imbalanced(src_load, dst_load, env))
2399 		goto unlock;
2400 
2401 assign:
2402 	/* Evaluate an idle CPU for a task numa move. */
2403 	if (!cur) {
2404 		int cpu = env->dst_stats.idle_cpu;
2405 
2406 		/* Nothing cached so current CPU went idle since the search. */
2407 		if (cpu < 0)
2408 			cpu = env->dst_cpu;
2409 
2410 		/*
2411 		 * If the CPU is no longer truly idle and the previous best CPU
2412 		 * is, keep using it.
2413 		 */
2414 		if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2415 		    idle_cpu(env->best_cpu)) {
2416 			cpu = env->best_cpu;
2417 		}
2418 
2419 		env->dst_cpu = cpu;
2420 	}
2421 
2422 	task_numa_assign(env, cur, imp);
2423 
2424 	/*
2425 	 * If a move to idle is allowed because there is capacity or load
2426 	 * balance improves then stop the search. While a better swap
2427 	 * candidate may exist, a search is not free.
2428 	 */
2429 	if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
2430 		stopsearch = true;
2431 
2432 	/*
2433 	 * If a swap candidate must be identified and the current best task
2434 	 * moves its preferred node then stop the search.
2435 	 */
2436 	if (!maymove && env->best_task &&
2437 	    env->best_task->numa_preferred_nid == env->src_nid) {
2438 		stopsearch = true;
2439 	}
2440 unlock:
2441 	rcu_read_unlock();
2442 
2443 	return stopsearch;
2444 }
2445 
task_numa_find_cpu(struct task_numa_env * env,long taskimp,long groupimp)2446 static void task_numa_find_cpu(struct task_numa_env *env,
2447 				long taskimp, long groupimp)
2448 {
2449 	bool maymove = false;
2450 	int cpu;
2451 
2452 	/*
2453 	 * If dst node has spare capacity, then check if there is an
2454 	 * imbalance that would be overruled by the load balancer.
2455 	 */
2456 	if (env->dst_stats.node_type == node_has_spare) {
2457 		unsigned int imbalance;
2458 		int src_running, dst_running;
2459 
2460 		/*
2461 		 * Would movement cause an imbalance? Note that if src has
2462 		 * more running tasks that the imbalance is ignored as the
2463 		 * move improves the imbalance from the perspective of the
2464 		 * CPU load balancer.
2465 		 * */
2466 		src_running = env->src_stats.nr_running - 1;
2467 		dst_running = env->dst_stats.nr_running + 1;
2468 		imbalance = max(0, dst_running - src_running);
2469 		imbalance = adjust_numa_imbalance(imbalance, dst_running,
2470 						  env->imb_numa_nr);
2471 
2472 		/* Use idle CPU if there is no imbalance */
2473 		if (!imbalance) {
2474 			maymove = true;
2475 			if (env->dst_stats.idle_cpu >= 0) {
2476 				env->dst_cpu = env->dst_stats.idle_cpu;
2477 				task_numa_assign(env, NULL, 0);
2478 				return;
2479 			}
2480 		}
2481 	} else {
2482 		long src_load, dst_load, load;
2483 		/*
2484 		 * If the improvement from just moving env->p direction is better
2485 		 * than swapping tasks around, check if a move is possible.
2486 		 */
2487 		load = task_h_load(env->p);
2488 		dst_load = env->dst_stats.load + load;
2489 		src_load = env->src_stats.load - load;
2490 		maymove = !load_too_imbalanced(src_load, dst_load, env);
2491 	}
2492 
2493 	for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2494 		/* Skip this CPU if the source task cannot migrate */
2495 		if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2496 			continue;
2497 
2498 		env->dst_cpu = cpu;
2499 		if (task_numa_compare(env, taskimp, groupimp, maymove))
2500 			break;
2501 	}
2502 }
2503 
task_numa_migrate(struct task_struct * p)2504 static int task_numa_migrate(struct task_struct *p)
2505 {
2506 	struct task_numa_env env = {
2507 		.p = p,
2508 
2509 		.src_cpu = task_cpu(p),
2510 		.src_nid = task_node(p),
2511 
2512 		.imbalance_pct = 112,
2513 
2514 		.best_task = NULL,
2515 		.best_imp = 0,
2516 		.best_cpu = -1,
2517 	};
2518 	unsigned long taskweight, groupweight;
2519 	struct sched_domain *sd;
2520 	long taskimp, groupimp;
2521 	struct numa_group *ng;
2522 	struct rq *best_rq;
2523 	int nid, ret, dist;
2524 
2525 	/*
2526 	 * Pick the lowest SD_NUMA domain, as that would have the smallest
2527 	 * imbalance and would be the first to start moving tasks about.
2528 	 *
2529 	 * And we want to avoid any moving of tasks about, as that would create
2530 	 * random movement of tasks -- counter the numa conditions we're trying
2531 	 * to satisfy here.
2532 	 */
2533 	rcu_read_lock();
2534 	sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2535 	if (sd) {
2536 		env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2537 		env.imb_numa_nr = sd->imb_numa_nr;
2538 	}
2539 	rcu_read_unlock();
2540 
2541 	/*
2542 	 * Cpusets can break the scheduler domain tree into smaller
2543 	 * balance domains, some of which do not cross NUMA boundaries.
2544 	 * Tasks that are "trapped" in such domains cannot be migrated
2545 	 * elsewhere, so there is no point in (re)trying.
2546 	 */
2547 	if (unlikely(!sd)) {
2548 		sched_setnuma(p, task_node(p));
2549 		return -EINVAL;
2550 	}
2551 
2552 	env.dst_nid = p->numa_preferred_nid;
2553 	dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2554 	taskweight = task_weight(p, env.src_nid, dist);
2555 	groupweight = group_weight(p, env.src_nid, dist);
2556 	update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2557 	taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2558 	groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2559 	update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2560 
2561 	/* Try to find a spot on the preferred nid. */
2562 	task_numa_find_cpu(&env, taskimp, groupimp);
2563 
2564 	/*
2565 	 * Look at other nodes in these cases:
2566 	 * - there is no space available on the preferred_nid
2567 	 * - the task is part of a numa_group that is interleaved across
2568 	 *   multiple NUMA nodes; in order to better consolidate the group,
2569 	 *   we need to check other locations.
2570 	 */
2571 	ng = deref_curr_numa_group(p);
2572 	if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2573 		for_each_node_state(nid, N_CPU) {
2574 			if (nid == env.src_nid || nid == p->numa_preferred_nid)
2575 				continue;
2576 
2577 			dist = node_distance(env.src_nid, env.dst_nid);
2578 			if (sched_numa_topology_type == NUMA_BACKPLANE &&
2579 						dist != env.dist) {
2580 				taskweight = task_weight(p, env.src_nid, dist);
2581 				groupweight = group_weight(p, env.src_nid, dist);
2582 			}
2583 
2584 			/* Only consider nodes where both task and groups benefit */
2585 			taskimp = task_weight(p, nid, dist) - taskweight;
2586 			groupimp = group_weight(p, nid, dist) - groupweight;
2587 			if (taskimp < 0 && groupimp < 0)
2588 				continue;
2589 
2590 			env.dist = dist;
2591 			env.dst_nid = nid;
2592 			update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2593 			task_numa_find_cpu(&env, taskimp, groupimp);
2594 		}
2595 	}
2596 
2597 	/*
2598 	 * If the task is part of a workload that spans multiple NUMA nodes,
2599 	 * and is migrating into one of the workload's active nodes, remember
2600 	 * this node as the task's preferred numa node, so the workload can
2601 	 * settle down.
2602 	 * A task that migrated to a second choice node will be better off
2603 	 * trying for a better one later. Do not set the preferred node here.
2604 	 */
2605 	if (ng) {
2606 		if (env.best_cpu == -1)
2607 			nid = env.src_nid;
2608 		else
2609 			nid = cpu_to_node(env.best_cpu);
2610 
2611 		if (nid != p->numa_preferred_nid)
2612 			sched_setnuma(p, nid);
2613 	}
2614 
2615 	/* No better CPU than the current one was found. */
2616 	if (env.best_cpu == -1) {
2617 		trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2618 		return -EAGAIN;
2619 	}
2620 
2621 	best_rq = cpu_rq(env.best_cpu);
2622 	if (env.best_task == NULL) {
2623 		ret = migrate_task_to(p, env.best_cpu);
2624 		WRITE_ONCE(best_rq->numa_migrate_on, 0);
2625 		if (ret != 0)
2626 			trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2627 		return ret;
2628 	}
2629 
2630 	ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2631 	WRITE_ONCE(best_rq->numa_migrate_on, 0);
2632 
2633 	if (ret != 0)
2634 		trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2635 	put_task_struct(env.best_task);
2636 	return ret;
2637 }
2638 
2639 /* Attempt to migrate a task to a CPU on the preferred node. */
numa_migrate_preferred(struct task_struct * p)2640 static void numa_migrate_preferred(struct task_struct *p)
2641 {
2642 	unsigned long interval = HZ;
2643 
2644 	/* This task has no NUMA fault statistics yet */
2645 	if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2646 		return;
2647 
2648 	/* Periodically retry migrating the task to the preferred node */
2649 	interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2650 	p->numa_migrate_retry = jiffies + interval;
2651 
2652 	/* Success if task is already running on preferred CPU */
2653 	if (task_node(p) == p->numa_preferred_nid)
2654 		return;
2655 
2656 	/* Otherwise, try migrate to a CPU on the preferred node */
2657 	task_numa_migrate(p);
2658 }
2659 
2660 /*
2661  * Find out how many nodes the workload is actively running on. Do this by
2662  * tracking the nodes from which NUMA hinting faults are triggered. This can
2663  * be different from the set of nodes where the workload's memory is currently
2664  * located.
2665  */
numa_group_count_active_nodes(struct numa_group * numa_group)2666 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2667 {
2668 	unsigned long faults, max_faults = 0;
2669 	int nid, active_nodes = 0;
2670 
2671 	for_each_node_state(nid, N_CPU) {
2672 		faults = group_faults_cpu(numa_group, nid);
2673 		if (faults > max_faults)
2674 			max_faults = faults;
2675 	}
2676 
2677 	for_each_node_state(nid, N_CPU) {
2678 		faults = group_faults_cpu(numa_group, nid);
2679 		if (faults * ACTIVE_NODE_FRACTION > max_faults)
2680 			active_nodes++;
2681 	}
2682 
2683 	numa_group->max_faults_cpu = max_faults;
2684 	numa_group->active_nodes = active_nodes;
2685 }
2686 
2687 /*
2688  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2689  * increments. The more local the fault statistics are, the higher the scan
2690  * period will be for the next scan window. If local/(local+remote) ratio is
2691  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2692  * the scan period will decrease. Aim for 70% local accesses.
2693  */
2694 #define NUMA_PERIOD_SLOTS 10
2695 #define NUMA_PERIOD_THRESHOLD 7
2696 
2697 /*
2698  * Increase the scan period (slow down scanning) if the majority of
2699  * our memory is already on our local node, or if the majority of
2700  * the page accesses are shared with other processes.
2701  * Otherwise, decrease the scan period.
2702  */
update_task_scan_period(struct task_struct * p,unsigned long shared,unsigned long private)2703 static void update_task_scan_period(struct task_struct *p,
2704 			unsigned long shared, unsigned long private)
2705 {
2706 	unsigned int period_slot;
2707 	int lr_ratio, ps_ratio;
2708 	int diff;
2709 
2710 	unsigned long remote = p->numa_faults_locality[0];
2711 	unsigned long local = p->numa_faults_locality[1];
2712 
2713 	/*
2714 	 * If there were no record hinting faults then either the task is
2715 	 * completely idle or all activity is in areas that are not of interest
2716 	 * to automatic numa balancing. Related to that, if there were failed
2717 	 * migration then it implies we are migrating too quickly or the local
2718 	 * node is overloaded. In either case, scan slower
2719 	 */
2720 	if (local + shared == 0 || p->numa_faults_locality[2]) {
2721 		p->numa_scan_period = min(p->numa_scan_period_max,
2722 			p->numa_scan_period << 1);
2723 
2724 		p->mm->numa_next_scan = jiffies +
2725 			msecs_to_jiffies(p->numa_scan_period);
2726 
2727 		return;
2728 	}
2729 
2730 	/*
2731 	 * Prepare to scale scan period relative to the current period.
2732 	 *	 == NUMA_PERIOD_THRESHOLD scan period stays the same
2733 	 *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2734 	 *	 >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2735 	 */
2736 	period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2737 	lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2738 	ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2739 
2740 	if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2741 		/*
2742 		 * Most memory accesses are local. There is no need to
2743 		 * do fast NUMA scanning, since memory is already local.
2744 		 */
2745 		int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2746 		if (!slot)
2747 			slot = 1;
2748 		diff = slot * period_slot;
2749 	} else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2750 		/*
2751 		 * Most memory accesses are shared with other tasks.
2752 		 * There is no point in continuing fast NUMA scanning,
2753 		 * since other tasks may just move the memory elsewhere.
2754 		 */
2755 		int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2756 		if (!slot)
2757 			slot = 1;
2758 		diff = slot * period_slot;
2759 	} else {
2760 		/*
2761 		 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2762 		 * yet they are not on the local NUMA node. Speed up
2763 		 * NUMA scanning to get the memory moved over.
2764 		 */
2765 		int ratio = max(lr_ratio, ps_ratio);
2766 		diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2767 	}
2768 
2769 	p->numa_scan_period = clamp(p->numa_scan_period + diff,
2770 			task_scan_min(p), task_scan_max(p));
2771 	memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2772 }
2773 
2774 /*
2775  * Get the fraction of time the task has been running since the last
2776  * NUMA placement cycle. The scheduler keeps similar statistics, but
2777  * decays those on a 32ms period, which is orders of magnitude off
2778  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2779  * stats only if the task is so new there are no NUMA statistics yet.
2780  */
numa_get_avg_runtime(struct task_struct * p,u64 * period)2781 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2782 {
2783 	u64 runtime, delta, now;
2784 	/* Use the start of this time slice to avoid calculations. */
2785 	now = p->se.exec_start;
2786 	runtime = p->se.sum_exec_runtime;
2787 
2788 	if (p->last_task_numa_placement) {
2789 		delta = runtime - p->last_sum_exec_runtime;
2790 		*period = now - p->last_task_numa_placement;
2791 
2792 		/* Avoid time going backwards, prevent potential divide error: */
2793 		if (unlikely((s64)*period < 0))
2794 			*period = 0;
2795 	} else {
2796 		delta = p->se.avg.load_sum;
2797 		*period = LOAD_AVG_MAX;
2798 	}
2799 
2800 	p->last_sum_exec_runtime = runtime;
2801 	p->last_task_numa_placement = now;
2802 
2803 	return delta;
2804 }
2805 
2806 /*
2807  * Determine the preferred nid for a task in a numa_group. This needs to
2808  * be done in a way that produces consistent results with group_weight,
2809  * otherwise workloads might not converge.
2810  */
preferred_group_nid(struct task_struct * p,int nid)2811 static int preferred_group_nid(struct task_struct *p, int nid)
2812 {
2813 	nodemask_t nodes;
2814 	int dist;
2815 
2816 	/* Direct connections between all NUMA nodes. */
2817 	if (sched_numa_topology_type == NUMA_DIRECT)
2818 		return nid;
2819 
2820 	/*
2821 	 * On a system with glueless mesh NUMA topology, group_weight
2822 	 * scores nodes according to the number of NUMA hinting faults on
2823 	 * both the node itself, and on nearby nodes.
2824 	 */
2825 	if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2826 		unsigned long score, max_score = 0;
2827 		int node, max_node = nid;
2828 
2829 		dist = sched_max_numa_distance;
2830 
2831 		for_each_node_state(node, N_CPU) {
2832 			score = group_weight(p, node, dist);
2833 			if (score > max_score) {
2834 				max_score = score;
2835 				max_node = node;
2836 			}
2837 		}
2838 		return max_node;
2839 	}
2840 
2841 	/*
2842 	 * Finding the preferred nid in a system with NUMA backplane
2843 	 * interconnect topology is more involved. The goal is to locate
2844 	 * tasks from numa_groups near each other in the system, and
2845 	 * untangle workloads from different sides of the system. This requires
2846 	 * searching down the hierarchy of node groups, recursively searching
2847 	 * inside the highest scoring group of nodes. The nodemask tricks
2848 	 * keep the complexity of the search down.
2849 	 */
2850 	nodes = node_states[N_CPU];
2851 	for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2852 		unsigned long max_faults = 0;
2853 		nodemask_t max_group = NODE_MASK_NONE;
2854 		int a, b;
2855 
2856 		/* Are there nodes at this distance from each other? */
2857 		if (!find_numa_distance(dist))
2858 			continue;
2859 
2860 		for_each_node_mask(a, nodes) {
2861 			unsigned long faults = 0;
2862 			nodemask_t this_group;
2863 			nodes_clear(this_group);
2864 
2865 			/* Sum group's NUMA faults; includes a==b case. */
2866 			for_each_node_mask(b, nodes) {
2867 				if (node_distance(a, b) < dist) {
2868 					faults += group_faults(p, b);
2869 					node_set(b, this_group);
2870 					node_clear(b, nodes);
2871 				}
2872 			}
2873 
2874 			/* Remember the top group. */
2875 			if (faults > max_faults) {
2876 				max_faults = faults;
2877 				max_group = this_group;
2878 				/*
2879 				 * subtle: at the smallest distance there is
2880 				 * just one node left in each "group", the
2881 				 * winner is the preferred nid.
2882 				 */
2883 				nid = a;
2884 			}
2885 		}
2886 		/* Next round, evaluate the nodes within max_group. */
2887 		if (!max_faults)
2888 			break;
2889 		nodes = max_group;
2890 	}
2891 	return nid;
2892 }
2893 
task_numa_placement(struct task_struct * p)2894 static void task_numa_placement(struct task_struct *p)
2895 {
2896 	int seq, nid, max_nid = NUMA_NO_NODE;
2897 	unsigned long max_faults = 0;
2898 	unsigned long fault_types[2] = { 0, 0 };
2899 	unsigned long total_faults;
2900 	u64 runtime, period;
2901 	spinlock_t *group_lock = NULL;
2902 	struct numa_group *ng;
2903 
2904 	/*
2905 	 * The p->mm->numa_scan_seq field gets updated without
2906 	 * exclusive access. Use READ_ONCE() here to ensure
2907 	 * that the field is read in a single access:
2908 	 */
2909 	seq = READ_ONCE(p->mm->numa_scan_seq);
2910 	if (p->numa_scan_seq == seq)
2911 		return;
2912 	p->numa_scan_seq = seq;
2913 	p->numa_scan_period_max = task_scan_max(p);
2914 
2915 	total_faults = p->numa_faults_locality[0] +
2916 		       p->numa_faults_locality[1];
2917 	runtime = numa_get_avg_runtime(p, &period);
2918 
2919 	/* If the task is part of a group prevent parallel updates to group stats */
2920 	ng = deref_curr_numa_group(p);
2921 	if (ng) {
2922 		group_lock = &ng->lock;
2923 		spin_lock_irq(group_lock);
2924 	}
2925 
2926 	/* Find the node with the highest number of faults */
2927 	for_each_online_node(nid) {
2928 		/* Keep track of the offsets in numa_faults array */
2929 		int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2930 		unsigned long faults = 0, group_faults = 0;
2931 		int priv;
2932 
2933 		for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2934 			long diff, f_diff, f_weight;
2935 
2936 			mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2937 			membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2938 			cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2939 			cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2940 
2941 			/* Decay existing window, copy faults since last scan */
2942 			diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2943 			fault_types[priv] += p->numa_faults[membuf_idx];
2944 			p->numa_faults[membuf_idx] = 0;
2945 
2946 			/*
2947 			 * Normalize the faults_from, so all tasks in a group
2948 			 * count according to CPU use, instead of by the raw
2949 			 * number of faults. Tasks with little runtime have
2950 			 * little over-all impact on throughput, and thus their
2951 			 * faults are less important.
2952 			 */
2953 			f_weight = div64_u64(runtime << 16, period + 1);
2954 			f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2955 				   (total_faults + 1);
2956 			f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2957 			p->numa_faults[cpubuf_idx] = 0;
2958 
2959 			p->numa_faults[mem_idx] += diff;
2960 			p->numa_faults[cpu_idx] += f_diff;
2961 			faults += p->numa_faults[mem_idx];
2962 			p->total_numa_faults += diff;
2963 			if (ng) {
2964 				/*
2965 				 * safe because we can only change our own group
2966 				 *
2967 				 * mem_idx represents the offset for a given
2968 				 * nid and priv in a specific region because it
2969 				 * is at the beginning of the numa_faults array.
2970 				 */
2971 				ng->faults[mem_idx] += diff;
2972 				ng->faults[cpu_idx] += f_diff;
2973 				ng->total_faults += diff;
2974 				group_faults += ng->faults[mem_idx];
2975 			}
2976 		}
2977 
2978 		if (!ng) {
2979 			if (faults > max_faults) {
2980 				max_faults = faults;
2981 				max_nid = nid;
2982 			}
2983 		} else if (group_faults > max_faults) {
2984 			max_faults = group_faults;
2985 			max_nid = nid;
2986 		}
2987 	}
2988 
2989 	/* Cannot migrate task to CPU-less node */
2990 	max_nid = numa_nearest_node(max_nid, N_CPU);
2991 
2992 	if (ng) {
2993 		numa_group_count_active_nodes(ng);
2994 		spin_unlock_irq(group_lock);
2995 		max_nid = preferred_group_nid(p, max_nid);
2996 	}
2997 
2998 	if (max_faults) {
2999 		/* Set the new preferred node */
3000 		if (max_nid != p->numa_preferred_nid)
3001 			sched_setnuma(p, max_nid);
3002 	}
3003 
3004 	update_task_scan_period(p, fault_types[0], fault_types[1]);
3005 }
3006 
get_numa_group(struct numa_group * grp)3007 static inline int get_numa_group(struct numa_group *grp)
3008 {
3009 	return refcount_inc_not_zero(&grp->refcount);
3010 }
3011 
put_numa_group(struct numa_group * grp)3012 static inline void put_numa_group(struct numa_group *grp)
3013 {
3014 	if (refcount_dec_and_test(&grp->refcount))
3015 		kfree_rcu(grp, rcu);
3016 }
3017 
task_numa_group(struct task_struct * p,int cpupid,int flags,int * priv)3018 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
3019 			int *priv)
3020 {
3021 	struct numa_group *grp, *my_grp;
3022 	struct task_struct *tsk;
3023 	bool join = false;
3024 	int cpu = cpupid_to_cpu(cpupid);
3025 	int i;
3026 
3027 	if (unlikely(!deref_curr_numa_group(p))) {
3028 		unsigned int size = sizeof(struct numa_group) +
3029 				    NR_NUMA_HINT_FAULT_STATS *
3030 				    nr_node_ids * sizeof(unsigned long);
3031 
3032 		grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
3033 		if (!grp)
3034 			return;
3035 
3036 		refcount_set(&grp->refcount, 1);
3037 		grp->active_nodes = 1;
3038 		grp->max_faults_cpu = 0;
3039 		spin_lock_init(&grp->lock);
3040 		grp->gid = p->pid;
3041 
3042 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3043 			grp->faults[i] = p->numa_faults[i];
3044 
3045 		grp->total_faults = p->total_numa_faults;
3046 
3047 		grp->nr_tasks++;
3048 		rcu_assign_pointer(p->numa_group, grp);
3049 	}
3050 
3051 	rcu_read_lock();
3052 	tsk = READ_ONCE(cpu_rq(cpu)->curr);
3053 
3054 	if (!cpupid_match_pid(tsk, cpupid))
3055 		goto no_join;
3056 
3057 	grp = rcu_dereference(tsk->numa_group);
3058 	if (!grp)
3059 		goto no_join;
3060 
3061 	my_grp = deref_curr_numa_group(p);
3062 	if (grp == my_grp)
3063 		goto no_join;
3064 
3065 	/*
3066 	 * Only join the other group if its bigger; if we're the bigger group,
3067 	 * the other task will join us.
3068 	 */
3069 	if (my_grp->nr_tasks > grp->nr_tasks)
3070 		goto no_join;
3071 
3072 	/*
3073 	 * Tie-break on the grp address.
3074 	 */
3075 	if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
3076 		goto no_join;
3077 
3078 	/* Always join threads in the same process. */
3079 	if (tsk->mm == current->mm)
3080 		join = true;
3081 
3082 	/* Simple filter to avoid false positives due to PID collisions */
3083 	if (flags & TNF_SHARED)
3084 		join = true;
3085 
3086 	/* Update priv based on whether false sharing was detected */
3087 	*priv = !join;
3088 
3089 	if (join && !get_numa_group(grp))
3090 		goto no_join;
3091 
3092 	rcu_read_unlock();
3093 
3094 	if (!join)
3095 		return;
3096 
3097 	WARN_ON_ONCE(irqs_disabled());
3098 	double_lock_irq(&my_grp->lock, &grp->lock);
3099 
3100 	for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
3101 		my_grp->faults[i] -= p->numa_faults[i];
3102 		grp->faults[i] += p->numa_faults[i];
3103 	}
3104 	my_grp->total_faults -= p->total_numa_faults;
3105 	grp->total_faults += p->total_numa_faults;
3106 
3107 	my_grp->nr_tasks--;
3108 	grp->nr_tasks++;
3109 
3110 	spin_unlock(&my_grp->lock);
3111 	spin_unlock_irq(&grp->lock);
3112 
3113 	rcu_assign_pointer(p->numa_group, grp);
3114 
3115 	put_numa_group(my_grp);
3116 	return;
3117 
3118 no_join:
3119 	rcu_read_unlock();
3120 	return;
3121 }
3122 
3123 /*
3124  * Get rid of NUMA statistics associated with a task (either current or dead).
3125  * If @final is set, the task is dead and has reached refcount zero, so we can
3126  * safely free all relevant data structures. Otherwise, there might be
3127  * concurrent reads from places like load balancing and procfs, and we should
3128  * reset the data back to default state without freeing ->numa_faults.
3129  */
task_numa_free(struct task_struct * p,bool final)3130 void task_numa_free(struct task_struct *p, bool final)
3131 {
3132 	/* safe: p either is current or is being freed by current */
3133 	struct numa_group *grp = rcu_dereference_raw(p->numa_group);
3134 	unsigned long *numa_faults = p->numa_faults;
3135 	unsigned long flags;
3136 	int i;
3137 
3138 	if (!numa_faults)
3139 		return;
3140 
3141 	if (grp) {
3142 		spin_lock_irqsave(&grp->lock, flags);
3143 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3144 			grp->faults[i] -= p->numa_faults[i];
3145 		grp->total_faults -= p->total_numa_faults;
3146 
3147 		grp->nr_tasks--;
3148 		spin_unlock_irqrestore(&grp->lock, flags);
3149 		RCU_INIT_POINTER(p->numa_group, NULL);
3150 		put_numa_group(grp);
3151 	}
3152 
3153 	if (final) {
3154 		p->numa_faults = NULL;
3155 		kfree(numa_faults);
3156 	} else {
3157 		p->total_numa_faults = 0;
3158 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3159 			numa_faults[i] = 0;
3160 	}
3161 }
3162 
3163 /*
3164  * Got a PROT_NONE fault for a page on @node.
3165  */
task_numa_fault(int last_cpupid,int mem_node,int pages,int flags)3166 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
3167 {
3168 	struct task_struct *p = current;
3169 	bool migrated = flags & TNF_MIGRATED;
3170 	int cpu_node = task_node(current);
3171 	int local = !!(flags & TNF_FAULT_LOCAL);
3172 	struct numa_group *ng;
3173 	int priv;
3174 
3175 	if (!static_branch_likely(&sched_numa_balancing))
3176 		return;
3177 
3178 	/* for example, ksmd faulting in a user's mm */
3179 	if (!p->mm)
3180 		return;
3181 
3182 	/*
3183 	 * NUMA faults statistics are unnecessary for the slow memory
3184 	 * node for memory tiering mode.
3185 	 */
3186 	if (!node_is_toptier(mem_node) &&
3187 	    (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
3188 	     !cpupid_valid(last_cpupid)))
3189 		return;
3190 
3191 	/* Allocate buffer to track faults on a per-node basis */
3192 	if (unlikely(!p->numa_faults)) {
3193 		int size = sizeof(*p->numa_faults) *
3194 			   NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
3195 
3196 		p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
3197 		if (!p->numa_faults)
3198 			return;
3199 
3200 		p->total_numa_faults = 0;
3201 		memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
3202 	}
3203 
3204 	/*
3205 	 * First accesses are treated as private, otherwise consider accesses
3206 	 * to be private if the accessing pid has not changed
3207 	 */
3208 	if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
3209 		priv = 1;
3210 	} else {
3211 		priv = cpupid_match_pid(p, last_cpupid);
3212 		if (!priv && !(flags & TNF_NO_GROUP))
3213 			task_numa_group(p, last_cpupid, flags, &priv);
3214 	}
3215 
3216 	/*
3217 	 * If a workload spans multiple NUMA nodes, a shared fault that
3218 	 * occurs wholly within the set of nodes that the workload is
3219 	 * actively using should be counted as local. This allows the
3220 	 * scan rate to slow down when a workload has settled down.
3221 	 */
3222 	ng = deref_curr_numa_group(p);
3223 	if (!priv && !local && ng && ng->active_nodes > 1 &&
3224 				numa_is_active_node(cpu_node, ng) &&
3225 				numa_is_active_node(mem_node, ng))
3226 		local = 1;
3227 
3228 	/*
3229 	 * Retry to migrate task to preferred node periodically, in case it
3230 	 * previously failed, or the scheduler moved us.
3231 	 */
3232 	if (time_after(jiffies, p->numa_migrate_retry)) {
3233 		task_numa_placement(p);
3234 		numa_migrate_preferred(p);
3235 	}
3236 
3237 	if (migrated)
3238 		p->numa_pages_migrated += pages;
3239 	if (flags & TNF_MIGRATE_FAIL)
3240 		p->numa_faults_locality[2] += pages;
3241 
3242 	p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
3243 	p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
3244 	p->numa_faults_locality[local] += pages;
3245 }
3246 
reset_ptenuma_scan(struct task_struct * p)3247 static void reset_ptenuma_scan(struct task_struct *p)
3248 {
3249 	/*
3250 	 * We only did a read acquisition of the mmap sem, so
3251 	 * p->mm->numa_scan_seq is written to without exclusive access
3252 	 * and the update is not guaranteed to be atomic. That's not
3253 	 * much of an issue though, since this is just used for
3254 	 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
3255 	 * expensive, to avoid any form of compiler optimizations:
3256 	 */
3257 	WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
3258 	p->mm->numa_scan_offset = 0;
3259 }
3260 
vma_is_accessed(struct mm_struct * mm,struct vm_area_struct * vma)3261 static bool vma_is_accessed(struct mm_struct *mm, struct vm_area_struct *vma)
3262 {
3263 	unsigned long pids;
3264 	/*
3265 	 * Allow unconditional access first two times, so that all the (pages)
3266 	 * of VMAs get prot_none fault introduced irrespective of accesses.
3267 	 * This is also done to avoid any side effect of task scanning
3268 	 * amplifying the unfairness of disjoint set of VMAs' access.
3269 	 */
3270 	if ((READ_ONCE(current->mm->numa_scan_seq) - vma->numab_state->start_scan_seq) < 2)
3271 		return true;
3272 
3273 	pids = vma->numab_state->pids_active[0] | vma->numab_state->pids_active[1];
3274 	if (test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids))
3275 		return true;
3276 
3277 	/*
3278 	 * Complete a scan that has already started regardless of PID access, or
3279 	 * some VMAs may never be scanned in multi-threaded applications:
3280 	 */
3281 	if (mm->numa_scan_offset > vma->vm_start) {
3282 		trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_IGNORE_PID);
3283 		return true;
3284 	}
3285 
3286 	/*
3287 	 * This vma has not been accessed for a while, and if the number
3288 	 * the threads in the same process is low, which means no other
3289 	 * threads can help scan this vma, force a vma scan.
3290 	 */
3291 	if (READ_ONCE(mm->numa_scan_seq) >
3292 	   (vma->numab_state->prev_scan_seq + get_nr_threads(current)))
3293 		return true;
3294 
3295 	return false;
3296 }
3297 
3298 #define VMA_PID_RESET_PERIOD (4 * sysctl_numa_balancing_scan_delay)
3299 
3300 /*
3301  * The expensive part of numa migration is done from task_work context.
3302  * Triggered from task_tick_numa().
3303  */
task_numa_work(struct callback_head * work)3304 static void task_numa_work(struct callback_head *work)
3305 {
3306 	unsigned long migrate, next_scan, now = jiffies;
3307 	struct task_struct *p = current;
3308 	struct mm_struct *mm = p->mm;
3309 	u64 runtime = p->se.sum_exec_runtime;
3310 	struct vm_area_struct *vma;
3311 	unsigned long start, end;
3312 	unsigned long nr_pte_updates = 0;
3313 	long pages, virtpages;
3314 	struct vma_iterator vmi;
3315 	bool vma_pids_skipped;
3316 	bool vma_pids_forced = false;
3317 
3318 	WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
3319 
3320 	work->next = work;
3321 	/*
3322 	 * Who cares about NUMA placement when they're dying.
3323 	 *
3324 	 * NOTE: make sure not to dereference p->mm before this check,
3325 	 * exit_task_work() happens _after_ exit_mm() so we could be called
3326 	 * without p->mm even though we still had it when we enqueued this
3327 	 * work.
3328 	 */
3329 	if (p->flags & PF_EXITING)
3330 		return;
3331 
3332 	if (!mm->numa_next_scan) {
3333 		mm->numa_next_scan = now +
3334 			msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3335 	}
3336 
3337 	/*
3338 	 * Enforce maximal scan/migration frequency..
3339 	 */
3340 	migrate = mm->numa_next_scan;
3341 	if (time_before(now, migrate))
3342 		return;
3343 
3344 	if (p->numa_scan_period == 0) {
3345 		p->numa_scan_period_max = task_scan_max(p);
3346 		p->numa_scan_period = task_scan_start(p);
3347 	}
3348 
3349 	next_scan = now + msecs_to_jiffies(p->numa_scan_period);
3350 	if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
3351 		return;
3352 
3353 	/*
3354 	 * Delay this task enough that another task of this mm will likely win
3355 	 * the next time around.
3356 	 */
3357 	p->node_stamp += 2 * TICK_NSEC;
3358 
3359 	pages = sysctl_numa_balancing_scan_size;
3360 	pages <<= 20 - PAGE_SHIFT; /* MB in pages */
3361 	virtpages = pages * 8;	   /* Scan up to this much virtual space */
3362 	if (!pages)
3363 		return;
3364 
3365 
3366 	if (!mmap_read_trylock(mm))
3367 		return;
3368 
3369 	/*
3370 	 * VMAs are skipped if the current PID has not trapped a fault within
3371 	 * the VMA recently. Allow scanning to be forced if there is no
3372 	 * suitable VMA remaining.
3373 	 */
3374 	vma_pids_skipped = false;
3375 
3376 retry_pids:
3377 	start = mm->numa_scan_offset;
3378 	vma_iter_init(&vmi, mm, start);
3379 	vma = vma_next(&vmi);
3380 	if (!vma) {
3381 		reset_ptenuma_scan(p);
3382 		start = 0;
3383 		vma_iter_set(&vmi, start);
3384 		vma = vma_next(&vmi);
3385 	}
3386 
3387 	for (; vma; vma = vma_next(&vmi)) {
3388 		if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
3389 			is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3390 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_UNSUITABLE);
3391 			continue;
3392 		}
3393 
3394 		/*
3395 		 * Shared library pages mapped by multiple processes are not
3396 		 * migrated as it is expected they are cache replicated. Avoid
3397 		 * hinting faults in read-only file-backed mappings or the vDSO
3398 		 * as migrating the pages will be of marginal benefit.
3399 		 */
3400 		if (!vma->vm_mm ||
3401 		    (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) {
3402 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SHARED_RO);
3403 			continue;
3404 		}
3405 
3406 		/*
3407 		 * Skip inaccessible VMAs to avoid any confusion between
3408 		 * PROT_NONE and NUMA hinting PTEs
3409 		 */
3410 		if (!vma_is_accessible(vma)) {
3411 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_INACCESSIBLE);
3412 			continue;
3413 		}
3414 
3415 		/* Initialise new per-VMA NUMAB state. */
3416 		if (!vma->numab_state) {
3417 			struct vma_numab_state *ptr;
3418 
3419 			ptr = kzalloc(sizeof(*ptr), GFP_KERNEL);
3420 			if (!ptr)
3421 				continue;
3422 
3423 			if (cmpxchg(&vma->numab_state, NULL, ptr)) {
3424 				kfree(ptr);
3425 				continue;
3426 			}
3427 
3428 			vma->numab_state->start_scan_seq = mm->numa_scan_seq;
3429 
3430 			vma->numab_state->next_scan = now +
3431 				msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3432 
3433 			/* Reset happens after 4 times scan delay of scan start */
3434 			vma->numab_state->pids_active_reset =  vma->numab_state->next_scan +
3435 				msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3436 
3437 			/*
3438 			 * Ensure prev_scan_seq does not match numa_scan_seq,
3439 			 * to prevent VMAs being skipped prematurely on the
3440 			 * first scan:
3441 			 */
3442 			 vma->numab_state->prev_scan_seq = mm->numa_scan_seq - 1;
3443 		}
3444 
3445 		/*
3446 		 * Scanning the VMAs of short lived tasks add more overhead. So
3447 		 * delay the scan for new VMAs.
3448 		 */
3449 		if (mm->numa_scan_seq && time_before(jiffies,
3450 						vma->numab_state->next_scan)) {
3451 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SCAN_DELAY);
3452 			continue;
3453 		}
3454 
3455 		/* RESET access PIDs regularly for old VMAs. */
3456 		if (mm->numa_scan_seq &&
3457 				time_after(jiffies, vma->numab_state->pids_active_reset)) {
3458 			vma->numab_state->pids_active_reset = vma->numab_state->pids_active_reset +
3459 				msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3460 			vma->numab_state->pids_active[0] = READ_ONCE(vma->numab_state->pids_active[1]);
3461 			vma->numab_state->pids_active[1] = 0;
3462 		}
3463 
3464 		/* Do not rescan VMAs twice within the same sequence. */
3465 		if (vma->numab_state->prev_scan_seq == mm->numa_scan_seq) {
3466 			mm->numa_scan_offset = vma->vm_end;
3467 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SEQ_COMPLETED);
3468 			continue;
3469 		}
3470 
3471 		/*
3472 		 * Do not scan the VMA if task has not accessed it, unless no other
3473 		 * VMA candidate exists.
3474 		 */
3475 		if (!vma_pids_forced && !vma_is_accessed(mm, vma)) {
3476 			vma_pids_skipped = true;
3477 			trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_PID_INACTIVE);
3478 			continue;
3479 		}
3480 
3481 		do {
3482 			start = max(start, vma->vm_start);
3483 			end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3484 			end = min(end, vma->vm_end);
3485 			nr_pte_updates = change_prot_numa(vma, start, end);
3486 
3487 			/*
3488 			 * Try to scan sysctl_numa_balancing_size worth of
3489 			 * hpages that have at least one present PTE that
3490 			 * is not already PTE-numa. If the VMA contains
3491 			 * areas that are unused or already full of prot_numa
3492 			 * PTEs, scan up to virtpages, to skip through those
3493 			 * areas faster.
3494 			 */
3495 			if (nr_pte_updates)
3496 				pages -= (end - start) >> PAGE_SHIFT;
3497 			virtpages -= (end - start) >> PAGE_SHIFT;
3498 
3499 			start = end;
3500 			if (pages <= 0 || virtpages <= 0)
3501 				goto out;
3502 
3503 			cond_resched();
3504 		} while (end != vma->vm_end);
3505 
3506 		/* VMA scan is complete, do not scan until next sequence. */
3507 		vma->numab_state->prev_scan_seq = mm->numa_scan_seq;
3508 
3509 		/*
3510 		 * Only force scan within one VMA at a time, to limit the
3511 		 * cost of scanning a potentially uninteresting VMA.
3512 		 */
3513 		if (vma_pids_forced)
3514 			break;
3515 	}
3516 
3517 	/*
3518 	 * If no VMAs are remaining and VMAs were skipped due to the PID
3519 	 * not accessing the VMA previously, then force a scan to ensure
3520 	 * forward progress:
3521 	 */
3522 	if (!vma && !vma_pids_forced && vma_pids_skipped) {
3523 		vma_pids_forced = true;
3524 		goto retry_pids;
3525 	}
3526 
3527 out:
3528 	/*
3529 	 * It is possible to reach the end of the VMA list but the last few
3530 	 * VMAs are not guaranteed to the vma_migratable. If they are not, we
3531 	 * would find the !migratable VMA on the next scan but not reset the
3532 	 * scanner to the start so check it now.
3533 	 */
3534 	if (vma)
3535 		mm->numa_scan_offset = start;
3536 	else
3537 		reset_ptenuma_scan(p);
3538 	mmap_read_unlock(mm);
3539 
3540 	/*
3541 	 * Make sure tasks use at least 32x as much time to run other code
3542 	 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3543 	 * Usually update_task_scan_period slows down scanning enough; on an
3544 	 * overloaded system we need to limit overhead on a per task basis.
3545 	 */
3546 	if (unlikely(p->se.sum_exec_runtime != runtime)) {
3547 		u64 diff = p->se.sum_exec_runtime - runtime;
3548 		p->node_stamp += 32 * diff;
3549 	}
3550 }
3551 
init_numa_balancing(unsigned long clone_flags,struct task_struct * p)3552 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3553 {
3554 	int mm_users = 0;
3555 	struct mm_struct *mm = p->mm;
3556 
3557 	if (mm) {
3558 		mm_users = atomic_read(&mm->mm_users);
3559 		if (mm_users == 1) {
3560 			mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3561 			mm->numa_scan_seq = 0;
3562 		}
3563 	}
3564 	p->node_stamp			= 0;
3565 	p->numa_scan_seq		= mm ? mm->numa_scan_seq : 0;
3566 	p->numa_scan_period		= sysctl_numa_balancing_scan_delay;
3567 	p->numa_migrate_retry		= 0;
3568 	/* Protect against double add, see task_tick_numa and task_numa_work */
3569 	p->numa_work.next		= &p->numa_work;
3570 	p->numa_faults			= NULL;
3571 	p->numa_pages_migrated		= 0;
3572 	p->total_numa_faults		= 0;
3573 	RCU_INIT_POINTER(p->numa_group, NULL);
3574 	p->last_task_numa_placement	= 0;
3575 	p->last_sum_exec_runtime	= 0;
3576 
3577 	init_task_work(&p->numa_work, task_numa_work);
3578 
3579 	/* New address space, reset the preferred nid */
3580 	if (!(clone_flags & CLONE_VM)) {
3581 		p->numa_preferred_nid = NUMA_NO_NODE;
3582 		return;
3583 	}
3584 
3585 	/*
3586 	 * New thread, keep existing numa_preferred_nid which should be copied
3587 	 * already by arch_dup_task_struct but stagger when scans start.
3588 	 */
3589 	if (mm) {
3590 		unsigned int delay;
3591 
3592 		delay = min_t(unsigned int, task_scan_max(current),
3593 			current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3594 		delay += 2 * TICK_NSEC;
3595 		p->node_stamp = delay;
3596 	}
3597 }
3598 
3599 /*
3600  * Drive the periodic memory faults..
3601  */
task_tick_numa(struct rq * rq,struct task_struct * curr)3602 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3603 {
3604 	struct callback_head *work = &curr->numa_work;
3605 	u64 period, now;
3606 
3607 	/*
3608 	 * We don't care about NUMA placement if we don't have memory.
3609 	 */
3610 	if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3611 		return;
3612 
3613 	/*
3614 	 * Using runtime rather than walltime has the dual advantage that
3615 	 * we (mostly) drive the selection from busy threads and that the
3616 	 * task needs to have done some actual work before we bother with
3617 	 * NUMA placement.
3618 	 */
3619 	now = curr->se.sum_exec_runtime;
3620 	period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3621 
3622 	if (now > curr->node_stamp + period) {
3623 		if (!curr->node_stamp)
3624 			curr->numa_scan_period = task_scan_start(curr);
3625 		curr->node_stamp += period;
3626 
3627 		if (!time_before(jiffies, curr->mm->numa_next_scan))
3628 			task_work_add(curr, work, TWA_RESUME);
3629 	}
3630 }
3631 
update_scan_period(struct task_struct * p,int new_cpu)3632 static void update_scan_period(struct task_struct *p, int new_cpu)
3633 {
3634 	int src_nid = cpu_to_node(task_cpu(p));
3635 	int dst_nid = cpu_to_node(new_cpu);
3636 
3637 	if (!static_branch_likely(&sched_numa_balancing))
3638 		return;
3639 
3640 	if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3641 		return;
3642 
3643 	if (src_nid == dst_nid)
3644 		return;
3645 
3646 	/*
3647 	 * Allow resets if faults have been trapped before one scan
3648 	 * has completed. This is most likely due to a new task that
3649 	 * is pulled cross-node due to wakeups or load balancing.
3650 	 */
3651 	if (p->numa_scan_seq) {
3652 		/*
3653 		 * Avoid scan adjustments if moving to the preferred
3654 		 * node or if the task was not previously running on
3655 		 * the preferred node.
3656 		 */
3657 		if (dst_nid == p->numa_preferred_nid ||
3658 		    (p->numa_preferred_nid != NUMA_NO_NODE &&
3659 			src_nid != p->numa_preferred_nid))
3660 			return;
3661 	}
3662 
3663 	p->numa_scan_period = task_scan_start(p);
3664 }
3665 
3666 #else
task_tick_numa(struct rq * rq,struct task_struct * curr)3667 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3668 {
3669 }
3670 
account_numa_enqueue(struct rq * rq,struct task_struct * p)3671 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3672 {
3673 }
3674 
account_numa_dequeue(struct rq * rq,struct task_struct * p)3675 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3676 {
3677 }
3678 
update_scan_period(struct task_struct * p,int new_cpu)3679 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3680 {
3681 }
3682 
3683 #endif /* CONFIG_NUMA_BALANCING */
3684 
3685 static void
account_entity_enqueue(struct cfs_rq * cfs_rq,struct sched_entity * se)3686 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3687 {
3688 	update_load_add(&cfs_rq->load, se->load.weight);
3689 #ifdef CONFIG_SMP
3690 	if (entity_is_task(se)) {
3691 		struct rq *rq = rq_of(cfs_rq);
3692 
3693 		account_numa_enqueue(rq, task_of(se));
3694 		list_add(&se->group_node, &rq->cfs_tasks);
3695 	}
3696 #endif
3697 	cfs_rq->nr_queued++;
3698 }
3699 
3700 static void
account_entity_dequeue(struct cfs_rq * cfs_rq,struct sched_entity * se)3701 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3702 {
3703 	update_load_sub(&cfs_rq->load, se->load.weight);
3704 #ifdef CONFIG_SMP
3705 	if (entity_is_task(se)) {
3706 		account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3707 		list_del_init(&se->group_node);
3708 	}
3709 #endif
3710 	cfs_rq->nr_queued--;
3711 }
3712 
3713 /*
3714  * Signed add and clamp on underflow.
3715  *
3716  * Explicitly do a load-store to ensure the intermediate value never hits
3717  * memory. This allows lockless observations without ever seeing the negative
3718  * values.
3719  */
3720 #define add_positive(_ptr, _val) do {                           \
3721 	typeof(_ptr) ptr = (_ptr);                              \
3722 	typeof(_val) val = (_val);                              \
3723 	typeof(*ptr) res, var = READ_ONCE(*ptr);                \
3724 								\
3725 	res = var + val;                                        \
3726 								\
3727 	if (val < 0 && res > var)                               \
3728 		res = 0;                                        \
3729 								\
3730 	WRITE_ONCE(*ptr, res);                                  \
3731 } while (0)
3732 
3733 /*
3734  * Unsigned subtract and clamp on underflow.
3735  *
3736  * Explicitly do a load-store to ensure the intermediate value never hits
3737  * memory. This allows lockless observations without ever seeing the negative
3738  * values.
3739  */
3740 #define sub_positive(_ptr, _val) do {				\
3741 	typeof(_ptr) ptr = (_ptr);				\
3742 	typeof(*ptr) val = (_val);				\
3743 	typeof(*ptr) res, var = READ_ONCE(*ptr);		\
3744 	res = var - val;					\
3745 	if (res > var)						\
3746 		res = 0;					\
3747 	WRITE_ONCE(*ptr, res);					\
3748 } while (0)
3749 
3750 /*
3751  * Remove and clamp on negative, from a local variable.
3752  *
3753  * A variant of sub_positive(), which does not use explicit load-store
3754  * and is thus optimized for local variable updates.
3755  */
3756 #define lsub_positive(_ptr, _val) do {				\
3757 	typeof(_ptr) ptr = (_ptr);				\
3758 	*ptr -= min_t(typeof(*ptr), *ptr, _val);		\
3759 } while (0)
3760 
3761 #ifdef CONFIG_SMP
3762 static inline void
enqueue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3763 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3764 {
3765 	cfs_rq->avg.load_avg += se->avg.load_avg;
3766 	cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3767 }
3768 
3769 static inline void
dequeue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3770 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3771 {
3772 	sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3773 	sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3774 	/* See update_cfs_rq_load_avg() */
3775 	cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3776 					  cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3777 }
3778 #else
3779 static inline void
enqueue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3780 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3781 static inline void
dequeue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3782 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3783 #endif
3784 
3785 static void place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags);
3786 
reweight_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,unsigned long weight)3787 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3788 			    unsigned long weight)
3789 {
3790 	bool curr = cfs_rq->curr == se;
3791 
3792 	if (se->on_rq) {
3793 		/* commit outstanding execution time */
3794 		update_curr(cfs_rq);
3795 		update_entity_lag(cfs_rq, se);
3796 		se->deadline -= se->vruntime;
3797 		se->rel_deadline = 1;
3798 		if (!curr)
3799 			__dequeue_entity(cfs_rq, se);
3800 		update_load_sub(&cfs_rq->load, se->load.weight);
3801 	}
3802 	dequeue_load_avg(cfs_rq, se);
3803 
3804 	/*
3805 	 * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i),
3806 	 * we need to scale se->vlag when w_i changes.
3807 	 */
3808 	se->vlag = div_s64(se->vlag * se->load.weight, weight);
3809 	if (se->rel_deadline)
3810 		se->deadline = div_s64(se->deadline * se->load.weight, weight);
3811 
3812 	update_load_set(&se->load, weight);
3813 
3814 #ifdef CONFIG_SMP
3815 	do {
3816 		u32 divider = get_pelt_divider(&se->avg);
3817 
3818 		se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3819 	} while (0);
3820 #endif
3821 
3822 	enqueue_load_avg(cfs_rq, se);
3823 	if (se->on_rq) {
3824 		update_load_add(&cfs_rq->load, se->load.weight);
3825 		place_entity(cfs_rq, se, 0);
3826 		if (!curr)
3827 			__enqueue_entity(cfs_rq, se);
3828 
3829 		/*
3830 		 * The entity's vruntime has been adjusted, so let's check
3831 		 * whether the rq-wide min_vruntime needs updated too. Since
3832 		 * the calculations above require stable min_vruntime rather
3833 		 * than up-to-date one, we do the update at the end of the
3834 		 * reweight process.
3835 		 */
3836 		update_min_vruntime(cfs_rq);
3837 	}
3838 }
3839 
reweight_task_fair(struct rq * rq,struct task_struct * p,const struct load_weight * lw)3840 static void reweight_task_fair(struct rq *rq, struct task_struct *p,
3841 			       const struct load_weight *lw)
3842 {
3843 	struct sched_entity *se = &p->se;
3844 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
3845 	struct load_weight *load = &se->load;
3846 
3847 	reweight_entity(cfs_rq, se, lw->weight);
3848 	load->inv_weight = lw->inv_weight;
3849 }
3850 
3851 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3852 
3853 #ifdef CONFIG_FAIR_GROUP_SCHED
3854 #ifdef CONFIG_SMP
3855 /*
3856  * All this does is approximate the hierarchical proportion which includes that
3857  * global sum we all love to hate.
3858  *
3859  * That is, the weight of a group entity, is the proportional share of the
3860  * group weight based on the group runqueue weights. That is:
3861  *
3862  *                     tg->weight * grq->load.weight
3863  *   ge->load.weight = -----------------------------               (1)
3864  *                       \Sum grq->load.weight
3865  *
3866  * Now, because computing that sum is prohibitively expensive to compute (been
3867  * there, done that) we approximate it with this average stuff. The average
3868  * moves slower and therefore the approximation is cheaper and more stable.
3869  *
3870  * So instead of the above, we substitute:
3871  *
3872  *   grq->load.weight -> grq->avg.load_avg                         (2)
3873  *
3874  * which yields the following:
3875  *
3876  *                     tg->weight * grq->avg.load_avg
3877  *   ge->load.weight = ------------------------------              (3)
3878  *                             tg->load_avg
3879  *
3880  * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3881  *
3882  * That is shares_avg, and it is right (given the approximation (2)).
3883  *
3884  * The problem with it is that because the average is slow -- it was designed
3885  * to be exactly that of course -- this leads to transients in boundary
3886  * conditions. In specific, the case where the group was idle and we start the
3887  * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3888  * yielding bad latency etc..
3889  *
3890  * Now, in that special case (1) reduces to:
3891  *
3892  *                     tg->weight * grq->load.weight
3893  *   ge->load.weight = ----------------------------- = tg->weight   (4)
3894  *                         grp->load.weight
3895  *
3896  * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3897  *
3898  * So what we do is modify our approximation (3) to approach (4) in the (near)
3899  * UP case, like:
3900  *
3901  *   ge->load.weight =
3902  *
3903  *              tg->weight * grq->load.weight
3904  *     ---------------------------------------------------         (5)
3905  *     tg->load_avg - grq->avg.load_avg + grq->load.weight
3906  *
3907  * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3908  * we need to use grq->avg.load_avg as its lower bound, which then gives:
3909  *
3910  *
3911  *                     tg->weight * grq->load.weight
3912  *   ge->load.weight = -----------------------------		   (6)
3913  *                             tg_load_avg'
3914  *
3915  * Where:
3916  *
3917  *   tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3918  *                  max(grq->load.weight, grq->avg.load_avg)
3919  *
3920  * And that is shares_weight and is icky. In the (near) UP case it approaches
3921  * (4) while in the normal case it approaches (3). It consistently
3922  * overestimates the ge->load.weight and therefore:
3923  *
3924  *   \Sum ge->load.weight >= tg->weight
3925  *
3926  * hence icky!
3927  */
calc_group_shares(struct cfs_rq * cfs_rq)3928 static long calc_group_shares(struct cfs_rq *cfs_rq)
3929 {
3930 	long tg_weight, tg_shares, load, shares;
3931 	struct task_group *tg = cfs_rq->tg;
3932 
3933 	tg_shares = READ_ONCE(tg->shares);
3934 
3935 	load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3936 
3937 	tg_weight = atomic_long_read(&tg->load_avg);
3938 
3939 	/* Ensure tg_weight >= load */
3940 	tg_weight -= cfs_rq->tg_load_avg_contrib;
3941 	tg_weight += load;
3942 
3943 	shares = (tg_shares * load);
3944 	if (tg_weight)
3945 		shares /= tg_weight;
3946 
3947 	/*
3948 	 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3949 	 * of a group with small tg->shares value. It is a floor value which is
3950 	 * assigned as a minimum load.weight to the sched_entity representing
3951 	 * the group on a CPU.
3952 	 *
3953 	 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3954 	 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3955 	 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3956 	 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3957 	 * instead of 0.
3958 	 */
3959 	return clamp_t(long, shares, MIN_SHARES, tg_shares);
3960 }
3961 #endif /* CONFIG_SMP */
3962 
3963 /*
3964  * Recomputes the group entity based on the current state of its group
3965  * runqueue.
3966  */
update_cfs_group(struct sched_entity * se)3967 static void update_cfs_group(struct sched_entity *se)
3968 {
3969 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3970 	long shares;
3971 
3972 	/*
3973 	 * When a group becomes empty, preserve its weight. This matters for
3974 	 * DELAY_DEQUEUE.
3975 	 */
3976 	if (!gcfs_rq || !gcfs_rq->load.weight)
3977 		return;
3978 
3979 	if (throttled_hierarchy(gcfs_rq))
3980 		return;
3981 
3982 #ifndef CONFIG_SMP
3983 	shares = READ_ONCE(gcfs_rq->tg->shares);
3984 #else
3985 	shares = calc_group_shares(gcfs_rq);
3986 #endif
3987 	if (unlikely(se->load.weight != shares))
3988 		reweight_entity(cfs_rq_of(se), se, shares);
3989 }
3990 
3991 #else /* CONFIG_FAIR_GROUP_SCHED */
update_cfs_group(struct sched_entity * se)3992 static inline void update_cfs_group(struct sched_entity *se)
3993 {
3994 }
3995 #endif /* CONFIG_FAIR_GROUP_SCHED */
3996 
cfs_rq_util_change(struct cfs_rq * cfs_rq,int flags)3997 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3998 {
3999 	struct rq *rq = rq_of(cfs_rq);
4000 
4001 	if (&rq->cfs == cfs_rq) {
4002 		/*
4003 		 * There are a few boundary cases this might miss but it should
4004 		 * get called often enough that that should (hopefully) not be
4005 		 * a real problem.
4006 		 *
4007 		 * It will not get called when we go idle, because the idle
4008 		 * thread is a different class (!fair), nor will the utilization
4009 		 * number include things like RT tasks.
4010 		 *
4011 		 * As is, the util number is not freq-invariant (we'd have to
4012 		 * implement arch_scale_freq_capacity() for that).
4013 		 *
4014 		 * See cpu_util_cfs().
4015 		 */
4016 		cpufreq_update_util(rq, flags);
4017 	}
4018 }
4019 
4020 #ifdef CONFIG_SMP
load_avg_is_decayed(struct sched_avg * sa)4021 static inline bool load_avg_is_decayed(struct sched_avg *sa)
4022 {
4023 	if (sa->load_sum)
4024 		return false;
4025 
4026 	if (sa->util_sum)
4027 		return false;
4028 
4029 	if (sa->runnable_sum)
4030 		return false;
4031 
4032 	/*
4033 	 * _avg must be null when _sum are null because _avg = _sum / divider
4034 	 * Make sure that rounding and/or propagation of PELT values never
4035 	 * break this.
4036 	 */
4037 	WARN_ON_ONCE(sa->load_avg ||
4038 		      sa->util_avg ||
4039 		      sa->runnable_avg);
4040 
4041 	return true;
4042 }
4043 
cfs_rq_last_update_time(struct cfs_rq * cfs_rq)4044 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
4045 {
4046 	return u64_u32_load_copy(cfs_rq->avg.last_update_time,
4047 				 cfs_rq->last_update_time_copy);
4048 }
4049 #ifdef CONFIG_FAIR_GROUP_SCHED
4050 /*
4051  * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
4052  * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
4053  * bottom-up, we only have to test whether the cfs_rq before us on the list
4054  * is our child.
4055  * If cfs_rq is not on the list, test whether a child needs its to be added to
4056  * connect a branch to the tree  * (see list_add_leaf_cfs_rq() for details).
4057  */
child_cfs_rq_on_list(struct cfs_rq * cfs_rq)4058 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
4059 {
4060 	struct cfs_rq *prev_cfs_rq;
4061 	struct list_head *prev;
4062 	struct rq *rq = rq_of(cfs_rq);
4063 
4064 	if (cfs_rq->on_list) {
4065 		prev = cfs_rq->leaf_cfs_rq_list.prev;
4066 	} else {
4067 		prev = rq->tmp_alone_branch;
4068 	}
4069 
4070 	if (prev == &rq->leaf_cfs_rq_list)
4071 		return false;
4072 
4073 	prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
4074 
4075 	return (prev_cfs_rq->tg->parent == cfs_rq->tg);
4076 }
4077 
cfs_rq_is_decayed(struct cfs_rq * cfs_rq)4078 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4079 {
4080 	if (cfs_rq->load.weight)
4081 		return false;
4082 
4083 	if (!load_avg_is_decayed(&cfs_rq->avg))
4084 		return false;
4085 
4086 	if (child_cfs_rq_on_list(cfs_rq))
4087 		return false;
4088 
4089 	return true;
4090 }
4091 
4092 /**
4093  * update_tg_load_avg - update the tg's load avg
4094  * @cfs_rq: the cfs_rq whose avg changed
4095  *
4096  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
4097  * However, because tg->load_avg is a global value there are performance
4098  * considerations.
4099  *
4100  * In order to avoid having to look at the other cfs_rq's, we use a
4101  * differential update where we store the last value we propagated. This in
4102  * turn allows skipping updates if the differential is 'small'.
4103  *
4104  * Updating tg's load_avg is necessary before update_cfs_share().
4105  */
update_tg_load_avg(struct cfs_rq * cfs_rq)4106 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
4107 {
4108 	long delta;
4109 	u64 now;
4110 
4111 	/*
4112 	 * No need to update load_avg for root_task_group as it is not used.
4113 	 */
4114 	if (cfs_rq->tg == &root_task_group)
4115 		return;
4116 
4117 	/* rq has been offline and doesn't contribute to the share anymore: */
4118 	if (!cpu_active(cpu_of(rq_of(cfs_rq))))
4119 		return;
4120 
4121 	/*
4122 	 * For migration heavy workloads, access to tg->load_avg can be
4123 	 * unbound. Limit the update rate to at most once per ms.
4124 	 */
4125 	now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
4126 	if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC)
4127 		return;
4128 
4129 	delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
4130 	if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
4131 		atomic_long_add(delta, &cfs_rq->tg->load_avg);
4132 		cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
4133 		cfs_rq->last_update_tg_load_avg = now;
4134 	}
4135 }
4136 
clear_tg_load_avg(struct cfs_rq * cfs_rq)4137 static inline void clear_tg_load_avg(struct cfs_rq *cfs_rq)
4138 {
4139 	long delta;
4140 	u64 now;
4141 
4142 	/*
4143 	 * No need to update load_avg for root_task_group, as it is not used.
4144 	 */
4145 	if (cfs_rq->tg == &root_task_group)
4146 		return;
4147 
4148 	now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
4149 	delta = 0 - cfs_rq->tg_load_avg_contrib;
4150 	atomic_long_add(delta, &cfs_rq->tg->load_avg);
4151 	cfs_rq->tg_load_avg_contrib = 0;
4152 	cfs_rq->last_update_tg_load_avg = now;
4153 }
4154 
4155 /* CPU offline callback: */
clear_tg_offline_cfs_rqs(struct rq * rq)4156 static void __maybe_unused clear_tg_offline_cfs_rqs(struct rq *rq)
4157 {
4158 	struct task_group *tg;
4159 
4160 	lockdep_assert_rq_held(rq);
4161 
4162 	/*
4163 	 * The rq clock has already been updated in
4164 	 * set_rq_offline(), so we should skip updating
4165 	 * the rq clock again in unthrottle_cfs_rq().
4166 	 */
4167 	rq_clock_start_loop_update(rq);
4168 
4169 	rcu_read_lock();
4170 	list_for_each_entry_rcu(tg, &task_groups, list) {
4171 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4172 
4173 		clear_tg_load_avg(cfs_rq);
4174 	}
4175 	rcu_read_unlock();
4176 
4177 	rq_clock_stop_loop_update(rq);
4178 }
4179 
4180 /*
4181  * Called within set_task_rq() right before setting a task's CPU. The
4182  * caller only guarantees p->pi_lock is held; no other assumptions,
4183  * including the state of rq->lock, should be made.
4184  */
set_task_rq_fair(struct sched_entity * se,struct cfs_rq * prev,struct cfs_rq * next)4185 void set_task_rq_fair(struct sched_entity *se,
4186 		      struct cfs_rq *prev, struct cfs_rq *next)
4187 {
4188 	u64 p_last_update_time;
4189 	u64 n_last_update_time;
4190 
4191 	if (!sched_feat(ATTACH_AGE_LOAD))
4192 		return;
4193 
4194 	/*
4195 	 * We are supposed to update the task to "current" time, then its up to
4196 	 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
4197 	 * getting what current time is, so simply throw away the out-of-date
4198 	 * time. This will result in the wakee task is less decayed, but giving
4199 	 * the wakee more load sounds not bad.
4200 	 */
4201 	if (!(se->avg.last_update_time && prev))
4202 		return;
4203 
4204 	p_last_update_time = cfs_rq_last_update_time(prev);
4205 	n_last_update_time = cfs_rq_last_update_time(next);
4206 
4207 	__update_load_avg_blocked_se(p_last_update_time, se);
4208 	se->avg.last_update_time = n_last_update_time;
4209 }
4210 
4211 /*
4212  * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
4213  * propagate its contribution. The key to this propagation is the invariant
4214  * that for each group:
4215  *
4216  *   ge->avg == grq->avg						(1)
4217  *
4218  * _IFF_ we look at the pure running and runnable sums. Because they
4219  * represent the very same entity, just at different points in the hierarchy.
4220  *
4221  * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
4222  * and simply copies the running/runnable sum over (but still wrong, because
4223  * the group entity and group rq do not have their PELT windows aligned).
4224  *
4225  * However, update_tg_cfs_load() is more complex. So we have:
4226  *
4227  *   ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg		(2)
4228  *
4229  * And since, like util, the runnable part should be directly transferable,
4230  * the following would _appear_ to be the straight forward approach:
4231  *
4232  *   grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg	(3)
4233  *
4234  * And per (1) we have:
4235  *
4236  *   ge->avg.runnable_avg == grq->avg.runnable_avg
4237  *
4238  * Which gives:
4239  *
4240  *                      ge->load.weight * grq->avg.load_avg
4241  *   ge->avg.load_avg = -----------------------------------		(4)
4242  *                               grq->load.weight
4243  *
4244  * Except that is wrong!
4245  *
4246  * Because while for entities historical weight is not important and we
4247  * really only care about our future and therefore can consider a pure
4248  * runnable sum, runqueues can NOT do this.
4249  *
4250  * We specifically want runqueues to have a load_avg that includes
4251  * historical weights. Those represent the blocked load, the load we expect
4252  * to (shortly) return to us. This only works by keeping the weights as
4253  * integral part of the sum. We therefore cannot decompose as per (3).
4254  *
4255  * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4256  * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
4257  * rq itself is runnable anywhere between 2/3 and 1 depending on how the
4258  * runnable section of these tasks overlap (or not). If they were to perfectly
4259  * align the rq as a whole would be runnable 2/3 of the time. If however we
4260  * always have at least 1 runnable task, the rq as a whole is always runnable.
4261  *
4262  * So we'll have to approximate.. :/
4263  *
4264  * Given the constraint:
4265  *
4266  *   ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4267  *
4268  * We can construct a rule that adds runnable to a rq by assuming minimal
4269  * overlap.
4270  *
4271  * On removal, we'll assume each task is equally runnable; which yields:
4272  *
4273  *   grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4274  *
4275  * XXX: only do this for the part of runnable > running ?
4276  *
4277  */
4278 static inline void
update_tg_cfs_util(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4279 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4280 {
4281 	long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
4282 	u32 new_sum, divider;
4283 
4284 	/* Nothing to update */
4285 	if (!delta_avg)
4286 		return;
4287 
4288 	/*
4289 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4290 	 * See ___update_load_avg() for details.
4291 	 */
4292 	divider = get_pelt_divider(&cfs_rq->avg);
4293 
4294 
4295 	/* Set new sched_entity's utilization */
4296 	se->avg.util_avg = gcfs_rq->avg.util_avg;
4297 	new_sum = se->avg.util_avg * divider;
4298 	delta_sum = (long)new_sum - (long)se->avg.util_sum;
4299 	se->avg.util_sum = new_sum;
4300 
4301 	/* Update parent cfs_rq utilization */
4302 	add_positive(&cfs_rq->avg.util_avg, delta_avg);
4303 	add_positive(&cfs_rq->avg.util_sum, delta_sum);
4304 
4305 	/* See update_cfs_rq_load_avg() */
4306 	cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4307 					  cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4308 }
4309 
4310 static inline void
update_tg_cfs_runnable(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4311 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4312 {
4313 	long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
4314 	u32 new_sum, divider;
4315 
4316 	/* Nothing to update */
4317 	if (!delta_avg)
4318 		return;
4319 
4320 	/*
4321 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4322 	 * See ___update_load_avg() for details.
4323 	 */
4324 	divider = get_pelt_divider(&cfs_rq->avg);
4325 
4326 	/* Set new sched_entity's runnable */
4327 	se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
4328 	new_sum = se->avg.runnable_avg * divider;
4329 	delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
4330 	se->avg.runnable_sum = new_sum;
4331 
4332 	/* Update parent cfs_rq runnable */
4333 	add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
4334 	add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
4335 	/* See update_cfs_rq_load_avg() */
4336 	cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4337 					      cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4338 }
4339 
4340 static inline void
update_tg_cfs_load(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4341 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4342 {
4343 	long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
4344 	unsigned long load_avg;
4345 	u64 load_sum = 0;
4346 	s64 delta_sum;
4347 	u32 divider;
4348 
4349 	if (!runnable_sum)
4350 		return;
4351 
4352 	gcfs_rq->prop_runnable_sum = 0;
4353 
4354 	/*
4355 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4356 	 * See ___update_load_avg() for details.
4357 	 */
4358 	divider = get_pelt_divider(&cfs_rq->avg);
4359 
4360 	if (runnable_sum >= 0) {
4361 		/*
4362 		 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
4363 		 * the CPU is saturated running == runnable.
4364 		 */
4365 		runnable_sum += se->avg.load_sum;
4366 		runnable_sum = min_t(long, runnable_sum, divider);
4367 	} else {
4368 		/*
4369 		 * Estimate the new unweighted runnable_sum of the gcfs_rq by
4370 		 * assuming all tasks are equally runnable.
4371 		 */
4372 		if (scale_load_down(gcfs_rq->load.weight)) {
4373 			load_sum = div_u64(gcfs_rq->avg.load_sum,
4374 				scale_load_down(gcfs_rq->load.weight));
4375 		}
4376 
4377 		/* But make sure to not inflate se's runnable */
4378 		runnable_sum = min(se->avg.load_sum, load_sum);
4379 	}
4380 
4381 	/*
4382 	 * runnable_sum can't be lower than running_sum
4383 	 * Rescale running sum to be in the same range as runnable sum
4384 	 * running_sum is in [0 : LOAD_AVG_MAX <<  SCHED_CAPACITY_SHIFT]
4385 	 * runnable_sum is in [0 : LOAD_AVG_MAX]
4386 	 */
4387 	running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
4388 	runnable_sum = max(runnable_sum, running_sum);
4389 
4390 	load_sum = se_weight(se) * runnable_sum;
4391 	load_avg = div_u64(load_sum, divider);
4392 
4393 	delta_avg = load_avg - se->avg.load_avg;
4394 	if (!delta_avg)
4395 		return;
4396 
4397 	delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
4398 
4399 	se->avg.load_sum = runnable_sum;
4400 	se->avg.load_avg = load_avg;
4401 	add_positive(&cfs_rq->avg.load_avg, delta_avg);
4402 	add_positive(&cfs_rq->avg.load_sum, delta_sum);
4403 	/* See update_cfs_rq_load_avg() */
4404 	cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
4405 					  cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
4406 }
4407 
add_tg_cfs_propagate(struct cfs_rq * cfs_rq,long runnable_sum)4408 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
4409 {
4410 	cfs_rq->propagate = 1;
4411 	cfs_rq->prop_runnable_sum += runnable_sum;
4412 }
4413 
4414 /* Update task and its cfs_rq load average */
propagate_entity_load_avg(struct sched_entity * se)4415 static inline int propagate_entity_load_avg(struct sched_entity *se)
4416 {
4417 	struct cfs_rq *cfs_rq, *gcfs_rq;
4418 
4419 	if (entity_is_task(se))
4420 		return 0;
4421 
4422 	gcfs_rq = group_cfs_rq(se);
4423 	if (!gcfs_rq->propagate)
4424 		return 0;
4425 
4426 	gcfs_rq->propagate = 0;
4427 
4428 	cfs_rq = cfs_rq_of(se);
4429 
4430 	add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
4431 
4432 	update_tg_cfs_util(cfs_rq, se, gcfs_rq);
4433 	update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
4434 	update_tg_cfs_load(cfs_rq, se, gcfs_rq);
4435 
4436 	trace_pelt_cfs_tp(cfs_rq);
4437 	trace_pelt_se_tp(se);
4438 
4439 	return 1;
4440 }
4441 
4442 /*
4443  * Check if we need to update the load and the utilization of a blocked
4444  * group_entity:
4445  */
skip_blocked_update(struct sched_entity * se)4446 static inline bool skip_blocked_update(struct sched_entity *se)
4447 {
4448 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
4449 
4450 	/*
4451 	 * If sched_entity still have not zero load or utilization, we have to
4452 	 * decay it:
4453 	 */
4454 	if (se->avg.load_avg || se->avg.util_avg)
4455 		return false;
4456 
4457 	/*
4458 	 * If there is a pending propagation, we have to update the load and
4459 	 * the utilization of the sched_entity:
4460 	 */
4461 	if (gcfs_rq->propagate)
4462 		return false;
4463 
4464 	/*
4465 	 * Otherwise, the load and the utilization of the sched_entity is
4466 	 * already zero and there is no pending propagation, so it will be a
4467 	 * waste of time to try to decay it:
4468 	 */
4469 	return true;
4470 }
4471 
4472 #else /* CONFIG_FAIR_GROUP_SCHED */
4473 
update_tg_load_avg(struct cfs_rq * cfs_rq)4474 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
4475 
clear_tg_offline_cfs_rqs(struct rq * rq)4476 static inline void clear_tg_offline_cfs_rqs(struct rq *rq) {}
4477 
propagate_entity_load_avg(struct sched_entity * se)4478 static inline int propagate_entity_load_avg(struct sched_entity *se)
4479 {
4480 	return 0;
4481 }
4482 
add_tg_cfs_propagate(struct cfs_rq * cfs_rq,long runnable_sum)4483 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
4484 
4485 #endif /* CONFIG_FAIR_GROUP_SCHED */
4486 
4487 #ifdef CONFIG_NO_HZ_COMMON
migrate_se_pelt_lag(struct sched_entity * se)4488 static inline void migrate_se_pelt_lag(struct sched_entity *se)
4489 {
4490 	u64 throttled = 0, now, lut;
4491 	struct cfs_rq *cfs_rq;
4492 	struct rq *rq;
4493 	bool is_idle;
4494 
4495 	if (load_avg_is_decayed(&se->avg))
4496 		return;
4497 
4498 	cfs_rq = cfs_rq_of(se);
4499 	rq = rq_of(cfs_rq);
4500 
4501 	rcu_read_lock();
4502 	is_idle = is_idle_task(rcu_dereference(rq->curr));
4503 	rcu_read_unlock();
4504 
4505 	/*
4506 	 * The lag estimation comes with a cost we don't want to pay all the
4507 	 * time. Hence, limiting to the case where the source CPU is idle and
4508 	 * we know we are at the greatest risk to have an outdated clock.
4509 	 */
4510 	if (!is_idle)
4511 		return;
4512 
4513 	/*
4514 	 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
4515 	 *
4516 	 *   last_update_time (the cfs_rq's last_update_time)
4517 	 *	= cfs_rq_clock_pelt()@cfs_rq_idle
4518 	 *      = rq_clock_pelt()@cfs_rq_idle
4519 	 *        - cfs->throttled_clock_pelt_time@cfs_rq_idle
4520 	 *
4521 	 *   cfs_idle_lag (delta between rq's update and cfs_rq's update)
4522 	 *      = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
4523 	 *
4524 	 *   rq_idle_lag (delta between now and rq's update)
4525 	 *      = sched_clock_cpu() - rq_clock()@rq_idle
4526 	 *
4527 	 * We can then write:
4528 	 *
4529 	 *    now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
4530 	 *          sched_clock_cpu() - rq_clock()@rq_idle
4531 	 * Where:
4532 	 *      rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
4533 	 *      rq_clock()@rq_idle      is rq->clock_idle
4534 	 *      cfs->throttled_clock_pelt_time@cfs_rq_idle
4535 	 *                              is cfs_rq->throttled_pelt_idle
4536 	 */
4537 
4538 #ifdef CONFIG_CFS_BANDWIDTH
4539 	throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
4540 	/* The clock has been stopped for throttling */
4541 	if (throttled == U64_MAX)
4542 		return;
4543 #endif
4544 	now = u64_u32_load(rq->clock_pelt_idle);
4545 	/*
4546 	 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
4547 	 * is observed the old clock_pelt_idle value and the new clock_idle,
4548 	 * which lead to an underestimation. The opposite would lead to an
4549 	 * overestimation.
4550 	 */
4551 	smp_rmb();
4552 	lut = cfs_rq_last_update_time(cfs_rq);
4553 
4554 	now -= throttled;
4555 	if (now < lut)
4556 		/*
4557 		 * cfs_rq->avg.last_update_time is more recent than our
4558 		 * estimation, let's use it.
4559 		 */
4560 		now = lut;
4561 	else
4562 		now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4563 
4564 	__update_load_avg_blocked_se(now, se);
4565 }
4566 #else
migrate_se_pelt_lag(struct sched_entity * se)4567 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4568 #endif
4569 
4570 /**
4571  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4572  * @now: current time, as per cfs_rq_clock_pelt()
4573  * @cfs_rq: cfs_rq to update
4574  *
4575  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4576  * avg. The immediate corollary is that all (fair) tasks must be attached.
4577  *
4578  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4579  *
4580  * Return: true if the load decayed or we removed load.
4581  *
4582  * Since both these conditions indicate a changed cfs_rq->avg.load we should
4583  * call update_tg_load_avg() when this function returns true.
4584  */
4585 static inline int
update_cfs_rq_load_avg(u64 now,struct cfs_rq * cfs_rq)4586 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4587 {
4588 	unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4589 	struct sched_avg *sa = &cfs_rq->avg;
4590 	int decayed = 0;
4591 
4592 	if (cfs_rq->removed.nr) {
4593 		unsigned long r;
4594 		u32 divider = get_pelt_divider(&cfs_rq->avg);
4595 
4596 		raw_spin_lock(&cfs_rq->removed.lock);
4597 		swap(cfs_rq->removed.util_avg, removed_util);
4598 		swap(cfs_rq->removed.load_avg, removed_load);
4599 		swap(cfs_rq->removed.runnable_avg, removed_runnable);
4600 		cfs_rq->removed.nr = 0;
4601 		raw_spin_unlock(&cfs_rq->removed.lock);
4602 
4603 		r = removed_load;
4604 		sub_positive(&sa->load_avg, r);
4605 		sub_positive(&sa->load_sum, r * divider);
4606 		/* See sa->util_sum below */
4607 		sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4608 
4609 		r = removed_util;
4610 		sub_positive(&sa->util_avg, r);
4611 		sub_positive(&sa->util_sum, r * divider);
4612 		/*
4613 		 * Because of rounding, se->util_sum might ends up being +1 more than
4614 		 * cfs->util_sum. Although this is not a problem by itself, detaching
4615 		 * a lot of tasks with the rounding problem between 2 updates of
4616 		 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4617 		 * cfs_util_avg is not.
4618 		 * Check that util_sum is still above its lower bound for the new
4619 		 * util_avg. Given that period_contrib might have moved since the last
4620 		 * sync, we are only sure that util_sum must be above or equal to
4621 		 *    util_avg * minimum possible divider
4622 		 */
4623 		sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4624 
4625 		r = removed_runnable;
4626 		sub_positive(&sa->runnable_avg, r);
4627 		sub_positive(&sa->runnable_sum, r * divider);
4628 		/* See sa->util_sum above */
4629 		sa->runnable_sum = max_t(u32, sa->runnable_sum,
4630 					      sa->runnable_avg * PELT_MIN_DIVIDER);
4631 
4632 		/*
4633 		 * removed_runnable is the unweighted version of removed_load so we
4634 		 * can use it to estimate removed_load_sum.
4635 		 */
4636 		add_tg_cfs_propagate(cfs_rq,
4637 			-(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4638 
4639 		decayed = 1;
4640 	}
4641 
4642 	decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4643 	u64_u32_store_copy(sa->last_update_time,
4644 			   cfs_rq->last_update_time_copy,
4645 			   sa->last_update_time);
4646 	return decayed;
4647 }
4648 
4649 /**
4650  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4651  * @cfs_rq: cfs_rq to attach to
4652  * @se: sched_entity to attach
4653  *
4654  * Must call update_cfs_rq_load_avg() before this, since we rely on
4655  * cfs_rq->avg.last_update_time being current.
4656  */
attach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)4657 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4658 {
4659 	/*
4660 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4661 	 * See ___update_load_avg() for details.
4662 	 */
4663 	u32 divider = get_pelt_divider(&cfs_rq->avg);
4664 
4665 	/*
4666 	 * When we attach the @se to the @cfs_rq, we must align the decay
4667 	 * window because without that, really weird and wonderful things can
4668 	 * happen.
4669 	 *
4670 	 * XXX illustrate
4671 	 */
4672 	se->avg.last_update_time = cfs_rq->avg.last_update_time;
4673 	se->avg.period_contrib = cfs_rq->avg.period_contrib;
4674 
4675 	/*
4676 	 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4677 	 * period_contrib. This isn't strictly correct, but since we're
4678 	 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4679 	 * _sum a little.
4680 	 */
4681 	se->avg.util_sum = se->avg.util_avg * divider;
4682 
4683 	se->avg.runnable_sum = se->avg.runnable_avg * divider;
4684 
4685 	se->avg.load_sum = se->avg.load_avg * divider;
4686 	if (se_weight(se) < se->avg.load_sum)
4687 		se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4688 	else
4689 		se->avg.load_sum = 1;
4690 
4691 	enqueue_load_avg(cfs_rq, se);
4692 	cfs_rq->avg.util_avg += se->avg.util_avg;
4693 	cfs_rq->avg.util_sum += se->avg.util_sum;
4694 	cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4695 	cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4696 
4697 	add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4698 
4699 	cfs_rq_util_change(cfs_rq, 0);
4700 
4701 	trace_pelt_cfs_tp(cfs_rq);
4702 }
4703 
4704 /**
4705  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4706  * @cfs_rq: cfs_rq to detach from
4707  * @se: sched_entity to detach
4708  *
4709  * Must call update_cfs_rq_load_avg() before this, since we rely on
4710  * cfs_rq->avg.last_update_time being current.
4711  */
detach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)4712 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4713 {
4714 	dequeue_load_avg(cfs_rq, se);
4715 	sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4716 	sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4717 	/* See update_cfs_rq_load_avg() */
4718 	cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4719 					  cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4720 
4721 	sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4722 	sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4723 	/* See update_cfs_rq_load_avg() */
4724 	cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4725 					      cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4726 
4727 	add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4728 
4729 	cfs_rq_util_change(cfs_rq, 0);
4730 
4731 	trace_pelt_cfs_tp(cfs_rq);
4732 }
4733 
4734 /*
4735  * Optional action to be done while updating the load average
4736  */
4737 #define UPDATE_TG	0x1
4738 #define SKIP_AGE_LOAD	0x2
4739 #define DO_ATTACH	0x4
4740 #define DO_DETACH	0x8
4741 
4742 /* Update task and its cfs_rq load average */
update_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)4743 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4744 {
4745 	u64 now = cfs_rq_clock_pelt(cfs_rq);
4746 	int decayed;
4747 
4748 	/*
4749 	 * Track task load average for carrying it to new CPU after migrated, and
4750 	 * track group sched_entity load average for task_h_load calculation in migration
4751 	 */
4752 	if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4753 		__update_load_avg_se(now, cfs_rq, se);
4754 
4755 	decayed  = update_cfs_rq_load_avg(now, cfs_rq);
4756 	decayed |= propagate_entity_load_avg(se);
4757 
4758 	if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4759 
4760 		/*
4761 		 * DO_ATTACH means we're here from enqueue_entity().
4762 		 * !last_update_time means we've passed through
4763 		 * migrate_task_rq_fair() indicating we migrated.
4764 		 *
4765 		 * IOW we're enqueueing a task on a new CPU.
4766 		 */
4767 		attach_entity_load_avg(cfs_rq, se);
4768 		update_tg_load_avg(cfs_rq);
4769 
4770 	} else if (flags & DO_DETACH) {
4771 		/*
4772 		 * DO_DETACH means we're here from dequeue_entity()
4773 		 * and we are migrating task out of the CPU.
4774 		 */
4775 		detach_entity_load_avg(cfs_rq, se);
4776 		update_tg_load_avg(cfs_rq);
4777 	} else if (decayed) {
4778 		cfs_rq_util_change(cfs_rq, 0);
4779 
4780 		if (flags & UPDATE_TG)
4781 			update_tg_load_avg(cfs_rq);
4782 	}
4783 }
4784 
4785 /*
4786  * Synchronize entity load avg of dequeued entity without locking
4787  * the previous rq.
4788  */
sync_entity_load_avg(struct sched_entity * se)4789 static void sync_entity_load_avg(struct sched_entity *se)
4790 {
4791 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4792 	u64 last_update_time;
4793 
4794 	last_update_time = cfs_rq_last_update_time(cfs_rq);
4795 	__update_load_avg_blocked_se(last_update_time, se);
4796 }
4797 
4798 /*
4799  * Task first catches up with cfs_rq, and then subtract
4800  * itself from the cfs_rq (task must be off the queue now).
4801  */
remove_entity_load_avg(struct sched_entity * se)4802 static void remove_entity_load_avg(struct sched_entity *se)
4803 {
4804 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4805 	unsigned long flags;
4806 
4807 	/*
4808 	 * tasks cannot exit without having gone through wake_up_new_task() ->
4809 	 * enqueue_task_fair() which will have added things to the cfs_rq,
4810 	 * so we can remove unconditionally.
4811 	 */
4812 
4813 	sync_entity_load_avg(se);
4814 
4815 	raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4816 	++cfs_rq->removed.nr;
4817 	cfs_rq->removed.util_avg	+= se->avg.util_avg;
4818 	cfs_rq->removed.load_avg	+= se->avg.load_avg;
4819 	cfs_rq->removed.runnable_avg	+= se->avg.runnable_avg;
4820 	raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4821 }
4822 
cfs_rq_runnable_avg(struct cfs_rq * cfs_rq)4823 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4824 {
4825 	return cfs_rq->avg.runnable_avg;
4826 }
4827 
cfs_rq_load_avg(struct cfs_rq * cfs_rq)4828 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4829 {
4830 	return cfs_rq->avg.load_avg;
4831 }
4832 
4833 static int sched_balance_newidle(struct rq *this_rq, struct rq_flags *rf);
4834 
task_util(struct task_struct * p)4835 static inline unsigned long task_util(struct task_struct *p)
4836 {
4837 	return READ_ONCE(p->se.avg.util_avg);
4838 }
4839 
task_runnable(struct task_struct * p)4840 static inline unsigned long task_runnable(struct task_struct *p)
4841 {
4842 	return READ_ONCE(p->se.avg.runnable_avg);
4843 }
4844 
_task_util_est(struct task_struct * p)4845 static inline unsigned long _task_util_est(struct task_struct *p)
4846 {
4847 	return READ_ONCE(p->se.avg.util_est) & ~UTIL_AVG_UNCHANGED;
4848 }
4849 
task_util_est(struct task_struct * p)4850 static inline unsigned long task_util_est(struct task_struct *p)
4851 {
4852 	return max(task_util(p), _task_util_est(p));
4853 }
4854 
util_est_enqueue(struct cfs_rq * cfs_rq,struct task_struct * p)4855 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4856 				    struct task_struct *p)
4857 {
4858 	unsigned int enqueued;
4859 
4860 	if (!sched_feat(UTIL_EST))
4861 		return;
4862 
4863 	/* Update root cfs_rq's estimated utilization */
4864 	enqueued  = cfs_rq->avg.util_est;
4865 	enqueued += _task_util_est(p);
4866 	WRITE_ONCE(cfs_rq->avg.util_est, enqueued);
4867 
4868 	trace_sched_util_est_cfs_tp(cfs_rq);
4869 }
4870 
util_est_dequeue(struct cfs_rq * cfs_rq,struct task_struct * p)4871 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4872 				    struct task_struct *p)
4873 {
4874 	unsigned int enqueued;
4875 
4876 	if (!sched_feat(UTIL_EST))
4877 		return;
4878 
4879 	/* Update root cfs_rq's estimated utilization */
4880 	enqueued  = cfs_rq->avg.util_est;
4881 	enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4882 	WRITE_ONCE(cfs_rq->avg.util_est, enqueued);
4883 
4884 	trace_sched_util_est_cfs_tp(cfs_rq);
4885 }
4886 
4887 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4888 
util_est_update(struct cfs_rq * cfs_rq,struct task_struct * p,bool task_sleep)4889 static inline void util_est_update(struct cfs_rq *cfs_rq,
4890 				   struct task_struct *p,
4891 				   bool task_sleep)
4892 {
4893 	unsigned int ewma, dequeued, last_ewma_diff;
4894 
4895 	if (!sched_feat(UTIL_EST))
4896 		return;
4897 
4898 	/*
4899 	 * Skip update of task's estimated utilization when the task has not
4900 	 * yet completed an activation, e.g. being migrated.
4901 	 */
4902 	if (!task_sleep)
4903 		return;
4904 
4905 	/* Get current estimate of utilization */
4906 	ewma = READ_ONCE(p->se.avg.util_est);
4907 
4908 	/*
4909 	 * If the PELT values haven't changed since enqueue time,
4910 	 * skip the util_est update.
4911 	 */
4912 	if (ewma & UTIL_AVG_UNCHANGED)
4913 		return;
4914 
4915 	/* Get utilization at dequeue */
4916 	dequeued = task_util(p);
4917 
4918 	/*
4919 	 * Reset EWMA on utilization increases, the moving average is used only
4920 	 * to smooth utilization decreases.
4921 	 */
4922 	if (ewma <= dequeued) {
4923 		ewma = dequeued;
4924 		goto done;
4925 	}
4926 
4927 	/*
4928 	 * Skip update of task's estimated utilization when its members are
4929 	 * already ~1% close to its last activation value.
4930 	 */
4931 	last_ewma_diff = ewma - dequeued;
4932 	if (last_ewma_diff < UTIL_EST_MARGIN)
4933 		goto done;
4934 
4935 	/*
4936 	 * To avoid overestimation of actual task utilization, skip updates if
4937 	 * we cannot grant there is idle time in this CPU.
4938 	 */
4939 	if (dequeued > arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))))
4940 		return;
4941 
4942 	/*
4943 	 * To avoid underestimate of task utilization, skip updates of EWMA if
4944 	 * we cannot grant that thread got all CPU time it wanted.
4945 	 */
4946 	if ((dequeued + UTIL_EST_MARGIN) < task_runnable(p))
4947 		goto done;
4948 
4949 
4950 	/*
4951 	 * Update Task's estimated utilization
4952 	 *
4953 	 * When *p completes an activation we can consolidate another sample
4954 	 * of the task size. This is done by using this value to update the
4955 	 * Exponential Weighted Moving Average (EWMA):
4956 	 *
4957 	 *  ewma(t) = w *  task_util(p) + (1-w) * ewma(t-1)
4958 	 *          = w *  task_util(p) +         ewma(t-1)  - w * ewma(t-1)
4959 	 *          = w * (task_util(p) -         ewma(t-1)) +     ewma(t-1)
4960 	 *          = w * (      -last_ewma_diff           ) +     ewma(t-1)
4961 	 *          = w * (-last_ewma_diff +  ewma(t-1) / w)
4962 	 *
4963 	 * Where 'w' is the weight of new samples, which is configured to be
4964 	 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4965 	 */
4966 	ewma <<= UTIL_EST_WEIGHT_SHIFT;
4967 	ewma  -= last_ewma_diff;
4968 	ewma >>= UTIL_EST_WEIGHT_SHIFT;
4969 done:
4970 	ewma |= UTIL_AVG_UNCHANGED;
4971 	WRITE_ONCE(p->se.avg.util_est, ewma);
4972 
4973 	trace_sched_util_est_se_tp(&p->se);
4974 }
4975 
get_actual_cpu_capacity(int cpu)4976 static inline unsigned long get_actual_cpu_capacity(int cpu)
4977 {
4978 	unsigned long capacity = arch_scale_cpu_capacity(cpu);
4979 
4980 	capacity -= max(hw_load_avg(cpu_rq(cpu)), cpufreq_get_pressure(cpu));
4981 
4982 	return capacity;
4983 }
4984 
util_fits_cpu(unsigned long util,unsigned long uclamp_min,unsigned long uclamp_max,int cpu)4985 static inline int util_fits_cpu(unsigned long util,
4986 				unsigned long uclamp_min,
4987 				unsigned long uclamp_max,
4988 				int cpu)
4989 {
4990 	unsigned long capacity = capacity_of(cpu);
4991 	unsigned long capacity_orig;
4992 	bool fits, uclamp_max_fits;
4993 
4994 	/*
4995 	 * Check if the real util fits without any uclamp boost/cap applied.
4996 	 */
4997 	fits = fits_capacity(util, capacity);
4998 
4999 	if (!uclamp_is_used())
5000 		return fits;
5001 
5002 	/*
5003 	 * We must use arch_scale_cpu_capacity() for comparing against uclamp_min and
5004 	 * uclamp_max. We only care about capacity pressure (by using
5005 	 * capacity_of()) for comparing against the real util.
5006 	 *
5007 	 * If a task is boosted to 1024 for example, we don't want a tiny
5008 	 * pressure to skew the check whether it fits a CPU or not.
5009 	 *
5010 	 * Similarly if a task is capped to arch_scale_cpu_capacity(little_cpu), it
5011 	 * should fit a little cpu even if there's some pressure.
5012 	 *
5013 	 * Only exception is for HW or cpufreq pressure since it has a direct impact
5014 	 * on available OPP of the system.
5015 	 *
5016 	 * We honour it for uclamp_min only as a drop in performance level
5017 	 * could result in not getting the requested minimum performance level.
5018 	 *
5019 	 * For uclamp_max, we can tolerate a drop in performance level as the
5020 	 * goal is to cap the task. So it's okay if it's getting less.
5021 	 */
5022 	capacity_orig = arch_scale_cpu_capacity(cpu);
5023 
5024 	/*
5025 	 * We want to force a task to fit a cpu as implied by uclamp_max.
5026 	 * But we do have some corner cases to cater for..
5027 	 *
5028 	 *
5029 	 *                                 C=z
5030 	 *   |                             ___
5031 	 *   |                  C=y       |   |
5032 	 *   |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _  uclamp_max
5033 	 *   |      C=x        |   |      |   |
5034 	 *   |      ___        |   |      |   |
5035 	 *   |     |   |       |   |      |   |    (util somewhere in this region)
5036 	 *   |     |   |       |   |      |   |
5037 	 *   |     |   |       |   |      |   |
5038 	 *   +----------------------------------------
5039 	 *         CPU0        CPU1       CPU2
5040 	 *
5041 	 *   In the above example if a task is capped to a specific performance
5042 	 *   point, y, then when:
5043 	 *
5044 	 *   * util = 80% of x then it does not fit on CPU0 and should migrate
5045 	 *     to CPU1
5046 	 *   * util = 80% of y then it is forced to fit on CPU1 to honour
5047 	 *     uclamp_max request.
5048 	 *
5049 	 *   which is what we're enforcing here. A task always fits if
5050 	 *   uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
5051 	 *   the normal upmigration rules should withhold still.
5052 	 *
5053 	 *   Only exception is when we are on max capacity, then we need to be
5054 	 *   careful not to block overutilized state. This is so because:
5055 	 *
5056 	 *     1. There's no concept of capping at max_capacity! We can't go
5057 	 *        beyond this performance level anyway.
5058 	 *     2. The system is being saturated when we're operating near
5059 	 *        max capacity, it doesn't make sense to block overutilized.
5060 	 */
5061 	uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
5062 	uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
5063 	fits = fits || uclamp_max_fits;
5064 
5065 	/*
5066 	 *
5067 	 *                                 C=z
5068 	 *   |                             ___       (region a, capped, util >= uclamp_max)
5069 	 *   |                  C=y       |   |
5070 	 *   |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
5071 	 *   |      C=x        |   |      |   |
5072 	 *   |      ___        |   |      |   |      (region b, uclamp_min <= util <= uclamp_max)
5073 	 *   |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
5074 	 *   |     |   |       |   |      |   |
5075 	 *   |     |   |       |   |      |   |      (region c, boosted, util < uclamp_min)
5076 	 *   +----------------------------------------
5077 	 *         CPU0        CPU1       CPU2
5078 	 *
5079 	 * a) If util > uclamp_max, then we're capped, we don't care about
5080 	 *    actual fitness value here. We only care if uclamp_max fits
5081 	 *    capacity without taking margin/pressure into account.
5082 	 *    See comment above.
5083 	 *
5084 	 * b) If uclamp_min <= util <= uclamp_max, then the normal
5085 	 *    fits_capacity() rules apply. Except we need to ensure that we
5086 	 *    enforce we remain within uclamp_max, see comment above.
5087 	 *
5088 	 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
5089 	 *    need to take into account the boosted value fits the CPU without
5090 	 *    taking margin/pressure into account.
5091 	 *
5092 	 * Cases (a) and (b) are handled in the 'fits' variable already. We
5093 	 * just need to consider an extra check for case (c) after ensuring we
5094 	 * handle the case uclamp_min > uclamp_max.
5095 	 */
5096 	uclamp_min = min(uclamp_min, uclamp_max);
5097 	if (fits && (util < uclamp_min) &&
5098 	    (uclamp_min > get_actual_cpu_capacity(cpu)))
5099 		return -1;
5100 
5101 	return fits;
5102 }
5103 
task_fits_cpu(struct task_struct * p,int cpu)5104 static inline int task_fits_cpu(struct task_struct *p, int cpu)
5105 {
5106 	unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
5107 	unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
5108 	unsigned long util = task_util_est(p);
5109 	/*
5110 	 * Return true only if the cpu fully fits the task requirements, which
5111 	 * include the utilization but also the performance hints.
5112 	 */
5113 	return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
5114 }
5115 
update_misfit_status(struct task_struct * p,struct rq * rq)5116 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
5117 {
5118 	int cpu = cpu_of(rq);
5119 
5120 	if (!sched_asym_cpucap_active())
5121 		return;
5122 
5123 	/*
5124 	 * Affinity allows us to go somewhere higher?  Or are we on biggest
5125 	 * available CPU already? Or do we fit into this CPU ?
5126 	 */
5127 	if (!p || (p->nr_cpus_allowed == 1) ||
5128 	    (arch_scale_cpu_capacity(cpu) == p->max_allowed_capacity) ||
5129 	    task_fits_cpu(p, cpu)) {
5130 
5131 		rq->misfit_task_load = 0;
5132 		return;
5133 	}
5134 
5135 	/*
5136 	 * Make sure that misfit_task_load will not be null even if
5137 	 * task_h_load() returns 0.
5138 	 */
5139 	rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
5140 }
5141 
5142 #else /* CONFIG_SMP */
5143 
cfs_rq_is_decayed(struct cfs_rq * cfs_rq)5144 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
5145 {
5146 	return !cfs_rq->nr_queued;
5147 }
5148 
5149 #define UPDATE_TG	0x0
5150 #define SKIP_AGE_LOAD	0x0
5151 #define DO_ATTACH	0x0
5152 #define DO_DETACH	0x0
5153 
update_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se,int not_used1)5154 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
5155 {
5156 	cfs_rq_util_change(cfs_rq, 0);
5157 }
5158 
remove_entity_load_avg(struct sched_entity * se)5159 static inline void remove_entity_load_avg(struct sched_entity *se) {}
5160 
5161 static inline void
attach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)5162 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5163 static inline void
detach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)5164 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5165 
sched_balance_newidle(struct rq * rq,struct rq_flags * rf)5166 static inline int sched_balance_newidle(struct rq *rq, struct rq_flags *rf)
5167 {
5168 	return 0;
5169 }
5170 
5171 static inline void
util_est_enqueue(struct cfs_rq * cfs_rq,struct task_struct * p)5172 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5173 
5174 static inline void
util_est_dequeue(struct cfs_rq * cfs_rq,struct task_struct * p)5175 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5176 
5177 static inline void
util_est_update(struct cfs_rq * cfs_rq,struct task_struct * p,bool task_sleep)5178 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
5179 		bool task_sleep) {}
update_misfit_status(struct task_struct * p,struct rq * rq)5180 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
5181 
5182 #endif /* CONFIG_SMP */
5183 
__setparam_fair(struct task_struct * p,const struct sched_attr * attr)5184 void __setparam_fair(struct task_struct *p, const struct sched_attr *attr)
5185 {
5186 	struct sched_entity *se = &p->se;
5187 
5188 	p->static_prio = NICE_TO_PRIO(attr->sched_nice);
5189 	if (attr->sched_runtime) {
5190 		se->custom_slice = 1;
5191 		se->slice = clamp_t(u64, attr->sched_runtime,
5192 				      NSEC_PER_MSEC/10,   /* HZ=1000 * 10 */
5193 				      NSEC_PER_MSEC*100); /* HZ=100  / 10 */
5194 	} else {
5195 		se->custom_slice = 0;
5196 		se->slice = sysctl_sched_base_slice;
5197 	}
5198 }
5199 
5200 static void
place_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5201 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5202 {
5203 	u64 vslice, vruntime = avg_vruntime(cfs_rq);
5204 	s64 lag = 0;
5205 
5206 	if (!se->custom_slice)
5207 		se->slice = sysctl_sched_base_slice;
5208 	vslice = calc_delta_fair(se->slice, se);
5209 
5210 	/*
5211 	 * Due to how V is constructed as the weighted average of entities,
5212 	 * adding tasks with positive lag, or removing tasks with negative lag
5213 	 * will move 'time' backwards, this can screw around with the lag of
5214 	 * other tasks.
5215 	 *
5216 	 * EEVDF: placement strategy #1 / #2
5217 	 */
5218 	if (sched_feat(PLACE_LAG) && cfs_rq->nr_queued && se->vlag) {
5219 		struct sched_entity *curr = cfs_rq->curr;
5220 		unsigned long load;
5221 
5222 		lag = se->vlag;
5223 
5224 		/*
5225 		 * If we want to place a task and preserve lag, we have to
5226 		 * consider the effect of the new entity on the weighted
5227 		 * average and compensate for this, otherwise lag can quickly
5228 		 * evaporate.
5229 		 *
5230 		 * Lag is defined as:
5231 		 *
5232 		 *   lag_i = S - s_i = w_i * (V - v_i)
5233 		 *
5234 		 * To avoid the 'w_i' term all over the place, we only track
5235 		 * the virtual lag:
5236 		 *
5237 		 *   vl_i = V - v_i <=> v_i = V - vl_i
5238 		 *
5239 		 * And we take V to be the weighted average of all v:
5240 		 *
5241 		 *   V = (\Sum w_j*v_j) / W
5242 		 *
5243 		 * Where W is: \Sum w_j
5244 		 *
5245 		 * Then, the weighted average after adding an entity with lag
5246 		 * vl_i is given by:
5247 		 *
5248 		 *   V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i)
5249 		 *      = (W*V + w_i*(V - vl_i)) / (W + w_i)
5250 		 *      = (W*V + w_i*V - w_i*vl_i) / (W + w_i)
5251 		 *      = (V*(W + w_i) - w_i*l) / (W + w_i)
5252 		 *      = V - w_i*vl_i / (W + w_i)
5253 		 *
5254 		 * And the actual lag after adding an entity with vl_i is:
5255 		 *
5256 		 *   vl'_i = V' - v_i
5257 		 *         = V - w_i*vl_i / (W + w_i) - (V - vl_i)
5258 		 *         = vl_i - w_i*vl_i / (W + w_i)
5259 		 *
5260 		 * Which is strictly less than vl_i. So in order to preserve lag
5261 		 * we should inflate the lag before placement such that the
5262 		 * effective lag after placement comes out right.
5263 		 *
5264 		 * As such, invert the above relation for vl'_i to get the vl_i
5265 		 * we need to use such that the lag after placement is the lag
5266 		 * we computed before dequeue.
5267 		 *
5268 		 *   vl'_i = vl_i - w_i*vl_i / (W + w_i)
5269 		 *         = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)
5270 		 *
5271 		 *   (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i
5272 		 *                   = W*vl_i
5273 		 *
5274 		 *   vl_i = (W + w_i)*vl'_i / W
5275 		 */
5276 		load = cfs_rq->avg_load;
5277 		if (curr && curr->on_rq)
5278 			load += scale_load_down(curr->load.weight);
5279 
5280 		lag *= load + scale_load_down(se->load.weight);
5281 		if (WARN_ON_ONCE(!load))
5282 			load = 1;
5283 		lag = div_s64(lag, load);
5284 	}
5285 
5286 	se->vruntime = vruntime - lag;
5287 
5288 	if (se->rel_deadline) {
5289 		se->deadline += se->vruntime;
5290 		se->rel_deadline = 0;
5291 		return;
5292 	}
5293 
5294 	/*
5295 	 * When joining the competition; the existing tasks will be,
5296 	 * on average, halfway through their slice, as such start tasks
5297 	 * off with half a slice to ease into the competition.
5298 	 */
5299 	if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL))
5300 		vslice /= 2;
5301 
5302 	/*
5303 	 * EEVDF: vd_i = ve_i + r_i/w_i
5304 	 */
5305 	se->deadline = se->vruntime + vslice;
5306 }
5307 
5308 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
5309 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
5310 
5311 static void
5312 requeue_delayed_entity(struct sched_entity *se);
5313 
5314 static void
enqueue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5315 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5316 {
5317 	bool curr = cfs_rq->curr == se;
5318 
5319 	/*
5320 	 * If we're the current task, we must renormalise before calling
5321 	 * update_curr().
5322 	 */
5323 	if (curr)
5324 		place_entity(cfs_rq, se, flags);
5325 
5326 	update_curr(cfs_rq);
5327 
5328 	/*
5329 	 * When enqueuing a sched_entity, we must:
5330 	 *   - Update loads to have both entity and cfs_rq synced with now.
5331 	 *   - For group_entity, update its runnable_weight to reflect the new
5332 	 *     h_nr_runnable of its group cfs_rq.
5333 	 *   - For group_entity, update its weight to reflect the new share of
5334 	 *     its group cfs_rq
5335 	 *   - Add its new weight to cfs_rq->load.weight
5336 	 */
5337 	update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
5338 	se_update_runnable(se);
5339 	/*
5340 	 * XXX update_load_avg() above will have attached us to the pelt sum;
5341 	 * but update_cfs_group() here will re-adjust the weight and have to
5342 	 * undo/redo all that. Seems wasteful.
5343 	 */
5344 	update_cfs_group(se);
5345 
5346 	/*
5347 	 * XXX now that the entity has been re-weighted, and it's lag adjusted,
5348 	 * we can place the entity.
5349 	 */
5350 	if (!curr)
5351 		place_entity(cfs_rq, se, flags);
5352 
5353 	account_entity_enqueue(cfs_rq, se);
5354 
5355 	/* Entity has migrated, no longer consider this task hot */
5356 	if (flags & ENQUEUE_MIGRATED)
5357 		se->exec_start = 0;
5358 
5359 	check_schedstat_required();
5360 	update_stats_enqueue_fair(cfs_rq, se, flags);
5361 	if (!curr)
5362 		__enqueue_entity(cfs_rq, se);
5363 	se->on_rq = 1;
5364 
5365 	if (cfs_rq->nr_queued == 1) {
5366 		check_enqueue_throttle(cfs_rq);
5367 		if (!throttled_hierarchy(cfs_rq)) {
5368 			list_add_leaf_cfs_rq(cfs_rq);
5369 		} else {
5370 #ifdef CONFIG_CFS_BANDWIDTH
5371 			struct rq *rq = rq_of(cfs_rq);
5372 
5373 			if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)
5374 				cfs_rq->throttled_clock = rq_clock(rq);
5375 			if (!cfs_rq->throttled_clock_self)
5376 				cfs_rq->throttled_clock_self = rq_clock(rq);
5377 #endif
5378 		}
5379 	}
5380 }
5381 
__clear_buddies_next(struct sched_entity * se)5382 static void __clear_buddies_next(struct sched_entity *se)
5383 {
5384 	for_each_sched_entity(se) {
5385 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
5386 		if (cfs_rq->next != se)
5387 			break;
5388 
5389 		cfs_rq->next = NULL;
5390 	}
5391 }
5392 
clear_buddies(struct cfs_rq * cfs_rq,struct sched_entity * se)5393 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
5394 {
5395 	if (cfs_rq->next == se)
5396 		__clear_buddies_next(se);
5397 }
5398 
5399 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5400 
set_delayed(struct sched_entity * se)5401 static void set_delayed(struct sched_entity *se)
5402 {
5403 	se->sched_delayed = 1;
5404 
5405 	/*
5406 	 * Delayed se of cfs_rq have no tasks queued on them.
5407 	 * Do not adjust h_nr_runnable since dequeue_entities()
5408 	 * will account it for blocked tasks.
5409 	 */
5410 	if (!entity_is_task(se))
5411 		return;
5412 
5413 	for_each_sched_entity(se) {
5414 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
5415 
5416 		cfs_rq->h_nr_runnable--;
5417 		if (cfs_rq_throttled(cfs_rq))
5418 			break;
5419 	}
5420 }
5421 
clear_delayed(struct sched_entity * se)5422 static void clear_delayed(struct sched_entity *se)
5423 {
5424 	se->sched_delayed = 0;
5425 
5426 	/*
5427 	 * Delayed se of cfs_rq have no tasks queued on them.
5428 	 * Do not adjust h_nr_runnable since a dequeue has
5429 	 * already accounted for it or an enqueue of a task
5430 	 * below it will account for it in enqueue_task_fair().
5431 	 */
5432 	if (!entity_is_task(se))
5433 		return;
5434 
5435 	for_each_sched_entity(se) {
5436 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
5437 
5438 		cfs_rq->h_nr_runnable++;
5439 		if (cfs_rq_throttled(cfs_rq))
5440 			break;
5441 	}
5442 }
5443 
finish_delayed_dequeue_entity(struct sched_entity * se)5444 static inline void finish_delayed_dequeue_entity(struct sched_entity *se)
5445 {
5446 	clear_delayed(se);
5447 	if (sched_feat(DELAY_ZERO) && se->vlag > 0)
5448 		se->vlag = 0;
5449 }
5450 
5451 static bool
dequeue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5452 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5453 {
5454 	bool sleep = flags & DEQUEUE_SLEEP;
5455 	int action = UPDATE_TG;
5456 
5457 	update_curr(cfs_rq);
5458 	clear_buddies(cfs_rq, se);
5459 
5460 	if (flags & DEQUEUE_DELAYED) {
5461 		WARN_ON_ONCE(!se->sched_delayed);
5462 	} else {
5463 		bool delay = sleep;
5464 		/*
5465 		 * DELAY_DEQUEUE relies on spurious wakeups, special task
5466 		 * states must not suffer spurious wakeups, excempt them.
5467 		 */
5468 		if (flags & DEQUEUE_SPECIAL)
5469 			delay = false;
5470 
5471 		WARN_ON_ONCE(delay && se->sched_delayed);
5472 
5473 		if (sched_feat(DELAY_DEQUEUE) && delay &&
5474 		    !entity_eligible(cfs_rq, se)) {
5475 			update_load_avg(cfs_rq, se, 0);
5476 			set_delayed(se);
5477 			return false;
5478 		}
5479 	}
5480 
5481 	if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
5482 		action |= DO_DETACH;
5483 
5484 	/*
5485 	 * When dequeuing a sched_entity, we must:
5486 	 *   - Update loads to have both entity and cfs_rq synced with now.
5487 	 *   - For group_entity, update its runnable_weight to reflect the new
5488 	 *     h_nr_runnable of its group cfs_rq.
5489 	 *   - Subtract its previous weight from cfs_rq->load.weight.
5490 	 *   - For group entity, update its weight to reflect the new share
5491 	 *     of its group cfs_rq.
5492 	 */
5493 	update_load_avg(cfs_rq, se, action);
5494 	se_update_runnable(se);
5495 
5496 	update_stats_dequeue_fair(cfs_rq, se, flags);
5497 
5498 	update_entity_lag(cfs_rq, se);
5499 	if (sched_feat(PLACE_REL_DEADLINE) && !sleep) {
5500 		se->deadline -= se->vruntime;
5501 		se->rel_deadline = 1;
5502 	}
5503 
5504 	if (se != cfs_rq->curr)
5505 		__dequeue_entity(cfs_rq, se);
5506 	se->on_rq = 0;
5507 	account_entity_dequeue(cfs_rq, se);
5508 
5509 	/* return excess runtime on last dequeue */
5510 	return_cfs_rq_runtime(cfs_rq);
5511 
5512 	update_cfs_group(se);
5513 
5514 	/*
5515 	 * Now advance min_vruntime if @se was the entity holding it back,
5516 	 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
5517 	 * put back on, and if we advance min_vruntime, we'll be placed back
5518 	 * further than we started -- i.e. we'll be penalized.
5519 	 */
5520 	if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
5521 		update_min_vruntime(cfs_rq);
5522 
5523 	if (flags & DEQUEUE_DELAYED)
5524 		finish_delayed_dequeue_entity(se);
5525 
5526 	if (cfs_rq->nr_queued == 0)
5527 		update_idle_cfs_rq_clock_pelt(cfs_rq);
5528 
5529 	return true;
5530 }
5531 
5532 static void
set_next_entity(struct cfs_rq * cfs_rq,struct sched_entity * se)5533 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
5534 {
5535 	clear_buddies(cfs_rq, se);
5536 
5537 	/* 'current' is not kept within the tree. */
5538 	if (se->on_rq) {
5539 		/*
5540 		 * Any task has to be enqueued before it get to execute on
5541 		 * a CPU. So account for the time it spent waiting on the
5542 		 * runqueue.
5543 		 */
5544 		update_stats_wait_end_fair(cfs_rq, se);
5545 		__dequeue_entity(cfs_rq, se);
5546 		update_load_avg(cfs_rq, se, UPDATE_TG);
5547 
5548 		set_protect_slice(se);
5549 	}
5550 
5551 	update_stats_curr_start(cfs_rq, se);
5552 	WARN_ON_ONCE(cfs_rq->curr);
5553 	cfs_rq->curr = se;
5554 
5555 	/*
5556 	 * Track our maximum slice length, if the CPU's load is at
5557 	 * least twice that of our own weight (i.e. don't track it
5558 	 * when there are only lesser-weight tasks around):
5559 	 */
5560 	if (schedstat_enabled() &&
5561 	    rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
5562 		struct sched_statistics *stats;
5563 
5564 		stats = __schedstats_from_se(se);
5565 		__schedstat_set(stats->slice_max,
5566 				max((u64)stats->slice_max,
5567 				    se->sum_exec_runtime - se->prev_sum_exec_runtime));
5568 	}
5569 
5570 	se->prev_sum_exec_runtime = se->sum_exec_runtime;
5571 }
5572 
5573 static int dequeue_entities(struct rq *rq, struct sched_entity *se, int flags);
5574 
5575 /*
5576  * Pick the next process, keeping these things in mind, in this order:
5577  * 1) keep things fair between processes/task groups
5578  * 2) pick the "next" process, since someone really wants that to run
5579  * 3) pick the "last" process, for cache locality
5580  * 4) do not run the "skip" process, if something else is available
5581  */
5582 static struct sched_entity *
pick_next_entity(struct rq * rq,struct cfs_rq * cfs_rq)5583 pick_next_entity(struct rq *rq, struct cfs_rq *cfs_rq)
5584 {
5585 	struct sched_entity *se;
5586 
5587 	/*
5588 	 * Picking the ->next buddy will affect latency but not fairness.
5589 	 */
5590 	if (sched_feat(PICK_BUDDY) &&
5591 	    cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next)) {
5592 		/* ->next will never be delayed */
5593 		WARN_ON_ONCE(cfs_rq->next->sched_delayed);
5594 		return cfs_rq->next;
5595 	}
5596 
5597 	se = pick_eevdf(cfs_rq);
5598 	if (se->sched_delayed) {
5599 		dequeue_entities(rq, se, DEQUEUE_SLEEP | DEQUEUE_DELAYED);
5600 		/*
5601 		 * Must not reference @se again, see __block_task().
5602 		 */
5603 		return NULL;
5604 	}
5605 	return se;
5606 }
5607 
5608 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5609 
put_prev_entity(struct cfs_rq * cfs_rq,struct sched_entity * prev)5610 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5611 {
5612 	/*
5613 	 * If still on the runqueue then deactivate_task()
5614 	 * was not called and update_curr() has to be done:
5615 	 */
5616 	if (prev->on_rq)
5617 		update_curr(cfs_rq);
5618 
5619 	/* throttle cfs_rqs exceeding runtime */
5620 	check_cfs_rq_runtime(cfs_rq);
5621 
5622 	if (prev->on_rq) {
5623 		update_stats_wait_start_fair(cfs_rq, prev);
5624 		/* Put 'current' back into the tree. */
5625 		__enqueue_entity(cfs_rq, prev);
5626 		/* in !on_rq case, update occurred at dequeue */
5627 		update_load_avg(cfs_rq, prev, 0);
5628 	}
5629 	WARN_ON_ONCE(cfs_rq->curr != prev);
5630 	cfs_rq->curr = NULL;
5631 }
5632 
5633 static void
entity_tick(struct cfs_rq * cfs_rq,struct sched_entity * curr,int queued)5634 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5635 {
5636 	/*
5637 	 * Update run-time statistics of the 'current'.
5638 	 */
5639 	update_curr(cfs_rq);
5640 
5641 	/*
5642 	 * Ensure that runnable average is periodically updated.
5643 	 */
5644 	update_load_avg(cfs_rq, curr, UPDATE_TG);
5645 	update_cfs_group(curr);
5646 
5647 #ifdef CONFIG_SCHED_HRTICK
5648 	/*
5649 	 * queued ticks are scheduled to match the slice, so don't bother
5650 	 * validating it and just reschedule.
5651 	 */
5652 	if (queued) {
5653 		resched_curr_lazy(rq_of(cfs_rq));
5654 		return;
5655 	}
5656 #endif
5657 }
5658 
5659 
5660 /**************************************************
5661  * CFS bandwidth control machinery
5662  */
5663 
5664 #ifdef CONFIG_CFS_BANDWIDTH
5665 
5666 #ifdef CONFIG_JUMP_LABEL
5667 static struct static_key __cfs_bandwidth_used;
5668 
cfs_bandwidth_used(void)5669 static inline bool cfs_bandwidth_used(void)
5670 {
5671 	return static_key_false(&__cfs_bandwidth_used);
5672 }
5673 
cfs_bandwidth_usage_inc(void)5674 void cfs_bandwidth_usage_inc(void)
5675 {
5676 	static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5677 }
5678 
cfs_bandwidth_usage_dec(void)5679 void cfs_bandwidth_usage_dec(void)
5680 {
5681 	static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5682 }
5683 #else /* CONFIG_JUMP_LABEL */
cfs_bandwidth_used(void)5684 static bool cfs_bandwidth_used(void)
5685 {
5686 	return true;
5687 }
5688 
cfs_bandwidth_usage_inc(void)5689 void cfs_bandwidth_usage_inc(void) {}
cfs_bandwidth_usage_dec(void)5690 void cfs_bandwidth_usage_dec(void) {}
5691 #endif /* CONFIG_JUMP_LABEL */
5692 
5693 /*
5694  * default period for cfs group bandwidth.
5695  * default: 0.1s, units: nanoseconds
5696  */
default_cfs_period(void)5697 static inline u64 default_cfs_period(void)
5698 {
5699 	return 100000000ULL;
5700 }
5701 
sched_cfs_bandwidth_slice(void)5702 static inline u64 sched_cfs_bandwidth_slice(void)
5703 {
5704 	return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5705 }
5706 
5707 /*
5708  * Replenish runtime according to assigned quota. We use sched_clock_cpu
5709  * directly instead of rq->clock to avoid adding additional synchronization
5710  * around rq->lock.
5711  *
5712  * requires cfs_b->lock
5713  */
__refill_cfs_bandwidth_runtime(struct cfs_bandwidth * cfs_b)5714 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5715 {
5716 	s64 runtime;
5717 
5718 	if (unlikely(cfs_b->quota == RUNTIME_INF))
5719 		return;
5720 
5721 	cfs_b->runtime += cfs_b->quota;
5722 	runtime = cfs_b->runtime_snap - cfs_b->runtime;
5723 	if (runtime > 0) {
5724 		cfs_b->burst_time += runtime;
5725 		cfs_b->nr_burst++;
5726 	}
5727 
5728 	cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5729 	cfs_b->runtime_snap = cfs_b->runtime;
5730 }
5731 
tg_cfs_bandwidth(struct task_group * tg)5732 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5733 {
5734 	return &tg->cfs_bandwidth;
5735 }
5736 
5737 /* returns 0 on failure to allocate runtime */
__assign_cfs_rq_runtime(struct cfs_bandwidth * cfs_b,struct cfs_rq * cfs_rq,u64 target_runtime)5738 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5739 				   struct cfs_rq *cfs_rq, u64 target_runtime)
5740 {
5741 	u64 min_amount, amount = 0;
5742 
5743 	lockdep_assert_held(&cfs_b->lock);
5744 
5745 	/* note: this is a positive sum as runtime_remaining <= 0 */
5746 	min_amount = target_runtime - cfs_rq->runtime_remaining;
5747 
5748 	if (cfs_b->quota == RUNTIME_INF)
5749 		amount = min_amount;
5750 	else {
5751 		start_cfs_bandwidth(cfs_b);
5752 
5753 		if (cfs_b->runtime > 0) {
5754 			amount = min(cfs_b->runtime, min_amount);
5755 			cfs_b->runtime -= amount;
5756 			cfs_b->idle = 0;
5757 		}
5758 	}
5759 
5760 	cfs_rq->runtime_remaining += amount;
5761 
5762 	return cfs_rq->runtime_remaining > 0;
5763 }
5764 
5765 /* returns 0 on failure to allocate runtime */
assign_cfs_rq_runtime(struct cfs_rq * cfs_rq)5766 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5767 {
5768 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5769 	int ret;
5770 
5771 	raw_spin_lock(&cfs_b->lock);
5772 	ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5773 	raw_spin_unlock(&cfs_b->lock);
5774 
5775 	return ret;
5776 }
5777 
__account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)5778 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5779 {
5780 	/* dock delta_exec before expiring quota (as it could span periods) */
5781 	cfs_rq->runtime_remaining -= delta_exec;
5782 
5783 	if (likely(cfs_rq->runtime_remaining > 0))
5784 		return;
5785 
5786 	if (cfs_rq->throttled)
5787 		return;
5788 	/*
5789 	 * if we're unable to extend our runtime we resched so that the active
5790 	 * hierarchy can be throttled
5791 	 */
5792 	if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5793 		resched_curr(rq_of(cfs_rq));
5794 }
5795 
5796 static __always_inline
account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)5797 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5798 {
5799 	if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5800 		return;
5801 
5802 	__account_cfs_rq_runtime(cfs_rq, delta_exec);
5803 }
5804 
cfs_rq_throttled(struct cfs_rq * cfs_rq)5805 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5806 {
5807 	return cfs_bandwidth_used() && cfs_rq->throttled;
5808 }
5809 
5810 /* check whether cfs_rq, or any parent, is throttled */
throttled_hierarchy(struct cfs_rq * cfs_rq)5811 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5812 {
5813 	return cfs_bandwidth_used() && cfs_rq->throttle_count;
5814 }
5815 
5816 /*
5817  * Ensure that neither of the group entities corresponding to src_cpu or
5818  * dest_cpu are members of a throttled hierarchy when performing group
5819  * load-balance operations.
5820  */
throttled_lb_pair(struct task_group * tg,int src_cpu,int dest_cpu)5821 static inline int throttled_lb_pair(struct task_group *tg,
5822 				    int src_cpu, int dest_cpu)
5823 {
5824 	struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5825 
5826 	src_cfs_rq = tg->cfs_rq[src_cpu];
5827 	dest_cfs_rq = tg->cfs_rq[dest_cpu];
5828 
5829 	return throttled_hierarchy(src_cfs_rq) ||
5830 	       throttled_hierarchy(dest_cfs_rq);
5831 }
5832 
tg_unthrottle_up(struct task_group * tg,void * data)5833 static int tg_unthrottle_up(struct task_group *tg, void *data)
5834 {
5835 	struct rq *rq = data;
5836 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5837 
5838 	cfs_rq->throttle_count--;
5839 	if (!cfs_rq->throttle_count) {
5840 		cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5841 					     cfs_rq->throttled_clock_pelt;
5842 
5843 		/* Add cfs_rq with load or one or more already running entities to the list */
5844 		if (!cfs_rq_is_decayed(cfs_rq))
5845 			list_add_leaf_cfs_rq(cfs_rq);
5846 
5847 		if (cfs_rq->throttled_clock_self) {
5848 			u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;
5849 
5850 			cfs_rq->throttled_clock_self = 0;
5851 
5852 			if (WARN_ON_ONCE((s64)delta < 0))
5853 				delta = 0;
5854 
5855 			cfs_rq->throttled_clock_self_time += delta;
5856 		}
5857 	}
5858 
5859 	return 0;
5860 }
5861 
tg_throttle_down(struct task_group * tg,void * data)5862 static int tg_throttle_down(struct task_group *tg, void *data)
5863 {
5864 	struct rq *rq = data;
5865 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5866 
5867 	/* group is entering throttled state, stop time */
5868 	if (!cfs_rq->throttle_count) {
5869 		cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5870 		list_del_leaf_cfs_rq(cfs_rq);
5871 
5872 		WARN_ON_ONCE(cfs_rq->throttled_clock_self);
5873 		if (cfs_rq->nr_queued)
5874 			cfs_rq->throttled_clock_self = rq_clock(rq);
5875 	}
5876 	cfs_rq->throttle_count++;
5877 
5878 	return 0;
5879 }
5880 
throttle_cfs_rq(struct cfs_rq * cfs_rq)5881 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5882 {
5883 	struct rq *rq = rq_of(cfs_rq);
5884 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5885 	struct sched_entity *se;
5886 	long queued_delta, runnable_delta, idle_delta, dequeue = 1;
5887 	long rq_h_nr_queued = rq->cfs.h_nr_queued;
5888 
5889 	raw_spin_lock(&cfs_b->lock);
5890 	/* This will start the period timer if necessary */
5891 	if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5892 		/*
5893 		 * We have raced with bandwidth becoming available, and if we
5894 		 * actually throttled the timer might not unthrottle us for an
5895 		 * entire period. We additionally needed to make sure that any
5896 		 * subsequent check_cfs_rq_runtime calls agree not to throttle
5897 		 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5898 		 * for 1ns of runtime rather than just check cfs_b.
5899 		 */
5900 		dequeue = 0;
5901 	} else {
5902 		list_add_tail_rcu(&cfs_rq->throttled_list,
5903 				  &cfs_b->throttled_cfs_rq);
5904 	}
5905 	raw_spin_unlock(&cfs_b->lock);
5906 
5907 	if (!dequeue)
5908 		return false;  /* Throttle no longer required. */
5909 
5910 	se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5911 
5912 	/* freeze hierarchy runnable averages while throttled */
5913 	rcu_read_lock();
5914 	walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5915 	rcu_read_unlock();
5916 
5917 	queued_delta = cfs_rq->h_nr_queued;
5918 	runnable_delta = cfs_rq->h_nr_runnable;
5919 	idle_delta = cfs_rq->h_nr_idle;
5920 	for_each_sched_entity(se) {
5921 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5922 		int flags;
5923 
5924 		/* throttled entity or throttle-on-deactivate */
5925 		if (!se->on_rq)
5926 			goto done;
5927 
5928 		/*
5929 		 * Abuse SPECIAL to avoid delayed dequeue in this instance.
5930 		 * This avoids teaching dequeue_entities() about throttled
5931 		 * entities and keeps things relatively simple.
5932 		 */
5933 		flags = DEQUEUE_SLEEP | DEQUEUE_SPECIAL;
5934 		if (se->sched_delayed)
5935 			flags |= DEQUEUE_DELAYED;
5936 		dequeue_entity(qcfs_rq, se, flags);
5937 
5938 		if (cfs_rq_is_idle(group_cfs_rq(se)))
5939 			idle_delta = cfs_rq->h_nr_queued;
5940 
5941 		qcfs_rq->h_nr_queued -= queued_delta;
5942 		qcfs_rq->h_nr_runnable -= runnable_delta;
5943 		qcfs_rq->h_nr_idle -= idle_delta;
5944 
5945 		if (qcfs_rq->load.weight) {
5946 			/* Avoid re-evaluating load for this entity: */
5947 			se = parent_entity(se);
5948 			break;
5949 		}
5950 	}
5951 
5952 	for_each_sched_entity(se) {
5953 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5954 		/* throttled entity or throttle-on-deactivate */
5955 		if (!se->on_rq)
5956 			goto done;
5957 
5958 		update_load_avg(qcfs_rq, se, 0);
5959 		se_update_runnable(se);
5960 
5961 		if (cfs_rq_is_idle(group_cfs_rq(se)))
5962 			idle_delta = cfs_rq->h_nr_queued;
5963 
5964 		qcfs_rq->h_nr_queued -= queued_delta;
5965 		qcfs_rq->h_nr_runnable -= runnable_delta;
5966 		qcfs_rq->h_nr_idle -= idle_delta;
5967 	}
5968 
5969 	/* At this point se is NULL and we are at root level*/
5970 	sub_nr_running(rq, queued_delta);
5971 
5972 	/* Stop the fair server if throttling resulted in no runnable tasks */
5973 	if (rq_h_nr_queued && !rq->cfs.h_nr_queued)
5974 		dl_server_stop(&rq->fair_server);
5975 done:
5976 	/*
5977 	 * Note: distribution will already see us throttled via the
5978 	 * throttled-list.  rq->lock protects completion.
5979 	 */
5980 	cfs_rq->throttled = 1;
5981 	WARN_ON_ONCE(cfs_rq->throttled_clock);
5982 	if (cfs_rq->nr_queued)
5983 		cfs_rq->throttled_clock = rq_clock(rq);
5984 	return true;
5985 }
5986 
unthrottle_cfs_rq(struct cfs_rq * cfs_rq)5987 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5988 {
5989 	struct rq *rq = rq_of(cfs_rq);
5990 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5991 	struct sched_entity *se;
5992 	long queued_delta, runnable_delta, idle_delta;
5993 	long rq_h_nr_queued = rq->cfs.h_nr_queued;
5994 
5995 	se = cfs_rq->tg->se[cpu_of(rq)];
5996 
5997 	cfs_rq->throttled = 0;
5998 
5999 	update_rq_clock(rq);
6000 
6001 	raw_spin_lock(&cfs_b->lock);
6002 	if (cfs_rq->throttled_clock) {
6003 		cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
6004 		cfs_rq->throttled_clock = 0;
6005 	}
6006 	list_del_rcu(&cfs_rq->throttled_list);
6007 	raw_spin_unlock(&cfs_b->lock);
6008 
6009 	/* update hierarchical throttle state */
6010 	walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
6011 
6012 	if (!cfs_rq->load.weight) {
6013 		if (!cfs_rq->on_list)
6014 			return;
6015 		/*
6016 		 * Nothing to run but something to decay (on_list)?
6017 		 * Complete the branch.
6018 		 */
6019 		for_each_sched_entity(se) {
6020 			if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
6021 				break;
6022 		}
6023 		goto unthrottle_throttle;
6024 	}
6025 
6026 	queued_delta = cfs_rq->h_nr_queued;
6027 	runnable_delta = cfs_rq->h_nr_runnable;
6028 	idle_delta = cfs_rq->h_nr_idle;
6029 	for_each_sched_entity(se) {
6030 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
6031 
6032 		/* Handle any unfinished DELAY_DEQUEUE business first. */
6033 		if (se->sched_delayed) {
6034 			int flags = DEQUEUE_SLEEP | DEQUEUE_DELAYED;
6035 
6036 			dequeue_entity(qcfs_rq, se, flags);
6037 		} else if (se->on_rq)
6038 			break;
6039 		enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
6040 
6041 		if (cfs_rq_is_idle(group_cfs_rq(se)))
6042 			idle_delta = cfs_rq->h_nr_queued;
6043 
6044 		qcfs_rq->h_nr_queued += queued_delta;
6045 		qcfs_rq->h_nr_runnable += runnable_delta;
6046 		qcfs_rq->h_nr_idle += idle_delta;
6047 
6048 		/* end evaluation on encountering a throttled cfs_rq */
6049 		if (cfs_rq_throttled(qcfs_rq))
6050 			goto unthrottle_throttle;
6051 	}
6052 
6053 	for_each_sched_entity(se) {
6054 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
6055 
6056 		update_load_avg(qcfs_rq, se, UPDATE_TG);
6057 		se_update_runnable(se);
6058 
6059 		if (cfs_rq_is_idle(group_cfs_rq(se)))
6060 			idle_delta = cfs_rq->h_nr_queued;
6061 
6062 		qcfs_rq->h_nr_queued += queued_delta;
6063 		qcfs_rq->h_nr_runnable += runnable_delta;
6064 		qcfs_rq->h_nr_idle += idle_delta;
6065 
6066 		/* end evaluation on encountering a throttled cfs_rq */
6067 		if (cfs_rq_throttled(qcfs_rq))
6068 			goto unthrottle_throttle;
6069 	}
6070 
6071 	/* Start the fair server if un-throttling resulted in new runnable tasks */
6072 	if (!rq_h_nr_queued && rq->cfs.h_nr_queued)
6073 		dl_server_start(&rq->fair_server);
6074 
6075 	/* At this point se is NULL and we are at root level*/
6076 	add_nr_running(rq, queued_delta);
6077 
6078 unthrottle_throttle:
6079 	assert_list_leaf_cfs_rq(rq);
6080 
6081 	/* Determine whether we need to wake up potentially idle CPU: */
6082 	if (rq->curr == rq->idle && rq->cfs.nr_queued)
6083 		resched_curr(rq);
6084 }
6085 
6086 #ifdef CONFIG_SMP
__cfsb_csd_unthrottle(void * arg)6087 static void __cfsb_csd_unthrottle(void *arg)
6088 {
6089 	struct cfs_rq *cursor, *tmp;
6090 	struct rq *rq = arg;
6091 	struct rq_flags rf;
6092 
6093 	rq_lock(rq, &rf);
6094 
6095 	/*
6096 	 * Iterating over the list can trigger several call to
6097 	 * update_rq_clock() in unthrottle_cfs_rq().
6098 	 * Do it once and skip the potential next ones.
6099 	 */
6100 	update_rq_clock(rq);
6101 	rq_clock_start_loop_update(rq);
6102 
6103 	/*
6104 	 * Since we hold rq lock we're safe from concurrent manipulation of
6105 	 * the CSD list. However, this RCU critical section annotates the
6106 	 * fact that we pair with sched_free_group_rcu(), so that we cannot
6107 	 * race with group being freed in the window between removing it
6108 	 * from the list and advancing to the next entry in the list.
6109 	 */
6110 	rcu_read_lock();
6111 
6112 	list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
6113 				 throttled_csd_list) {
6114 		list_del_init(&cursor->throttled_csd_list);
6115 
6116 		if (cfs_rq_throttled(cursor))
6117 			unthrottle_cfs_rq(cursor);
6118 	}
6119 
6120 	rcu_read_unlock();
6121 
6122 	rq_clock_stop_loop_update(rq);
6123 	rq_unlock(rq, &rf);
6124 }
6125 
__unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)6126 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
6127 {
6128 	struct rq *rq = rq_of(cfs_rq);
6129 	bool first;
6130 
6131 	if (rq == this_rq()) {
6132 		unthrottle_cfs_rq(cfs_rq);
6133 		return;
6134 	}
6135 
6136 	/* Already enqueued */
6137 	if (WARN_ON_ONCE(!list_empty(&cfs_rq->throttled_csd_list)))
6138 		return;
6139 
6140 	first = list_empty(&rq->cfsb_csd_list);
6141 	list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
6142 	if (first)
6143 		smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
6144 }
6145 #else
__unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)6146 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
6147 {
6148 	unthrottle_cfs_rq(cfs_rq);
6149 }
6150 #endif
6151 
unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)6152 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
6153 {
6154 	lockdep_assert_rq_held(rq_of(cfs_rq));
6155 
6156 	if (WARN_ON_ONCE(!cfs_rq_throttled(cfs_rq) ||
6157 	    cfs_rq->runtime_remaining <= 0))
6158 		return;
6159 
6160 	__unthrottle_cfs_rq_async(cfs_rq);
6161 }
6162 
distribute_cfs_runtime(struct cfs_bandwidth * cfs_b)6163 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
6164 {
6165 	int this_cpu = smp_processor_id();
6166 	u64 runtime, remaining = 1;
6167 	bool throttled = false;
6168 	struct cfs_rq *cfs_rq, *tmp;
6169 	struct rq_flags rf;
6170 	struct rq *rq;
6171 	LIST_HEAD(local_unthrottle);
6172 
6173 	rcu_read_lock();
6174 	list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
6175 				throttled_list) {
6176 		rq = rq_of(cfs_rq);
6177 
6178 		if (!remaining) {
6179 			throttled = true;
6180 			break;
6181 		}
6182 
6183 		rq_lock_irqsave(rq, &rf);
6184 		if (!cfs_rq_throttled(cfs_rq))
6185 			goto next;
6186 
6187 		/* Already queued for async unthrottle */
6188 		if (!list_empty(&cfs_rq->throttled_csd_list))
6189 			goto next;
6190 
6191 		/* By the above checks, this should never be true */
6192 		WARN_ON_ONCE(cfs_rq->runtime_remaining > 0);
6193 
6194 		raw_spin_lock(&cfs_b->lock);
6195 		runtime = -cfs_rq->runtime_remaining + 1;
6196 		if (runtime > cfs_b->runtime)
6197 			runtime = cfs_b->runtime;
6198 		cfs_b->runtime -= runtime;
6199 		remaining = cfs_b->runtime;
6200 		raw_spin_unlock(&cfs_b->lock);
6201 
6202 		cfs_rq->runtime_remaining += runtime;
6203 
6204 		/* we check whether we're throttled above */
6205 		if (cfs_rq->runtime_remaining > 0) {
6206 			if (cpu_of(rq) != this_cpu) {
6207 				unthrottle_cfs_rq_async(cfs_rq);
6208 			} else {
6209 				/*
6210 				 * We currently only expect to be unthrottling
6211 				 * a single cfs_rq locally.
6212 				 */
6213 				WARN_ON_ONCE(!list_empty(&local_unthrottle));
6214 				list_add_tail(&cfs_rq->throttled_csd_list,
6215 					      &local_unthrottle);
6216 			}
6217 		} else {
6218 			throttled = true;
6219 		}
6220 
6221 next:
6222 		rq_unlock_irqrestore(rq, &rf);
6223 	}
6224 
6225 	list_for_each_entry_safe(cfs_rq, tmp, &local_unthrottle,
6226 				 throttled_csd_list) {
6227 		struct rq *rq = rq_of(cfs_rq);
6228 
6229 		rq_lock_irqsave(rq, &rf);
6230 
6231 		list_del_init(&cfs_rq->throttled_csd_list);
6232 
6233 		if (cfs_rq_throttled(cfs_rq))
6234 			unthrottle_cfs_rq(cfs_rq);
6235 
6236 		rq_unlock_irqrestore(rq, &rf);
6237 	}
6238 	WARN_ON_ONCE(!list_empty(&local_unthrottle));
6239 
6240 	rcu_read_unlock();
6241 
6242 	return throttled;
6243 }
6244 
6245 /*
6246  * Responsible for refilling a task_group's bandwidth and unthrottling its
6247  * cfs_rqs as appropriate. If there has been no activity within the last
6248  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
6249  * used to track this state.
6250  */
do_sched_cfs_period_timer(struct cfs_bandwidth * cfs_b,int overrun,unsigned long flags)6251 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
6252 {
6253 	int throttled;
6254 
6255 	/* no need to continue the timer with no bandwidth constraint */
6256 	if (cfs_b->quota == RUNTIME_INF)
6257 		goto out_deactivate;
6258 
6259 	throttled = !list_empty(&cfs_b->throttled_cfs_rq);
6260 	cfs_b->nr_periods += overrun;
6261 
6262 	/* Refill extra burst quota even if cfs_b->idle */
6263 	__refill_cfs_bandwidth_runtime(cfs_b);
6264 
6265 	/*
6266 	 * idle depends on !throttled (for the case of a large deficit), and if
6267 	 * we're going inactive then everything else can be deferred
6268 	 */
6269 	if (cfs_b->idle && !throttled)
6270 		goto out_deactivate;
6271 
6272 	if (!throttled) {
6273 		/* mark as potentially idle for the upcoming period */
6274 		cfs_b->idle = 1;
6275 		return 0;
6276 	}
6277 
6278 	/* account preceding periods in which throttling occurred */
6279 	cfs_b->nr_throttled += overrun;
6280 
6281 	/*
6282 	 * This check is repeated as we release cfs_b->lock while we unthrottle.
6283 	 */
6284 	while (throttled && cfs_b->runtime > 0) {
6285 		raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6286 		/* we can't nest cfs_b->lock while distributing bandwidth */
6287 		throttled = distribute_cfs_runtime(cfs_b);
6288 		raw_spin_lock_irqsave(&cfs_b->lock, flags);
6289 	}
6290 
6291 	/*
6292 	 * While we are ensured activity in the period following an
6293 	 * unthrottle, this also covers the case in which the new bandwidth is
6294 	 * insufficient to cover the existing bandwidth deficit.  (Forcing the
6295 	 * timer to remain active while there are any throttled entities.)
6296 	 */
6297 	cfs_b->idle = 0;
6298 
6299 	return 0;
6300 
6301 out_deactivate:
6302 	return 1;
6303 }
6304 
6305 /* a cfs_rq won't donate quota below this amount */
6306 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
6307 /* minimum remaining period time to redistribute slack quota */
6308 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
6309 /* how long we wait to gather additional slack before distributing */
6310 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
6311 
6312 /*
6313  * Are we near the end of the current quota period?
6314  *
6315  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
6316  * hrtimer base being cleared by hrtimer_start. In the case of
6317  * migrate_hrtimers, base is never cleared, so we are fine.
6318  */
runtime_refresh_within(struct cfs_bandwidth * cfs_b,u64 min_expire)6319 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
6320 {
6321 	struct hrtimer *refresh_timer = &cfs_b->period_timer;
6322 	s64 remaining;
6323 
6324 	/* if the call-back is running a quota refresh is already occurring */
6325 	if (hrtimer_callback_running(refresh_timer))
6326 		return 1;
6327 
6328 	/* is a quota refresh about to occur? */
6329 	remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
6330 	if (remaining < (s64)min_expire)
6331 		return 1;
6332 
6333 	return 0;
6334 }
6335 
start_cfs_slack_bandwidth(struct cfs_bandwidth * cfs_b)6336 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
6337 {
6338 	u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
6339 
6340 	/* if there's a quota refresh soon don't bother with slack */
6341 	if (runtime_refresh_within(cfs_b, min_left))
6342 		return;
6343 
6344 	/* don't push forwards an existing deferred unthrottle */
6345 	if (cfs_b->slack_started)
6346 		return;
6347 	cfs_b->slack_started = true;
6348 
6349 	hrtimer_start(&cfs_b->slack_timer,
6350 			ns_to_ktime(cfs_bandwidth_slack_period),
6351 			HRTIMER_MODE_REL);
6352 }
6353 
6354 /* we know any runtime found here is valid as update_curr() precedes return */
__return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6355 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6356 {
6357 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6358 	s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
6359 
6360 	if (slack_runtime <= 0)
6361 		return;
6362 
6363 	raw_spin_lock(&cfs_b->lock);
6364 	if (cfs_b->quota != RUNTIME_INF) {
6365 		cfs_b->runtime += slack_runtime;
6366 
6367 		/* we are under rq->lock, defer unthrottling using a timer */
6368 		if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
6369 		    !list_empty(&cfs_b->throttled_cfs_rq))
6370 			start_cfs_slack_bandwidth(cfs_b);
6371 	}
6372 	raw_spin_unlock(&cfs_b->lock);
6373 
6374 	/* even if it's not valid for return we don't want to try again */
6375 	cfs_rq->runtime_remaining -= slack_runtime;
6376 }
6377 
return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6378 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6379 {
6380 	if (!cfs_bandwidth_used())
6381 		return;
6382 
6383 	if (!cfs_rq->runtime_enabled || cfs_rq->nr_queued)
6384 		return;
6385 
6386 	__return_cfs_rq_runtime(cfs_rq);
6387 }
6388 
6389 /*
6390  * This is done with a timer (instead of inline with bandwidth return) since
6391  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6392  */
do_sched_cfs_slack_timer(struct cfs_bandwidth * cfs_b)6393 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
6394 {
6395 	u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
6396 	unsigned long flags;
6397 
6398 	/* confirm we're still not at a refresh boundary */
6399 	raw_spin_lock_irqsave(&cfs_b->lock, flags);
6400 	cfs_b->slack_started = false;
6401 
6402 	if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
6403 		raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6404 		return;
6405 	}
6406 
6407 	if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
6408 		runtime = cfs_b->runtime;
6409 
6410 	raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6411 
6412 	if (!runtime)
6413 		return;
6414 
6415 	distribute_cfs_runtime(cfs_b);
6416 }
6417 
6418 /*
6419  * When a group wakes up we want to make sure that its quota is not already
6420  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
6421  * runtime as update_curr() throttling can not trigger until it's on-rq.
6422  */
check_enqueue_throttle(struct cfs_rq * cfs_rq)6423 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
6424 {
6425 	if (!cfs_bandwidth_used())
6426 		return;
6427 
6428 	/* an active group must be handled by the update_curr()->put() path */
6429 	if (!cfs_rq->runtime_enabled || cfs_rq->curr)
6430 		return;
6431 
6432 	/* ensure the group is not already throttled */
6433 	if (cfs_rq_throttled(cfs_rq))
6434 		return;
6435 
6436 	/* update runtime allocation */
6437 	account_cfs_rq_runtime(cfs_rq, 0);
6438 	if (cfs_rq->runtime_remaining <= 0)
6439 		throttle_cfs_rq(cfs_rq);
6440 }
6441 
sync_throttle(struct task_group * tg,int cpu)6442 static void sync_throttle(struct task_group *tg, int cpu)
6443 {
6444 	struct cfs_rq *pcfs_rq, *cfs_rq;
6445 
6446 	if (!cfs_bandwidth_used())
6447 		return;
6448 
6449 	if (!tg->parent)
6450 		return;
6451 
6452 	cfs_rq = tg->cfs_rq[cpu];
6453 	pcfs_rq = tg->parent->cfs_rq[cpu];
6454 
6455 	cfs_rq->throttle_count = pcfs_rq->throttle_count;
6456 	cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
6457 }
6458 
6459 /* conditionally throttle active cfs_rq's from put_prev_entity() */
check_cfs_rq_runtime(struct cfs_rq * cfs_rq)6460 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6461 {
6462 	if (!cfs_bandwidth_used())
6463 		return false;
6464 
6465 	if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
6466 		return false;
6467 
6468 	/*
6469 	 * it's possible for a throttled entity to be forced into a running
6470 	 * state (e.g. set_curr_task), in this case we're finished.
6471 	 */
6472 	if (cfs_rq_throttled(cfs_rq))
6473 		return true;
6474 
6475 	return throttle_cfs_rq(cfs_rq);
6476 }
6477 
sched_cfs_slack_timer(struct hrtimer * timer)6478 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
6479 {
6480 	struct cfs_bandwidth *cfs_b =
6481 		container_of(timer, struct cfs_bandwidth, slack_timer);
6482 
6483 	do_sched_cfs_slack_timer(cfs_b);
6484 
6485 	return HRTIMER_NORESTART;
6486 }
6487 
6488 extern const u64 max_cfs_quota_period;
6489 
sched_cfs_period_timer(struct hrtimer * timer)6490 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
6491 {
6492 	struct cfs_bandwidth *cfs_b =
6493 		container_of(timer, struct cfs_bandwidth, period_timer);
6494 	unsigned long flags;
6495 	int overrun;
6496 	int idle = 0;
6497 	int count = 0;
6498 
6499 	raw_spin_lock_irqsave(&cfs_b->lock, flags);
6500 	for (;;) {
6501 		overrun = hrtimer_forward_now(timer, cfs_b->period);
6502 		if (!overrun)
6503 			break;
6504 
6505 		idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
6506 
6507 		if (++count > 3) {
6508 			u64 new, old = ktime_to_ns(cfs_b->period);
6509 
6510 			/*
6511 			 * Grow period by a factor of 2 to avoid losing precision.
6512 			 * Precision loss in the quota/period ratio can cause __cfs_schedulable
6513 			 * to fail.
6514 			 */
6515 			new = old * 2;
6516 			if (new < max_cfs_quota_period) {
6517 				cfs_b->period = ns_to_ktime(new);
6518 				cfs_b->quota *= 2;
6519 				cfs_b->burst *= 2;
6520 
6521 				pr_warn_ratelimited(
6522 	"cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6523 					smp_processor_id(),
6524 					div_u64(new, NSEC_PER_USEC),
6525 					div_u64(cfs_b->quota, NSEC_PER_USEC));
6526 			} else {
6527 				pr_warn_ratelimited(
6528 	"cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6529 					smp_processor_id(),
6530 					div_u64(old, NSEC_PER_USEC),
6531 					div_u64(cfs_b->quota, NSEC_PER_USEC));
6532 			}
6533 
6534 			/* reset count so we don't come right back in here */
6535 			count = 0;
6536 		}
6537 	}
6538 	if (idle)
6539 		cfs_b->period_active = 0;
6540 	raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6541 
6542 	return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
6543 }
6544 
init_cfs_bandwidth(struct cfs_bandwidth * cfs_b,struct cfs_bandwidth * parent)6545 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent)
6546 {
6547 	raw_spin_lock_init(&cfs_b->lock);
6548 	cfs_b->runtime = 0;
6549 	cfs_b->quota = RUNTIME_INF;
6550 	cfs_b->period = ns_to_ktime(default_cfs_period());
6551 	cfs_b->burst = 0;
6552 	cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;
6553 
6554 	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
6555 	hrtimer_setup(&cfs_b->period_timer, sched_cfs_period_timer, CLOCK_MONOTONIC,
6556 		      HRTIMER_MODE_ABS_PINNED);
6557 
6558 	/* Add a random offset so that timers interleave */
6559 	hrtimer_set_expires(&cfs_b->period_timer,
6560 			    get_random_u32_below(cfs_b->period));
6561 	hrtimer_setup(&cfs_b->slack_timer, sched_cfs_slack_timer, CLOCK_MONOTONIC,
6562 		      HRTIMER_MODE_REL);
6563 	cfs_b->slack_started = false;
6564 }
6565 
init_cfs_rq_runtime(struct cfs_rq * cfs_rq)6566 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6567 {
6568 	cfs_rq->runtime_enabled = 0;
6569 	INIT_LIST_HEAD(&cfs_rq->throttled_list);
6570 	INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
6571 }
6572 
start_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6573 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6574 {
6575 	lockdep_assert_held(&cfs_b->lock);
6576 
6577 	if (cfs_b->period_active)
6578 		return;
6579 
6580 	cfs_b->period_active = 1;
6581 	hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
6582 	hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
6583 }
6584 
destroy_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6585 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6586 {
6587 	int __maybe_unused i;
6588 
6589 	/* init_cfs_bandwidth() was not called */
6590 	if (!cfs_b->throttled_cfs_rq.next)
6591 		return;
6592 
6593 	hrtimer_cancel(&cfs_b->period_timer);
6594 	hrtimer_cancel(&cfs_b->slack_timer);
6595 
6596 	/*
6597 	 * It is possible that we still have some cfs_rq's pending on a CSD
6598 	 * list, though this race is very rare. In order for this to occur, we
6599 	 * must have raced with the last task leaving the group while there
6600 	 * exist throttled cfs_rq(s), and the period_timer must have queued the
6601 	 * CSD item but the remote cpu has not yet processed it. To handle this,
6602 	 * we can simply flush all pending CSD work inline here. We're
6603 	 * guaranteed at this point that no additional cfs_rq of this group can
6604 	 * join a CSD list.
6605 	 */
6606 #ifdef CONFIG_SMP
6607 	for_each_possible_cpu(i) {
6608 		struct rq *rq = cpu_rq(i);
6609 		unsigned long flags;
6610 
6611 		if (list_empty(&rq->cfsb_csd_list))
6612 			continue;
6613 
6614 		local_irq_save(flags);
6615 		__cfsb_csd_unthrottle(rq);
6616 		local_irq_restore(flags);
6617 	}
6618 #endif
6619 }
6620 
6621 /*
6622  * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
6623  *
6624  * The race is harmless, since modifying bandwidth settings of unhooked group
6625  * bits doesn't do much.
6626  */
6627 
6628 /* cpu online callback */
update_runtime_enabled(struct rq * rq)6629 static void __maybe_unused update_runtime_enabled(struct rq *rq)
6630 {
6631 	struct task_group *tg;
6632 
6633 	lockdep_assert_rq_held(rq);
6634 
6635 	rcu_read_lock();
6636 	list_for_each_entry_rcu(tg, &task_groups, list) {
6637 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6638 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6639 
6640 		raw_spin_lock(&cfs_b->lock);
6641 		cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6642 		raw_spin_unlock(&cfs_b->lock);
6643 	}
6644 	rcu_read_unlock();
6645 }
6646 
6647 /* cpu offline callback */
unthrottle_offline_cfs_rqs(struct rq * rq)6648 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6649 {
6650 	struct task_group *tg;
6651 
6652 	lockdep_assert_rq_held(rq);
6653 
6654 	// Do not unthrottle for an active CPU
6655 	if (cpumask_test_cpu(cpu_of(rq), cpu_active_mask))
6656 		return;
6657 
6658 	/*
6659 	 * The rq clock has already been updated in the
6660 	 * set_rq_offline(), so we should skip updating
6661 	 * the rq clock again in unthrottle_cfs_rq().
6662 	 */
6663 	rq_clock_start_loop_update(rq);
6664 
6665 	rcu_read_lock();
6666 	list_for_each_entry_rcu(tg, &task_groups, list) {
6667 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6668 
6669 		if (!cfs_rq->runtime_enabled)
6670 			continue;
6671 
6672 		/*
6673 		 * Offline rq is schedulable till CPU is completely disabled
6674 		 * in take_cpu_down(), so we prevent new cfs throttling here.
6675 		 */
6676 		cfs_rq->runtime_enabled = 0;
6677 
6678 		if (!cfs_rq_throttled(cfs_rq))
6679 			continue;
6680 
6681 		/*
6682 		 * clock_task is not advancing so we just need to make sure
6683 		 * there's some valid quota amount
6684 		 */
6685 		cfs_rq->runtime_remaining = 1;
6686 		unthrottle_cfs_rq(cfs_rq);
6687 	}
6688 	rcu_read_unlock();
6689 
6690 	rq_clock_stop_loop_update(rq);
6691 }
6692 
cfs_task_bw_constrained(struct task_struct * p)6693 bool cfs_task_bw_constrained(struct task_struct *p)
6694 {
6695 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
6696 
6697 	if (!cfs_bandwidth_used())
6698 		return false;
6699 
6700 	if (cfs_rq->runtime_enabled ||
6701 	    tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)
6702 		return true;
6703 
6704 	return false;
6705 }
6706 
6707 #ifdef CONFIG_NO_HZ_FULL
6708 /* called from pick_next_task_fair() */
sched_fair_update_stop_tick(struct rq * rq,struct task_struct * p)6709 static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
6710 {
6711 	int cpu = cpu_of(rq);
6712 
6713 	if (!cfs_bandwidth_used())
6714 		return;
6715 
6716 	if (!tick_nohz_full_cpu(cpu))
6717 		return;
6718 
6719 	if (rq->nr_running != 1)
6720 		return;
6721 
6722 	/*
6723 	 *  We know there is only one task runnable and we've just picked it. The
6724 	 *  normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
6725 	 *  be otherwise able to stop the tick. Just need to check if we are using
6726 	 *  bandwidth control.
6727 	 */
6728 	if (cfs_task_bw_constrained(p))
6729 		tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
6730 }
6731 #endif
6732 
6733 #else /* CONFIG_CFS_BANDWIDTH */
6734 
account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)6735 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
check_cfs_rq_runtime(struct cfs_rq * cfs_rq)6736 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
check_enqueue_throttle(struct cfs_rq * cfs_rq)6737 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
sync_throttle(struct task_group * tg,int cpu)6738 static inline void sync_throttle(struct task_group *tg, int cpu) {}
return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6739 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6740 
cfs_rq_throttled(struct cfs_rq * cfs_rq)6741 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6742 {
6743 	return 0;
6744 }
6745 
throttled_hierarchy(struct cfs_rq * cfs_rq)6746 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6747 {
6748 	return 0;
6749 }
6750 
throttled_lb_pair(struct task_group * tg,int src_cpu,int dest_cpu)6751 static inline int throttled_lb_pair(struct task_group *tg,
6752 				    int src_cpu, int dest_cpu)
6753 {
6754 	return 0;
6755 }
6756 
6757 #ifdef CONFIG_FAIR_GROUP_SCHED
init_cfs_bandwidth(struct cfs_bandwidth * cfs_b,struct cfs_bandwidth * parent)6758 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {}
init_cfs_rq_runtime(struct cfs_rq * cfs_rq)6759 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6760 #endif
6761 
tg_cfs_bandwidth(struct task_group * tg)6762 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6763 {
6764 	return NULL;
6765 }
destroy_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6766 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
update_runtime_enabled(struct rq * rq)6767 static inline void update_runtime_enabled(struct rq *rq) {}
unthrottle_offline_cfs_rqs(struct rq * rq)6768 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6769 #ifdef CONFIG_CGROUP_SCHED
cfs_task_bw_constrained(struct task_struct * p)6770 bool cfs_task_bw_constrained(struct task_struct *p)
6771 {
6772 	return false;
6773 }
6774 #endif
6775 #endif /* CONFIG_CFS_BANDWIDTH */
6776 
6777 #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL)
sched_fair_update_stop_tick(struct rq * rq,struct task_struct * p)6778 static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {}
6779 #endif
6780 
6781 /**************************************************
6782  * CFS operations on tasks:
6783  */
6784 
6785 #ifdef CONFIG_SCHED_HRTICK
hrtick_start_fair(struct rq * rq,struct task_struct * p)6786 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6787 {
6788 	struct sched_entity *se = &p->se;
6789 
6790 	WARN_ON_ONCE(task_rq(p) != rq);
6791 
6792 	if (rq->cfs.h_nr_queued > 1) {
6793 		u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6794 		u64 slice = se->slice;
6795 		s64 delta = slice - ran;
6796 
6797 		if (delta < 0) {
6798 			if (task_current_donor(rq, p))
6799 				resched_curr(rq);
6800 			return;
6801 		}
6802 		hrtick_start(rq, delta);
6803 	}
6804 }
6805 
6806 /*
6807  * called from enqueue/dequeue and updates the hrtick when the
6808  * current task is from our class and nr_running is low enough
6809  * to matter.
6810  */
hrtick_update(struct rq * rq)6811 static void hrtick_update(struct rq *rq)
6812 {
6813 	struct task_struct *donor = rq->donor;
6814 
6815 	if (!hrtick_enabled_fair(rq) || donor->sched_class != &fair_sched_class)
6816 		return;
6817 
6818 	hrtick_start_fair(rq, donor);
6819 }
6820 #else /* !CONFIG_SCHED_HRTICK */
6821 static inline void
hrtick_start_fair(struct rq * rq,struct task_struct * p)6822 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6823 {
6824 }
6825 
hrtick_update(struct rq * rq)6826 static inline void hrtick_update(struct rq *rq)
6827 {
6828 }
6829 #endif
6830 
6831 #ifdef CONFIG_SMP
cpu_overutilized(int cpu)6832 static inline bool cpu_overutilized(int cpu)
6833 {
6834 	unsigned long  rq_util_min, rq_util_max;
6835 
6836 	if (!sched_energy_enabled())
6837 		return false;
6838 
6839 	rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6840 	rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6841 
6842 	/* Return true only if the utilization doesn't fit CPU's capacity */
6843 	return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6844 }
6845 
6846 /*
6847  * overutilized value make sense only if EAS is enabled
6848  */
is_rd_overutilized(struct root_domain * rd)6849 static inline bool is_rd_overutilized(struct root_domain *rd)
6850 {
6851 	return !sched_energy_enabled() || READ_ONCE(rd->overutilized);
6852 }
6853 
set_rd_overutilized(struct root_domain * rd,bool flag)6854 static inline void set_rd_overutilized(struct root_domain *rd, bool flag)
6855 {
6856 	if (!sched_energy_enabled())
6857 		return;
6858 
6859 	WRITE_ONCE(rd->overutilized, flag);
6860 	trace_sched_overutilized_tp(rd, flag);
6861 }
6862 
check_update_overutilized_status(struct rq * rq)6863 static inline void check_update_overutilized_status(struct rq *rq)
6864 {
6865 	/*
6866 	 * overutilized field is used for load balancing decisions only
6867 	 * if energy aware scheduler is being used
6868 	 */
6869 
6870 	if (!is_rd_overutilized(rq->rd) && cpu_overutilized(rq->cpu))
6871 		set_rd_overutilized(rq->rd, 1);
6872 }
6873 #else
check_update_overutilized_status(struct rq * rq)6874 static inline void check_update_overutilized_status(struct rq *rq) { }
6875 #endif
6876 
6877 /* Runqueue only has SCHED_IDLE tasks enqueued */
sched_idle_rq(struct rq * rq)6878 static int sched_idle_rq(struct rq *rq)
6879 {
6880 	return unlikely(rq->nr_running == rq->cfs.h_nr_idle &&
6881 			rq->nr_running);
6882 }
6883 
6884 #ifdef CONFIG_SMP
sched_idle_cpu(int cpu)6885 static int sched_idle_cpu(int cpu)
6886 {
6887 	return sched_idle_rq(cpu_rq(cpu));
6888 }
6889 #endif
6890 
6891 static void
requeue_delayed_entity(struct sched_entity * se)6892 requeue_delayed_entity(struct sched_entity *se)
6893 {
6894 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
6895 
6896 	/*
6897 	 * se->sched_delayed should imply: se->on_rq == 1.
6898 	 * Because a delayed entity is one that is still on
6899 	 * the runqueue competing until elegibility.
6900 	 */
6901 	WARN_ON_ONCE(!se->sched_delayed);
6902 	WARN_ON_ONCE(!se->on_rq);
6903 
6904 	if (sched_feat(DELAY_ZERO)) {
6905 		update_entity_lag(cfs_rq, se);
6906 		if (se->vlag > 0) {
6907 			cfs_rq->nr_queued--;
6908 			if (se != cfs_rq->curr)
6909 				__dequeue_entity(cfs_rq, se);
6910 			se->vlag = 0;
6911 			place_entity(cfs_rq, se, 0);
6912 			if (se != cfs_rq->curr)
6913 				__enqueue_entity(cfs_rq, se);
6914 			cfs_rq->nr_queued++;
6915 		}
6916 	}
6917 
6918 	update_load_avg(cfs_rq, se, 0);
6919 	clear_delayed(se);
6920 }
6921 
6922 /*
6923  * The enqueue_task method is called before nr_running is
6924  * increased. Here we update the fair scheduling stats and
6925  * then put the task into the rbtree:
6926  */
6927 static void
enqueue_task_fair(struct rq * rq,struct task_struct * p,int flags)6928 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6929 {
6930 	struct cfs_rq *cfs_rq;
6931 	struct sched_entity *se = &p->se;
6932 	int h_nr_idle = task_has_idle_policy(p);
6933 	int h_nr_runnable = 1;
6934 	int task_new = !(flags & ENQUEUE_WAKEUP);
6935 	int rq_h_nr_queued = rq->cfs.h_nr_queued;
6936 	u64 slice = 0;
6937 
6938 	/*
6939 	 * The code below (indirectly) updates schedutil which looks at
6940 	 * the cfs_rq utilization to select a frequency.
6941 	 * Let's add the task's estimated utilization to the cfs_rq's
6942 	 * estimated utilization, before we update schedutil.
6943 	 */
6944 	if (!(p->se.sched_delayed && (task_on_rq_migrating(p) || (flags & ENQUEUE_RESTORE))))
6945 		util_est_enqueue(&rq->cfs, p);
6946 
6947 	if (flags & ENQUEUE_DELAYED) {
6948 		requeue_delayed_entity(se);
6949 		return;
6950 	}
6951 
6952 	/*
6953 	 * If in_iowait is set, the code below may not trigger any cpufreq
6954 	 * utilization updates, so do it here explicitly with the IOWAIT flag
6955 	 * passed.
6956 	 */
6957 	if (p->in_iowait)
6958 		cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6959 
6960 	if (task_new && se->sched_delayed)
6961 		h_nr_runnable = 0;
6962 
6963 	for_each_sched_entity(se) {
6964 		if (se->on_rq) {
6965 			if (se->sched_delayed)
6966 				requeue_delayed_entity(se);
6967 			break;
6968 		}
6969 		cfs_rq = cfs_rq_of(se);
6970 
6971 		/*
6972 		 * Basically set the slice of group entries to the min_slice of
6973 		 * their respective cfs_rq. This ensures the group can service
6974 		 * its entities in the desired time-frame.
6975 		 */
6976 		if (slice) {
6977 			se->slice = slice;
6978 			se->custom_slice = 1;
6979 		}
6980 		enqueue_entity(cfs_rq, se, flags);
6981 		slice = cfs_rq_min_slice(cfs_rq);
6982 
6983 		cfs_rq->h_nr_runnable += h_nr_runnable;
6984 		cfs_rq->h_nr_queued++;
6985 		cfs_rq->h_nr_idle += h_nr_idle;
6986 
6987 		if (cfs_rq_is_idle(cfs_rq))
6988 			h_nr_idle = 1;
6989 
6990 		/* end evaluation on encountering a throttled cfs_rq */
6991 		if (cfs_rq_throttled(cfs_rq))
6992 			goto enqueue_throttle;
6993 
6994 		flags = ENQUEUE_WAKEUP;
6995 	}
6996 
6997 	for_each_sched_entity(se) {
6998 		cfs_rq = cfs_rq_of(se);
6999 
7000 		update_load_avg(cfs_rq, se, UPDATE_TG);
7001 		se_update_runnable(se);
7002 		update_cfs_group(se);
7003 
7004 		se->slice = slice;
7005 		if (se != cfs_rq->curr)
7006 			min_vruntime_cb_propagate(&se->run_node, NULL);
7007 		slice = cfs_rq_min_slice(cfs_rq);
7008 
7009 		cfs_rq->h_nr_runnable += h_nr_runnable;
7010 		cfs_rq->h_nr_queued++;
7011 		cfs_rq->h_nr_idle += h_nr_idle;
7012 
7013 		if (cfs_rq_is_idle(cfs_rq))
7014 			h_nr_idle = 1;
7015 
7016 		/* end evaluation on encountering a throttled cfs_rq */
7017 		if (cfs_rq_throttled(cfs_rq))
7018 			goto enqueue_throttle;
7019 	}
7020 
7021 	if (!rq_h_nr_queued && rq->cfs.h_nr_queued) {
7022 		/* Account for idle runtime */
7023 		if (!rq->nr_running)
7024 			dl_server_update_idle_time(rq, rq->curr);
7025 		dl_server_start(&rq->fair_server);
7026 	}
7027 
7028 	/* At this point se is NULL and we are at root level*/
7029 	add_nr_running(rq, 1);
7030 
7031 	/*
7032 	 * Since new tasks are assigned an initial util_avg equal to
7033 	 * half of the spare capacity of their CPU, tiny tasks have the
7034 	 * ability to cross the overutilized threshold, which will
7035 	 * result in the load balancer ruining all the task placement
7036 	 * done by EAS. As a way to mitigate that effect, do not account
7037 	 * for the first enqueue operation of new tasks during the
7038 	 * overutilized flag detection.
7039 	 *
7040 	 * A better way of solving this problem would be to wait for
7041 	 * the PELT signals of tasks to converge before taking them
7042 	 * into account, but that is not straightforward to implement,
7043 	 * and the following generally works well enough in practice.
7044 	 */
7045 	if (!task_new)
7046 		check_update_overutilized_status(rq);
7047 
7048 enqueue_throttle:
7049 	assert_list_leaf_cfs_rq(rq);
7050 
7051 	hrtick_update(rq);
7052 }
7053 
7054 static void set_next_buddy(struct sched_entity *se);
7055 
7056 /*
7057  * Basically dequeue_task_fair(), except it can deal with dequeue_entity()
7058  * failing half-way through and resume the dequeue later.
7059  *
7060  * Returns:
7061  * -1 - dequeue delayed
7062  *  0 - dequeue throttled
7063  *  1 - dequeue complete
7064  */
dequeue_entities(struct rq * rq,struct sched_entity * se,int flags)7065 static int dequeue_entities(struct rq *rq, struct sched_entity *se, int flags)
7066 {
7067 	bool was_sched_idle = sched_idle_rq(rq);
7068 	int rq_h_nr_queued = rq->cfs.h_nr_queued;
7069 	bool task_sleep = flags & DEQUEUE_SLEEP;
7070 	bool task_delayed = flags & DEQUEUE_DELAYED;
7071 	struct task_struct *p = NULL;
7072 	int h_nr_idle = 0;
7073 	int h_nr_queued = 0;
7074 	int h_nr_runnable = 0;
7075 	struct cfs_rq *cfs_rq;
7076 	u64 slice = 0;
7077 
7078 	if (entity_is_task(se)) {
7079 		p = task_of(se);
7080 		h_nr_queued = 1;
7081 		h_nr_idle = task_has_idle_policy(p);
7082 		if (task_sleep || task_delayed || !se->sched_delayed)
7083 			h_nr_runnable = 1;
7084 	} else {
7085 		cfs_rq = group_cfs_rq(se);
7086 		slice = cfs_rq_min_slice(cfs_rq);
7087 	}
7088 
7089 	for_each_sched_entity(se) {
7090 		cfs_rq = cfs_rq_of(se);
7091 
7092 		if (!dequeue_entity(cfs_rq, se, flags)) {
7093 			if (p && &p->se == se)
7094 				return -1;
7095 
7096 			break;
7097 		}
7098 
7099 		cfs_rq->h_nr_runnable -= h_nr_runnable;
7100 		cfs_rq->h_nr_queued -= h_nr_queued;
7101 		cfs_rq->h_nr_idle -= h_nr_idle;
7102 
7103 		if (cfs_rq_is_idle(cfs_rq))
7104 			h_nr_idle = h_nr_queued;
7105 
7106 		/* end evaluation on encountering a throttled cfs_rq */
7107 		if (cfs_rq_throttled(cfs_rq))
7108 			return 0;
7109 
7110 		/* Don't dequeue parent if it has other entities besides us */
7111 		if (cfs_rq->load.weight) {
7112 			slice = cfs_rq_min_slice(cfs_rq);
7113 
7114 			/* Avoid re-evaluating load for this entity: */
7115 			se = parent_entity(se);
7116 			/*
7117 			 * Bias pick_next to pick a task from this cfs_rq, as
7118 			 * p is sleeping when it is within its sched_slice.
7119 			 */
7120 			if (task_sleep && se && !throttled_hierarchy(cfs_rq))
7121 				set_next_buddy(se);
7122 			break;
7123 		}
7124 		flags |= DEQUEUE_SLEEP;
7125 		flags &= ~(DEQUEUE_DELAYED | DEQUEUE_SPECIAL);
7126 	}
7127 
7128 	for_each_sched_entity(se) {
7129 		cfs_rq = cfs_rq_of(se);
7130 
7131 		update_load_avg(cfs_rq, se, UPDATE_TG);
7132 		se_update_runnable(se);
7133 		update_cfs_group(se);
7134 
7135 		se->slice = slice;
7136 		if (se != cfs_rq->curr)
7137 			min_vruntime_cb_propagate(&se->run_node, NULL);
7138 		slice = cfs_rq_min_slice(cfs_rq);
7139 
7140 		cfs_rq->h_nr_runnable -= h_nr_runnable;
7141 		cfs_rq->h_nr_queued -= h_nr_queued;
7142 		cfs_rq->h_nr_idle -= h_nr_idle;
7143 
7144 		if (cfs_rq_is_idle(cfs_rq))
7145 			h_nr_idle = h_nr_queued;
7146 
7147 		/* end evaluation on encountering a throttled cfs_rq */
7148 		if (cfs_rq_throttled(cfs_rq))
7149 			return 0;
7150 	}
7151 
7152 	sub_nr_running(rq, h_nr_queued);
7153 
7154 	if (rq_h_nr_queued && !rq->cfs.h_nr_queued)
7155 		dl_server_stop(&rq->fair_server);
7156 
7157 	/* balance early to pull high priority tasks */
7158 	if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
7159 		rq->next_balance = jiffies;
7160 
7161 	if (p && task_delayed) {
7162 		WARN_ON_ONCE(!task_sleep);
7163 		WARN_ON_ONCE(p->on_rq != 1);
7164 
7165 		/* Fix-up what dequeue_task_fair() skipped */
7166 		hrtick_update(rq);
7167 
7168 		/*
7169 		 * Fix-up what block_task() skipped.
7170 		 *
7171 		 * Must be last, @p might not be valid after this.
7172 		 */
7173 		__block_task(rq, p);
7174 	}
7175 
7176 	return 1;
7177 }
7178 
7179 /*
7180  * The dequeue_task method is called before nr_running is
7181  * decreased. We remove the task from the rbtree and
7182  * update the fair scheduling stats:
7183  */
dequeue_task_fair(struct rq * rq,struct task_struct * p,int flags)7184 static bool dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
7185 {
7186 	if (!(p->se.sched_delayed && (task_on_rq_migrating(p) || (flags & DEQUEUE_SAVE))))
7187 		util_est_dequeue(&rq->cfs, p);
7188 
7189 	util_est_update(&rq->cfs, p, flags & DEQUEUE_SLEEP);
7190 	if (dequeue_entities(rq, &p->se, flags) < 0)
7191 		return false;
7192 
7193 	/*
7194 	 * Must not reference @p after dequeue_entities(DEQUEUE_DELAYED).
7195 	 */
7196 
7197 	hrtick_update(rq);
7198 	return true;
7199 }
7200 
7201 #ifdef CONFIG_SMP
7202 
7203 /* Working cpumask for: sched_balance_rq(), sched_balance_newidle(). */
7204 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7205 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
7206 static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask);
7207 
7208 #ifdef CONFIG_NO_HZ_COMMON
7209 
7210 static struct {
7211 	cpumask_var_t idle_cpus_mask;
7212 	atomic_t nr_cpus;
7213 	int has_blocked;		/* Idle CPUS has blocked load */
7214 	int needs_update;		/* Newly idle CPUs need their next_balance collated */
7215 	unsigned long next_balance;     /* in jiffy units */
7216 	unsigned long next_blocked;	/* Next update of blocked load in jiffies */
7217 } nohz ____cacheline_aligned;
7218 
7219 #endif /* CONFIG_NO_HZ_COMMON */
7220 
cpu_load(struct rq * rq)7221 static unsigned long cpu_load(struct rq *rq)
7222 {
7223 	return cfs_rq_load_avg(&rq->cfs);
7224 }
7225 
7226 /*
7227  * cpu_load_without - compute CPU load without any contributions from *p
7228  * @cpu: the CPU which load is requested
7229  * @p: the task which load should be discounted
7230  *
7231  * The load of a CPU is defined by the load of tasks currently enqueued on that
7232  * CPU as well as tasks which are currently sleeping after an execution on that
7233  * CPU.
7234  *
7235  * This method returns the load of the specified CPU by discounting the load of
7236  * the specified task, whenever the task is currently contributing to the CPU
7237  * load.
7238  */
cpu_load_without(struct rq * rq,struct task_struct * p)7239 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
7240 {
7241 	struct cfs_rq *cfs_rq;
7242 	unsigned int load;
7243 
7244 	/* Task has no contribution or is new */
7245 	if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7246 		return cpu_load(rq);
7247 
7248 	cfs_rq = &rq->cfs;
7249 	load = READ_ONCE(cfs_rq->avg.load_avg);
7250 
7251 	/* Discount task's util from CPU's util */
7252 	lsub_positive(&load, task_h_load(p));
7253 
7254 	return load;
7255 }
7256 
cpu_runnable(struct rq * rq)7257 static unsigned long cpu_runnable(struct rq *rq)
7258 {
7259 	return cfs_rq_runnable_avg(&rq->cfs);
7260 }
7261 
cpu_runnable_without(struct rq * rq,struct task_struct * p)7262 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
7263 {
7264 	struct cfs_rq *cfs_rq;
7265 	unsigned int runnable;
7266 
7267 	/* Task has no contribution or is new */
7268 	if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7269 		return cpu_runnable(rq);
7270 
7271 	cfs_rq = &rq->cfs;
7272 	runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
7273 
7274 	/* Discount task's runnable from CPU's runnable */
7275 	lsub_positive(&runnable, p->se.avg.runnable_avg);
7276 
7277 	return runnable;
7278 }
7279 
capacity_of(int cpu)7280 static unsigned long capacity_of(int cpu)
7281 {
7282 	return cpu_rq(cpu)->cpu_capacity;
7283 }
7284 
record_wakee(struct task_struct * p)7285 static void record_wakee(struct task_struct *p)
7286 {
7287 	/*
7288 	 * Only decay a single time; tasks that have less then 1 wakeup per
7289 	 * jiffy will not have built up many flips.
7290 	 */
7291 	if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
7292 		current->wakee_flips >>= 1;
7293 		current->wakee_flip_decay_ts = jiffies;
7294 	}
7295 
7296 	if (current->last_wakee != p) {
7297 		current->last_wakee = p;
7298 		current->wakee_flips++;
7299 	}
7300 }
7301 
7302 /*
7303  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
7304  *
7305  * A waker of many should wake a different task than the one last awakened
7306  * at a frequency roughly N times higher than one of its wakees.
7307  *
7308  * In order to determine whether we should let the load spread vs consolidating
7309  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
7310  * partner, and a factor of lls_size higher frequency in the other.
7311  *
7312  * With both conditions met, we can be relatively sure that the relationship is
7313  * non-monogamous, with partner count exceeding socket size.
7314  *
7315  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
7316  * whatever is irrelevant, spread criteria is apparent partner count exceeds
7317  * socket size.
7318  */
wake_wide(struct task_struct * p)7319 static int wake_wide(struct task_struct *p)
7320 {
7321 	unsigned int master = current->wakee_flips;
7322 	unsigned int slave = p->wakee_flips;
7323 	int factor = __this_cpu_read(sd_llc_size);
7324 
7325 	if (master < slave)
7326 		swap(master, slave);
7327 	if (slave < factor || master < slave * factor)
7328 		return 0;
7329 	return 1;
7330 }
7331 
7332 /*
7333  * The purpose of wake_affine() is to quickly determine on which CPU we can run
7334  * soonest. For the purpose of speed we only consider the waking and previous
7335  * CPU.
7336  *
7337  * wake_affine_idle() - only considers 'now', it check if the waking CPU is
7338  *			cache-affine and is (or	will be) idle.
7339  *
7340  * wake_affine_weight() - considers the weight to reflect the average
7341  *			  scheduling latency of the CPUs. This seems to work
7342  *			  for the overloaded case.
7343  */
7344 static int
wake_affine_idle(int this_cpu,int prev_cpu,int sync)7345 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
7346 {
7347 	/*
7348 	 * If this_cpu is idle, it implies the wakeup is from interrupt
7349 	 * context. Only allow the move if cache is shared. Otherwise an
7350 	 * interrupt intensive workload could force all tasks onto one
7351 	 * node depending on the IO topology or IRQ affinity settings.
7352 	 *
7353 	 * If the prev_cpu is idle and cache affine then avoid a migration.
7354 	 * There is no guarantee that the cache hot data from an interrupt
7355 	 * is more important than cache hot data on the prev_cpu and from
7356 	 * a cpufreq perspective, it's better to have higher utilisation
7357 	 * on one CPU.
7358 	 */
7359 	if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
7360 		return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
7361 
7362 	if (sync && cpu_rq(this_cpu)->nr_running == 1)
7363 		return this_cpu;
7364 
7365 	if (available_idle_cpu(prev_cpu))
7366 		return prev_cpu;
7367 
7368 	return nr_cpumask_bits;
7369 }
7370 
7371 static int
wake_affine_weight(struct sched_domain * sd,struct task_struct * p,int this_cpu,int prev_cpu,int sync)7372 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
7373 		   int this_cpu, int prev_cpu, int sync)
7374 {
7375 	s64 this_eff_load, prev_eff_load;
7376 	unsigned long task_load;
7377 
7378 	this_eff_load = cpu_load(cpu_rq(this_cpu));
7379 
7380 	if (sync) {
7381 		unsigned long current_load = task_h_load(current);
7382 
7383 		if (current_load > this_eff_load)
7384 			return this_cpu;
7385 
7386 		this_eff_load -= current_load;
7387 	}
7388 
7389 	task_load = task_h_load(p);
7390 
7391 	this_eff_load += task_load;
7392 	if (sched_feat(WA_BIAS))
7393 		this_eff_load *= 100;
7394 	this_eff_load *= capacity_of(prev_cpu);
7395 
7396 	prev_eff_load = cpu_load(cpu_rq(prev_cpu));
7397 	prev_eff_load -= task_load;
7398 	if (sched_feat(WA_BIAS))
7399 		prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
7400 	prev_eff_load *= capacity_of(this_cpu);
7401 
7402 	/*
7403 	 * If sync, adjust the weight of prev_eff_load such that if
7404 	 * prev_eff == this_eff that select_idle_sibling() will consider
7405 	 * stacking the wakee on top of the waker if no other CPU is
7406 	 * idle.
7407 	 */
7408 	if (sync)
7409 		prev_eff_load += 1;
7410 
7411 	return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
7412 }
7413 
wake_affine(struct sched_domain * sd,struct task_struct * p,int this_cpu,int prev_cpu,int sync)7414 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
7415 		       int this_cpu, int prev_cpu, int sync)
7416 {
7417 	int target = nr_cpumask_bits;
7418 
7419 	if (sched_feat(WA_IDLE))
7420 		target = wake_affine_idle(this_cpu, prev_cpu, sync);
7421 
7422 	if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
7423 		target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
7424 
7425 	schedstat_inc(p->stats.nr_wakeups_affine_attempts);
7426 	if (target != this_cpu)
7427 		return prev_cpu;
7428 
7429 	schedstat_inc(sd->ttwu_move_affine);
7430 	schedstat_inc(p->stats.nr_wakeups_affine);
7431 	return target;
7432 }
7433 
7434 static struct sched_group *
7435 sched_balance_find_dst_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
7436 
7437 /*
7438  * sched_balance_find_dst_group_cpu - find the idlest CPU among the CPUs in the group.
7439  */
7440 static int
sched_balance_find_dst_group_cpu(struct sched_group * group,struct task_struct * p,int this_cpu)7441 sched_balance_find_dst_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
7442 {
7443 	unsigned long load, min_load = ULONG_MAX;
7444 	unsigned int min_exit_latency = UINT_MAX;
7445 	u64 latest_idle_timestamp = 0;
7446 	int least_loaded_cpu = this_cpu;
7447 	int shallowest_idle_cpu = -1;
7448 	int i;
7449 
7450 	/* Check if we have any choice: */
7451 	if (group->group_weight == 1)
7452 		return cpumask_first(sched_group_span(group));
7453 
7454 	/* Traverse only the allowed CPUs */
7455 	for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
7456 		struct rq *rq = cpu_rq(i);
7457 
7458 		if (!sched_core_cookie_match(rq, p))
7459 			continue;
7460 
7461 		if (sched_idle_cpu(i))
7462 			return i;
7463 
7464 		if (available_idle_cpu(i)) {
7465 			struct cpuidle_state *idle = idle_get_state(rq);
7466 			if (idle && idle->exit_latency < min_exit_latency) {
7467 				/*
7468 				 * We give priority to a CPU whose idle state
7469 				 * has the smallest exit latency irrespective
7470 				 * of any idle timestamp.
7471 				 */
7472 				min_exit_latency = idle->exit_latency;
7473 				latest_idle_timestamp = rq->idle_stamp;
7474 				shallowest_idle_cpu = i;
7475 			} else if ((!idle || idle->exit_latency == min_exit_latency) &&
7476 				   rq->idle_stamp > latest_idle_timestamp) {
7477 				/*
7478 				 * If equal or no active idle state, then
7479 				 * the most recently idled CPU might have
7480 				 * a warmer cache.
7481 				 */
7482 				latest_idle_timestamp = rq->idle_stamp;
7483 				shallowest_idle_cpu = i;
7484 			}
7485 		} else if (shallowest_idle_cpu == -1) {
7486 			load = cpu_load(cpu_rq(i));
7487 			if (load < min_load) {
7488 				min_load = load;
7489 				least_loaded_cpu = i;
7490 			}
7491 		}
7492 	}
7493 
7494 	return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
7495 }
7496 
sched_balance_find_dst_cpu(struct sched_domain * sd,struct task_struct * p,int cpu,int prev_cpu,int sd_flag)7497 static inline int sched_balance_find_dst_cpu(struct sched_domain *sd, struct task_struct *p,
7498 				  int cpu, int prev_cpu, int sd_flag)
7499 {
7500 	int new_cpu = cpu;
7501 
7502 	if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
7503 		return prev_cpu;
7504 
7505 	/*
7506 	 * We need task's util for cpu_util_without, sync it up to
7507 	 * prev_cpu's last_update_time.
7508 	 */
7509 	if (!(sd_flag & SD_BALANCE_FORK))
7510 		sync_entity_load_avg(&p->se);
7511 
7512 	while (sd) {
7513 		struct sched_group *group;
7514 		struct sched_domain *tmp;
7515 		int weight;
7516 
7517 		if (!(sd->flags & sd_flag)) {
7518 			sd = sd->child;
7519 			continue;
7520 		}
7521 
7522 		group = sched_balance_find_dst_group(sd, p, cpu);
7523 		if (!group) {
7524 			sd = sd->child;
7525 			continue;
7526 		}
7527 
7528 		new_cpu = sched_balance_find_dst_group_cpu(group, p, cpu);
7529 		if (new_cpu == cpu) {
7530 			/* Now try balancing at a lower domain level of 'cpu': */
7531 			sd = sd->child;
7532 			continue;
7533 		}
7534 
7535 		/* Now try balancing at a lower domain level of 'new_cpu': */
7536 		cpu = new_cpu;
7537 		weight = sd->span_weight;
7538 		sd = NULL;
7539 		for_each_domain(cpu, tmp) {
7540 			if (weight <= tmp->span_weight)
7541 				break;
7542 			if (tmp->flags & sd_flag)
7543 				sd = tmp;
7544 		}
7545 	}
7546 
7547 	return new_cpu;
7548 }
7549 
__select_idle_cpu(int cpu,struct task_struct * p)7550 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
7551 {
7552 	if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
7553 	    sched_cpu_cookie_match(cpu_rq(cpu), p))
7554 		return cpu;
7555 
7556 	return -1;
7557 }
7558 
7559 #ifdef CONFIG_SCHED_SMT
7560 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7561 EXPORT_SYMBOL_GPL(sched_smt_present);
7562 
set_idle_cores(int cpu,int val)7563 static inline void set_idle_cores(int cpu, int val)
7564 {
7565 	struct sched_domain_shared *sds;
7566 
7567 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7568 	if (sds)
7569 		WRITE_ONCE(sds->has_idle_cores, val);
7570 }
7571 
test_idle_cores(int cpu)7572 static inline bool test_idle_cores(int cpu)
7573 {
7574 	struct sched_domain_shared *sds;
7575 
7576 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7577 	if (sds)
7578 		return READ_ONCE(sds->has_idle_cores);
7579 
7580 	return false;
7581 }
7582 
7583 /*
7584  * Scans the local SMT mask to see if the entire core is idle, and records this
7585  * information in sd_llc_shared->has_idle_cores.
7586  *
7587  * Since SMT siblings share all cache levels, inspecting this limited remote
7588  * state should be fairly cheap.
7589  */
__update_idle_core(struct rq * rq)7590 void __update_idle_core(struct rq *rq)
7591 {
7592 	int core = cpu_of(rq);
7593 	int cpu;
7594 
7595 	rcu_read_lock();
7596 	if (test_idle_cores(core))
7597 		goto unlock;
7598 
7599 	for_each_cpu(cpu, cpu_smt_mask(core)) {
7600 		if (cpu == core)
7601 			continue;
7602 
7603 		if (!available_idle_cpu(cpu))
7604 			goto unlock;
7605 	}
7606 
7607 	set_idle_cores(core, 1);
7608 unlock:
7609 	rcu_read_unlock();
7610 }
7611 
7612 /*
7613  * Scan the entire LLC domain for idle cores; this dynamically switches off if
7614  * there are no idle cores left in the system; tracked through
7615  * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7616  */
select_idle_core(struct task_struct * p,int core,struct cpumask * cpus,int * idle_cpu)7617 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7618 {
7619 	bool idle = true;
7620 	int cpu;
7621 
7622 	for_each_cpu(cpu, cpu_smt_mask(core)) {
7623 		if (!available_idle_cpu(cpu)) {
7624 			idle = false;
7625 			if (*idle_cpu == -1) {
7626 				if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, cpus)) {
7627 					*idle_cpu = cpu;
7628 					break;
7629 				}
7630 				continue;
7631 			}
7632 			break;
7633 		}
7634 		if (*idle_cpu == -1 && cpumask_test_cpu(cpu, cpus))
7635 			*idle_cpu = cpu;
7636 	}
7637 
7638 	if (idle)
7639 		return core;
7640 
7641 	cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
7642 	return -1;
7643 }
7644 
7645 /*
7646  * Scan the local SMT mask for idle CPUs.
7647  */
select_idle_smt(struct task_struct * p,struct sched_domain * sd,int target)7648 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7649 {
7650 	int cpu;
7651 
7652 	for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
7653 		if (cpu == target)
7654 			continue;
7655 		/*
7656 		 * Check if the CPU is in the LLC scheduling domain of @target.
7657 		 * Due to isolcpus, there is no guarantee that all the siblings are in the domain.
7658 		 */
7659 		if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7660 			continue;
7661 		if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
7662 			return cpu;
7663 	}
7664 
7665 	return -1;
7666 }
7667 
7668 #else /* CONFIG_SCHED_SMT */
7669 
set_idle_cores(int cpu,int val)7670 static inline void set_idle_cores(int cpu, int val)
7671 {
7672 }
7673 
test_idle_cores(int cpu)7674 static inline bool test_idle_cores(int cpu)
7675 {
7676 	return false;
7677 }
7678 
select_idle_core(struct task_struct * p,int core,struct cpumask * cpus,int * idle_cpu)7679 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7680 {
7681 	return __select_idle_cpu(core, p);
7682 }
7683 
select_idle_smt(struct task_struct * p,struct sched_domain * sd,int target)7684 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7685 {
7686 	return -1;
7687 }
7688 
7689 #endif /* CONFIG_SCHED_SMT */
7690 
7691 /*
7692  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
7693  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7694  * average idle time for this rq (as found in rq->avg_idle).
7695  */
select_idle_cpu(struct task_struct * p,struct sched_domain * sd,bool has_idle_core,int target)7696 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
7697 {
7698 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7699 	int i, cpu, idle_cpu = -1, nr = INT_MAX;
7700 	struct sched_domain_shared *sd_share;
7701 
7702 	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7703 
7704 	if (sched_feat(SIS_UTIL)) {
7705 		sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
7706 		if (sd_share) {
7707 			/* because !--nr is the condition to stop scan */
7708 			nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
7709 			/* overloaded LLC is unlikely to have idle cpu/core */
7710 			if (nr == 1)
7711 				return -1;
7712 		}
7713 	}
7714 
7715 	if (static_branch_unlikely(&sched_cluster_active)) {
7716 		struct sched_group *sg = sd->groups;
7717 
7718 		if (sg->flags & SD_CLUSTER) {
7719 			for_each_cpu_wrap(cpu, sched_group_span(sg), target + 1) {
7720 				if (!cpumask_test_cpu(cpu, cpus))
7721 					continue;
7722 
7723 				if (has_idle_core) {
7724 					i = select_idle_core(p, cpu, cpus, &idle_cpu);
7725 					if ((unsigned int)i < nr_cpumask_bits)
7726 						return i;
7727 				} else {
7728 					if (--nr <= 0)
7729 						return -1;
7730 					idle_cpu = __select_idle_cpu(cpu, p);
7731 					if ((unsigned int)idle_cpu < nr_cpumask_bits)
7732 						return idle_cpu;
7733 				}
7734 			}
7735 			cpumask_andnot(cpus, cpus, sched_group_span(sg));
7736 		}
7737 	}
7738 
7739 	for_each_cpu_wrap(cpu, cpus, target + 1) {
7740 		if (has_idle_core) {
7741 			i = select_idle_core(p, cpu, cpus, &idle_cpu);
7742 			if ((unsigned int)i < nr_cpumask_bits)
7743 				return i;
7744 
7745 		} else {
7746 			if (--nr <= 0)
7747 				return -1;
7748 			idle_cpu = __select_idle_cpu(cpu, p);
7749 			if ((unsigned int)idle_cpu < nr_cpumask_bits)
7750 				break;
7751 		}
7752 	}
7753 
7754 	if (has_idle_core)
7755 		set_idle_cores(target, false);
7756 
7757 	return idle_cpu;
7758 }
7759 
7760 /*
7761  * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
7762  * the task fits. If no CPU is big enough, but there are idle ones, try to
7763  * maximize capacity.
7764  */
7765 static int
select_idle_capacity(struct task_struct * p,struct sched_domain * sd,int target)7766 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
7767 {
7768 	unsigned long task_util, util_min, util_max, best_cap = 0;
7769 	int fits, best_fits = 0;
7770 	int cpu, best_cpu = -1;
7771 	struct cpumask *cpus;
7772 
7773 	cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7774 	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7775 
7776 	task_util = task_util_est(p);
7777 	util_min = uclamp_eff_value(p, UCLAMP_MIN);
7778 	util_max = uclamp_eff_value(p, UCLAMP_MAX);
7779 
7780 	for_each_cpu_wrap(cpu, cpus, target) {
7781 		unsigned long cpu_cap = capacity_of(cpu);
7782 
7783 		if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
7784 			continue;
7785 
7786 		fits = util_fits_cpu(task_util, util_min, util_max, cpu);
7787 
7788 		/* This CPU fits with all requirements */
7789 		if (fits > 0)
7790 			return cpu;
7791 		/*
7792 		 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
7793 		 * Look for the CPU with best capacity.
7794 		 */
7795 		else if (fits < 0)
7796 			cpu_cap = get_actual_cpu_capacity(cpu);
7797 
7798 		/*
7799 		 * First, select CPU which fits better (-1 being better than 0).
7800 		 * Then, select the one with best capacity at same level.
7801 		 */
7802 		if ((fits < best_fits) ||
7803 		    ((fits == best_fits) && (cpu_cap > best_cap))) {
7804 			best_cap = cpu_cap;
7805 			best_cpu = cpu;
7806 			best_fits = fits;
7807 		}
7808 	}
7809 
7810 	return best_cpu;
7811 }
7812 
asym_fits_cpu(unsigned long util,unsigned long util_min,unsigned long util_max,int cpu)7813 static inline bool asym_fits_cpu(unsigned long util,
7814 				 unsigned long util_min,
7815 				 unsigned long util_max,
7816 				 int cpu)
7817 {
7818 	if (sched_asym_cpucap_active())
7819 		/*
7820 		 * Return true only if the cpu fully fits the task requirements
7821 		 * which include the utilization and the performance hints.
7822 		 */
7823 		return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7824 
7825 	return true;
7826 }
7827 
7828 /*
7829  * Try and locate an idle core/thread in the LLC cache domain.
7830  */
select_idle_sibling(struct task_struct * p,int prev,int target)7831 static int select_idle_sibling(struct task_struct *p, int prev, int target)
7832 {
7833 	bool has_idle_core = false;
7834 	struct sched_domain *sd;
7835 	unsigned long task_util, util_min, util_max;
7836 	int i, recent_used_cpu, prev_aff = -1;
7837 
7838 	/*
7839 	 * On asymmetric system, update task utilization because we will check
7840 	 * that the task fits with CPU's capacity.
7841 	 */
7842 	if (sched_asym_cpucap_active()) {
7843 		sync_entity_load_avg(&p->se);
7844 		task_util = task_util_est(p);
7845 		util_min = uclamp_eff_value(p, UCLAMP_MIN);
7846 		util_max = uclamp_eff_value(p, UCLAMP_MAX);
7847 	}
7848 
7849 	/*
7850 	 * per-cpu select_rq_mask usage
7851 	 */
7852 	lockdep_assert_irqs_disabled();
7853 
7854 	if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7855 	    asym_fits_cpu(task_util, util_min, util_max, target))
7856 		return target;
7857 
7858 	/*
7859 	 * If the previous CPU is cache affine and idle, don't be stupid:
7860 	 */
7861 	if (prev != target && cpus_share_cache(prev, target) &&
7862 	    (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7863 	    asym_fits_cpu(task_util, util_min, util_max, prev)) {
7864 
7865 		if (!static_branch_unlikely(&sched_cluster_active) ||
7866 		    cpus_share_resources(prev, target))
7867 			return prev;
7868 
7869 		prev_aff = prev;
7870 	}
7871 
7872 	/*
7873 	 * Allow a per-cpu kthread to stack with the wakee if the
7874 	 * kworker thread and the tasks previous CPUs are the same.
7875 	 * The assumption is that the wakee queued work for the
7876 	 * per-cpu kthread that is now complete and the wakeup is
7877 	 * essentially a sync wakeup. An obvious example of this
7878 	 * pattern is IO completions.
7879 	 */
7880 	if (is_per_cpu_kthread(current) &&
7881 	    in_task() &&
7882 	    prev == smp_processor_id() &&
7883 	    this_rq()->nr_running <= 1 &&
7884 	    asym_fits_cpu(task_util, util_min, util_max, prev)) {
7885 		return prev;
7886 	}
7887 
7888 	/* Check a recently used CPU as a potential idle candidate: */
7889 	recent_used_cpu = p->recent_used_cpu;
7890 	p->recent_used_cpu = prev;
7891 	if (recent_used_cpu != prev &&
7892 	    recent_used_cpu != target &&
7893 	    cpus_share_cache(recent_used_cpu, target) &&
7894 	    (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7895 	    cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&
7896 	    asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7897 
7898 		if (!static_branch_unlikely(&sched_cluster_active) ||
7899 		    cpus_share_resources(recent_used_cpu, target))
7900 			return recent_used_cpu;
7901 
7902 	} else {
7903 		recent_used_cpu = -1;
7904 	}
7905 
7906 	/*
7907 	 * For asymmetric CPU capacity systems, our domain of interest is
7908 	 * sd_asym_cpucapacity rather than sd_llc.
7909 	 */
7910 	if (sched_asym_cpucap_active()) {
7911 		sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7912 		/*
7913 		 * On an asymmetric CPU capacity system where an exclusive
7914 		 * cpuset defines a symmetric island (i.e. one unique
7915 		 * capacity_orig value through the cpuset), the key will be set
7916 		 * but the CPUs within that cpuset will not have a domain with
7917 		 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7918 		 * capacity path.
7919 		 */
7920 		if (sd) {
7921 			i = select_idle_capacity(p, sd, target);
7922 			return ((unsigned)i < nr_cpumask_bits) ? i : target;
7923 		}
7924 	}
7925 
7926 	sd = rcu_dereference(per_cpu(sd_llc, target));
7927 	if (!sd)
7928 		return target;
7929 
7930 	if (sched_smt_active()) {
7931 		has_idle_core = test_idle_cores(target);
7932 
7933 		if (!has_idle_core && cpus_share_cache(prev, target)) {
7934 			i = select_idle_smt(p, sd, prev);
7935 			if ((unsigned int)i < nr_cpumask_bits)
7936 				return i;
7937 		}
7938 	}
7939 
7940 	i = select_idle_cpu(p, sd, has_idle_core, target);
7941 	if ((unsigned)i < nr_cpumask_bits)
7942 		return i;
7943 
7944 	/*
7945 	 * For cluster machines which have lower sharing cache like L2 or
7946 	 * LLC Tag, we tend to find an idle CPU in the target's cluster
7947 	 * first. But prev_cpu or recent_used_cpu may also be a good candidate,
7948 	 * use them if possible when no idle CPU found in select_idle_cpu().
7949 	 */
7950 	if ((unsigned int)prev_aff < nr_cpumask_bits)
7951 		return prev_aff;
7952 	if ((unsigned int)recent_used_cpu < nr_cpumask_bits)
7953 		return recent_used_cpu;
7954 
7955 	return target;
7956 }
7957 
7958 /**
7959  * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7960  * @cpu: the CPU to get the utilization for
7961  * @p: task for which the CPU utilization should be predicted or NULL
7962  * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7963  * @boost: 1 to enable boosting, otherwise 0
7964  *
7965  * The unit of the return value must be the same as the one of CPU capacity
7966  * so that CPU utilization can be compared with CPU capacity.
7967  *
7968  * CPU utilization is the sum of running time of runnable tasks plus the
7969  * recent utilization of currently non-runnable tasks on that CPU.
7970  * It represents the amount of CPU capacity currently used by CFS tasks in
7971  * the range [0..max CPU capacity] with max CPU capacity being the CPU
7972  * capacity at f_max.
7973  *
7974  * The estimated CPU utilization is defined as the maximum between CPU
7975  * utilization and sum of the estimated utilization of the currently
7976  * runnable tasks on that CPU. It preserves a utilization "snapshot" of
7977  * previously-executed tasks, which helps better deduce how busy a CPU will
7978  * be when a long-sleeping task wakes up. The contribution to CPU utilization
7979  * of such a task would be significantly decayed at this point of time.
7980  *
7981  * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization).
7982  * CPU contention for CFS tasks can be detected by CPU runnable > CPU
7983  * utilization. Boosting is implemented in cpu_util() so that internal
7984  * users (e.g. EAS) can use it next to external users (e.g. schedutil),
7985  * latter via cpu_util_cfs_boost().
7986  *
7987  * CPU utilization can be higher than the current CPU capacity
7988  * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because
7989  * of rounding errors as well as task migrations or wakeups of new tasks.
7990  * CPU utilization has to be capped to fit into the [0..max CPU capacity]
7991  * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%)
7992  * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7993  * capacity. CPU utilization is allowed to overshoot current CPU capacity
7994  * though since this is useful for predicting the CPU capacity required
7995  * after task migrations (scheduler-driven DVFS).
7996  *
7997  * Return: (Boosted) (estimated) utilization for the specified CPU.
7998  */
7999 static unsigned long
cpu_util(int cpu,struct task_struct * p,int dst_cpu,int boost)8000 cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
8001 {
8002 	struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
8003 	unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
8004 	unsigned long runnable;
8005 
8006 	if (boost) {
8007 		runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
8008 		util = max(util, runnable);
8009 	}
8010 
8011 	/*
8012 	 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
8013 	 * contribution. If @p migrates from another CPU to @cpu add its
8014 	 * contribution. In all the other cases @cpu is not impacted by the
8015 	 * migration so its util_avg is already correct.
8016 	 */
8017 	if (p && task_cpu(p) == cpu && dst_cpu != cpu)
8018 		lsub_positive(&util, task_util(p));
8019 	else if (p && task_cpu(p) != cpu && dst_cpu == cpu)
8020 		util += task_util(p);
8021 
8022 	if (sched_feat(UTIL_EST)) {
8023 		unsigned long util_est;
8024 
8025 		util_est = READ_ONCE(cfs_rq->avg.util_est);
8026 
8027 		/*
8028 		 * During wake-up @p isn't enqueued yet and doesn't contribute
8029 		 * to any cpu_rq(cpu)->cfs.avg.util_est.
8030 		 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
8031 		 * has been enqueued.
8032 		 *
8033 		 * During exec (@dst_cpu = -1) @p is enqueued and does
8034 		 * contribute to cpu_rq(cpu)->cfs.util_est.
8035 		 * Remove it to "simulate" cpu_util without @p's contribution.
8036 		 *
8037 		 * Despite the task_on_rq_queued(@p) check there is still a
8038 		 * small window for a possible race when an exec
8039 		 * select_task_rq_fair() races with LB's detach_task().
8040 		 *
8041 		 *   detach_task()
8042 		 *     deactivate_task()
8043 		 *       p->on_rq = TASK_ON_RQ_MIGRATING;
8044 		 *       -------------------------------- A
8045 		 *       dequeue_task()                    \
8046 		 *         dequeue_task_fair()              + Race Time
8047 		 *           util_est_dequeue()            /
8048 		 *       -------------------------------- B
8049 		 *
8050 		 * The additional check "current == p" is required to further
8051 		 * reduce the race window.
8052 		 */
8053 		if (dst_cpu == cpu)
8054 			util_est += _task_util_est(p);
8055 		else if (p && unlikely(task_on_rq_queued(p) || current == p))
8056 			lsub_positive(&util_est, _task_util_est(p));
8057 
8058 		util = max(util, util_est);
8059 	}
8060 
8061 	return min(util, arch_scale_cpu_capacity(cpu));
8062 }
8063 
cpu_util_cfs(int cpu)8064 unsigned long cpu_util_cfs(int cpu)
8065 {
8066 	return cpu_util(cpu, NULL, -1, 0);
8067 }
8068 
cpu_util_cfs_boost(int cpu)8069 unsigned long cpu_util_cfs_boost(int cpu)
8070 {
8071 	return cpu_util(cpu, NULL, -1, 1);
8072 }
8073 
8074 /*
8075  * cpu_util_without: compute cpu utilization without any contributions from *p
8076  * @cpu: the CPU which utilization is requested
8077  * @p: the task which utilization should be discounted
8078  *
8079  * The utilization of a CPU is defined by the utilization of tasks currently
8080  * enqueued on that CPU as well as tasks which are currently sleeping after an
8081  * execution on that CPU.
8082  *
8083  * This method returns the utilization of the specified CPU by discounting the
8084  * utilization of the specified task, whenever the task is currently
8085  * contributing to the CPU utilization.
8086  */
cpu_util_without(int cpu,struct task_struct * p)8087 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
8088 {
8089 	/* Task has no contribution or is new */
8090 	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8091 		p = NULL;
8092 
8093 	return cpu_util(cpu, p, -1, 0);
8094 }
8095 
8096 /*
8097  * This function computes an effective utilization for the given CPU, to be
8098  * used for frequency selection given the linear relation: f = u * f_max.
8099  *
8100  * The scheduler tracks the following metrics:
8101  *
8102  *   cpu_util_{cfs,rt,dl,irq}()
8103  *   cpu_bw_dl()
8104  *
8105  * Where the cfs,rt and dl util numbers are tracked with the same metric and
8106  * synchronized windows and are thus directly comparable.
8107  *
8108  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
8109  * which excludes things like IRQ and steal-time. These latter are then accrued
8110  * in the IRQ utilization.
8111  *
8112  * The DL bandwidth number OTOH is not a measured metric but a value computed
8113  * based on the task model parameters and gives the minimal utilization
8114  * required to meet deadlines.
8115  */
effective_cpu_util(int cpu,unsigned long util_cfs,unsigned long * min,unsigned long * max)8116 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
8117 				 unsigned long *min,
8118 				 unsigned long *max)
8119 {
8120 	unsigned long util, irq, scale;
8121 	struct rq *rq = cpu_rq(cpu);
8122 
8123 	scale = arch_scale_cpu_capacity(cpu);
8124 
8125 	/*
8126 	 * Early check to see if IRQ/steal time saturates the CPU, can be
8127 	 * because of inaccuracies in how we track these -- see
8128 	 * update_irq_load_avg().
8129 	 */
8130 	irq = cpu_util_irq(rq);
8131 	if (unlikely(irq >= scale)) {
8132 		if (min)
8133 			*min = scale;
8134 		if (max)
8135 			*max = scale;
8136 		return scale;
8137 	}
8138 
8139 	if (min) {
8140 		/*
8141 		 * The minimum utilization returns the highest level between:
8142 		 * - the computed DL bandwidth needed with the IRQ pressure which
8143 		 *   steals time to the deadline task.
8144 		 * - The minimum performance requirement for CFS and/or RT.
8145 		 */
8146 		*min = max(irq + cpu_bw_dl(rq), uclamp_rq_get(rq, UCLAMP_MIN));
8147 
8148 		/*
8149 		 * When an RT task is runnable and uclamp is not used, we must
8150 		 * ensure that the task will run at maximum compute capacity.
8151 		 */
8152 		if (!uclamp_is_used() && rt_rq_is_runnable(&rq->rt))
8153 			*min = max(*min, scale);
8154 	}
8155 
8156 	/*
8157 	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
8158 	 * CFS tasks and we use the same metric to track the effective
8159 	 * utilization (PELT windows are synchronized) we can directly add them
8160 	 * to obtain the CPU's actual utilization.
8161 	 */
8162 	util = util_cfs + cpu_util_rt(rq);
8163 	util += cpu_util_dl(rq);
8164 
8165 	/*
8166 	 * The maximum hint is a soft bandwidth requirement, which can be lower
8167 	 * than the actual utilization because of uclamp_max requirements.
8168 	 */
8169 	if (max)
8170 		*max = min(scale, uclamp_rq_get(rq, UCLAMP_MAX));
8171 
8172 	if (util >= scale)
8173 		return scale;
8174 
8175 	/*
8176 	 * There is still idle time; further improve the number by using the
8177 	 * IRQ metric. Because IRQ/steal time is hidden from the task clock we
8178 	 * need to scale the task numbers:
8179 	 *
8180 	 *              max - irq
8181 	 *   U' = irq + --------- * U
8182 	 *                 max
8183 	 */
8184 	util = scale_irq_capacity(util, irq, scale);
8185 	util += irq;
8186 
8187 	return min(scale, util);
8188 }
8189 
sched_cpu_util(int cpu)8190 unsigned long sched_cpu_util(int cpu)
8191 {
8192 	return effective_cpu_util(cpu, cpu_util_cfs(cpu), NULL, NULL);
8193 }
8194 
8195 /*
8196  * energy_env - Utilization landscape for energy estimation.
8197  * @task_busy_time: Utilization contribution by the task for which we test the
8198  *                  placement. Given by eenv_task_busy_time().
8199  * @pd_busy_time:   Utilization of the whole perf domain without the task
8200  *                  contribution. Given by eenv_pd_busy_time().
8201  * @cpu_cap:        Maximum CPU capacity for the perf domain.
8202  * @pd_cap:         Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
8203  */
8204 struct energy_env {
8205 	unsigned long task_busy_time;
8206 	unsigned long pd_busy_time;
8207 	unsigned long cpu_cap;
8208 	unsigned long pd_cap;
8209 };
8210 
8211 /*
8212  * Compute the task busy time for compute_energy(). This time cannot be
8213  * injected directly into effective_cpu_util() because of the IRQ scaling.
8214  * The latter only makes sense with the most recent CPUs where the task has
8215  * run.
8216  */
eenv_task_busy_time(struct energy_env * eenv,struct task_struct * p,int prev_cpu)8217 static inline void eenv_task_busy_time(struct energy_env *eenv,
8218 				       struct task_struct *p, int prev_cpu)
8219 {
8220 	unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
8221 	unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
8222 
8223 	if (unlikely(irq >= max_cap))
8224 		busy_time = max_cap;
8225 	else
8226 		busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
8227 
8228 	eenv->task_busy_time = busy_time;
8229 }
8230 
8231 /*
8232  * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
8233  * utilization for each @pd_cpus, it however doesn't take into account
8234  * clamping since the ratio (utilization / cpu_capacity) is already enough to
8235  * scale the EM reported power consumption at the (eventually clamped)
8236  * cpu_capacity.
8237  *
8238  * The contribution of the task @p for which we want to estimate the
8239  * energy cost is removed (by cpu_util()) and must be calculated
8240  * separately (see eenv_task_busy_time). This ensures:
8241  *
8242  *   - A stable PD utilization, no matter which CPU of that PD we want to place
8243  *     the task on.
8244  *
8245  *   - A fair comparison between CPUs as the task contribution (task_util())
8246  *     will always be the same no matter which CPU utilization we rely on
8247  *     (util_avg or util_est).
8248  *
8249  * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
8250  * exceed @eenv->pd_cap.
8251  */
eenv_pd_busy_time(struct energy_env * eenv,struct cpumask * pd_cpus,struct task_struct * p)8252 static inline void eenv_pd_busy_time(struct energy_env *eenv,
8253 				     struct cpumask *pd_cpus,
8254 				     struct task_struct *p)
8255 {
8256 	unsigned long busy_time = 0;
8257 	int cpu;
8258 
8259 	for_each_cpu(cpu, pd_cpus) {
8260 		unsigned long util = cpu_util(cpu, p, -1, 0);
8261 
8262 		busy_time += effective_cpu_util(cpu, util, NULL, NULL);
8263 	}
8264 
8265 	eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
8266 }
8267 
8268 /*
8269  * Compute the maximum utilization for compute_energy() when the task @p
8270  * is placed on the cpu @dst_cpu.
8271  *
8272  * Returns the maximum utilization among @eenv->cpus. This utilization can't
8273  * exceed @eenv->cpu_cap.
8274  */
8275 static inline unsigned long
eenv_pd_max_util(struct energy_env * eenv,struct cpumask * pd_cpus,struct task_struct * p,int dst_cpu)8276 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
8277 		 struct task_struct *p, int dst_cpu)
8278 {
8279 	unsigned long max_util = 0;
8280 	int cpu;
8281 
8282 	for_each_cpu(cpu, pd_cpus) {
8283 		struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
8284 		unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
8285 		unsigned long eff_util, min, max;
8286 
8287 		/*
8288 		 * Performance domain frequency: utilization clamping
8289 		 * must be considered since it affects the selection
8290 		 * of the performance domain frequency.
8291 		 * NOTE: in case RT tasks are running, by default the min
8292 		 * utilization can be max OPP.
8293 		 */
8294 		eff_util = effective_cpu_util(cpu, util, &min, &max);
8295 
8296 		/* Task's uclamp can modify min and max value */
8297 		if (tsk && uclamp_is_used()) {
8298 			min = max(min, uclamp_eff_value(p, UCLAMP_MIN));
8299 
8300 			/*
8301 			 * If there is no active max uclamp constraint,
8302 			 * directly use task's one, otherwise keep max.
8303 			 */
8304 			if (uclamp_rq_is_idle(cpu_rq(cpu)))
8305 				max = uclamp_eff_value(p, UCLAMP_MAX);
8306 			else
8307 				max = max(max, uclamp_eff_value(p, UCLAMP_MAX));
8308 		}
8309 
8310 		eff_util = sugov_effective_cpu_perf(cpu, eff_util, min, max);
8311 		max_util = max(max_util, eff_util);
8312 	}
8313 
8314 	return min(max_util, eenv->cpu_cap);
8315 }
8316 
8317 /*
8318  * compute_energy(): Use the Energy Model to estimate the energy that @pd would
8319  * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
8320  * contribution is ignored.
8321  */
8322 static inline unsigned long
compute_energy(struct energy_env * eenv,struct perf_domain * pd,struct cpumask * pd_cpus,struct task_struct * p,int dst_cpu)8323 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
8324 	       struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
8325 {
8326 	unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
8327 	unsigned long busy_time = eenv->pd_busy_time;
8328 	unsigned long energy;
8329 
8330 	if (dst_cpu >= 0)
8331 		busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
8332 
8333 	energy = em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
8334 
8335 	trace_sched_compute_energy_tp(p, dst_cpu, energy, max_util, busy_time);
8336 
8337 	return energy;
8338 }
8339 
8340 /*
8341  * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
8342  * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
8343  * spare capacity in each performance domain and uses it as a potential
8344  * candidate to execute the task. Then, it uses the Energy Model to figure
8345  * out which of the CPU candidates is the most energy-efficient.
8346  *
8347  * The rationale for this heuristic is as follows. In a performance domain,
8348  * all the most energy efficient CPU candidates (according to the Energy
8349  * Model) are those for which we'll request a low frequency. When there are
8350  * several CPUs for which the frequency request will be the same, we don't
8351  * have enough data to break the tie between them, because the Energy Model
8352  * only includes active power costs. With this model, if we assume that
8353  * frequency requests follow utilization (e.g. using schedutil), the CPU with
8354  * the maximum spare capacity in a performance domain is guaranteed to be among
8355  * the best candidates of the performance domain.
8356  *
8357  * In practice, it could be preferable from an energy standpoint to pack
8358  * small tasks on a CPU in order to let other CPUs go in deeper idle states,
8359  * but that could also hurt our chances to go cluster idle, and we have no
8360  * ways to tell with the current Energy Model if this is actually a good
8361  * idea or not. So, find_energy_efficient_cpu() basically favors
8362  * cluster-packing, and spreading inside a cluster. That should at least be
8363  * a good thing for latency, and this is consistent with the idea that most
8364  * of the energy savings of EAS come from the asymmetry of the system, and
8365  * not so much from breaking the tie between identical CPUs. That's also the
8366  * reason why EAS is enabled in the topology code only for systems where
8367  * SD_ASYM_CPUCAPACITY is set.
8368  *
8369  * NOTE: Forkees are not accepted in the energy-aware wake-up path because
8370  * they don't have any useful utilization data yet and it's not possible to
8371  * forecast their impact on energy consumption. Consequently, they will be
8372  * placed by sched_balance_find_dst_cpu() on the least loaded CPU, which might turn out
8373  * to be energy-inefficient in some use-cases. The alternative would be to
8374  * bias new tasks towards specific types of CPUs first, or to try to infer
8375  * their util_avg from the parent task, but those heuristics could hurt
8376  * other use-cases too. So, until someone finds a better way to solve this,
8377  * let's keep things simple by re-using the existing slow path.
8378  */
find_energy_efficient_cpu(struct task_struct * p,int prev_cpu)8379 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
8380 {
8381 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
8382 	unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
8383 	unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
8384 	unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
8385 	struct root_domain *rd = this_rq()->rd;
8386 	int cpu, best_energy_cpu, target = -1;
8387 	int prev_fits = -1, best_fits = -1;
8388 	unsigned long best_actual_cap = 0;
8389 	unsigned long prev_actual_cap = 0;
8390 	struct sched_domain *sd;
8391 	struct perf_domain *pd;
8392 	struct energy_env eenv;
8393 
8394 	rcu_read_lock();
8395 	pd = rcu_dereference(rd->pd);
8396 	if (!pd)
8397 		goto unlock;
8398 
8399 	/*
8400 	 * Energy-aware wake-up happens on the lowest sched_domain starting
8401 	 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
8402 	 */
8403 	sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
8404 	while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
8405 		sd = sd->parent;
8406 	if (!sd)
8407 		goto unlock;
8408 
8409 	target = prev_cpu;
8410 
8411 	sync_entity_load_avg(&p->se);
8412 	if (!task_util_est(p) && p_util_min == 0)
8413 		goto unlock;
8414 
8415 	eenv_task_busy_time(&eenv, p, prev_cpu);
8416 
8417 	for (; pd; pd = pd->next) {
8418 		unsigned long util_min = p_util_min, util_max = p_util_max;
8419 		unsigned long cpu_cap, cpu_actual_cap, util;
8420 		long prev_spare_cap = -1, max_spare_cap = -1;
8421 		unsigned long rq_util_min, rq_util_max;
8422 		unsigned long cur_delta, base_energy;
8423 		int max_spare_cap_cpu = -1;
8424 		int fits, max_fits = -1;
8425 
8426 		cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
8427 
8428 		if (cpumask_empty(cpus))
8429 			continue;
8430 
8431 		/* Account external pressure for the energy estimation */
8432 		cpu = cpumask_first(cpus);
8433 		cpu_actual_cap = get_actual_cpu_capacity(cpu);
8434 
8435 		eenv.cpu_cap = cpu_actual_cap;
8436 		eenv.pd_cap = 0;
8437 
8438 		for_each_cpu(cpu, cpus) {
8439 			struct rq *rq = cpu_rq(cpu);
8440 
8441 			eenv.pd_cap += cpu_actual_cap;
8442 
8443 			if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
8444 				continue;
8445 
8446 			if (!cpumask_test_cpu(cpu, p->cpus_ptr))
8447 				continue;
8448 
8449 			util = cpu_util(cpu, p, cpu, 0);
8450 			cpu_cap = capacity_of(cpu);
8451 
8452 			/*
8453 			 * Skip CPUs that cannot satisfy the capacity request.
8454 			 * IOW, placing the task there would make the CPU
8455 			 * overutilized. Take uclamp into account to see how
8456 			 * much capacity we can get out of the CPU; this is
8457 			 * aligned with sched_cpu_util().
8458 			 */
8459 			if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
8460 				/*
8461 				 * Open code uclamp_rq_util_with() except for
8462 				 * the clamp() part. I.e.: apply max aggregation
8463 				 * only. util_fits_cpu() logic requires to
8464 				 * operate on non clamped util but must use the
8465 				 * max-aggregated uclamp_{min, max}.
8466 				 */
8467 				rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
8468 				rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
8469 
8470 				util_min = max(rq_util_min, p_util_min);
8471 				util_max = max(rq_util_max, p_util_max);
8472 			}
8473 
8474 			fits = util_fits_cpu(util, util_min, util_max, cpu);
8475 			if (!fits)
8476 				continue;
8477 
8478 			lsub_positive(&cpu_cap, util);
8479 
8480 			if (cpu == prev_cpu) {
8481 				/* Always use prev_cpu as a candidate. */
8482 				prev_spare_cap = cpu_cap;
8483 				prev_fits = fits;
8484 			} else if ((fits > max_fits) ||
8485 				   ((fits == max_fits) && ((long)cpu_cap > max_spare_cap))) {
8486 				/*
8487 				 * Find the CPU with the maximum spare capacity
8488 				 * among the remaining CPUs in the performance
8489 				 * domain.
8490 				 */
8491 				max_spare_cap = cpu_cap;
8492 				max_spare_cap_cpu = cpu;
8493 				max_fits = fits;
8494 			}
8495 		}
8496 
8497 		if (max_spare_cap_cpu < 0 && prev_spare_cap < 0)
8498 			continue;
8499 
8500 		eenv_pd_busy_time(&eenv, cpus, p);
8501 		/* Compute the 'base' energy of the pd, without @p */
8502 		base_energy = compute_energy(&eenv, pd, cpus, p, -1);
8503 
8504 		/* Evaluate the energy impact of using prev_cpu. */
8505 		if (prev_spare_cap > -1) {
8506 			prev_delta = compute_energy(&eenv, pd, cpus, p,
8507 						    prev_cpu);
8508 			/* CPU utilization has changed */
8509 			if (prev_delta < base_energy)
8510 				goto unlock;
8511 			prev_delta -= base_energy;
8512 			prev_actual_cap = cpu_actual_cap;
8513 			best_delta = min(best_delta, prev_delta);
8514 		}
8515 
8516 		/* Evaluate the energy impact of using max_spare_cap_cpu. */
8517 		if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
8518 			/* Current best energy cpu fits better */
8519 			if (max_fits < best_fits)
8520 				continue;
8521 
8522 			/*
8523 			 * Both don't fit performance hint (i.e. uclamp_min)
8524 			 * but best energy cpu has better capacity.
8525 			 */
8526 			if ((max_fits < 0) &&
8527 			    (cpu_actual_cap <= best_actual_cap))
8528 				continue;
8529 
8530 			cur_delta = compute_energy(&eenv, pd, cpus, p,
8531 						   max_spare_cap_cpu);
8532 			/* CPU utilization has changed */
8533 			if (cur_delta < base_energy)
8534 				goto unlock;
8535 			cur_delta -= base_energy;
8536 
8537 			/*
8538 			 * Both fit for the task but best energy cpu has lower
8539 			 * energy impact.
8540 			 */
8541 			if ((max_fits > 0) && (best_fits > 0) &&
8542 			    (cur_delta >= best_delta))
8543 				continue;
8544 
8545 			best_delta = cur_delta;
8546 			best_energy_cpu = max_spare_cap_cpu;
8547 			best_fits = max_fits;
8548 			best_actual_cap = cpu_actual_cap;
8549 		}
8550 	}
8551 	rcu_read_unlock();
8552 
8553 	if ((best_fits > prev_fits) ||
8554 	    ((best_fits > 0) && (best_delta < prev_delta)) ||
8555 	    ((best_fits < 0) && (best_actual_cap > prev_actual_cap)))
8556 		target = best_energy_cpu;
8557 
8558 	return target;
8559 
8560 unlock:
8561 	rcu_read_unlock();
8562 
8563 	return target;
8564 }
8565 
8566 /*
8567  * select_task_rq_fair: Select target runqueue for the waking task in domains
8568  * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
8569  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
8570  *
8571  * Balances load by selecting the idlest CPU in the idlest group, or under
8572  * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
8573  *
8574  * Returns the target CPU number.
8575  */
8576 static int
select_task_rq_fair(struct task_struct * p,int prev_cpu,int wake_flags)8577 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
8578 {
8579 	int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
8580 	struct sched_domain *tmp, *sd = NULL;
8581 	int cpu = smp_processor_id();
8582 	int new_cpu = prev_cpu;
8583 	int want_affine = 0;
8584 	/* SD_flags and WF_flags share the first nibble */
8585 	int sd_flag = wake_flags & 0xF;
8586 
8587 	/*
8588 	 * required for stable ->cpus_allowed
8589 	 */
8590 	lockdep_assert_held(&p->pi_lock);
8591 	if (wake_flags & WF_TTWU) {
8592 		record_wakee(p);
8593 
8594 		if ((wake_flags & WF_CURRENT_CPU) &&
8595 		    cpumask_test_cpu(cpu, p->cpus_ptr))
8596 			return cpu;
8597 
8598 		if (!is_rd_overutilized(this_rq()->rd)) {
8599 			new_cpu = find_energy_efficient_cpu(p, prev_cpu);
8600 			if (new_cpu >= 0)
8601 				return new_cpu;
8602 			new_cpu = prev_cpu;
8603 		}
8604 
8605 		want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
8606 	}
8607 
8608 	rcu_read_lock();
8609 	for_each_domain(cpu, tmp) {
8610 		/*
8611 		 * If both 'cpu' and 'prev_cpu' are part of this domain,
8612 		 * cpu is a valid SD_WAKE_AFFINE target.
8613 		 */
8614 		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
8615 		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
8616 			if (cpu != prev_cpu)
8617 				new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
8618 
8619 			sd = NULL; /* Prefer wake_affine over balance flags */
8620 			break;
8621 		}
8622 
8623 		/*
8624 		 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
8625 		 * usually do not have SD_BALANCE_WAKE set. That means wakeup
8626 		 * will usually go to the fast path.
8627 		 */
8628 		if (tmp->flags & sd_flag)
8629 			sd = tmp;
8630 		else if (!want_affine)
8631 			break;
8632 	}
8633 
8634 	if (unlikely(sd)) {
8635 		/* Slow path */
8636 		new_cpu = sched_balance_find_dst_cpu(sd, p, cpu, prev_cpu, sd_flag);
8637 	} else if (wake_flags & WF_TTWU) { /* XXX always ? */
8638 		/* Fast path */
8639 		new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
8640 	}
8641 	rcu_read_unlock();
8642 
8643 	return new_cpu;
8644 }
8645 
8646 /*
8647  * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
8648  * cfs_rq_of(p) references at time of call are still valid and identify the
8649  * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
8650  */
migrate_task_rq_fair(struct task_struct * p,int new_cpu)8651 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
8652 {
8653 	struct sched_entity *se = &p->se;
8654 
8655 	if (!task_on_rq_migrating(p)) {
8656 		remove_entity_load_avg(se);
8657 
8658 		/*
8659 		 * Here, the task's PELT values have been updated according to
8660 		 * the current rq's clock. But if that clock hasn't been
8661 		 * updated in a while, a substantial idle time will be missed,
8662 		 * leading to an inflation after wake-up on the new rq.
8663 		 *
8664 		 * Estimate the missing time from the cfs_rq last_update_time
8665 		 * and update sched_avg to improve the PELT continuity after
8666 		 * migration.
8667 		 */
8668 		migrate_se_pelt_lag(se);
8669 	}
8670 
8671 	/* Tell new CPU we are migrated */
8672 	se->avg.last_update_time = 0;
8673 
8674 	update_scan_period(p, new_cpu);
8675 }
8676 
task_dead_fair(struct task_struct * p)8677 static void task_dead_fair(struct task_struct *p)
8678 {
8679 	struct sched_entity *se = &p->se;
8680 
8681 	if (se->sched_delayed) {
8682 		struct rq_flags rf;
8683 		struct rq *rq;
8684 
8685 		rq = task_rq_lock(p, &rf);
8686 		if (se->sched_delayed) {
8687 			update_rq_clock(rq);
8688 			dequeue_entities(rq, se, DEQUEUE_SLEEP | DEQUEUE_DELAYED);
8689 		}
8690 		task_rq_unlock(rq, p, &rf);
8691 	}
8692 
8693 	remove_entity_load_avg(se);
8694 }
8695 
8696 /*
8697  * Set the max capacity the task is allowed to run at for misfit detection.
8698  */
set_task_max_allowed_capacity(struct task_struct * p)8699 static void set_task_max_allowed_capacity(struct task_struct *p)
8700 {
8701 	struct asym_cap_data *entry;
8702 
8703 	if (!sched_asym_cpucap_active())
8704 		return;
8705 
8706 	rcu_read_lock();
8707 	list_for_each_entry_rcu(entry, &asym_cap_list, link) {
8708 		cpumask_t *cpumask;
8709 
8710 		cpumask = cpu_capacity_span(entry);
8711 		if (!cpumask_intersects(p->cpus_ptr, cpumask))
8712 			continue;
8713 
8714 		p->max_allowed_capacity = entry->capacity;
8715 		break;
8716 	}
8717 	rcu_read_unlock();
8718 }
8719 
set_cpus_allowed_fair(struct task_struct * p,struct affinity_context * ctx)8720 static void set_cpus_allowed_fair(struct task_struct *p, struct affinity_context *ctx)
8721 {
8722 	set_cpus_allowed_common(p, ctx);
8723 	set_task_max_allowed_capacity(p);
8724 }
8725 
8726 static int
balance_fair(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)8727 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8728 {
8729 	if (sched_fair_runnable(rq))
8730 		return 1;
8731 
8732 	return sched_balance_newidle(rq, rf) != 0;
8733 }
8734 #else
set_task_max_allowed_capacity(struct task_struct * p)8735 static inline void set_task_max_allowed_capacity(struct task_struct *p) {}
8736 #endif /* CONFIG_SMP */
8737 
set_next_buddy(struct sched_entity * se)8738 static void set_next_buddy(struct sched_entity *se)
8739 {
8740 	for_each_sched_entity(se) {
8741 		if (WARN_ON_ONCE(!se->on_rq))
8742 			return;
8743 		if (se_is_idle(se))
8744 			return;
8745 		cfs_rq_of(se)->next = se;
8746 	}
8747 }
8748 
8749 /*
8750  * Preempt the current task with a newly woken task if needed:
8751  */
check_preempt_wakeup_fair(struct rq * rq,struct task_struct * p,int wake_flags)8752 static void check_preempt_wakeup_fair(struct rq *rq, struct task_struct *p, int wake_flags)
8753 {
8754 	struct task_struct *donor = rq->donor;
8755 	struct sched_entity *se = &donor->se, *pse = &p->se;
8756 	struct cfs_rq *cfs_rq = task_cfs_rq(donor);
8757 	int cse_is_idle, pse_is_idle;
8758 
8759 	if (unlikely(se == pse))
8760 		return;
8761 
8762 	/*
8763 	 * This is possible from callers such as attach_tasks(), in which we
8764 	 * unconditionally wakeup_preempt() after an enqueue (which may have
8765 	 * lead to a throttle).  This both saves work and prevents false
8766 	 * next-buddy nomination below.
8767 	 */
8768 	if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
8769 		return;
8770 
8771 	if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK) && !pse->sched_delayed) {
8772 		set_next_buddy(pse);
8773 	}
8774 
8775 	/*
8776 	 * We can come here with TIF_NEED_RESCHED already set from new task
8777 	 * wake up path.
8778 	 *
8779 	 * Note: this also catches the edge-case of curr being in a throttled
8780 	 * group (e.g. via set_curr_task), since update_curr() (in the
8781 	 * enqueue of curr) will have resulted in resched being set.  This
8782 	 * prevents us from potentially nominating it as a false LAST_BUDDY
8783 	 * below.
8784 	 */
8785 	if (test_tsk_need_resched(rq->curr))
8786 		return;
8787 
8788 	if (!sched_feat(WAKEUP_PREEMPTION))
8789 		return;
8790 
8791 	find_matching_se(&se, &pse);
8792 	WARN_ON_ONCE(!pse);
8793 
8794 	cse_is_idle = se_is_idle(se);
8795 	pse_is_idle = se_is_idle(pse);
8796 
8797 	/*
8798 	 * Preempt an idle entity in favor of a non-idle entity (and don't preempt
8799 	 * in the inverse case).
8800 	 */
8801 	if (cse_is_idle && !pse_is_idle) {
8802 		/*
8803 		 * When non-idle entity preempt an idle entity,
8804 		 * don't give idle entity slice protection.
8805 		 */
8806 		cancel_protect_slice(se);
8807 		goto preempt;
8808 	}
8809 
8810 	if (cse_is_idle != pse_is_idle)
8811 		return;
8812 
8813 	/*
8814 	 * BATCH and IDLE tasks do not preempt others.
8815 	 */
8816 	if (unlikely(!normal_policy(p->policy)))
8817 		return;
8818 
8819 	cfs_rq = cfs_rq_of(se);
8820 	update_curr(cfs_rq);
8821 	/*
8822 	 * If @p has a shorter slice than current and @p is eligible, override
8823 	 * current's slice protection in order to allow preemption.
8824 	 *
8825 	 * Note that even if @p does not turn out to be the most eligible
8826 	 * task at this moment, current's slice protection will be lost.
8827 	 */
8828 	if (do_preempt_short(cfs_rq, pse, se))
8829 		cancel_protect_slice(se);
8830 
8831 	/*
8832 	 * If @p has become the most eligible task, force preemption.
8833 	 */
8834 	if (pick_eevdf(cfs_rq) == pse)
8835 		goto preempt;
8836 
8837 	return;
8838 
8839 preempt:
8840 	resched_curr_lazy(rq);
8841 }
8842 
pick_task_fair(struct rq * rq)8843 static struct task_struct *pick_task_fair(struct rq *rq)
8844 {
8845 	struct sched_entity *se;
8846 	struct cfs_rq *cfs_rq;
8847 
8848 again:
8849 	cfs_rq = &rq->cfs;
8850 	if (!cfs_rq->nr_queued)
8851 		return NULL;
8852 
8853 	do {
8854 		/* Might not have done put_prev_entity() */
8855 		if (cfs_rq->curr && cfs_rq->curr->on_rq)
8856 			update_curr(cfs_rq);
8857 
8858 		if (unlikely(check_cfs_rq_runtime(cfs_rq)))
8859 			goto again;
8860 
8861 		se = pick_next_entity(rq, cfs_rq);
8862 		if (!se)
8863 			goto again;
8864 		cfs_rq = group_cfs_rq(se);
8865 	} while (cfs_rq);
8866 
8867 	return task_of(se);
8868 }
8869 
8870 static void __set_next_task_fair(struct rq *rq, struct task_struct *p, bool first);
8871 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first);
8872 
8873 struct task_struct *
pick_next_task_fair(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)8874 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8875 {
8876 	struct sched_entity *se;
8877 	struct task_struct *p;
8878 	int new_tasks;
8879 
8880 again:
8881 	p = pick_task_fair(rq);
8882 	if (!p)
8883 		goto idle;
8884 	se = &p->se;
8885 
8886 #ifdef CONFIG_FAIR_GROUP_SCHED
8887 	if (prev->sched_class != &fair_sched_class)
8888 		goto simple;
8889 
8890 	__put_prev_set_next_dl_server(rq, prev, p);
8891 
8892 	/*
8893 	 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
8894 	 * likely that a next task is from the same cgroup as the current.
8895 	 *
8896 	 * Therefore attempt to avoid putting and setting the entire cgroup
8897 	 * hierarchy, only change the part that actually changes.
8898 	 *
8899 	 * Since we haven't yet done put_prev_entity and if the selected task
8900 	 * is a different task than we started out with, try and touch the
8901 	 * least amount of cfs_rqs.
8902 	 */
8903 	if (prev != p) {
8904 		struct sched_entity *pse = &prev->se;
8905 		struct cfs_rq *cfs_rq;
8906 
8907 		while (!(cfs_rq = is_same_group(se, pse))) {
8908 			int se_depth = se->depth;
8909 			int pse_depth = pse->depth;
8910 
8911 			if (se_depth <= pse_depth) {
8912 				put_prev_entity(cfs_rq_of(pse), pse);
8913 				pse = parent_entity(pse);
8914 			}
8915 			if (se_depth >= pse_depth) {
8916 				set_next_entity(cfs_rq_of(se), se);
8917 				se = parent_entity(se);
8918 			}
8919 		}
8920 
8921 		put_prev_entity(cfs_rq, pse);
8922 		set_next_entity(cfs_rq, se);
8923 
8924 		__set_next_task_fair(rq, p, true);
8925 	}
8926 
8927 	return p;
8928 
8929 simple:
8930 #endif
8931 	put_prev_set_next_task(rq, prev, p);
8932 	return p;
8933 
8934 idle:
8935 	if (!rf)
8936 		return NULL;
8937 
8938 	new_tasks = sched_balance_newidle(rq, rf);
8939 
8940 	/*
8941 	 * Because sched_balance_newidle() releases (and re-acquires) rq->lock, it is
8942 	 * possible for any higher priority task to appear. In that case we
8943 	 * must re-start the pick_next_entity() loop.
8944 	 */
8945 	if (new_tasks < 0)
8946 		return RETRY_TASK;
8947 
8948 	if (new_tasks > 0)
8949 		goto again;
8950 
8951 	/*
8952 	 * rq is about to be idle, check if we need to update the
8953 	 * lost_idle_time of clock_pelt
8954 	 */
8955 	update_idle_rq_clock_pelt(rq);
8956 
8957 	return NULL;
8958 }
8959 
__pick_next_task_fair(struct rq * rq,struct task_struct * prev)8960 static struct task_struct *__pick_next_task_fair(struct rq *rq, struct task_struct *prev)
8961 {
8962 	return pick_next_task_fair(rq, prev, NULL);
8963 }
8964 
fair_server_has_tasks(struct sched_dl_entity * dl_se)8965 static bool fair_server_has_tasks(struct sched_dl_entity *dl_se)
8966 {
8967 	return !!dl_se->rq->cfs.nr_queued;
8968 }
8969 
fair_server_pick_task(struct sched_dl_entity * dl_se)8970 static struct task_struct *fair_server_pick_task(struct sched_dl_entity *dl_se)
8971 {
8972 	return pick_task_fair(dl_se->rq);
8973 }
8974 
fair_server_init(struct rq * rq)8975 void fair_server_init(struct rq *rq)
8976 {
8977 	struct sched_dl_entity *dl_se = &rq->fair_server;
8978 
8979 	init_dl_entity(dl_se);
8980 
8981 	dl_server_init(dl_se, rq, fair_server_has_tasks, fair_server_pick_task);
8982 }
8983 
8984 /*
8985  * Account for a descheduled task:
8986  */
put_prev_task_fair(struct rq * rq,struct task_struct * prev,struct task_struct * next)8987 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev, struct task_struct *next)
8988 {
8989 	struct sched_entity *se = &prev->se;
8990 	struct cfs_rq *cfs_rq;
8991 
8992 	for_each_sched_entity(se) {
8993 		cfs_rq = cfs_rq_of(se);
8994 		put_prev_entity(cfs_rq, se);
8995 	}
8996 }
8997 
8998 /*
8999  * sched_yield() is very simple
9000  */
yield_task_fair(struct rq * rq)9001 static void yield_task_fair(struct rq *rq)
9002 {
9003 	struct task_struct *curr = rq->curr;
9004 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
9005 	struct sched_entity *se = &curr->se;
9006 
9007 	/*
9008 	 * Are we the only task in the tree?
9009 	 */
9010 	if (unlikely(rq->nr_running == 1))
9011 		return;
9012 
9013 	clear_buddies(cfs_rq, se);
9014 
9015 	update_rq_clock(rq);
9016 	/*
9017 	 * Update run-time statistics of the 'current'.
9018 	 */
9019 	update_curr(cfs_rq);
9020 	/*
9021 	 * Tell update_rq_clock() that we've just updated,
9022 	 * so we don't do microscopic update in schedule()
9023 	 * and double the fastpath cost.
9024 	 */
9025 	rq_clock_skip_update(rq);
9026 
9027 	se->deadline += calc_delta_fair(se->slice, se);
9028 }
9029 
yield_to_task_fair(struct rq * rq,struct task_struct * p)9030 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
9031 {
9032 	struct sched_entity *se = &p->se;
9033 
9034 	/* throttled hierarchies are not runnable */
9035 	if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
9036 		return false;
9037 
9038 	/* Tell the scheduler that we'd really like se to run next. */
9039 	set_next_buddy(se);
9040 
9041 	yield_task_fair(rq);
9042 
9043 	return true;
9044 }
9045 
9046 #ifdef CONFIG_SMP
9047 /**************************************************
9048  * Fair scheduling class load-balancing methods.
9049  *
9050  * BASICS
9051  *
9052  * The purpose of load-balancing is to achieve the same basic fairness the
9053  * per-CPU scheduler provides, namely provide a proportional amount of compute
9054  * time to each task. This is expressed in the following equation:
9055  *
9056  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
9057  *
9058  * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
9059  * W_i,0 is defined as:
9060  *
9061  *   W_i,0 = \Sum_j w_i,j                                             (2)
9062  *
9063  * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
9064  * is derived from the nice value as per sched_prio_to_weight[].
9065  *
9066  * The weight average is an exponential decay average of the instantaneous
9067  * weight:
9068  *
9069  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
9070  *
9071  * C_i is the compute capacity of CPU i, typically it is the
9072  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
9073  * can also include other factors [XXX].
9074  *
9075  * To achieve this balance we define a measure of imbalance which follows
9076  * directly from (1):
9077  *
9078  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
9079  *
9080  * We them move tasks around to minimize the imbalance. In the continuous
9081  * function space it is obvious this converges, in the discrete case we get
9082  * a few fun cases generally called infeasible weight scenarios.
9083  *
9084  * [XXX expand on:
9085  *     - infeasible weights;
9086  *     - local vs global optima in the discrete case. ]
9087  *
9088  *
9089  * SCHED DOMAINS
9090  *
9091  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
9092  * for all i,j solution, we create a tree of CPUs that follows the hardware
9093  * topology where each level pairs two lower groups (or better). This results
9094  * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
9095  * tree to only the first of the previous level and we decrease the frequency
9096  * of load-balance at each level inversely proportional to the number of CPUs in
9097  * the groups.
9098  *
9099  * This yields:
9100  *
9101  *     log_2 n     1     n
9102  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
9103  *     i = 0      2^i   2^i
9104  *                               `- size of each group
9105  *         |         |     `- number of CPUs doing load-balance
9106  *         |         `- freq
9107  *         `- sum over all levels
9108  *
9109  * Coupled with a limit on how many tasks we can migrate every balance pass,
9110  * this makes (5) the runtime complexity of the balancer.
9111  *
9112  * An important property here is that each CPU is still (indirectly) connected
9113  * to every other CPU in at most O(log n) steps:
9114  *
9115  * The adjacency matrix of the resulting graph is given by:
9116  *
9117  *             log_2 n
9118  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
9119  *             k = 0
9120  *
9121  * And you'll find that:
9122  *
9123  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
9124  *
9125  * Showing there's indeed a path between every CPU in at most O(log n) steps.
9126  * The task movement gives a factor of O(m), giving a convergence complexity
9127  * of:
9128  *
9129  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
9130  *
9131  *
9132  * WORK CONSERVING
9133  *
9134  * In order to avoid CPUs going idle while there's still work to do, new idle
9135  * balancing is more aggressive and has the newly idle CPU iterate up the domain
9136  * tree itself instead of relying on other CPUs to bring it work.
9137  *
9138  * This adds some complexity to both (5) and (8) but it reduces the total idle
9139  * time.
9140  *
9141  * [XXX more?]
9142  *
9143  *
9144  * CGROUPS
9145  *
9146  * Cgroups make a horror show out of (2), instead of a simple sum we get:
9147  *
9148  *                                s_k,i
9149  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
9150  *                                 S_k
9151  *
9152  * Where
9153  *
9154  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
9155  *
9156  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
9157  *
9158  * The big problem is S_k, its a global sum needed to compute a local (W_i)
9159  * property.
9160  *
9161  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
9162  *      rewrite all of this once again.]
9163  */
9164 
9165 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
9166 
9167 enum fbq_type { regular, remote, all };
9168 
9169 /*
9170  * 'group_type' describes the group of CPUs at the moment of load balancing.
9171  *
9172  * The enum is ordered by pulling priority, with the group with lowest priority
9173  * first so the group_type can simply be compared when selecting the busiest
9174  * group. See update_sd_pick_busiest().
9175  */
9176 enum group_type {
9177 	/* The group has spare capacity that can be used to run more tasks.  */
9178 	group_has_spare = 0,
9179 	/*
9180 	 * The group is fully used and the tasks don't compete for more CPU
9181 	 * cycles. Nevertheless, some tasks might wait before running.
9182 	 */
9183 	group_fully_busy,
9184 	/*
9185 	 * One task doesn't fit with CPU's capacity and must be migrated to a
9186 	 * more powerful CPU.
9187 	 */
9188 	group_misfit_task,
9189 	/*
9190 	 * Balance SMT group that's fully busy. Can benefit from migration
9191 	 * a task on SMT with busy sibling to another CPU on idle core.
9192 	 */
9193 	group_smt_balance,
9194 	/*
9195 	 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
9196 	 * and the task should be migrated to it instead of running on the
9197 	 * current CPU.
9198 	 */
9199 	group_asym_packing,
9200 	/*
9201 	 * The tasks' affinity constraints previously prevented the scheduler
9202 	 * from balancing the load across the system.
9203 	 */
9204 	group_imbalanced,
9205 	/*
9206 	 * The CPU is overloaded and can't provide expected CPU cycles to all
9207 	 * tasks.
9208 	 */
9209 	group_overloaded
9210 };
9211 
9212 enum migration_type {
9213 	migrate_load = 0,
9214 	migrate_util,
9215 	migrate_task,
9216 	migrate_misfit
9217 };
9218 
9219 #define LBF_ALL_PINNED	0x01
9220 #define LBF_NEED_BREAK	0x02
9221 #define LBF_DST_PINNED  0x04
9222 #define LBF_SOME_PINNED	0x08
9223 #define LBF_ACTIVE_LB	0x10
9224 
9225 struct lb_env {
9226 	struct sched_domain	*sd;
9227 
9228 	struct rq		*src_rq;
9229 	int			src_cpu;
9230 
9231 	int			dst_cpu;
9232 	struct rq		*dst_rq;
9233 
9234 	struct cpumask		*dst_grpmask;
9235 	int			new_dst_cpu;
9236 	enum cpu_idle_type	idle;
9237 	long			imbalance;
9238 	/* The set of CPUs under consideration for load-balancing */
9239 	struct cpumask		*cpus;
9240 
9241 	unsigned int		flags;
9242 
9243 	unsigned int		loop;
9244 	unsigned int		loop_break;
9245 	unsigned int		loop_max;
9246 
9247 	enum fbq_type		fbq_type;
9248 	enum migration_type	migration_type;
9249 	struct list_head	tasks;
9250 };
9251 
9252 /*
9253  * Is this task likely cache-hot:
9254  */
task_hot(struct task_struct * p,struct lb_env * env)9255 static int task_hot(struct task_struct *p, struct lb_env *env)
9256 {
9257 	s64 delta;
9258 
9259 	lockdep_assert_rq_held(env->src_rq);
9260 
9261 	if (p->sched_class != &fair_sched_class)
9262 		return 0;
9263 
9264 	if (unlikely(task_has_idle_policy(p)))
9265 		return 0;
9266 
9267 	/* SMT siblings share cache */
9268 	if (env->sd->flags & SD_SHARE_CPUCAPACITY)
9269 		return 0;
9270 
9271 	/*
9272 	 * Buddy candidates are cache hot:
9273 	 */
9274 	if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
9275 	    (&p->se == cfs_rq_of(&p->se)->next))
9276 		return 1;
9277 
9278 	if (sysctl_sched_migration_cost == -1)
9279 		return 1;
9280 
9281 	/*
9282 	 * Don't migrate task if the task's cookie does not match
9283 	 * with the destination CPU's core cookie.
9284 	 */
9285 	if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
9286 		return 1;
9287 
9288 	if (sysctl_sched_migration_cost == 0)
9289 		return 0;
9290 
9291 	delta = rq_clock_task(env->src_rq) - p->se.exec_start;
9292 
9293 	return delta < (s64)sysctl_sched_migration_cost;
9294 }
9295 
9296 #ifdef CONFIG_NUMA_BALANCING
9297 /*
9298  * Returns a positive value, if task migration degrades locality.
9299  * Returns 0, if task migration is not affected by locality.
9300  * Returns a negative value, if task migration improves locality i.e migration preferred.
9301  */
migrate_degrades_locality(struct task_struct * p,struct lb_env * env)9302 static long migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
9303 {
9304 	struct numa_group *numa_group = rcu_dereference(p->numa_group);
9305 	unsigned long src_weight, dst_weight;
9306 	int src_nid, dst_nid, dist;
9307 
9308 	if (!static_branch_likely(&sched_numa_balancing))
9309 		return 0;
9310 
9311 	if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
9312 		return 0;
9313 
9314 	src_nid = cpu_to_node(env->src_cpu);
9315 	dst_nid = cpu_to_node(env->dst_cpu);
9316 
9317 	if (src_nid == dst_nid)
9318 		return 0;
9319 
9320 	/* Migrating away from the preferred node is always bad. */
9321 	if (src_nid == p->numa_preferred_nid) {
9322 		if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
9323 			return 1;
9324 		else
9325 			return 0;
9326 	}
9327 
9328 	/* Encourage migration to the preferred node. */
9329 	if (dst_nid == p->numa_preferred_nid)
9330 		return -1;
9331 
9332 	/* Leaving a core idle is often worse than degrading locality. */
9333 	if (env->idle == CPU_IDLE)
9334 		return 0;
9335 
9336 	dist = node_distance(src_nid, dst_nid);
9337 	if (numa_group) {
9338 		src_weight = group_weight(p, src_nid, dist);
9339 		dst_weight = group_weight(p, dst_nid, dist);
9340 	} else {
9341 		src_weight = task_weight(p, src_nid, dist);
9342 		dst_weight = task_weight(p, dst_nid, dist);
9343 	}
9344 
9345 	return src_weight - dst_weight;
9346 }
9347 
9348 #else
migrate_degrades_locality(struct task_struct * p,struct lb_env * env)9349 static inline long migrate_degrades_locality(struct task_struct *p,
9350 					     struct lb_env *env)
9351 {
9352 	return 0;
9353 }
9354 #endif
9355 
9356 /*
9357  * Check whether the task is ineligible on the destination cpu
9358  *
9359  * When the PLACE_LAG scheduling feature is enabled and
9360  * dst_cfs_rq->nr_queued is greater than 1, if the task
9361  * is ineligible, it will also be ineligible when
9362  * it is migrated to the destination cpu.
9363  */
task_is_ineligible_on_dst_cpu(struct task_struct * p,int dest_cpu)9364 static inline int task_is_ineligible_on_dst_cpu(struct task_struct *p, int dest_cpu)
9365 {
9366 	struct cfs_rq *dst_cfs_rq;
9367 
9368 #ifdef CONFIG_FAIR_GROUP_SCHED
9369 	dst_cfs_rq = task_group(p)->cfs_rq[dest_cpu];
9370 #else
9371 	dst_cfs_rq = &cpu_rq(dest_cpu)->cfs;
9372 #endif
9373 	if (sched_feat(PLACE_LAG) && dst_cfs_rq->nr_queued &&
9374 	    !entity_eligible(task_cfs_rq(p), &p->se))
9375 		return 1;
9376 
9377 	return 0;
9378 }
9379 
9380 /*
9381  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
9382  */
9383 static
can_migrate_task(struct task_struct * p,struct lb_env * env)9384 int can_migrate_task(struct task_struct *p, struct lb_env *env)
9385 {
9386 	long degrades, hot;
9387 
9388 	lockdep_assert_rq_held(env->src_rq);
9389 	if (p->sched_task_hot)
9390 		p->sched_task_hot = 0;
9391 
9392 	/*
9393 	 * We do not migrate tasks that are:
9394 	 * 1) delayed dequeued unless we migrate load, or
9395 	 * 2) throttled_lb_pair, or
9396 	 * 3) cannot be migrated to this CPU due to cpus_ptr, or
9397 	 * 4) running (obviously), or
9398 	 * 5) are cache-hot on their current CPU.
9399 	 */
9400 	if ((p->se.sched_delayed) && (env->migration_type != migrate_load))
9401 		return 0;
9402 
9403 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
9404 		return 0;
9405 
9406 	/*
9407 	 * We want to prioritize the migration of eligible tasks.
9408 	 * For ineligible tasks we soft-limit them and only allow
9409 	 * them to migrate when nr_balance_failed is non-zero to
9410 	 * avoid load-balancing trying very hard to balance the load.
9411 	 */
9412 	if (!env->sd->nr_balance_failed &&
9413 	    task_is_ineligible_on_dst_cpu(p, env->dst_cpu))
9414 		return 0;
9415 
9416 	/* Disregard percpu kthreads; they are where they need to be. */
9417 	if (kthread_is_per_cpu(p))
9418 		return 0;
9419 
9420 	if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
9421 		int cpu;
9422 
9423 		schedstat_inc(p->stats.nr_failed_migrations_affine);
9424 
9425 		env->flags |= LBF_SOME_PINNED;
9426 
9427 		/*
9428 		 * Remember if this task can be migrated to any other CPU in
9429 		 * our sched_group. We may want to revisit it if we couldn't
9430 		 * meet load balance goals by pulling other tasks on src_cpu.
9431 		 *
9432 		 * Avoid computing new_dst_cpu
9433 		 * - for NEWLY_IDLE
9434 		 * - if we have already computed one in current iteration
9435 		 * - if it's an active balance
9436 		 */
9437 		if (env->idle == CPU_NEWLY_IDLE ||
9438 		    env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
9439 			return 0;
9440 
9441 		/* Prevent to re-select dst_cpu via env's CPUs: */
9442 		cpu = cpumask_first_and_and(env->dst_grpmask, env->cpus, p->cpus_ptr);
9443 
9444 		if (cpu < nr_cpu_ids) {
9445 			env->flags |= LBF_DST_PINNED;
9446 			env->new_dst_cpu = cpu;
9447 		}
9448 
9449 		return 0;
9450 	}
9451 
9452 	/* Record that we found at least one task that could run on dst_cpu */
9453 	env->flags &= ~LBF_ALL_PINNED;
9454 
9455 	if (task_on_cpu(env->src_rq, p)) {
9456 		schedstat_inc(p->stats.nr_failed_migrations_running);
9457 		return 0;
9458 	}
9459 
9460 	/*
9461 	 * Aggressive migration if:
9462 	 * 1) active balance
9463 	 * 2) destination numa is preferred
9464 	 * 3) task is cache cold, or
9465 	 * 4) too many balance attempts have failed.
9466 	 */
9467 	if (env->flags & LBF_ACTIVE_LB)
9468 		return 1;
9469 
9470 	degrades = migrate_degrades_locality(p, env);
9471 	if (!degrades)
9472 		hot = task_hot(p, env);
9473 	else
9474 		hot = degrades > 0;
9475 
9476 	if (!hot || env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
9477 		if (hot)
9478 			p->sched_task_hot = 1;
9479 		return 1;
9480 	}
9481 
9482 	schedstat_inc(p->stats.nr_failed_migrations_hot);
9483 	return 0;
9484 }
9485 
9486 /*
9487  * detach_task() -- detach the task for the migration specified in env
9488  */
detach_task(struct task_struct * p,struct lb_env * env)9489 static void detach_task(struct task_struct *p, struct lb_env *env)
9490 {
9491 	lockdep_assert_rq_held(env->src_rq);
9492 
9493 	if (p->sched_task_hot) {
9494 		p->sched_task_hot = 0;
9495 		schedstat_inc(env->sd->lb_hot_gained[env->idle]);
9496 		schedstat_inc(p->stats.nr_forced_migrations);
9497 	}
9498 
9499 	deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
9500 	set_task_cpu(p, env->dst_cpu);
9501 }
9502 
9503 /*
9504  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
9505  * part of active balancing operations within "domain".
9506  *
9507  * Returns a task if successful and NULL otherwise.
9508  */
detach_one_task(struct lb_env * env)9509 static struct task_struct *detach_one_task(struct lb_env *env)
9510 {
9511 	struct task_struct *p;
9512 
9513 	lockdep_assert_rq_held(env->src_rq);
9514 
9515 	list_for_each_entry_reverse(p,
9516 			&env->src_rq->cfs_tasks, se.group_node) {
9517 		if (!can_migrate_task(p, env))
9518 			continue;
9519 
9520 		detach_task(p, env);
9521 
9522 		/*
9523 		 * Right now, this is only the second place where
9524 		 * lb_gained[env->idle] is updated (other is detach_tasks)
9525 		 * so we can safely collect stats here rather than
9526 		 * inside detach_tasks().
9527 		 */
9528 		schedstat_inc(env->sd->lb_gained[env->idle]);
9529 		return p;
9530 	}
9531 	return NULL;
9532 }
9533 
9534 /*
9535  * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
9536  * busiest_rq, as part of a balancing operation within domain "sd".
9537  *
9538  * Returns number of detached tasks if successful and 0 otherwise.
9539  */
detach_tasks(struct lb_env * env)9540 static int detach_tasks(struct lb_env *env)
9541 {
9542 	struct list_head *tasks = &env->src_rq->cfs_tasks;
9543 	unsigned long util, load;
9544 	struct task_struct *p;
9545 	int detached = 0;
9546 
9547 	lockdep_assert_rq_held(env->src_rq);
9548 
9549 	/*
9550 	 * Source run queue has been emptied by another CPU, clear
9551 	 * LBF_ALL_PINNED flag as we will not test any task.
9552 	 */
9553 	if (env->src_rq->nr_running <= 1) {
9554 		env->flags &= ~LBF_ALL_PINNED;
9555 		return 0;
9556 	}
9557 
9558 	if (env->imbalance <= 0)
9559 		return 0;
9560 
9561 	while (!list_empty(tasks)) {
9562 		/*
9563 		 * We don't want to steal all, otherwise we may be treated likewise,
9564 		 * which could at worst lead to a livelock crash.
9565 		 */
9566 		if (env->idle && env->src_rq->nr_running <= 1)
9567 			break;
9568 
9569 		env->loop++;
9570 		/* We've more or less seen every task there is, call it quits */
9571 		if (env->loop > env->loop_max)
9572 			break;
9573 
9574 		/* take a breather every nr_migrate tasks */
9575 		if (env->loop > env->loop_break) {
9576 			env->loop_break += SCHED_NR_MIGRATE_BREAK;
9577 			env->flags |= LBF_NEED_BREAK;
9578 			break;
9579 		}
9580 
9581 		p = list_last_entry(tasks, struct task_struct, se.group_node);
9582 
9583 		if (!can_migrate_task(p, env))
9584 			goto next;
9585 
9586 		switch (env->migration_type) {
9587 		case migrate_load:
9588 			/*
9589 			 * Depending of the number of CPUs and tasks and the
9590 			 * cgroup hierarchy, task_h_load() can return a null
9591 			 * value. Make sure that env->imbalance decreases
9592 			 * otherwise detach_tasks() will stop only after
9593 			 * detaching up to loop_max tasks.
9594 			 */
9595 			load = max_t(unsigned long, task_h_load(p), 1);
9596 
9597 			if (sched_feat(LB_MIN) &&
9598 			    load < 16 && !env->sd->nr_balance_failed)
9599 				goto next;
9600 
9601 			/*
9602 			 * Make sure that we don't migrate too much load.
9603 			 * Nevertheless, let relax the constraint if
9604 			 * scheduler fails to find a good waiting task to
9605 			 * migrate.
9606 			 */
9607 			if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
9608 				goto next;
9609 
9610 			env->imbalance -= load;
9611 			break;
9612 
9613 		case migrate_util:
9614 			util = task_util_est(p);
9615 
9616 			if (shr_bound(util, env->sd->nr_balance_failed) > env->imbalance)
9617 				goto next;
9618 
9619 			env->imbalance -= util;
9620 			break;
9621 
9622 		case migrate_task:
9623 			env->imbalance--;
9624 			break;
9625 
9626 		case migrate_misfit:
9627 			/* This is not a misfit task */
9628 			if (task_fits_cpu(p, env->src_cpu))
9629 				goto next;
9630 
9631 			env->imbalance = 0;
9632 			break;
9633 		}
9634 
9635 		detach_task(p, env);
9636 		list_add(&p->se.group_node, &env->tasks);
9637 
9638 		detached++;
9639 
9640 #ifdef CONFIG_PREEMPTION
9641 		/*
9642 		 * NEWIDLE balancing is a source of latency, so preemptible
9643 		 * kernels will stop after the first task is detached to minimize
9644 		 * the critical section.
9645 		 */
9646 		if (env->idle == CPU_NEWLY_IDLE)
9647 			break;
9648 #endif
9649 
9650 		/*
9651 		 * We only want to steal up to the prescribed amount of
9652 		 * load/util/tasks.
9653 		 */
9654 		if (env->imbalance <= 0)
9655 			break;
9656 
9657 		continue;
9658 next:
9659 		if (p->sched_task_hot)
9660 			schedstat_inc(p->stats.nr_failed_migrations_hot);
9661 
9662 		list_move(&p->se.group_node, tasks);
9663 	}
9664 
9665 	/*
9666 	 * Right now, this is one of only two places we collect this stat
9667 	 * so we can safely collect detach_one_task() stats here rather
9668 	 * than inside detach_one_task().
9669 	 */
9670 	schedstat_add(env->sd->lb_gained[env->idle], detached);
9671 
9672 	return detached;
9673 }
9674 
9675 /*
9676  * attach_task() -- attach the task detached by detach_task() to its new rq.
9677  */
attach_task(struct rq * rq,struct task_struct * p)9678 static void attach_task(struct rq *rq, struct task_struct *p)
9679 {
9680 	lockdep_assert_rq_held(rq);
9681 
9682 	WARN_ON_ONCE(task_rq(p) != rq);
9683 	activate_task(rq, p, ENQUEUE_NOCLOCK);
9684 	wakeup_preempt(rq, p, 0);
9685 }
9686 
9687 /*
9688  * attach_one_task() -- attaches the task returned from detach_one_task() to
9689  * its new rq.
9690  */
attach_one_task(struct rq * rq,struct task_struct * p)9691 static void attach_one_task(struct rq *rq, struct task_struct *p)
9692 {
9693 	struct rq_flags rf;
9694 
9695 	rq_lock(rq, &rf);
9696 	update_rq_clock(rq);
9697 	attach_task(rq, p);
9698 	rq_unlock(rq, &rf);
9699 }
9700 
9701 /*
9702  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
9703  * new rq.
9704  */
attach_tasks(struct lb_env * env)9705 static void attach_tasks(struct lb_env *env)
9706 {
9707 	struct list_head *tasks = &env->tasks;
9708 	struct task_struct *p;
9709 	struct rq_flags rf;
9710 
9711 	rq_lock(env->dst_rq, &rf);
9712 	update_rq_clock(env->dst_rq);
9713 
9714 	while (!list_empty(tasks)) {
9715 		p = list_first_entry(tasks, struct task_struct, se.group_node);
9716 		list_del_init(&p->se.group_node);
9717 
9718 		attach_task(env->dst_rq, p);
9719 	}
9720 
9721 	rq_unlock(env->dst_rq, &rf);
9722 }
9723 
9724 #ifdef CONFIG_NO_HZ_COMMON
cfs_rq_has_blocked(struct cfs_rq * cfs_rq)9725 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
9726 {
9727 	if (cfs_rq->avg.load_avg)
9728 		return true;
9729 
9730 	if (cfs_rq->avg.util_avg)
9731 		return true;
9732 
9733 	return false;
9734 }
9735 
others_have_blocked(struct rq * rq)9736 static inline bool others_have_blocked(struct rq *rq)
9737 {
9738 	if (cpu_util_rt(rq))
9739 		return true;
9740 
9741 	if (cpu_util_dl(rq))
9742 		return true;
9743 
9744 	if (hw_load_avg(rq))
9745 		return true;
9746 
9747 	if (cpu_util_irq(rq))
9748 		return true;
9749 
9750 	return false;
9751 }
9752 
update_blocked_load_tick(struct rq * rq)9753 static inline void update_blocked_load_tick(struct rq *rq)
9754 {
9755 	WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
9756 }
9757 
update_blocked_load_status(struct rq * rq,bool has_blocked)9758 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
9759 {
9760 	if (!has_blocked)
9761 		rq->has_blocked_load = 0;
9762 }
9763 #else
cfs_rq_has_blocked(struct cfs_rq * cfs_rq)9764 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
others_have_blocked(struct rq * rq)9765 static inline bool others_have_blocked(struct rq *rq) { return false; }
update_blocked_load_tick(struct rq * rq)9766 static inline void update_blocked_load_tick(struct rq *rq) {}
update_blocked_load_status(struct rq * rq,bool has_blocked)9767 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
9768 #endif
9769 
__update_blocked_others(struct rq * rq,bool * done)9770 static bool __update_blocked_others(struct rq *rq, bool *done)
9771 {
9772 	bool updated;
9773 
9774 	/*
9775 	 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
9776 	 * DL and IRQ signals have been updated before updating CFS.
9777 	 */
9778 	updated = update_other_load_avgs(rq);
9779 
9780 	if (others_have_blocked(rq))
9781 		*done = false;
9782 
9783 	return updated;
9784 }
9785 
9786 #ifdef CONFIG_FAIR_GROUP_SCHED
9787 
__update_blocked_fair(struct rq * rq,bool * done)9788 static bool __update_blocked_fair(struct rq *rq, bool *done)
9789 {
9790 	struct cfs_rq *cfs_rq, *pos;
9791 	bool decayed = false;
9792 	int cpu = cpu_of(rq);
9793 
9794 	/*
9795 	 * Iterates the task_group tree in a bottom up fashion, see
9796 	 * list_add_leaf_cfs_rq() for details.
9797 	 */
9798 	for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
9799 		struct sched_entity *se;
9800 
9801 		if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
9802 			update_tg_load_avg(cfs_rq);
9803 
9804 			if (cfs_rq->nr_queued == 0)
9805 				update_idle_cfs_rq_clock_pelt(cfs_rq);
9806 
9807 			if (cfs_rq == &rq->cfs)
9808 				decayed = true;
9809 		}
9810 
9811 		/* Propagate pending load changes to the parent, if any: */
9812 		se = cfs_rq->tg->se[cpu];
9813 		if (se && !skip_blocked_update(se))
9814 			update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9815 
9816 		/*
9817 		 * There can be a lot of idle CPU cgroups.  Don't let fully
9818 		 * decayed cfs_rqs linger on the list.
9819 		 */
9820 		if (cfs_rq_is_decayed(cfs_rq))
9821 			list_del_leaf_cfs_rq(cfs_rq);
9822 
9823 		/* Don't need periodic decay once load/util_avg are null */
9824 		if (cfs_rq_has_blocked(cfs_rq))
9825 			*done = false;
9826 	}
9827 
9828 	return decayed;
9829 }
9830 
9831 /*
9832  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
9833  * This needs to be done in a top-down fashion because the load of a child
9834  * group is a fraction of its parents load.
9835  */
update_cfs_rq_h_load(struct cfs_rq * cfs_rq)9836 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
9837 {
9838 	struct rq *rq = rq_of(cfs_rq);
9839 	struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
9840 	unsigned long now = jiffies;
9841 	unsigned long load;
9842 
9843 	if (cfs_rq->last_h_load_update == now)
9844 		return;
9845 
9846 	WRITE_ONCE(cfs_rq->h_load_next, NULL);
9847 	for_each_sched_entity(se) {
9848 		cfs_rq = cfs_rq_of(se);
9849 		WRITE_ONCE(cfs_rq->h_load_next, se);
9850 		if (cfs_rq->last_h_load_update == now)
9851 			break;
9852 	}
9853 
9854 	if (!se) {
9855 		cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
9856 		cfs_rq->last_h_load_update = now;
9857 	}
9858 
9859 	while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
9860 		load = cfs_rq->h_load;
9861 		load = div64_ul(load * se->avg.load_avg,
9862 			cfs_rq_load_avg(cfs_rq) + 1);
9863 		cfs_rq = group_cfs_rq(se);
9864 		cfs_rq->h_load = load;
9865 		cfs_rq->last_h_load_update = now;
9866 	}
9867 }
9868 
task_h_load(struct task_struct * p)9869 static unsigned long task_h_load(struct task_struct *p)
9870 {
9871 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
9872 
9873 	update_cfs_rq_h_load(cfs_rq);
9874 	return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
9875 			cfs_rq_load_avg(cfs_rq) + 1);
9876 }
9877 #else
__update_blocked_fair(struct rq * rq,bool * done)9878 static bool __update_blocked_fair(struct rq *rq, bool *done)
9879 {
9880 	struct cfs_rq *cfs_rq = &rq->cfs;
9881 	bool decayed;
9882 
9883 	decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
9884 	if (cfs_rq_has_blocked(cfs_rq))
9885 		*done = false;
9886 
9887 	return decayed;
9888 }
9889 
task_h_load(struct task_struct * p)9890 static unsigned long task_h_load(struct task_struct *p)
9891 {
9892 	return p->se.avg.load_avg;
9893 }
9894 #endif
9895 
sched_balance_update_blocked_averages(int cpu)9896 static void sched_balance_update_blocked_averages(int cpu)
9897 {
9898 	bool decayed = false, done = true;
9899 	struct rq *rq = cpu_rq(cpu);
9900 	struct rq_flags rf;
9901 
9902 	rq_lock_irqsave(rq, &rf);
9903 	update_blocked_load_tick(rq);
9904 	update_rq_clock(rq);
9905 
9906 	decayed |= __update_blocked_others(rq, &done);
9907 	decayed |= __update_blocked_fair(rq, &done);
9908 
9909 	update_blocked_load_status(rq, !done);
9910 	if (decayed)
9911 		cpufreq_update_util(rq, 0);
9912 	rq_unlock_irqrestore(rq, &rf);
9913 }
9914 
9915 /********** Helpers for sched_balance_find_src_group ************************/
9916 
9917 /*
9918  * sg_lb_stats - stats of a sched_group required for load-balancing:
9919  */
9920 struct sg_lb_stats {
9921 	unsigned long avg_load;			/* Avg load            over the CPUs of the group */
9922 	unsigned long group_load;		/* Total load          over the CPUs of the group */
9923 	unsigned long group_capacity;		/* Capacity            over the CPUs of the group */
9924 	unsigned long group_util;		/* Total utilization   over the CPUs of the group */
9925 	unsigned long group_runnable;		/* Total runnable time over the CPUs of the group */
9926 	unsigned int sum_nr_running;		/* Nr of all tasks running in the group */
9927 	unsigned int sum_h_nr_running;		/* Nr of CFS tasks running in the group */
9928 	unsigned int idle_cpus;                 /* Nr of idle CPUs         in the group */
9929 	unsigned int group_weight;
9930 	enum group_type group_type;
9931 	unsigned int group_asym_packing;	/* Tasks should be moved to preferred CPU */
9932 	unsigned int group_smt_balance;		/* Task on busy SMT be moved */
9933 	unsigned long group_misfit_task_load;	/* A CPU has a task too big for its capacity */
9934 #ifdef CONFIG_NUMA_BALANCING
9935 	unsigned int nr_numa_running;
9936 	unsigned int nr_preferred_running;
9937 #endif
9938 };
9939 
9940 /*
9941  * sd_lb_stats - stats of a sched_domain required for load-balancing:
9942  */
9943 struct sd_lb_stats {
9944 	struct sched_group *busiest;		/* Busiest group in this sd */
9945 	struct sched_group *local;		/* Local group in this sd */
9946 	unsigned long total_load;		/* Total load of all groups in sd */
9947 	unsigned long total_capacity;		/* Total capacity of all groups in sd */
9948 	unsigned long avg_load;			/* Average load across all groups in sd */
9949 	unsigned int prefer_sibling;		/* Tasks should go to sibling first */
9950 
9951 	struct sg_lb_stats busiest_stat;	/* Statistics of the busiest group */
9952 	struct sg_lb_stats local_stat;		/* Statistics of the local group */
9953 };
9954 
init_sd_lb_stats(struct sd_lb_stats * sds)9955 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
9956 {
9957 	/*
9958 	 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
9959 	 * local_stat because update_sg_lb_stats() does a full clear/assignment.
9960 	 * We must however set busiest_stat::group_type and
9961 	 * busiest_stat::idle_cpus to the worst busiest group because
9962 	 * update_sd_pick_busiest() reads these before assignment.
9963 	 */
9964 	*sds = (struct sd_lb_stats){
9965 		.busiest = NULL,
9966 		.local = NULL,
9967 		.total_load = 0UL,
9968 		.total_capacity = 0UL,
9969 		.busiest_stat = {
9970 			.idle_cpus = UINT_MAX,
9971 			.group_type = group_has_spare,
9972 		},
9973 	};
9974 }
9975 
scale_rt_capacity(int cpu)9976 static unsigned long scale_rt_capacity(int cpu)
9977 {
9978 	unsigned long max = get_actual_cpu_capacity(cpu);
9979 	struct rq *rq = cpu_rq(cpu);
9980 	unsigned long used, free;
9981 	unsigned long irq;
9982 
9983 	irq = cpu_util_irq(rq);
9984 
9985 	if (unlikely(irq >= max))
9986 		return 1;
9987 
9988 	/*
9989 	 * avg_rt.util_avg and avg_dl.util_avg track binary signals
9990 	 * (running and not running) with weights 0 and 1024 respectively.
9991 	 */
9992 	used = cpu_util_rt(rq);
9993 	used += cpu_util_dl(rq);
9994 
9995 	if (unlikely(used >= max))
9996 		return 1;
9997 
9998 	free = max - used;
9999 
10000 	return scale_irq_capacity(free, irq, max);
10001 }
10002 
update_cpu_capacity(struct sched_domain * sd,int cpu)10003 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
10004 {
10005 	unsigned long capacity = scale_rt_capacity(cpu);
10006 	struct sched_group *sdg = sd->groups;
10007 
10008 	if (!capacity)
10009 		capacity = 1;
10010 
10011 	cpu_rq(cpu)->cpu_capacity = capacity;
10012 	trace_sched_cpu_capacity_tp(cpu_rq(cpu));
10013 
10014 	sdg->sgc->capacity = capacity;
10015 	sdg->sgc->min_capacity = capacity;
10016 	sdg->sgc->max_capacity = capacity;
10017 }
10018 
update_group_capacity(struct sched_domain * sd,int cpu)10019 void update_group_capacity(struct sched_domain *sd, int cpu)
10020 {
10021 	struct sched_domain *child = sd->child;
10022 	struct sched_group *group, *sdg = sd->groups;
10023 	unsigned long capacity, min_capacity, max_capacity;
10024 	unsigned long interval;
10025 
10026 	interval = msecs_to_jiffies(sd->balance_interval);
10027 	interval = clamp(interval, 1UL, max_load_balance_interval);
10028 	sdg->sgc->next_update = jiffies + interval;
10029 
10030 	if (!child) {
10031 		update_cpu_capacity(sd, cpu);
10032 		return;
10033 	}
10034 
10035 	capacity = 0;
10036 	min_capacity = ULONG_MAX;
10037 	max_capacity = 0;
10038 
10039 	if (child->flags & SD_OVERLAP) {
10040 		/*
10041 		 * SD_OVERLAP domains cannot assume that child groups
10042 		 * span the current group.
10043 		 */
10044 
10045 		for_each_cpu(cpu, sched_group_span(sdg)) {
10046 			unsigned long cpu_cap = capacity_of(cpu);
10047 
10048 			capacity += cpu_cap;
10049 			min_capacity = min(cpu_cap, min_capacity);
10050 			max_capacity = max(cpu_cap, max_capacity);
10051 		}
10052 	} else  {
10053 		/*
10054 		 * !SD_OVERLAP domains can assume that child groups
10055 		 * span the current group.
10056 		 */
10057 
10058 		group = child->groups;
10059 		do {
10060 			struct sched_group_capacity *sgc = group->sgc;
10061 
10062 			capacity += sgc->capacity;
10063 			min_capacity = min(sgc->min_capacity, min_capacity);
10064 			max_capacity = max(sgc->max_capacity, max_capacity);
10065 			group = group->next;
10066 		} while (group != child->groups);
10067 	}
10068 
10069 	sdg->sgc->capacity = capacity;
10070 	sdg->sgc->min_capacity = min_capacity;
10071 	sdg->sgc->max_capacity = max_capacity;
10072 }
10073 
10074 /*
10075  * Check whether the capacity of the rq has been noticeably reduced by side
10076  * activity. The imbalance_pct is used for the threshold.
10077  * Return true is the capacity is reduced
10078  */
10079 static inline int
check_cpu_capacity(struct rq * rq,struct sched_domain * sd)10080 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
10081 {
10082 	return ((rq->cpu_capacity * sd->imbalance_pct) <
10083 				(arch_scale_cpu_capacity(cpu_of(rq)) * 100));
10084 }
10085 
10086 /* Check if the rq has a misfit task */
check_misfit_status(struct rq * rq)10087 static inline bool check_misfit_status(struct rq *rq)
10088 {
10089 	return rq->misfit_task_load;
10090 }
10091 
10092 /*
10093  * Group imbalance indicates (and tries to solve) the problem where balancing
10094  * groups is inadequate due to ->cpus_ptr constraints.
10095  *
10096  * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
10097  * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
10098  * Something like:
10099  *
10100  *	{ 0 1 2 3 } { 4 5 6 7 }
10101  *	        *     * * *
10102  *
10103  * If we were to balance group-wise we'd place two tasks in the first group and
10104  * two tasks in the second group. Clearly this is undesired as it will overload
10105  * cpu 3 and leave one of the CPUs in the second group unused.
10106  *
10107  * The current solution to this issue is detecting the skew in the first group
10108  * by noticing the lower domain failed to reach balance and had difficulty
10109  * moving tasks due to affinity constraints.
10110  *
10111  * When this is so detected; this group becomes a candidate for busiest; see
10112  * update_sd_pick_busiest(). And calculate_imbalance() and
10113  * sched_balance_find_src_group() avoid some of the usual balance conditions to allow it
10114  * to create an effective group imbalance.
10115  *
10116  * This is a somewhat tricky proposition since the next run might not find the
10117  * group imbalance and decide the groups need to be balanced again. A most
10118  * subtle and fragile situation.
10119  */
10120 
sg_imbalanced(struct sched_group * group)10121 static inline int sg_imbalanced(struct sched_group *group)
10122 {
10123 	return group->sgc->imbalance;
10124 }
10125 
10126 /*
10127  * group_has_capacity returns true if the group has spare capacity that could
10128  * be used by some tasks.
10129  * We consider that a group has spare capacity if the number of task is
10130  * smaller than the number of CPUs or if the utilization is lower than the
10131  * available capacity for CFS tasks.
10132  * For the latter, we use a threshold to stabilize the state, to take into
10133  * account the variance of the tasks' load and to return true if the available
10134  * capacity in meaningful for the load balancer.
10135  * As an example, an available capacity of 1% can appear but it doesn't make
10136  * any benefit for the load balance.
10137  */
10138 static inline bool
group_has_capacity(unsigned int imbalance_pct,struct sg_lb_stats * sgs)10139 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
10140 {
10141 	if (sgs->sum_nr_running < sgs->group_weight)
10142 		return true;
10143 
10144 	if ((sgs->group_capacity * imbalance_pct) <
10145 			(sgs->group_runnable * 100))
10146 		return false;
10147 
10148 	if ((sgs->group_capacity * 100) >
10149 			(sgs->group_util * imbalance_pct))
10150 		return true;
10151 
10152 	return false;
10153 }
10154 
10155 /*
10156  *  group_is_overloaded returns true if the group has more tasks than it can
10157  *  handle.
10158  *  group_is_overloaded is not equals to !group_has_capacity because a group
10159  *  with the exact right number of tasks, has no more spare capacity but is not
10160  *  overloaded so both group_has_capacity and group_is_overloaded return
10161  *  false.
10162  */
10163 static inline bool
group_is_overloaded(unsigned int imbalance_pct,struct sg_lb_stats * sgs)10164 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
10165 {
10166 	if (sgs->sum_nr_running <= sgs->group_weight)
10167 		return false;
10168 
10169 	if ((sgs->group_capacity * 100) <
10170 			(sgs->group_util * imbalance_pct))
10171 		return true;
10172 
10173 	if ((sgs->group_capacity * imbalance_pct) <
10174 			(sgs->group_runnable * 100))
10175 		return true;
10176 
10177 	return false;
10178 }
10179 
10180 static inline enum
group_classify(unsigned int imbalance_pct,struct sched_group * group,struct sg_lb_stats * sgs)10181 group_type group_classify(unsigned int imbalance_pct,
10182 			  struct sched_group *group,
10183 			  struct sg_lb_stats *sgs)
10184 {
10185 	if (group_is_overloaded(imbalance_pct, sgs))
10186 		return group_overloaded;
10187 
10188 	if (sg_imbalanced(group))
10189 		return group_imbalanced;
10190 
10191 	if (sgs->group_asym_packing)
10192 		return group_asym_packing;
10193 
10194 	if (sgs->group_smt_balance)
10195 		return group_smt_balance;
10196 
10197 	if (sgs->group_misfit_task_load)
10198 		return group_misfit_task;
10199 
10200 	if (!group_has_capacity(imbalance_pct, sgs))
10201 		return group_fully_busy;
10202 
10203 	return group_has_spare;
10204 }
10205 
10206 /**
10207  * sched_use_asym_prio - Check whether asym_packing priority must be used
10208  * @sd:		The scheduling domain of the load balancing
10209  * @cpu:	A CPU
10210  *
10211  * Always use CPU priority when balancing load between SMT siblings. When
10212  * balancing load between cores, it is not sufficient that @cpu is idle. Only
10213  * use CPU priority if the whole core is idle.
10214  *
10215  * Returns: True if the priority of @cpu must be followed. False otherwise.
10216  */
sched_use_asym_prio(struct sched_domain * sd,int cpu)10217 static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
10218 {
10219 	if (!(sd->flags & SD_ASYM_PACKING))
10220 		return false;
10221 
10222 	if (!sched_smt_active())
10223 		return true;
10224 
10225 	return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);
10226 }
10227 
sched_asym(struct sched_domain * sd,int dst_cpu,int src_cpu)10228 static inline bool sched_asym(struct sched_domain *sd, int dst_cpu, int src_cpu)
10229 {
10230 	/*
10231 	 * First check if @dst_cpu can do asym_packing load balance. Only do it
10232 	 * if it has higher priority than @src_cpu.
10233 	 */
10234 	return sched_use_asym_prio(sd, dst_cpu) &&
10235 		sched_asym_prefer(dst_cpu, src_cpu);
10236 }
10237 
10238 /**
10239  * sched_group_asym - Check if the destination CPU can do asym_packing balance
10240  * @env:	The load balancing environment
10241  * @sgs:	Load-balancing statistics of the candidate busiest group
10242  * @group:	The candidate busiest group
10243  *
10244  * @env::dst_cpu can do asym_packing if it has higher priority than the
10245  * preferred CPU of @group.
10246  *
10247  * Return: true if @env::dst_cpu can do with asym_packing load balance. False
10248  * otherwise.
10249  */
10250 static inline bool
sched_group_asym(struct lb_env * env,struct sg_lb_stats * sgs,struct sched_group * group)10251 sched_group_asym(struct lb_env *env, struct sg_lb_stats *sgs, struct sched_group *group)
10252 {
10253 	/*
10254 	 * CPU priorities do not make sense for SMT cores with more than one
10255 	 * busy sibling.
10256 	 */
10257 	if ((group->flags & SD_SHARE_CPUCAPACITY) &&
10258 	    (sgs->group_weight - sgs->idle_cpus != 1))
10259 		return false;
10260 
10261 	return sched_asym(env->sd, env->dst_cpu, group->asym_prefer_cpu);
10262 }
10263 
10264 /* One group has more than one SMT CPU while the other group does not */
smt_vs_nonsmt_groups(struct sched_group * sg1,struct sched_group * sg2)10265 static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1,
10266 				    struct sched_group *sg2)
10267 {
10268 	if (!sg1 || !sg2)
10269 		return false;
10270 
10271 	return (sg1->flags & SD_SHARE_CPUCAPACITY) !=
10272 		(sg2->flags & SD_SHARE_CPUCAPACITY);
10273 }
10274 
smt_balance(struct lb_env * env,struct sg_lb_stats * sgs,struct sched_group * group)10275 static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs,
10276 			       struct sched_group *group)
10277 {
10278 	if (!env->idle)
10279 		return false;
10280 
10281 	/*
10282 	 * For SMT source group, it is better to move a task
10283 	 * to a CPU that doesn't have multiple tasks sharing its CPU capacity.
10284 	 * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY
10285 	 * will not be on.
10286 	 */
10287 	if (group->flags & SD_SHARE_CPUCAPACITY &&
10288 	    sgs->sum_h_nr_running > 1)
10289 		return true;
10290 
10291 	return false;
10292 }
10293 
sibling_imbalance(struct lb_env * env,struct sd_lb_stats * sds,struct sg_lb_stats * busiest,struct sg_lb_stats * local)10294 static inline long sibling_imbalance(struct lb_env *env,
10295 				    struct sd_lb_stats *sds,
10296 				    struct sg_lb_stats *busiest,
10297 				    struct sg_lb_stats *local)
10298 {
10299 	int ncores_busiest, ncores_local;
10300 	long imbalance;
10301 
10302 	if (!env->idle || !busiest->sum_nr_running)
10303 		return 0;
10304 
10305 	ncores_busiest = sds->busiest->cores;
10306 	ncores_local = sds->local->cores;
10307 
10308 	if (ncores_busiest == ncores_local) {
10309 		imbalance = busiest->sum_nr_running;
10310 		lsub_positive(&imbalance, local->sum_nr_running);
10311 		return imbalance;
10312 	}
10313 
10314 	/* Balance such that nr_running/ncores ratio are same on both groups */
10315 	imbalance = ncores_local * busiest->sum_nr_running;
10316 	lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);
10317 	/* Normalize imbalance and do rounding on normalization */
10318 	imbalance = 2 * imbalance + ncores_local + ncores_busiest;
10319 	imbalance /= ncores_local + ncores_busiest;
10320 
10321 	/* Take advantage of resource in an empty sched group */
10322 	if (imbalance <= 1 && local->sum_nr_running == 0 &&
10323 	    busiest->sum_nr_running > 1)
10324 		imbalance = 2;
10325 
10326 	return imbalance;
10327 }
10328 
10329 static inline bool
sched_reduced_capacity(struct rq * rq,struct sched_domain * sd)10330 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
10331 {
10332 	/*
10333 	 * When there is more than 1 task, the group_overloaded case already
10334 	 * takes care of cpu with reduced capacity
10335 	 */
10336 	if (rq->cfs.h_nr_runnable != 1)
10337 		return false;
10338 
10339 	return check_cpu_capacity(rq, sd);
10340 }
10341 
10342 /**
10343  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
10344  * @env: The load balancing environment.
10345  * @sds: Load-balancing data with statistics of the local group.
10346  * @group: sched_group whose statistics are to be updated.
10347  * @sgs: variable to hold the statistics for this group.
10348  * @sg_overloaded: sched_group is overloaded
10349  * @sg_overutilized: sched_group is overutilized
10350  */
update_sg_lb_stats(struct lb_env * env,struct sd_lb_stats * sds,struct sched_group * group,struct sg_lb_stats * sgs,bool * sg_overloaded,bool * sg_overutilized)10351 static inline void update_sg_lb_stats(struct lb_env *env,
10352 				      struct sd_lb_stats *sds,
10353 				      struct sched_group *group,
10354 				      struct sg_lb_stats *sgs,
10355 				      bool *sg_overloaded,
10356 				      bool *sg_overutilized)
10357 {
10358 	int i, nr_running, local_group, sd_flags = env->sd->flags;
10359 	bool balancing_at_rd = !env->sd->parent;
10360 
10361 	memset(sgs, 0, sizeof(*sgs));
10362 
10363 	local_group = group == sds->local;
10364 
10365 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10366 		struct rq *rq = cpu_rq(i);
10367 		unsigned long load = cpu_load(rq);
10368 
10369 		sgs->group_load += load;
10370 		sgs->group_util += cpu_util_cfs(i);
10371 		sgs->group_runnable += cpu_runnable(rq);
10372 		sgs->sum_h_nr_running += rq->cfs.h_nr_runnable;
10373 
10374 		nr_running = rq->nr_running;
10375 		sgs->sum_nr_running += nr_running;
10376 
10377 		if (cpu_overutilized(i))
10378 			*sg_overutilized = 1;
10379 
10380 		/*
10381 		 * No need to call idle_cpu() if nr_running is not 0
10382 		 */
10383 		if (!nr_running && idle_cpu(i)) {
10384 			sgs->idle_cpus++;
10385 			/* Idle cpu can't have misfit task */
10386 			continue;
10387 		}
10388 
10389 		/* Overload indicator is only updated at root domain */
10390 		if (balancing_at_rd && nr_running > 1)
10391 			*sg_overloaded = 1;
10392 
10393 #ifdef CONFIG_NUMA_BALANCING
10394 		/* Only fbq_classify_group() uses this to classify NUMA groups */
10395 		if (sd_flags & SD_NUMA) {
10396 			sgs->nr_numa_running += rq->nr_numa_running;
10397 			sgs->nr_preferred_running += rq->nr_preferred_running;
10398 		}
10399 #endif
10400 		if (local_group)
10401 			continue;
10402 
10403 		if (sd_flags & SD_ASYM_CPUCAPACITY) {
10404 			/* Check for a misfit task on the cpu */
10405 			if (sgs->group_misfit_task_load < rq->misfit_task_load) {
10406 				sgs->group_misfit_task_load = rq->misfit_task_load;
10407 				*sg_overloaded = 1;
10408 			}
10409 		} else if (env->idle && sched_reduced_capacity(rq, env->sd)) {
10410 			/* Check for a task running on a CPU with reduced capacity */
10411 			if (sgs->group_misfit_task_load < load)
10412 				sgs->group_misfit_task_load = load;
10413 		}
10414 	}
10415 
10416 	sgs->group_capacity = group->sgc->capacity;
10417 
10418 	sgs->group_weight = group->group_weight;
10419 
10420 	/* Check if dst CPU is idle and preferred to this group */
10421 	if (!local_group && env->idle && sgs->sum_h_nr_running &&
10422 	    sched_group_asym(env, sgs, group))
10423 		sgs->group_asym_packing = 1;
10424 
10425 	/* Check for loaded SMT group to be balanced to dst CPU */
10426 	if (!local_group && smt_balance(env, sgs, group))
10427 		sgs->group_smt_balance = 1;
10428 
10429 	sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
10430 
10431 	/* Computing avg_load makes sense only when group is overloaded */
10432 	if (sgs->group_type == group_overloaded)
10433 		sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10434 				sgs->group_capacity;
10435 }
10436 
10437 /**
10438  * update_sd_pick_busiest - return 1 on busiest group
10439  * @env: The load balancing environment.
10440  * @sds: sched_domain statistics
10441  * @sg: sched_group candidate to be checked for being the busiest
10442  * @sgs: sched_group statistics
10443  *
10444  * Determine if @sg is a busier group than the previously selected
10445  * busiest group.
10446  *
10447  * Return: %true if @sg is a busier group than the previously selected
10448  * busiest group. %false otherwise.
10449  */
update_sd_pick_busiest(struct lb_env * env,struct sd_lb_stats * sds,struct sched_group * sg,struct sg_lb_stats * sgs)10450 static bool update_sd_pick_busiest(struct lb_env *env,
10451 				   struct sd_lb_stats *sds,
10452 				   struct sched_group *sg,
10453 				   struct sg_lb_stats *sgs)
10454 {
10455 	struct sg_lb_stats *busiest = &sds->busiest_stat;
10456 
10457 	/* Make sure that there is at least one task to pull */
10458 	if (!sgs->sum_h_nr_running)
10459 		return false;
10460 
10461 	/*
10462 	 * Don't try to pull misfit tasks we can't help.
10463 	 * We can use max_capacity here as reduction in capacity on some
10464 	 * CPUs in the group should either be possible to resolve
10465 	 * internally or be covered by avg_load imbalance (eventually).
10466 	 */
10467 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
10468 	    (sgs->group_type == group_misfit_task) &&
10469 	    (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
10470 	     sds->local_stat.group_type != group_has_spare))
10471 		return false;
10472 
10473 	if (sgs->group_type > busiest->group_type)
10474 		return true;
10475 
10476 	if (sgs->group_type < busiest->group_type)
10477 		return false;
10478 
10479 	/*
10480 	 * The candidate and the current busiest group are the same type of
10481 	 * group. Let check which one is the busiest according to the type.
10482 	 */
10483 
10484 	switch (sgs->group_type) {
10485 	case group_overloaded:
10486 		/* Select the overloaded group with highest avg_load. */
10487 		return sgs->avg_load > busiest->avg_load;
10488 
10489 	case group_imbalanced:
10490 		/*
10491 		 * Select the 1st imbalanced group as we don't have any way to
10492 		 * choose one more than another.
10493 		 */
10494 		return false;
10495 
10496 	case group_asym_packing:
10497 		/* Prefer to move from lowest priority CPU's work */
10498 		return sched_asym_prefer(sds->busiest->asym_prefer_cpu, sg->asym_prefer_cpu);
10499 
10500 	case group_misfit_task:
10501 		/*
10502 		 * If we have more than one misfit sg go with the biggest
10503 		 * misfit.
10504 		 */
10505 		return sgs->group_misfit_task_load > busiest->group_misfit_task_load;
10506 
10507 	case group_smt_balance:
10508 		/*
10509 		 * Check if we have spare CPUs on either SMT group to
10510 		 * choose has spare or fully busy handling.
10511 		 */
10512 		if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)
10513 			goto has_spare;
10514 
10515 		fallthrough;
10516 
10517 	case group_fully_busy:
10518 		/*
10519 		 * Select the fully busy group with highest avg_load. In
10520 		 * theory, there is no need to pull task from such kind of
10521 		 * group because tasks have all compute capacity that they need
10522 		 * but we can still improve the overall throughput by reducing
10523 		 * contention when accessing shared HW resources.
10524 		 *
10525 		 * XXX for now avg_load is not computed and always 0 so we
10526 		 * select the 1st one, except if @sg is composed of SMT
10527 		 * siblings.
10528 		 */
10529 
10530 		if (sgs->avg_load < busiest->avg_load)
10531 			return false;
10532 
10533 		if (sgs->avg_load == busiest->avg_load) {
10534 			/*
10535 			 * SMT sched groups need more help than non-SMT groups.
10536 			 * If @sg happens to also be SMT, either choice is good.
10537 			 */
10538 			if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)
10539 				return false;
10540 		}
10541 
10542 		break;
10543 
10544 	case group_has_spare:
10545 		/*
10546 		 * Do not pick sg with SMT CPUs over sg with pure CPUs,
10547 		 * as we do not want to pull task off SMT core with one task
10548 		 * and make the core idle.
10549 		 */
10550 		if (smt_vs_nonsmt_groups(sds->busiest, sg)) {
10551 			if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)
10552 				return false;
10553 			else
10554 				return true;
10555 		}
10556 has_spare:
10557 
10558 		/*
10559 		 * Select not overloaded group with lowest number of idle CPUs
10560 		 * and highest number of running tasks. We could also compare
10561 		 * the spare capacity which is more stable but it can end up
10562 		 * that the group has less spare capacity but finally more idle
10563 		 * CPUs which means less opportunity to pull tasks.
10564 		 */
10565 		if (sgs->idle_cpus > busiest->idle_cpus)
10566 			return false;
10567 		else if ((sgs->idle_cpus == busiest->idle_cpus) &&
10568 			 (sgs->sum_nr_running <= busiest->sum_nr_running))
10569 			return false;
10570 
10571 		break;
10572 	}
10573 
10574 	/*
10575 	 * Candidate sg has no more than one task per CPU and has higher
10576 	 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
10577 	 * throughput. Maximize throughput, power/energy consequences are not
10578 	 * considered.
10579 	 */
10580 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
10581 	    (sgs->group_type <= group_fully_busy) &&
10582 	    (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
10583 		return false;
10584 
10585 	return true;
10586 }
10587 
10588 #ifdef CONFIG_NUMA_BALANCING
fbq_classify_group(struct sg_lb_stats * sgs)10589 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10590 {
10591 	if (sgs->sum_h_nr_running > sgs->nr_numa_running)
10592 		return regular;
10593 	if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
10594 		return remote;
10595 	return all;
10596 }
10597 
fbq_classify_rq(struct rq * rq)10598 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10599 {
10600 	if (rq->nr_running > rq->nr_numa_running)
10601 		return regular;
10602 	if (rq->nr_running > rq->nr_preferred_running)
10603 		return remote;
10604 	return all;
10605 }
10606 #else
fbq_classify_group(struct sg_lb_stats * sgs)10607 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10608 {
10609 	return all;
10610 }
10611 
fbq_classify_rq(struct rq * rq)10612 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10613 {
10614 	return regular;
10615 }
10616 #endif /* CONFIG_NUMA_BALANCING */
10617 
10618 
10619 struct sg_lb_stats;
10620 
10621 /*
10622  * task_running_on_cpu - return 1 if @p is running on @cpu.
10623  */
10624 
task_running_on_cpu(int cpu,struct task_struct * p)10625 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
10626 {
10627 	/* Task has no contribution or is new */
10628 	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
10629 		return 0;
10630 
10631 	if (task_on_rq_queued(p))
10632 		return 1;
10633 
10634 	return 0;
10635 }
10636 
10637 /**
10638  * idle_cpu_without - would a given CPU be idle without p ?
10639  * @cpu: the processor on which idleness is tested.
10640  * @p: task which should be ignored.
10641  *
10642  * Return: 1 if the CPU would be idle. 0 otherwise.
10643  */
idle_cpu_without(int cpu,struct task_struct * p)10644 static int idle_cpu_without(int cpu, struct task_struct *p)
10645 {
10646 	struct rq *rq = cpu_rq(cpu);
10647 
10648 	if (rq->curr != rq->idle && rq->curr != p)
10649 		return 0;
10650 
10651 	/*
10652 	 * rq->nr_running can't be used but an updated version without the
10653 	 * impact of p on cpu must be used instead. The updated nr_running
10654 	 * be computed and tested before calling idle_cpu_without().
10655 	 */
10656 
10657 	if (rq->ttwu_pending)
10658 		return 0;
10659 
10660 	return 1;
10661 }
10662 
10663 /*
10664  * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
10665  * @sd: The sched_domain level to look for idlest group.
10666  * @group: sched_group whose statistics are to be updated.
10667  * @sgs: variable to hold the statistics for this group.
10668  * @p: The task for which we look for the idlest group/CPU.
10669  */
update_sg_wakeup_stats(struct sched_domain * sd,struct sched_group * group,struct sg_lb_stats * sgs,struct task_struct * p)10670 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
10671 					  struct sched_group *group,
10672 					  struct sg_lb_stats *sgs,
10673 					  struct task_struct *p)
10674 {
10675 	int i, nr_running;
10676 
10677 	memset(sgs, 0, sizeof(*sgs));
10678 
10679 	/* Assume that task can't fit any CPU of the group */
10680 	if (sd->flags & SD_ASYM_CPUCAPACITY)
10681 		sgs->group_misfit_task_load = 1;
10682 
10683 	for_each_cpu(i, sched_group_span(group)) {
10684 		struct rq *rq = cpu_rq(i);
10685 		unsigned int local;
10686 
10687 		sgs->group_load += cpu_load_without(rq, p);
10688 		sgs->group_util += cpu_util_without(i, p);
10689 		sgs->group_runnable += cpu_runnable_without(rq, p);
10690 		local = task_running_on_cpu(i, p);
10691 		sgs->sum_h_nr_running += rq->cfs.h_nr_runnable - local;
10692 
10693 		nr_running = rq->nr_running - local;
10694 		sgs->sum_nr_running += nr_running;
10695 
10696 		/*
10697 		 * No need to call idle_cpu_without() if nr_running is not 0
10698 		 */
10699 		if (!nr_running && idle_cpu_without(i, p))
10700 			sgs->idle_cpus++;
10701 
10702 		/* Check if task fits in the CPU */
10703 		if (sd->flags & SD_ASYM_CPUCAPACITY &&
10704 		    sgs->group_misfit_task_load &&
10705 		    task_fits_cpu(p, i))
10706 			sgs->group_misfit_task_load = 0;
10707 
10708 	}
10709 
10710 	sgs->group_capacity = group->sgc->capacity;
10711 
10712 	sgs->group_weight = group->group_weight;
10713 
10714 	sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
10715 
10716 	/*
10717 	 * Computing avg_load makes sense only when group is fully busy or
10718 	 * overloaded
10719 	 */
10720 	if (sgs->group_type == group_fully_busy ||
10721 		sgs->group_type == group_overloaded)
10722 		sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10723 				sgs->group_capacity;
10724 }
10725 
update_pick_idlest(struct sched_group * idlest,struct sg_lb_stats * idlest_sgs,struct sched_group * group,struct sg_lb_stats * sgs)10726 static bool update_pick_idlest(struct sched_group *idlest,
10727 			       struct sg_lb_stats *idlest_sgs,
10728 			       struct sched_group *group,
10729 			       struct sg_lb_stats *sgs)
10730 {
10731 	if (sgs->group_type < idlest_sgs->group_type)
10732 		return true;
10733 
10734 	if (sgs->group_type > idlest_sgs->group_type)
10735 		return false;
10736 
10737 	/*
10738 	 * The candidate and the current idlest group are the same type of
10739 	 * group. Let check which one is the idlest according to the type.
10740 	 */
10741 
10742 	switch (sgs->group_type) {
10743 	case group_overloaded:
10744 	case group_fully_busy:
10745 		/* Select the group with lowest avg_load. */
10746 		if (idlest_sgs->avg_load <= sgs->avg_load)
10747 			return false;
10748 		break;
10749 
10750 	case group_imbalanced:
10751 	case group_asym_packing:
10752 	case group_smt_balance:
10753 		/* Those types are not used in the slow wakeup path */
10754 		return false;
10755 
10756 	case group_misfit_task:
10757 		/* Select group with the highest max capacity */
10758 		if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
10759 			return false;
10760 		break;
10761 
10762 	case group_has_spare:
10763 		/* Select group with most idle CPUs */
10764 		if (idlest_sgs->idle_cpus > sgs->idle_cpus)
10765 			return false;
10766 
10767 		/* Select group with lowest group_util */
10768 		if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
10769 			idlest_sgs->group_util <= sgs->group_util)
10770 			return false;
10771 
10772 		break;
10773 	}
10774 
10775 	return true;
10776 }
10777 
10778 /*
10779  * sched_balance_find_dst_group() finds and returns the least busy CPU group within the
10780  * domain.
10781  *
10782  * Assumes p is allowed on at least one CPU in sd.
10783  */
10784 static struct sched_group *
sched_balance_find_dst_group(struct sched_domain * sd,struct task_struct * p,int this_cpu)10785 sched_balance_find_dst_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
10786 {
10787 	struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
10788 	struct sg_lb_stats local_sgs, tmp_sgs;
10789 	struct sg_lb_stats *sgs;
10790 	unsigned long imbalance;
10791 	struct sg_lb_stats idlest_sgs = {
10792 			.avg_load = UINT_MAX,
10793 			.group_type = group_overloaded,
10794 	};
10795 
10796 	do {
10797 		int local_group;
10798 
10799 		/* Skip over this group if it has no CPUs allowed */
10800 		if (!cpumask_intersects(sched_group_span(group),
10801 					p->cpus_ptr))
10802 			continue;
10803 
10804 		/* Skip over this group if no cookie matched */
10805 		if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
10806 			continue;
10807 
10808 		local_group = cpumask_test_cpu(this_cpu,
10809 					       sched_group_span(group));
10810 
10811 		if (local_group) {
10812 			sgs = &local_sgs;
10813 			local = group;
10814 		} else {
10815 			sgs = &tmp_sgs;
10816 		}
10817 
10818 		update_sg_wakeup_stats(sd, group, sgs, p);
10819 
10820 		if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
10821 			idlest = group;
10822 			idlest_sgs = *sgs;
10823 		}
10824 
10825 	} while (group = group->next, group != sd->groups);
10826 
10827 
10828 	/* There is no idlest group to push tasks to */
10829 	if (!idlest)
10830 		return NULL;
10831 
10832 	/* The local group has been skipped because of CPU affinity */
10833 	if (!local)
10834 		return idlest;
10835 
10836 	/*
10837 	 * If the local group is idler than the selected idlest group
10838 	 * don't try and push the task.
10839 	 */
10840 	if (local_sgs.group_type < idlest_sgs.group_type)
10841 		return NULL;
10842 
10843 	/*
10844 	 * If the local group is busier than the selected idlest group
10845 	 * try and push the task.
10846 	 */
10847 	if (local_sgs.group_type > idlest_sgs.group_type)
10848 		return idlest;
10849 
10850 	switch (local_sgs.group_type) {
10851 	case group_overloaded:
10852 	case group_fully_busy:
10853 
10854 		/* Calculate allowed imbalance based on load */
10855 		imbalance = scale_load_down(NICE_0_LOAD) *
10856 				(sd->imbalance_pct-100) / 100;
10857 
10858 		/*
10859 		 * When comparing groups across NUMA domains, it's possible for
10860 		 * the local domain to be very lightly loaded relative to the
10861 		 * remote domains but "imbalance" skews the comparison making
10862 		 * remote CPUs look much more favourable. When considering
10863 		 * cross-domain, add imbalance to the load on the remote node
10864 		 * and consider staying local.
10865 		 */
10866 
10867 		if ((sd->flags & SD_NUMA) &&
10868 		    ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
10869 			return NULL;
10870 
10871 		/*
10872 		 * If the local group is less loaded than the selected
10873 		 * idlest group don't try and push any tasks.
10874 		 */
10875 		if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
10876 			return NULL;
10877 
10878 		if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
10879 			return NULL;
10880 		break;
10881 
10882 	case group_imbalanced:
10883 	case group_asym_packing:
10884 	case group_smt_balance:
10885 		/* Those type are not used in the slow wakeup path */
10886 		return NULL;
10887 
10888 	case group_misfit_task:
10889 		/* Select group with the highest max capacity */
10890 		if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
10891 			return NULL;
10892 		break;
10893 
10894 	case group_has_spare:
10895 #ifdef CONFIG_NUMA
10896 		if (sd->flags & SD_NUMA) {
10897 			int imb_numa_nr = sd->imb_numa_nr;
10898 #ifdef CONFIG_NUMA_BALANCING
10899 			int idlest_cpu;
10900 			/*
10901 			 * If there is spare capacity at NUMA, try to select
10902 			 * the preferred node
10903 			 */
10904 			if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
10905 				return NULL;
10906 
10907 			idlest_cpu = cpumask_first(sched_group_span(idlest));
10908 			if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
10909 				return idlest;
10910 #endif /* CONFIG_NUMA_BALANCING */
10911 			/*
10912 			 * Otherwise, keep the task close to the wakeup source
10913 			 * and improve locality if the number of running tasks
10914 			 * would remain below threshold where an imbalance is
10915 			 * allowed while accounting for the possibility the
10916 			 * task is pinned to a subset of CPUs. If there is a
10917 			 * real need of migration, periodic load balance will
10918 			 * take care of it.
10919 			 */
10920 			if (p->nr_cpus_allowed != NR_CPUS) {
10921 				struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
10922 
10923 				cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
10924 				imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
10925 			}
10926 
10927 			imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
10928 			if (!adjust_numa_imbalance(imbalance,
10929 						   local_sgs.sum_nr_running + 1,
10930 						   imb_numa_nr)) {
10931 				return NULL;
10932 			}
10933 		}
10934 #endif /* CONFIG_NUMA */
10935 
10936 		/*
10937 		 * Select group with highest number of idle CPUs. We could also
10938 		 * compare the utilization which is more stable but it can end
10939 		 * up that the group has less spare capacity but finally more
10940 		 * idle CPUs which means more opportunity to run task.
10941 		 */
10942 		if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
10943 			return NULL;
10944 		break;
10945 	}
10946 
10947 	return idlest;
10948 }
10949 
update_idle_cpu_scan(struct lb_env * env,unsigned long sum_util)10950 static void update_idle_cpu_scan(struct lb_env *env,
10951 				 unsigned long sum_util)
10952 {
10953 	struct sched_domain_shared *sd_share;
10954 	int llc_weight, pct;
10955 	u64 x, y, tmp;
10956 	/*
10957 	 * Update the number of CPUs to scan in LLC domain, which could
10958 	 * be used as a hint in select_idle_cpu(). The update of sd_share
10959 	 * could be expensive because it is within a shared cache line.
10960 	 * So the write of this hint only occurs during periodic load
10961 	 * balancing, rather than CPU_NEWLY_IDLE, because the latter
10962 	 * can fire way more frequently than the former.
10963 	 */
10964 	if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
10965 		return;
10966 
10967 	llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
10968 	if (env->sd->span_weight != llc_weight)
10969 		return;
10970 
10971 	sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
10972 	if (!sd_share)
10973 		return;
10974 
10975 	/*
10976 	 * The number of CPUs to search drops as sum_util increases, when
10977 	 * sum_util hits 85% or above, the scan stops.
10978 	 * The reason to choose 85% as the threshold is because this is the
10979 	 * imbalance_pct(117) when a LLC sched group is overloaded.
10980 	 *
10981 	 * let y = SCHED_CAPACITY_SCALE - p * x^2                       [1]
10982 	 * and y'= y / SCHED_CAPACITY_SCALE
10983 	 *
10984 	 * x is the ratio of sum_util compared to the CPU capacity:
10985 	 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
10986 	 * y' is the ratio of CPUs to be scanned in the LLC domain,
10987 	 * and the number of CPUs to scan is calculated by:
10988 	 *
10989 	 * nr_scan = llc_weight * y'                                    [2]
10990 	 *
10991 	 * When x hits the threshold of overloaded, AKA, when
10992 	 * x = 100 / pct, y drops to 0. According to [1],
10993 	 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
10994 	 *
10995 	 * Scale x by SCHED_CAPACITY_SCALE:
10996 	 * x' = sum_util / llc_weight;                                  [3]
10997 	 *
10998 	 * and finally [1] becomes:
10999 	 * y = SCHED_CAPACITY_SCALE -
11000 	 *     x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE)            [4]
11001 	 *
11002 	 */
11003 	/* equation [3] */
11004 	x = sum_util;
11005 	do_div(x, llc_weight);
11006 
11007 	/* equation [4] */
11008 	pct = env->sd->imbalance_pct;
11009 	tmp = x * x * pct * pct;
11010 	do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
11011 	tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
11012 	y = SCHED_CAPACITY_SCALE - tmp;
11013 
11014 	/* equation [2] */
11015 	y *= llc_weight;
11016 	do_div(y, SCHED_CAPACITY_SCALE);
11017 	if ((int)y != sd_share->nr_idle_scan)
11018 		WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
11019 }
11020 
11021 /**
11022  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
11023  * @env: The load balancing environment.
11024  * @sds: variable to hold the statistics for this sched_domain.
11025  */
11026 
update_sd_lb_stats(struct lb_env * env,struct sd_lb_stats * sds)11027 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
11028 {
11029 	struct sched_group *sg = env->sd->groups;
11030 	struct sg_lb_stats *local = &sds->local_stat;
11031 	struct sg_lb_stats tmp_sgs;
11032 	unsigned long sum_util = 0;
11033 	bool sg_overloaded = 0, sg_overutilized = 0;
11034 
11035 	do {
11036 		struct sg_lb_stats *sgs = &tmp_sgs;
11037 		int local_group;
11038 
11039 		local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
11040 		if (local_group) {
11041 			sds->local = sg;
11042 			sgs = local;
11043 
11044 			if (env->idle != CPU_NEWLY_IDLE ||
11045 			    time_after_eq(jiffies, sg->sgc->next_update))
11046 				update_group_capacity(env->sd, env->dst_cpu);
11047 		}
11048 
11049 		update_sg_lb_stats(env, sds, sg, sgs, &sg_overloaded, &sg_overutilized);
11050 
11051 		if (!local_group && update_sd_pick_busiest(env, sds, sg, sgs)) {
11052 			sds->busiest = sg;
11053 			sds->busiest_stat = *sgs;
11054 		}
11055 
11056 		/* Now, start updating sd_lb_stats */
11057 		sds->total_load += sgs->group_load;
11058 		sds->total_capacity += sgs->group_capacity;
11059 
11060 		sum_util += sgs->group_util;
11061 		sg = sg->next;
11062 	} while (sg != env->sd->groups);
11063 
11064 	/*
11065 	 * Indicate that the child domain of the busiest group prefers tasks
11066 	 * go to a child's sibling domains first. NB the flags of a sched group
11067 	 * are those of the child domain.
11068 	 */
11069 	if (sds->busiest)
11070 		sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);
11071 
11072 
11073 	if (env->sd->flags & SD_NUMA)
11074 		env->fbq_type = fbq_classify_group(&sds->busiest_stat);
11075 
11076 	if (!env->sd->parent) {
11077 		/* update overload indicator if we are at root domain */
11078 		set_rd_overloaded(env->dst_rq->rd, sg_overloaded);
11079 
11080 		/* Update over-utilization (tipping point, U >= 0) indicator */
11081 		set_rd_overutilized(env->dst_rq->rd, sg_overutilized);
11082 	} else if (sg_overutilized) {
11083 		set_rd_overutilized(env->dst_rq->rd, sg_overutilized);
11084 	}
11085 
11086 	update_idle_cpu_scan(env, sum_util);
11087 }
11088 
11089 /**
11090  * calculate_imbalance - Calculate the amount of imbalance present within the
11091  *			 groups of a given sched_domain during load balance.
11092  * @env: load balance environment
11093  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
11094  */
calculate_imbalance(struct lb_env * env,struct sd_lb_stats * sds)11095 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
11096 {
11097 	struct sg_lb_stats *local, *busiest;
11098 
11099 	local = &sds->local_stat;
11100 	busiest = &sds->busiest_stat;
11101 
11102 	if (busiest->group_type == group_misfit_task) {
11103 		if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
11104 			/* Set imbalance to allow misfit tasks to be balanced. */
11105 			env->migration_type = migrate_misfit;
11106 			env->imbalance = 1;
11107 		} else {
11108 			/*
11109 			 * Set load imbalance to allow moving task from cpu
11110 			 * with reduced capacity.
11111 			 */
11112 			env->migration_type = migrate_load;
11113 			env->imbalance = busiest->group_misfit_task_load;
11114 		}
11115 		return;
11116 	}
11117 
11118 	if (busiest->group_type == group_asym_packing) {
11119 		/*
11120 		 * In case of asym capacity, we will try to migrate all load to
11121 		 * the preferred CPU.
11122 		 */
11123 		env->migration_type = migrate_task;
11124 		env->imbalance = busiest->sum_h_nr_running;
11125 		return;
11126 	}
11127 
11128 	if (busiest->group_type == group_smt_balance) {
11129 		/* Reduce number of tasks sharing CPU capacity */
11130 		env->migration_type = migrate_task;
11131 		env->imbalance = 1;
11132 		return;
11133 	}
11134 
11135 	if (busiest->group_type == group_imbalanced) {
11136 		/*
11137 		 * In the group_imb case we cannot rely on group-wide averages
11138 		 * to ensure CPU-load equilibrium, try to move any task to fix
11139 		 * the imbalance. The next load balance will take care of
11140 		 * balancing back the system.
11141 		 */
11142 		env->migration_type = migrate_task;
11143 		env->imbalance = 1;
11144 		return;
11145 	}
11146 
11147 	/*
11148 	 * Try to use spare capacity of local group without overloading it or
11149 	 * emptying busiest.
11150 	 */
11151 	if (local->group_type == group_has_spare) {
11152 		if ((busiest->group_type > group_fully_busy) &&
11153 		    !(env->sd->flags & SD_SHARE_LLC)) {
11154 			/*
11155 			 * If busiest is overloaded, try to fill spare
11156 			 * capacity. This might end up creating spare capacity
11157 			 * in busiest or busiest still being overloaded but
11158 			 * there is no simple way to directly compute the
11159 			 * amount of load to migrate in order to balance the
11160 			 * system.
11161 			 */
11162 			env->migration_type = migrate_util;
11163 			env->imbalance = max(local->group_capacity, local->group_util) -
11164 					 local->group_util;
11165 
11166 			/*
11167 			 * In some cases, the group's utilization is max or even
11168 			 * higher than capacity because of migrations but the
11169 			 * local CPU is (newly) idle. There is at least one
11170 			 * waiting task in this overloaded busiest group. Let's
11171 			 * try to pull it.
11172 			 */
11173 			if (env->idle && env->imbalance == 0) {
11174 				env->migration_type = migrate_task;
11175 				env->imbalance = 1;
11176 			}
11177 
11178 			return;
11179 		}
11180 
11181 		if (busiest->group_weight == 1 || sds->prefer_sibling) {
11182 			/*
11183 			 * When prefer sibling, evenly spread running tasks on
11184 			 * groups.
11185 			 */
11186 			env->migration_type = migrate_task;
11187 			env->imbalance = sibling_imbalance(env, sds, busiest, local);
11188 		} else {
11189 
11190 			/*
11191 			 * If there is no overload, we just want to even the number of
11192 			 * idle CPUs.
11193 			 */
11194 			env->migration_type = migrate_task;
11195 			env->imbalance = max_t(long, 0,
11196 					       (local->idle_cpus - busiest->idle_cpus));
11197 		}
11198 
11199 #ifdef CONFIG_NUMA
11200 		/* Consider allowing a small imbalance between NUMA groups */
11201 		if (env->sd->flags & SD_NUMA) {
11202 			env->imbalance = adjust_numa_imbalance(env->imbalance,
11203 							       local->sum_nr_running + 1,
11204 							       env->sd->imb_numa_nr);
11205 		}
11206 #endif
11207 
11208 		/* Number of tasks to move to restore balance */
11209 		env->imbalance >>= 1;
11210 
11211 		return;
11212 	}
11213 
11214 	/*
11215 	 * Local is fully busy but has to take more load to relieve the
11216 	 * busiest group
11217 	 */
11218 	if (local->group_type < group_overloaded) {
11219 		/*
11220 		 * Local will become overloaded so the avg_load metrics are
11221 		 * finally needed.
11222 		 */
11223 
11224 		local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
11225 				  local->group_capacity;
11226 
11227 		/*
11228 		 * If the local group is more loaded than the selected
11229 		 * busiest group don't try to pull any tasks.
11230 		 */
11231 		if (local->avg_load >= busiest->avg_load) {
11232 			env->imbalance = 0;
11233 			return;
11234 		}
11235 
11236 		sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
11237 				sds->total_capacity;
11238 
11239 		/*
11240 		 * If the local group is more loaded than the average system
11241 		 * load, don't try to pull any tasks.
11242 		 */
11243 		if (local->avg_load >= sds->avg_load) {
11244 			env->imbalance = 0;
11245 			return;
11246 		}
11247 
11248 	}
11249 
11250 	/*
11251 	 * Both group are or will become overloaded and we're trying to get all
11252 	 * the CPUs to the average_load, so we don't want to push ourselves
11253 	 * above the average load, nor do we wish to reduce the max loaded CPU
11254 	 * below the average load. At the same time, we also don't want to
11255 	 * reduce the group load below the group capacity. Thus we look for
11256 	 * the minimum possible imbalance.
11257 	 */
11258 	env->migration_type = migrate_load;
11259 	env->imbalance = min(
11260 		(busiest->avg_load - sds->avg_load) * busiest->group_capacity,
11261 		(sds->avg_load - local->avg_load) * local->group_capacity
11262 	) / SCHED_CAPACITY_SCALE;
11263 }
11264 
11265 /******* sched_balance_find_src_group() helpers end here *********************/
11266 
11267 /*
11268  * Decision matrix according to the local and busiest group type:
11269  *
11270  * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
11271  * has_spare        nr_idle   balanced   N/A    N/A  balanced   balanced
11272  * fully_busy       nr_idle   nr_idle    N/A    N/A  balanced   balanced
11273  * misfit_task      force     N/A        N/A    N/A  N/A        N/A
11274  * asym_packing     force     force      N/A    N/A  force      force
11275  * imbalanced       force     force      N/A    N/A  force      force
11276  * overloaded       force     force      N/A    N/A  force      avg_load
11277  *
11278  * N/A :      Not Applicable because already filtered while updating
11279  *            statistics.
11280  * balanced : The system is balanced for these 2 groups.
11281  * force :    Calculate the imbalance as load migration is probably needed.
11282  * avg_load : Only if imbalance is significant enough.
11283  * nr_idle :  dst_cpu is not busy and the number of idle CPUs is quite
11284  *            different in groups.
11285  */
11286 
11287 /**
11288  * sched_balance_find_src_group - Returns the busiest group within the sched_domain
11289  * if there is an imbalance.
11290  * @env: The load balancing environment.
11291  *
11292  * Also calculates the amount of runnable load which should be moved
11293  * to restore balance.
11294  *
11295  * Return:	- The busiest group if imbalance exists.
11296  */
sched_balance_find_src_group(struct lb_env * env)11297 static struct sched_group *sched_balance_find_src_group(struct lb_env *env)
11298 {
11299 	struct sg_lb_stats *local, *busiest;
11300 	struct sd_lb_stats sds;
11301 
11302 	init_sd_lb_stats(&sds);
11303 
11304 	/*
11305 	 * Compute the various statistics relevant for load balancing at
11306 	 * this level.
11307 	 */
11308 	update_sd_lb_stats(env, &sds);
11309 
11310 	/* There is no busy sibling group to pull tasks from */
11311 	if (!sds.busiest)
11312 		goto out_balanced;
11313 
11314 	busiest = &sds.busiest_stat;
11315 
11316 	/* Misfit tasks should be dealt with regardless of the avg load */
11317 	if (busiest->group_type == group_misfit_task)
11318 		goto force_balance;
11319 
11320 	if (!is_rd_overutilized(env->dst_rq->rd) &&
11321 	    rcu_dereference(env->dst_rq->rd->pd))
11322 		goto out_balanced;
11323 
11324 	/* ASYM feature bypasses nice load balance check */
11325 	if (busiest->group_type == group_asym_packing)
11326 		goto force_balance;
11327 
11328 	/*
11329 	 * If the busiest group is imbalanced the below checks don't
11330 	 * work because they assume all things are equal, which typically
11331 	 * isn't true due to cpus_ptr constraints and the like.
11332 	 */
11333 	if (busiest->group_type == group_imbalanced)
11334 		goto force_balance;
11335 
11336 	local = &sds.local_stat;
11337 	/*
11338 	 * If the local group is busier than the selected busiest group
11339 	 * don't try and pull any tasks.
11340 	 */
11341 	if (local->group_type > busiest->group_type)
11342 		goto out_balanced;
11343 
11344 	/*
11345 	 * When groups are overloaded, use the avg_load to ensure fairness
11346 	 * between tasks.
11347 	 */
11348 	if (local->group_type == group_overloaded) {
11349 		/*
11350 		 * If the local group is more loaded than the selected
11351 		 * busiest group don't try to pull any tasks.
11352 		 */
11353 		if (local->avg_load >= busiest->avg_load)
11354 			goto out_balanced;
11355 
11356 		/* XXX broken for overlapping NUMA groups */
11357 		sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
11358 				sds.total_capacity;
11359 
11360 		/*
11361 		 * Don't pull any tasks if this group is already above the
11362 		 * domain average load.
11363 		 */
11364 		if (local->avg_load >= sds.avg_load)
11365 			goto out_balanced;
11366 
11367 		/*
11368 		 * If the busiest group is more loaded, use imbalance_pct to be
11369 		 * conservative.
11370 		 */
11371 		if (100 * busiest->avg_load <=
11372 				env->sd->imbalance_pct * local->avg_load)
11373 			goto out_balanced;
11374 	}
11375 
11376 	/*
11377 	 * Try to move all excess tasks to a sibling domain of the busiest
11378 	 * group's child domain.
11379 	 */
11380 	if (sds.prefer_sibling && local->group_type == group_has_spare &&
11381 	    sibling_imbalance(env, &sds, busiest, local) > 1)
11382 		goto force_balance;
11383 
11384 	if (busiest->group_type != group_overloaded) {
11385 		if (!env->idle) {
11386 			/*
11387 			 * If the busiest group is not overloaded (and as a
11388 			 * result the local one too) but this CPU is already
11389 			 * busy, let another idle CPU try to pull task.
11390 			 */
11391 			goto out_balanced;
11392 		}
11393 
11394 		if (busiest->group_type == group_smt_balance &&
11395 		    smt_vs_nonsmt_groups(sds.local, sds.busiest)) {
11396 			/* Let non SMT CPU pull from SMT CPU sharing with sibling */
11397 			goto force_balance;
11398 		}
11399 
11400 		if (busiest->group_weight > 1 &&
11401 		    local->idle_cpus <= (busiest->idle_cpus + 1)) {
11402 			/*
11403 			 * If the busiest group is not overloaded
11404 			 * and there is no imbalance between this and busiest
11405 			 * group wrt idle CPUs, it is balanced. The imbalance
11406 			 * becomes significant if the diff is greater than 1
11407 			 * otherwise we might end up to just move the imbalance
11408 			 * on another group. Of course this applies only if
11409 			 * there is more than 1 CPU per group.
11410 			 */
11411 			goto out_balanced;
11412 		}
11413 
11414 		if (busiest->sum_h_nr_running == 1) {
11415 			/*
11416 			 * busiest doesn't have any tasks waiting to run
11417 			 */
11418 			goto out_balanced;
11419 		}
11420 	}
11421 
11422 force_balance:
11423 	/* Looks like there is an imbalance. Compute it */
11424 	calculate_imbalance(env, &sds);
11425 	return env->imbalance ? sds.busiest : NULL;
11426 
11427 out_balanced:
11428 	env->imbalance = 0;
11429 	return NULL;
11430 }
11431 
11432 /*
11433  * sched_balance_find_src_rq - find the busiest runqueue among the CPUs in the group.
11434  */
sched_balance_find_src_rq(struct lb_env * env,struct sched_group * group)11435 static struct rq *sched_balance_find_src_rq(struct lb_env *env,
11436 				     struct sched_group *group)
11437 {
11438 	struct rq *busiest = NULL, *rq;
11439 	unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
11440 	unsigned int busiest_nr = 0;
11441 	int i;
11442 
11443 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
11444 		unsigned long capacity, load, util;
11445 		unsigned int nr_running;
11446 		enum fbq_type rt;
11447 
11448 		rq = cpu_rq(i);
11449 		rt = fbq_classify_rq(rq);
11450 
11451 		/*
11452 		 * We classify groups/runqueues into three groups:
11453 		 *  - regular: there are !numa tasks
11454 		 *  - remote:  there are numa tasks that run on the 'wrong' node
11455 		 *  - all:     there is no distinction
11456 		 *
11457 		 * In order to avoid migrating ideally placed numa tasks,
11458 		 * ignore those when there's better options.
11459 		 *
11460 		 * If we ignore the actual busiest queue to migrate another
11461 		 * task, the next balance pass can still reduce the busiest
11462 		 * queue by moving tasks around inside the node.
11463 		 *
11464 		 * If we cannot move enough load due to this classification
11465 		 * the next pass will adjust the group classification and
11466 		 * allow migration of more tasks.
11467 		 *
11468 		 * Both cases only affect the total convergence complexity.
11469 		 */
11470 		if (rt > env->fbq_type)
11471 			continue;
11472 
11473 		nr_running = rq->cfs.h_nr_runnable;
11474 		if (!nr_running)
11475 			continue;
11476 
11477 		capacity = capacity_of(i);
11478 
11479 		/*
11480 		 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
11481 		 * eventually lead to active_balancing high->low capacity.
11482 		 * Higher per-CPU capacity is considered better than balancing
11483 		 * average load.
11484 		 */
11485 		if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
11486 		    !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
11487 		    nr_running == 1)
11488 			continue;
11489 
11490 		/*
11491 		 * Make sure we only pull tasks from a CPU of lower priority
11492 		 * when balancing between SMT siblings.
11493 		 *
11494 		 * If balancing between cores, let lower priority CPUs help
11495 		 * SMT cores with more than one busy sibling.
11496 		 */
11497 		if (sched_asym(env->sd, i, env->dst_cpu) && nr_running == 1)
11498 			continue;
11499 
11500 		switch (env->migration_type) {
11501 		case migrate_load:
11502 			/*
11503 			 * When comparing with load imbalance, use cpu_load()
11504 			 * which is not scaled with the CPU capacity.
11505 			 */
11506 			load = cpu_load(rq);
11507 
11508 			if (nr_running == 1 && load > env->imbalance &&
11509 			    !check_cpu_capacity(rq, env->sd))
11510 				break;
11511 
11512 			/*
11513 			 * For the load comparisons with the other CPUs,
11514 			 * consider the cpu_load() scaled with the CPU
11515 			 * capacity, so that the load can be moved away
11516 			 * from the CPU that is potentially running at a
11517 			 * lower capacity.
11518 			 *
11519 			 * Thus we're looking for max(load_i / capacity_i),
11520 			 * crosswise multiplication to rid ourselves of the
11521 			 * division works out to:
11522 			 * load_i * capacity_j > load_j * capacity_i;
11523 			 * where j is our previous maximum.
11524 			 */
11525 			if (load * busiest_capacity > busiest_load * capacity) {
11526 				busiest_load = load;
11527 				busiest_capacity = capacity;
11528 				busiest = rq;
11529 			}
11530 			break;
11531 
11532 		case migrate_util:
11533 			util = cpu_util_cfs_boost(i);
11534 
11535 			/*
11536 			 * Don't try to pull utilization from a CPU with one
11537 			 * running task. Whatever its utilization, we will fail
11538 			 * detach the task.
11539 			 */
11540 			if (nr_running <= 1)
11541 				continue;
11542 
11543 			if (busiest_util < util) {
11544 				busiest_util = util;
11545 				busiest = rq;
11546 			}
11547 			break;
11548 
11549 		case migrate_task:
11550 			if (busiest_nr < nr_running) {
11551 				busiest_nr = nr_running;
11552 				busiest = rq;
11553 			}
11554 			break;
11555 
11556 		case migrate_misfit:
11557 			/*
11558 			 * For ASYM_CPUCAPACITY domains with misfit tasks we
11559 			 * simply seek the "biggest" misfit task.
11560 			 */
11561 			if (rq->misfit_task_load > busiest_load) {
11562 				busiest_load = rq->misfit_task_load;
11563 				busiest = rq;
11564 			}
11565 
11566 			break;
11567 
11568 		}
11569 	}
11570 
11571 	return busiest;
11572 }
11573 
11574 /*
11575  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
11576  * so long as it is large enough.
11577  */
11578 #define MAX_PINNED_INTERVAL	512
11579 
11580 static inline bool
asym_active_balance(struct lb_env * env)11581 asym_active_balance(struct lb_env *env)
11582 {
11583 	/*
11584 	 * ASYM_PACKING needs to force migrate tasks from busy but lower
11585 	 * priority CPUs in order to pack all tasks in the highest priority
11586 	 * CPUs. When done between cores, do it only if the whole core if the
11587 	 * whole core is idle.
11588 	 *
11589 	 * If @env::src_cpu is an SMT core with busy siblings, let
11590 	 * the lower priority @env::dst_cpu help it. Do not follow
11591 	 * CPU priority.
11592 	 */
11593 	return env->idle && sched_use_asym_prio(env->sd, env->dst_cpu) &&
11594 	       (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
11595 		!sched_use_asym_prio(env->sd, env->src_cpu));
11596 }
11597 
11598 static inline bool
imbalanced_active_balance(struct lb_env * env)11599 imbalanced_active_balance(struct lb_env *env)
11600 {
11601 	struct sched_domain *sd = env->sd;
11602 
11603 	/*
11604 	 * The imbalanced case includes the case of pinned tasks preventing a fair
11605 	 * distribution of the load on the system but also the even distribution of the
11606 	 * threads on a system with spare capacity
11607 	 */
11608 	if ((env->migration_type == migrate_task) &&
11609 	    (sd->nr_balance_failed > sd->cache_nice_tries+2))
11610 		return 1;
11611 
11612 	return 0;
11613 }
11614 
need_active_balance(struct lb_env * env)11615 static int need_active_balance(struct lb_env *env)
11616 {
11617 	struct sched_domain *sd = env->sd;
11618 
11619 	if (asym_active_balance(env))
11620 		return 1;
11621 
11622 	if (imbalanced_active_balance(env))
11623 		return 1;
11624 
11625 	/*
11626 	 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
11627 	 * It's worth migrating the task if the src_cpu's capacity is reduced
11628 	 * because of other sched_class or IRQs if more capacity stays
11629 	 * available on dst_cpu.
11630 	 */
11631 	if (env->idle &&
11632 	    (env->src_rq->cfs.h_nr_runnable == 1)) {
11633 		if ((check_cpu_capacity(env->src_rq, sd)) &&
11634 		    (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
11635 			return 1;
11636 	}
11637 
11638 	if (env->migration_type == migrate_misfit)
11639 		return 1;
11640 
11641 	return 0;
11642 }
11643 
11644 static int active_load_balance_cpu_stop(void *data);
11645 
should_we_balance(struct lb_env * env)11646 static int should_we_balance(struct lb_env *env)
11647 {
11648 	struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask);
11649 	struct sched_group *sg = env->sd->groups;
11650 	int cpu, idle_smt = -1;
11651 
11652 	/*
11653 	 * Ensure the balancing environment is consistent; can happen
11654 	 * when the softirq triggers 'during' hotplug.
11655 	 */
11656 	if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
11657 		return 0;
11658 
11659 	/*
11660 	 * In the newly idle case, we will allow all the CPUs
11661 	 * to do the newly idle load balance.
11662 	 *
11663 	 * However, we bail out if we already have tasks or a wakeup pending,
11664 	 * to optimize wakeup latency.
11665 	 */
11666 	if (env->idle == CPU_NEWLY_IDLE) {
11667 		if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
11668 			return 0;
11669 		return 1;
11670 	}
11671 
11672 	cpumask_copy(swb_cpus, group_balance_mask(sg));
11673 	/* Try to find first idle CPU */
11674 	for_each_cpu_and(cpu, swb_cpus, env->cpus) {
11675 		if (!idle_cpu(cpu))
11676 			continue;
11677 
11678 		/*
11679 		 * Don't balance to idle SMT in busy core right away when
11680 		 * balancing cores, but remember the first idle SMT CPU for
11681 		 * later consideration.  Find CPU on an idle core first.
11682 		 */
11683 		if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) {
11684 			if (idle_smt == -1)
11685 				idle_smt = cpu;
11686 			/*
11687 			 * If the core is not idle, and first SMT sibling which is
11688 			 * idle has been found, then its not needed to check other
11689 			 * SMT siblings for idleness:
11690 			 */
11691 #ifdef CONFIG_SCHED_SMT
11692 			cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu));
11693 #endif
11694 			continue;
11695 		}
11696 
11697 		/*
11698 		 * Are we the first idle core in a non-SMT domain or higher,
11699 		 * or the first idle CPU in a SMT domain?
11700 		 */
11701 		return cpu == env->dst_cpu;
11702 	}
11703 
11704 	/* Are we the first idle CPU with busy siblings? */
11705 	if (idle_smt != -1)
11706 		return idle_smt == env->dst_cpu;
11707 
11708 	/* Are we the first CPU of this group ? */
11709 	return group_balance_cpu(sg) == env->dst_cpu;
11710 }
11711 
update_lb_imbalance_stat(struct lb_env * env,struct sched_domain * sd,enum cpu_idle_type idle)11712 static void update_lb_imbalance_stat(struct lb_env *env, struct sched_domain *sd,
11713 				     enum cpu_idle_type idle)
11714 {
11715 	if (!schedstat_enabled())
11716 		return;
11717 
11718 	switch (env->migration_type) {
11719 	case migrate_load:
11720 		__schedstat_add(sd->lb_imbalance_load[idle], env->imbalance);
11721 		break;
11722 	case migrate_util:
11723 		__schedstat_add(sd->lb_imbalance_util[idle], env->imbalance);
11724 		break;
11725 	case migrate_task:
11726 		__schedstat_add(sd->lb_imbalance_task[idle], env->imbalance);
11727 		break;
11728 	case migrate_misfit:
11729 		__schedstat_add(sd->lb_imbalance_misfit[idle], env->imbalance);
11730 		break;
11731 	}
11732 }
11733 
11734 /*
11735  * Check this_cpu to ensure it is balanced within domain. Attempt to move
11736  * tasks if there is an imbalance.
11737  */
sched_balance_rq(int this_cpu,struct rq * this_rq,struct sched_domain * sd,enum cpu_idle_type idle,int * continue_balancing)11738 static int sched_balance_rq(int this_cpu, struct rq *this_rq,
11739 			struct sched_domain *sd, enum cpu_idle_type idle,
11740 			int *continue_balancing)
11741 {
11742 	int ld_moved, cur_ld_moved, active_balance = 0;
11743 	struct sched_domain *sd_parent = sd->parent;
11744 	struct sched_group *group;
11745 	struct rq *busiest;
11746 	struct rq_flags rf;
11747 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
11748 	struct lb_env env = {
11749 		.sd		= sd,
11750 		.dst_cpu	= this_cpu,
11751 		.dst_rq		= this_rq,
11752 		.dst_grpmask    = group_balance_mask(sd->groups),
11753 		.idle		= idle,
11754 		.loop_break	= SCHED_NR_MIGRATE_BREAK,
11755 		.cpus		= cpus,
11756 		.fbq_type	= all,
11757 		.tasks		= LIST_HEAD_INIT(env.tasks),
11758 	};
11759 
11760 	cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
11761 
11762 	schedstat_inc(sd->lb_count[idle]);
11763 
11764 redo:
11765 	if (!should_we_balance(&env)) {
11766 		*continue_balancing = 0;
11767 		goto out_balanced;
11768 	}
11769 
11770 	group = sched_balance_find_src_group(&env);
11771 	if (!group) {
11772 		schedstat_inc(sd->lb_nobusyg[idle]);
11773 		goto out_balanced;
11774 	}
11775 
11776 	busiest = sched_balance_find_src_rq(&env, group);
11777 	if (!busiest) {
11778 		schedstat_inc(sd->lb_nobusyq[idle]);
11779 		goto out_balanced;
11780 	}
11781 
11782 	WARN_ON_ONCE(busiest == env.dst_rq);
11783 
11784 	update_lb_imbalance_stat(&env, sd, idle);
11785 
11786 	env.src_cpu = busiest->cpu;
11787 	env.src_rq = busiest;
11788 
11789 	ld_moved = 0;
11790 	/* Clear this flag as soon as we find a pullable task */
11791 	env.flags |= LBF_ALL_PINNED;
11792 	if (busiest->nr_running > 1) {
11793 		/*
11794 		 * Attempt to move tasks. If sched_balance_find_src_group has found
11795 		 * an imbalance but busiest->nr_running <= 1, the group is
11796 		 * still unbalanced. ld_moved simply stays zero, so it is
11797 		 * correctly treated as an imbalance.
11798 		 */
11799 		env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
11800 
11801 more_balance:
11802 		rq_lock_irqsave(busiest, &rf);
11803 		update_rq_clock(busiest);
11804 
11805 		/*
11806 		 * cur_ld_moved - load moved in current iteration
11807 		 * ld_moved     - cumulative load moved across iterations
11808 		 */
11809 		cur_ld_moved = detach_tasks(&env);
11810 
11811 		/*
11812 		 * We've detached some tasks from busiest_rq. Every
11813 		 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
11814 		 * unlock busiest->lock, and we are able to be sure
11815 		 * that nobody can manipulate the tasks in parallel.
11816 		 * See task_rq_lock() family for the details.
11817 		 */
11818 
11819 		rq_unlock(busiest, &rf);
11820 
11821 		if (cur_ld_moved) {
11822 			attach_tasks(&env);
11823 			ld_moved += cur_ld_moved;
11824 		}
11825 
11826 		local_irq_restore(rf.flags);
11827 
11828 		if (env.flags & LBF_NEED_BREAK) {
11829 			env.flags &= ~LBF_NEED_BREAK;
11830 			goto more_balance;
11831 		}
11832 
11833 		/*
11834 		 * Revisit (affine) tasks on src_cpu that couldn't be moved to
11835 		 * us and move them to an alternate dst_cpu in our sched_group
11836 		 * where they can run. The upper limit on how many times we
11837 		 * iterate on same src_cpu is dependent on number of CPUs in our
11838 		 * sched_group.
11839 		 *
11840 		 * This changes load balance semantics a bit on who can move
11841 		 * load to a given_cpu. In addition to the given_cpu itself
11842 		 * (or a ilb_cpu acting on its behalf where given_cpu is
11843 		 * nohz-idle), we now have balance_cpu in a position to move
11844 		 * load to given_cpu. In rare situations, this may cause
11845 		 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
11846 		 * _independently_ and at _same_ time to move some load to
11847 		 * given_cpu) causing excess load to be moved to given_cpu.
11848 		 * This however should not happen so much in practice and
11849 		 * moreover subsequent load balance cycles should correct the
11850 		 * excess load moved.
11851 		 */
11852 		if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
11853 
11854 			/* Prevent to re-select dst_cpu via env's CPUs */
11855 			__cpumask_clear_cpu(env.dst_cpu, env.cpus);
11856 
11857 			env.dst_rq	 = cpu_rq(env.new_dst_cpu);
11858 			env.dst_cpu	 = env.new_dst_cpu;
11859 			env.flags	&= ~LBF_DST_PINNED;
11860 			env.loop	 = 0;
11861 			env.loop_break	 = SCHED_NR_MIGRATE_BREAK;
11862 
11863 			/*
11864 			 * Go back to "more_balance" rather than "redo" since we
11865 			 * need to continue with same src_cpu.
11866 			 */
11867 			goto more_balance;
11868 		}
11869 
11870 		/*
11871 		 * We failed to reach balance because of affinity.
11872 		 */
11873 		if (sd_parent) {
11874 			int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11875 
11876 			if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
11877 				*group_imbalance = 1;
11878 		}
11879 
11880 		/* All tasks on this runqueue were pinned by CPU affinity */
11881 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
11882 			__cpumask_clear_cpu(cpu_of(busiest), cpus);
11883 			/*
11884 			 * Attempting to continue load balancing at the current
11885 			 * sched_domain level only makes sense if there are
11886 			 * active CPUs remaining as possible busiest CPUs to
11887 			 * pull load from which are not contained within the
11888 			 * destination group that is receiving any migrated
11889 			 * load.
11890 			 */
11891 			if (!cpumask_subset(cpus, env.dst_grpmask)) {
11892 				env.loop = 0;
11893 				env.loop_break = SCHED_NR_MIGRATE_BREAK;
11894 				goto redo;
11895 			}
11896 			goto out_all_pinned;
11897 		}
11898 	}
11899 
11900 	if (!ld_moved) {
11901 		schedstat_inc(sd->lb_failed[idle]);
11902 		/*
11903 		 * Increment the failure counter only on periodic balance.
11904 		 * We do not want newidle balance, which can be very
11905 		 * frequent, pollute the failure counter causing
11906 		 * excessive cache_hot migrations and active balances.
11907 		 *
11908 		 * Similarly for migration_misfit which is not related to
11909 		 * load/util migration, don't pollute nr_balance_failed.
11910 		 */
11911 		if (idle != CPU_NEWLY_IDLE &&
11912 		    env.migration_type != migrate_misfit)
11913 			sd->nr_balance_failed++;
11914 
11915 		if (need_active_balance(&env)) {
11916 			unsigned long flags;
11917 
11918 			raw_spin_rq_lock_irqsave(busiest, flags);
11919 
11920 			/*
11921 			 * Don't kick the active_load_balance_cpu_stop,
11922 			 * if the curr task on busiest CPU can't be
11923 			 * moved to this_cpu:
11924 			 */
11925 			if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
11926 				raw_spin_rq_unlock_irqrestore(busiest, flags);
11927 				goto out_one_pinned;
11928 			}
11929 
11930 			/* Record that we found at least one task that could run on this_cpu */
11931 			env.flags &= ~LBF_ALL_PINNED;
11932 
11933 			/*
11934 			 * ->active_balance synchronizes accesses to
11935 			 * ->active_balance_work.  Once set, it's cleared
11936 			 * only after active load balance is finished.
11937 			 */
11938 			if (!busiest->active_balance) {
11939 				busiest->active_balance = 1;
11940 				busiest->push_cpu = this_cpu;
11941 				active_balance = 1;
11942 			}
11943 
11944 			preempt_disable();
11945 			raw_spin_rq_unlock_irqrestore(busiest, flags);
11946 			if (active_balance) {
11947 				stop_one_cpu_nowait(cpu_of(busiest),
11948 					active_load_balance_cpu_stop, busiest,
11949 					&busiest->active_balance_work);
11950 			}
11951 			preempt_enable();
11952 		}
11953 	} else {
11954 		sd->nr_balance_failed = 0;
11955 	}
11956 
11957 	if (likely(!active_balance) || need_active_balance(&env)) {
11958 		/* We were unbalanced, so reset the balancing interval */
11959 		sd->balance_interval = sd->min_interval;
11960 	}
11961 
11962 	goto out;
11963 
11964 out_balanced:
11965 	/*
11966 	 * We reach balance although we may have faced some affinity
11967 	 * constraints. Clear the imbalance flag only if other tasks got
11968 	 * a chance to move and fix the imbalance.
11969 	 */
11970 	if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
11971 		int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11972 
11973 		if (*group_imbalance)
11974 			*group_imbalance = 0;
11975 	}
11976 
11977 out_all_pinned:
11978 	/*
11979 	 * We reach balance because all tasks are pinned at this level so
11980 	 * we can't migrate them. Let the imbalance flag set so parent level
11981 	 * can try to migrate them.
11982 	 */
11983 	schedstat_inc(sd->lb_balanced[idle]);
11984 
11985 	sd->nr_balance_failed = 0;
11986 
11987 out_one_pinned:
11988 	ld_moved = 0;
11989 
11990 	/*
11991 	 * sched_balance_newidle() disregards balance intervals, so we could
11992 	 * repeatedly reach this code, which would lead to balance_interval
11993 	 * skyrocketing in a short amount of time. Skip the balance_interval
11994 	 * increase logic to avoid that.
11995 	 *
11996 	 * Similarly misfit migration which is not necessarily an indication of
11997 	 * the system being busy and requires lb to backoff to let it settle
11998 	 * down.
11999 	 */
12000 	if (env.idle == CPU_NEWLY_IDLE ||
12001 	    env.migration_type == migrate_misfit)
12002 		goto out;
12003 
12004 	/* tune up the balancing interval */
12005 	if ((env.flags & LBF_ALL_PINNED &&
12006 	     sd->balance_interval < MAX_PINNED_INTERVAL) ||
12007 	    sd->balance_interval < sd->max_interval)
12008 		sd->balance_interval *= 2;
12009 out:
12010 	return ld_moved;
12011 }
12012 
12013 static inline unsigned long
get_sd_balance_interval(struct sched_domain * sd,int cpu_busy)12014 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
12015 {
12016 	unsigned long interval = sd->balance_interval;
12017 
12018 	if (cpu_busy)
12019 		interval *= sd->busy_factor;
12020 
12021 	/* scale ms to jiffies */
12022 	interval = msecs_to_jiffies(interval);
12023 
12024 	/*
12025 	 * Reduce likelihood of busy balancing at higher domains racing with
12026 	 * balancing at lower domains by preventing their balancing periods
12027 	 * from being multiples of each other.
12028 	 */
12029 	if (cpu_busy)
12030 		interval -= 1;
12031 
12032 	interval = clamp(interval, 1UL, max_load_balance_interval);
12033 
12034 	return interval;
12035 }
12036 
12037 static inline void
update_next_balance(struct sched_domain * sd,unsigned long * next_balance)12038 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
12039 {
12040 	unsigned long interval, next;
12041 
12042 	/* used by idle balance, so cpu_busy = 0 */
12043 	interval = get_sd_balance_interval(sd, 0);
12044 	next = sd->last_balance + interval;
12045 
12046 	if (time_after(*next_balance, next))
12047 		*next_balance = next;
12048 }
12049 
12050 /*
12051  * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
12052  * running tasks off the busiest CPU onto idle CPUs. It requires at
12053  * least 1 task to be running on each physical CPU where possible, and
12054  * avoids physical / logical imbalances.
12055  */
active_load_balance_cpu_stop(void * data)12056 static int active_load_balance_cpu_stop(void *data)
12057 {
12058 	struct rq *busiest_rq = data;
12059 	int busiest_cpu = cpu_of(busiest_rq);
12060 	int target_cpu = busiest_rq->push_cpu;
12061 	struct rq *target_rq = cpu_rq(target_cpu);
12062 	struct sched_domain *sd;
12063 	struct task_struct *p = NULL;
12064 	struct rq_flags rf;
12065 
12066 	rq_lock_irq(busiest_rq, &rf);
12067 	/*
12068 	 * Between queueing the stop-work and running it is a hole in which
12069 	 * CPUs can become inactive. We should not move tasks from or to
12070 	 * inactive CPUs.
12071 	 */
12072 	if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
12073 		goto out_unlock;
12074 
12075 	/* Make sure the requested CPU hasn't gone down in the meantime: */
12076 	if (unlikely(busiest_cpu != smp_processor_id() ||
12077 		     !busiest_rq->active_balance))
12078 		goto out_unlock;
12079 
12080 	/* Is there any task to move? */
12081 	if (busiest_rq->nr_running <= 1)
12082 		goto out_unlock;
12083 
12084 	/*
12085 	 * This condition is "impossible", if it occurs
12086 	 * we need to fix it. Originally reported by
12087 	 * Bjorn Helgaas on a 128-CPU setup.
12088 	 */
12089 	WARN_ON_ONCE(busiest_rq == target_rq);
12090 
12091 	/* Search for an sd spanning us and the target CPU. */
12092 	rcu_read_lock();
12093 	for_each_domain(target_cpu, sd) {
12094 		if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
12095 			break;
12096 	}
12097 
12098 	if (likely(sd)) {
12099 		struct lb_env env = {
12100 			.sd		= sd,
12101 			.dst_cpu	= target_cpu,
12102 			.dst_rq		= target_rq,
12103 			.src_cpu	= busiest_rq->cpu,
12104 			.src_rq		= busiest_rq,
12105 			.idle		= CPU_IDLE,
12106 			.flags		= LBF_ACTIVE_LB,
12107 		};
12108 
12109 		schedstat_inc(sd->alb_count);
12110 		update_rq_clock(busiest_rq);
12111 
12112 		p = detach_one_task(&env);
12113 		if (p) {
12114 			schedstat_inc(sd->alb_pushed);
12115 			/* Active balancing done, reset the failure counter. */
12116 			sd->nr_balance_failed = 0;
12117 		} else {
12118 			schedstat_inc(sd->alb_failed);
12119 		}
12120 	}
12121 	rcu_read_unlock();
12122 out_unlock:
12123 	busiest_rq->active_balance = 0;
12124 	rq_unlock(busiest_rq, &rf);
12125 
12126 	if (p)
12127 		attach_one_task(target_rq, p);
12128 
12129 	local_irq_enable();
12130 
12131 	return 0;
12132 }
12133 
12134 /*
12135  * This flag serializes load-balancing passes over large domains
12136  * (above the NODE topology level) - only one load-balancing instance
12137  * may run at a time, to reduce overhead on very large systems with
12138  * lots of CPUs and large NUMA distances.
12139  *
12140  * - Note that load-balancing passes triggered while another one
12141  *   is executing are skipped and not re-tried.
12142  *
12143  * - Also note that this does not serialize rebalance_domains()
12144  *   execution, as non-SD_SERIALIZE domains will still be
12145  *   load-balanced in parallel.
12146  */
12147 static atomic_t sched_balance_running = ATOMIC_INIT(0);
12148 
12149 /*
12150  * Scale the max sched_balance_rq interval with the number of CPUs in the system.
12151  * This trades load-balance latency on larger machines for less cross talk.
12152  */
update_max_interval(void)12153 void update_max_interval(void)
12154 {
12155 	max_load_balance_interval = HZ*num_online_cpus()/10;
12156 }
12157 
update_newidle_cost(struct sched_domain * sd,u64 cost)12158 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
12159 {
12160 	if (cost > sd->max_newidle_lb_cost) {
12161 		/*
12162 		 * Track max cost of a domain to make sure to not delay the
12163 		 * next wakeup on the CPU.
12164 		 */
12165 		sd->max_newidle_lb_cost = cost;
12166 		sd->last_decay_max_lb_cost = jiffies;
12167 	} else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
12168 		/*
12169 		 * Decay the newidle max times by ~1% per second to ensure that
12170 		 * it is not outdated and the current max cost is actually
12171 		 * shorter.
12172 		 */
12173 		sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
12174 		sd->last_decay_max_lb_cost = jiffies;
12175 
12176 		return true;
12177 	}
12178 
12179 	return false;
12180 }
12181 
12182 /*
12183  * It checks each scheduling domain to see if it is due to be balanced,
12184  * and initiates a balancing operation if so.
12185  *
12186  * Balancing parameters are set up in init_sched_domains.
12187  */
sched_balance_domains(struct rq * rq,enum cpu_idle_type idle)12188 static void sched_balance_domains(struct rq *rq, enum cpu_idle_type idle)
12189 {
12190 	int continue_balancing = 1;
12191 	int cpu = rq->cpu;
12192 	int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
12193 	unsigned long interval;
12194 	struct sched_domain *sd;
12195 	/* Earliest time when we have to do rebalance again */
12196 	unsigned long next_balance = jiffies + 60*HZ;
12197 	int update_next_balance = 0;
12198 	int need_serialize, need_decay = 0;
12199 	u64 max_cost = 0;
12200 
12201 	rcu_read_lock();
12202 	for_each_domain(cpu, sd) {
12203 		/*
12204 		 * Decay the newidle max times here because this is a regular
12205 		 * visit to all the domains.
12206 		 */
12207 		need_decay = update_newidle_cost(sd, 0);
12208 		max_cost += sd->max_newidle_lb_cost;
12209 
12210 		/*
12211 		 * Stop the load balance at this level. There is another
12212 		 * CPU in our sched group which is doing load balancing more
12213 		 * actively.
12214 		 */
12215 		if (!continue_balancing) {
12216 			if (need_decay)
12217 				continue;
12218 			break;
12219 		}
12220 
12221 		interval = get_sd_balance_interval(sd, busy);
12222 
12223 		need_serialize = sd->flags & SD_SERIALIZE;
12224 		if (need_serialize) {
12225 			if (atomic_cmpxchg_acquire(&sched_balance_running, 0, 1))
12226 				goto out;
12227 		}
12228 
12229 		if (time_after_eq(jiffies, sd->last_balance + interval)) {
12230 			if (sched_balance_rq(cpu, rq, sd, idle, &continue_balancing)) {
12231 				/*
12232 				 * The LBF_DST_PINNED logic could have changed
12233 				 * env->dst_cpu, so we can't know our idle
12234 				 * state even if we migrated tasks. Update it.
12235 				 */
12236 				idle = idle_cpu(cpu);
12237 				busy = !idle && !sched_idle_cpu(cpu);
12238 			}
12239 			sd->last_balance = jiffies;
12240 			interval = get_sd_balance_interval(sd, busy);
12241 		}
12242 		if (need_serialize)
12243 			atomic_set_release(&sched_balance_running, 0);
12244 out:
12245 		if (time_after(next_balance, sd->last_balance + interval)) {
12246 			next_balance = sd->last_balance + interval;
12247 			update_next_balance = 1;
12248 		}
12249 	}
12250 	if (need_decay) {
12251 		/*
12252 		 * Ensure the rq-wide value also decays but keep it at a
12253 		 * reasonable floor to avoid funnies with rq->avg_idle.
12254 		 */
12255 		rq->max_idle_balance_cost =
12256 			max((u64)sysctl_sched_migration_cost, max_cost);
12257 	}
12258 	rcu_read_unlock();
12259 
12260 	/*
12261 	 * next_balance will be updated only when there is a need.
12262 	 * When the cpu is attached to null domain for ex, it will not be
12263 	 * updated.
12264 	 */
12265 	if (likely(update_next_balance))
12266 		rq->next_balance = next_balance;
12267 
12268 }
12269 
on_null_domain(struct rq * rq)12270 static inline int on_null_domain(struct rq *rq)
12271 {
12272 	return unlikely(!rcu_dereference_sched(rq->sd));
12273 }
12274 
12275 #ifdef CONFIG_NO_HZ_COMMON
12276 /*
12277  * NOHZ idle load balancing (ILB) details:
12278  *
12279  * - When one of the busy CPUs notices that there may be an idle rebalancing
12280  *   needed, they will kick the idle load balancer, which then does idle
12281  *   load balancing for all the idle CPUs.
12282  */
find_new_ilb(void)12283 static inline int find_new_ilb(void)
12284 {
12285 	const struct cpumask *hk_mask;
12286 	int ilb_cpu;
12287 
12288 	hk_mask = housekeeping_cpumask(HK_TYPE_KERNEL_NOISE);
12289 
12290 	for_each_cpu_and(ilb_cpu, nohz.idle_cpus_mask, hk_mask) {
12291 
12292 		if (ilb_cpu == smp_processor_id())
12293 			continue;
12294 
12295 		if (idle_cpu(ilb_cpu))
12296 			return ilb_cpu;
12297 	}
12298 
12299 	return -1;
12300 }
12301 
12302 /*
12303  * Kick a CPU to do the NOHZ balancing, if it is time for it, via a cross-CPU
12304  * SMP function call (IPI).
12305  *
12306  * We pick the first idle CPU in the HK_TYPE_KERNEL_NOISE housekeeping set
12307  * (if there is one).
12308  */
kick_ilb(unsigned int flags)12309 static void kick_ilb(unsigned int flags)
12310 {
12311 	int ilb_cpu;
12312 
12313 	/*
12314 	 * Increase nohz.next_balance only when if full ilb is triggered but
12315 	 * not if we only update stats.
12316 	 */
12317 	if (flags & NOHZ_BALANCE_KICK)
12318 		nohz.next_balance = jiffies+1;
12319 
12320 	ilb_cpu = find_new_ilb();
12321 	if (ilb_cpu < 0)
12322 		return;
12323 
12324 	/*
12325 	 * Don't bother if no new NOHZ balance work items for ilb_cpu,
12326 	 * i.e. all bits in flags are already set in ilb_cpu.
12327 	 */
12328 	if ((atomic_read(nohz_flags(ilb_cpu)) & flags) == flags)
12329 		return;
12330 
12331 	/*
12332 	 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
12333 	 * the first flag owns it; cleared by nohz_csd_func().
12334 	 */
12335 	flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
12336 	if (flags & NOHZ_KICK_MASK)
12337 		return;
12338 
12339 	/*
12340 	 * This way we generate an IPI on the target CPU which
12341 	 * is idle, and the softirq performing NOHZ idle load balancing
12342 	 * will be run before returning from the IPI.
12343 	 */
12344 	smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
12345 }
12346 
12347 /*
12348  * Current decision point for kicking the idle load balancer in the presence
12349  * of idle CPUs in the system.
12350  */
nohz_balancer_kick(struct rq * rq)12351 static void nohz_balancer_kick(struct rq *rq)
12352 {
12353 	unsigned long now = jiffies;
12354 	struct sched_domain_shared *sds;
12355 	struct sched_domain *sd;
12356 	int nr_busy, i, cpu = rq->cpu;
12357 	unsigned int flags = 0;
12358 
12359 	if (unlikely(rq->idle_balance))
12360 		return;
12361 
12362 	/*
12363 	 * We may be recently in ticked or tickless idle mode. At the first
12364 	 * busy tick after returning from idle, we will update the busy stats.
12365 	 */
12366 	nohz_balance_exit_idle(rq);
12367 
12368 	/*
12369 	 * None are in tickless mode and hence no need for NOHZ idle load
12370 	 * balancing:
12371 	 */
12372 	if (likely(!atomic_read(&nohz.nr_cpus)))
12373 		return;
12374 
12375 	if (READ_ONCE(nohz.has_blocked) &&
12376 	    time_after(now, READ_ONCE(nohz.next_blocked)))
12377 		flags = NOHZ_STATS_KICK;
12378 
12379 	if (time_before(now, nohz.next_balance))
12380 		goto out;
12381 
12382 	if (rq->nr_running >= 2) {
12383 		flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
12384 		goto out;
12385 	}
12386 
12387 	rcu_read_lock();
12388 
12389 	sd = rcu_dereference(rq->sd);
12390 	if (sd) {
12391 		/*
12392 		 * If there's a runnable CFS task and the current CPU has reduced
12393 		 * capacity, kick the ILB to see if there's a better CPU to run on:
12394 		 */
12395 		if (rq->cfs.h_nr_runnable >= 1 && check_cpu_capacity(rq, sd)) {
12396 			flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
12397 			goto unlock;
12398 		}
12399 	}
12400 
12401 	sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
12402 	if (sd) {
12403 		/*
12404 		 * When ASYM_PACKING; see if there's a more preferred CPU
12405 		 * currently idle; in which case, kick the ILB to move tasks
12406 		 * around.
12407 		 *
12408 		 * When balancing between cores, all the SMT siblings of the
12409 		 * preferred CPU must be idle.
12410 		 */
12411 		for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
12412 			if (sched_asym(sd, i, cpu)) {
12413 				flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
12414 				goto unlock;
12415 			}
12416 		}
12417 	}
12418 
12419 	sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
12420 	if (sd) {
12421 		/*
12422 		 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
12423 		 * to run the misfit task on.
12424 		 */
12425 		if (check_misfit_status(rq)) {
12426 			flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
12427 			goto unlock;
12428 		}
12429 
12430 		/*
12431 		 * For asymmetric systems, we do not want to nicely balance
12432 		 * cache use, instead we want to embrace asymmetry and only
12433 		 * ensure tasks have enough CPU capacity.
12434 		 *
12435 		 * Skip the LLC logic because it's not relevant in that case.
12436 		 */
12437 		goto unlock;
12438 	}
12439 
12440 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
12441 	if (sds) {
12442 		/*
12443 		 * If there is an imbalance between LLC domains (IOW we could
12444 		 * increase the overall cache utilization), we need a less-loaded LLC
12445 		 * domain to pull some load from. Likewise, we may need to spread
12446 		 * load within the current LLC domain (e.g. packed SMT cores but
12447 		 * other CPUs are idle). We can't really know from here how busy
12448 		 * the others are - so just get a NOHZ balance going if it looks
12449 		 * like this LLC domain has tasks we could move.
12450 		 */
12451 		nr_busy = atomic_read(&sds->nr_busy_cpus);
12452 		if (nr_busy > 1) {
12453 			flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
12454 			goto unlock;
12455 		}
12456 	}
12457 unlock:
12458 	rcu_read_unlock();
12459 out:
12460 	if (READ_ONCE(nohz.needs_update))
12461 		flags |= NOHZ_NEXT_KICK;
12462 
12463 	if (flags)
12464 		kick_ilb(flags);
12465 }
12466 
set_cpu_sd_state_busy(int cpu)12467 static void set_cpu_sd_state_busy(int cpu)
12468 {
12469 	struct sched_domain *sd;
12470 
12471 	rcu_read_lock();
12472 	sd = rcu_dereference(per_cpu(sd_llc, cpu));
12473 
12474 	if (!sd || !sd->nohz_idle)
12475 		goto unlock;
12476 	sd->nohz_idle = 0;
12477 
12478 	atomic_inc(&sd->shared->nr_busy_cpus);
12479 unlock:
12480 	rcu_read_unlock();
12481 }
12482 
nohz_balance_exit_idle(struct rq * rq)12483 void nohz_balance_exit_idle(struct rq *rq)
12484 {
12485 	WARN_ON_ONCE(rq != this_rq());
12486 
12487 	if (likely(!rq->nohz_tick_stopped))
12488 		return;
12489 
12490 	rq->nohz_tick_stopped = 0;
12491 	cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
12492 	atomic_dec(&nohz.nr_cpus);
12493 
12494 	set_cpu_sd_state_busy(rq->cpu);
12495 }
12496 
set_cpu_sd_state_idle(int cpu)12497 static void set_cpu_sd_state_idle(int cpu)
12498 {
12499 	struct sched_domain *sd;
12500 
12501 	rcu_read_lock();
12502 	sd = rcu_dereference(per_cpu(sd_llc, cpu));
12503 
12504 	if (!sd || sd->nohz_idle)
12505 		goto unlock;
12506 	sd->nohz_idle = 1;
12507 
12508 	atomic_dec(&sd->shared->nr_busy_cpus);
12509 unlock:
12510 	rcu_read_unlock();
12511 }
12512 
12513 /*
12514  * This routine will record that the CPU is going idle with tick stopped.
12515  * This info will be used in performing idle load balancing in the future.
12516  */
nohz_balance_enter_idle(int cpu)12517 void nohz_balance_enter_idle(int cpu)
12518 {
12519 	struct rq *rq = cpu_rq(cpu);
12520 
12521 	WARN_ON_ONCE(cpu != smp_processor_id());
12522 
12523 	/* If this CPU is going down, then nothing needs to be done: */
12524 	if (!cpu_active(cpu))
12525 		return;
12526 
12527 	/*
12528 	 * Can be set safely without rq->lock held
12529 	 * If a clear happens, it will have evaluated last additions because
12530 	 * rq->lock is held during the check and the clear
12531 	 */
12532 	rq->has_blocked_load = 1;
12533 
12534 	/*
12535 	 * The tick is still stopped but load could have been added in the
12536 	 * meantime. We set the nohz.has_blocked flag to trig a check of the
12537 	 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
12538 	 * of nohz.has_blocked can only happen after checking the new load
12539 	 */
12540 	if (rq->nohz_tick_stopped)
12541 		goto out;
12542 
12543 	/* If we're a completely isolated CPU, we don't play: */
12544 	if (on_null_domain(rq))
12545 		return;
12546 
12547 	rq->nohz_tick_stopped = 1;
12548 
12549 	cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
12550 	atomic_inc(&nohz.nr_cpus);
12551 
12552 	/*
12553 	 * Ensures that if nohz_idle_balance() fails to observe our
12554 	 * @idle_cpus_mask store, it must observe the @has_blocked
12555 	 * and @needs_update stores.
12556 	 */
12557 	smp_mb__after_atomic();
12558 
12559 	set_cpu_sd_state_idle(cpu);
12560 
12561 	WRITE_ONCE(nohz.needs_update, 1);
12562 out:
12563 	/*
12564 	 * Each time a cpu enter idle, we assume that it has blocked load and
12565 	 * enable the periodic update of the load of idle CPUs
12566 	 */
12567 	WRITE_ONCE(nohz.has_blocked, 1);
12568 }
12569 
update_nohz_stats(struct rq * rq)12570 static bool update_nohz_stats(struct rq *rq)
12571 {
12572 	unsigned int cpu = rq->cpu;
12573 
12574 	if (!rq->has_blocked_load)
12575 		return false;
12576 
12577 	if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
12578 		return false;
12579 
12580 	if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
12581 		return true;
12582 
12583 	sched_balance_update_blocked_averages(cpu);
12584 
12585 	return rq->has_blocked_load;
12586 }
12587 
12588 /*
12589  * Internal function that runs load balance for all idle CPUs. The load balance
12590  * can be a simple update of blocked load or a complete load balance with
12591  * tasks movement depending of flags.
12592  */
_nohz_idle_balance(struct rq * this_rq,unsigned int flags)12593 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
12594 {
12595 	/* Earliest time when we have to do rebalance again */
12596 	unsigned long now = jiffies;
12597 	unsigned long next_balance = now + 60*HZ;
12598 	bool has_blocked_load = false;
12599 	int update_next_balance = 0;
12600 	int this_cpu = this_rq->cpu;
12601 	int balance_cpu;
12602 	struct rq *rq;
12603 
12604 	WARN_ON_ONCE((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
12605 
12606 	/*
12607 	 * We assume there will be no idle load after this update and clear
12608 	 * the has_blocked flag. If a cpu enters idle in the mean time, it will
12609 	 * set the has_blocked flag and trigger another update of idle load.
12610 	 * Because a cpu that becomes idle, is added to idle_cpus_mask before
12611 	 * setting the flag, we are sure to not clear the state and not
12612 	 * check the load of an idle cpu.
12613 	 *
12614 	 * Same applies to idle_cpus_mask vs needs_update.
12615 	 */
12616 	if (flags & NOHZ_STATS_KICK)
12617 		WRITE_ONCE(nohz.has_blocked, 0);
12618 	if (flags & NOHZ_NEXT_KICK)
12619 		WRITE_ONCE(nohz.needs_update, 0);
12620 
12621 	/*
12622 	 * Ensures that if we miss the CPU, we must see the has_blocked
12623 	 * store from nohz_balance_enter_idle().
12624 	 */
12625 	smp_mb();
12626 
12627 	/*
12628 	 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
12629 	 * chance for other idle cpu to pull load.
12630 	 */
12631 	for_each_cpu_wrap(balance_cpu,  nohz.idle_cpus_mask, this_cpu+1) {
12632 		if (!idle_cpu(balance_cpu))
12633 			continue;
12634 
12635 		/*
12636 		 * If this CPU gets work to do, stop the load balancing
12637 		 * work being done for other CPUs. Next load
12638 		 * balancing owner will pick it up.
12639 		 */
12640 		if (!idle_cpu(this_cpu) && need_resched()) {
12641 			if (flags & NOHZ_STATS_KICK)
12642 				has_blocked_load = true;
12643 			if (flags & NOHZ_NEXT_KICK)
12644 				WRITE_ONCE(nohz.needs_update, 1);
12645 			goto abort;
12646 		}
12647 
12648 		rq = cpu_rq(balance_cpu);
12649 
12650 		if (flags & NOHZ_STATS_KICK)
12651 			has_blocked_load |= update_nohz_stats(rq);
12652 
12653 		/*
12654 		 * If time for next balance is due,
12655 		 * do the balance.
12656 		 */
12657 		if (time_after_eq(jiffies, rq->next_balance)) {
12658 			struct rq_flags rf;
12659 
12660 			rq_lock_irqsave(rq, &rf);
12661 			update_rq_clock(rq);
12662 			rq_unlock_irqrestore(rq, &rf);
12663 
12664 			if (flags & NOHZ_BALANCE_KICK)
12665 				sched_balance_domains(rq, CPU_IDLE);
12666 		}
12667 
12668 		if (time_after(next_balance, rq->next_balance)) {
12669 			next_balance = rq->next_balance;
12670 			update_next_balance = 1;
12671 		}
12672 	}
12673 
12674 	/*
12675 	 * next_balance will be updated only when there is a need.
12676 	 * When the CPU is attached to null domain for ex, it will not be
12677 	 * updated.
12678 	 */
12679 	if (likely(update_next_balance))
12680 		nohz.next_balance = next_balance;
12681 
12682 	if (flags & NOHZ_STATS_KICK)
12683 		WRITE_ONCE(nohz.next_blocked,
12684 			   now + msecs_to_jiffies(LOAD_AVG_PERIOD));
12685 
12686 abort:
12687 	/* There is still blocked load, enable periodic update */
12688 	if (has_blocked_load)
12689 		WRITE_ONCE(nohz.has_blocked, 1);
12690 }
12691 
12692 /*
12693  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
12694  * rebalancing for all the CPUs for whom scheduler ticks are stopped.
12695  */
nohz_idle_balance(struct rq * this_rq,enum cpu_idle_type idle)12696 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12697 {
12698 	unsigned int flags = this_rq->nohz_idle_balance;
12699 
12700 	if (!flags)
12701 		return false;
12702 
12703 	this_rq->nohz_idle_balance = 0;
12704 
12705 	if (idle != CPU_IDLE)
12706 		return false;
12707 
12708 	_nohz_idle_balance(this_rq, flags);
12709 
12710 	return true;
12711 }
12712 
12713 /*
12714  * Check if we need to directly run the ILB for updating blocked load before
12715  * entering idle state. Here we run ILB directly without issuing IPIs.
12716  *
12717  * Note that when this function is called, the tick may not yet be stopped on
12718  * this CPU yet. nohz.idle_cpus_mask is updated only when tick is stopped and
12719  * cleared on the next busy tick. In other words, nohz.idle_cpus_mask updates
12720  * don't align with CPUs enter/exit idle to avoid bottlenecks due to high idle
12721  * entry/exit rate (usec). So it is possible that _nohz_idle_balance() is
12722  * called from this function on (this) CPU that's not yet in the mask. That's
12723  * OK because the goal of nohz_run_idle_balance() is to run ILB only for
12724  * updating the blocked load of already idle CPUs without waking up one of
12725  * those idle CPUs and outside the preempt disable / IRQ off phase of the local
12726  * cpu about to enter idle, because it can take a long time.
12727  */
nohz_run_idle_balance(int cpu)12728 void nohz_run_idle_balance(int cpu)
12729 {
12730 	unsigned int flags;
12731 
12732 	flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
12733 
12734 	/*
12735 	 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
12736 	 * (i.e. NOHZ_STATS_KICK set) and will do the same.
12737 	 */
12738 	if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
12739 		_nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
12740 }
12741 
nohz_newidle_balance(struct rq * this_rq)12742 static void nohz_newidle_balance(struct rq *this_rq)
12743 {
12744 	int this_cpu = this_rq->cpu;
12745 
12746 	/* Will wake up very soon. No time for doing anything else*/
12747 	if (this_rq->avg_idle < sysctl_sched_migration_cost)
12748 		return;
12749 
12750 	/* Don't need to update blocked load of idle CPUs*/
12751 	if (!READ_ONCE(nohz.has_blocked) ||
12752 	    time_before(jiffies, READ_ONCE(nohz.next_blocked)))
12753 		return;
12754 
12755 	/*
12756 	 * Set the need to trigger ILB in order to update blocked load
12757 	 * before entering idle state.
12758 	 */
12759 	atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
12760 }
12761 
12762 #else /* !CONFIG_NO_HZ_COMMON */
nohz_balancer_kick(struct rq * rq)12763 static inline void nohz_balancer_kick(struct rq *rq) { }
12764 
nohz_idle_balance(struct rq * this_rq,enum cpu_idle_type idle)12765 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12766 {
12767 	return false;
12768 }
12769 
nohz_newidle_balance(struct rq * this_rq)12770 static inline void nohz_newidle_balance(struct rq *this_rq) { }
12771 #endif /* CONFIG_NO_HZ_COMMON */
12772 
12773 /*
12774  * sched_balance_newidle is called by schedule() if this_cpu is about to become
12775  * idle. Attempts to pull tasks from other CPUs.
12776  *
12777  * Returns:
12778  *   < 0 - we released the lock and there are !fair tasks present
12779  *     0 - failed, no new tasks
12780  *   > 0 - success, new (fair) tasks present
12781  */
sched_balance_newidle(struct rq * this_rq,struct rq_flags * rf)12782 static int sched_balance_newidle(struct rq *this_rq, struct rq_flags *rf)
12783 {
12784 	unsigned long next_balance = jiffies + HZ;
12785 	int this_cpu = this_rq->cpu;
12786 	int continue_balancing = 1;
12787 	u64 t0, t1, curr_cost = 0;
12788 	struct sched_domain *sd;
12789 	int pulled_task = 0;
12790 
12791 	update_misfit_status(NULL, this_rq);
12792 
12793 	/*
12794 	 * There is a task waiting to run. No need to search for one.
12795 	 * Return 0; the task will be enqueued when switching to idle.
12796 	 */
12797 	if (this_rq->ttwu_pending)
12798 		return 0;
12799 
12800 	/*
12801 	 * We must set idle_stamp _before_ calling sched_balance_rq()
12802 	 * for CPU_NEWLY_IDLE, such that we measure the this duration
12803 	 * as idle time.
12804 	 */
12805 	this_rq->idle_stamp = rq_clock(this_rq);
12806 
12807 	/*
12808 	 * Do not pull tasks towards !active CPUs...
12809 	 */
12810 	if (!cpu_active(this_cpu))
12811 		return 0;
12812 
12813 	/*
12814 	 * This is OK, because current is on_cpu, which avoids it being picked
12815 	 * for load-balance and preemption/IRQs are still disabled avoiding
12816 	 * further scheduler activity on it and we're being very careful to
12817 	 * re-start the picking loop.
12818 	 */
12819 	rq_unpin_lock(this_rq, rf);
12820 
12821 	rcu_read_lock();
12822 	sd = rcu_dereference_check_sched_domain(this_rq->sd);
12823 
12824 	if (!get_rd_overloaded(this_rq->rd) ||
12825 	    (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
12826 
12827 		if (sd)
12828 			update_next_balance(sd, &next_balance);
12829 		rcu_read_unlock();
12830 
12831 		goto out;
12832 	}
12833 	rcu_read_unlock();
12834 
12835 	raw_spin_rq_unlock(this_rq);
12836 
12837 	t0 = sched_clock_cpu(this_cpu);
12838 	sched_balance_update_blocked_averages(this_cpu);
12839 
12840 	rcu_read_lock();
12841 	for_each_domain(this_cpu, sd) {
12842 		u64 domain_cost;
12843 
12844 		update_next_balance(sd, &next_balance);
12845 
12846 		if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
12847 			break;
12848 
12849 		if (sd->flags & SD_BALANCE_NEWIDLE) {
12850 
12851 			pulled_task = sched_balance_rq(this_cpu, this_rq,
12852 						   sd, CPU_NEWLY_IDLE,
12853 						   &continue_balancing);
12854 
12855 			t1 = sched_clock_cpu(this_cpu);
12856 			domain_cost = t1 - t0;
12857 			update_newidle_cost(sd, domain_cost);
12858 
12859 			curr_cost += domain_cost;
12860 			t0 = t1;
12861 		}
12862 
12863 		/*
12864 		 * Stop searching for tasks to pull if there are
12865 		 * now runnable tasks on this rq.
12866 		 */
12867 		if (pulled_task || !continue_balancing)
12868 			break;
12869 	}
12870 	rcu_read_unlock();
12871 
12872 	raw_spin_rq_lock(this_rq);
12873 
12874 	if (curr_cost > this_rq->max_idle_balance_cost)
12875 		this_rq->max_idle_balance_cost = curr_cost;
12876 
12877 	/*
12878 	 * While browsing the domains, we released the rq lock, a task could
12879 	 * have been enqueued in the meantime. Since we're not going idle,
12880 	 * pretend we pulled a task.
12881 	 */
12882 	if (this_rq->cfs.h_nr_queued && !pulled_task)
12883 		pulled_task = 1;
12884 
12885 	/* Is there a task of a high priority class? */
12886 	if (this_rq->nr_running != this_rq->cfs.h_nr_queued)
12887 		pulled_task = -1;
12888 
12889 out:
12890 	/* Move the next balance forward */
12891 	if (time_after(this_rq->next_balance, next_balance))
12892 		this_rq->next_balance = next_balance;
12893 
12894 	if (pulled_task)
12895 		this_rq->idle_stamp = 0;
12896 	else
12897 		nohz_newidle_balance(this_rq);
12898 
12899 	rq_repin_lock(this_rq, rf);
12900 
12901 	return pulled_task;
12902 }
12903 
12904 /*
12905  * This softirq handler is triggered via SCHED_SOFTIRQ from two places:
12906  *
12907  * - directly from the local sched_tick() for periodic load balancing
12908  *
12909  * - indirectly from a remote sched_tick() for NOHZ idle balancing
12910  *   through the SMP cross-call nohz_csd_func()
12911  */
sched_balance_softirq(void)12912 static __latent_entropy void sched_balance_softirq(void)
12913 {
12914 	struct rq *this_rq = this_rq();
12915 	enum cpu_idle_type idle = this_rq->idle_balance;
12916 	/*
12917 	 * If this CPU has a pending NOHZ_BALANCE_KICK, then do the
12918 	 * balancing on behalf of the other idle CPUs whose ticks are
12919 	 * stopped. Do nohz_idle_balance *before* sched_balance_domains to
12920 	 * give the idle CPUs a chance to load balance. Else we may
12921 	 * load balance only within the local sched_domain hierarchy
12922 	 * and abort nohz_idle_balance altogether if we pull some load.
12923 	 */
12924 	if (nohz_idle_balance(this_rq, idle))
12925 		return;
12926 
12927 	/* normal load balance */
12928 	sched_balance_update_blocked_averages(this_rq->cpu);
12929 	sched_balance_domains(this_rq, idle);
12930 }
12931 
12932 /*
12933  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
12934  */
sched_balance_trigger(struct rq * rq)12935 void sched_balance_trigger(struct rq *rq)
12936 {
12937 	/*
12938 	 * Don't need to rebalance while attached to NULL domain or
12939 	 * runqueue CPU is not active
12940 	 */
12941 	if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
12942 		return;
12943 
12944 	if (time_after_eq(jiffies, rq->next_balance))
12945 		raise_softirq(SCHED_SOFTIRQ);
12946 
12947 	nohz_balancer_kick(rq);
12948 }
12949 
rq_online_fair(struct rq * rq)12950 static void rq_online_fair(struct rq *rq)
12951 {
12952 	update_sysctl();
12953 
12954 	update_runtime_enabled(rq);
12955 }
12956 
rq_offline_fair(struct rq * rq)12957 static void rq_offline_fair(struct rq *rq)
12958 {
12959 	update_sysctl();
12960 
12961 	/* Ensure any throttled groups are reachable by pick_next_task */
12962 	unthrottle_offline_cfs_rqs(rq);
12963 
12964 	/* Ensure that we remove rq contribution to group share: */
12965 	clear_tg_offline_cfs_rqs(rq);
12966 }
12967 
12968 #endif /* CONFIG_SMP */
12969 
12970 #ifdef CONFIG_SCHED_CORE
12971 static inline bool
__entity_slice_used(struct sched_entity * se,int min_nr_tasks)12972 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
12973 {
12974 	u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
12975 	u64 slice = se->slice;
12976 
12977 	return (rtime * min_nr_tasks > slice);
12978 }
12979 
12980 #define MIN_NR_TASKS_DURING_FORCEIDLE	2
task_tick_core(struct rq * rq,struct task_struct * curr)12981 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
12982 {
12983 	if (!sched_core_enabled(rq))
12984 		return;
12985 
12986 	/*
12987 	 * If runqueue has only one task which used up its slice and
12988 	 * if the sibling is forced idle, then trigger schedule to
12989 	 * give forced idle task a chance.
12990 	 *
12991 	 * sched_slice() considers only this active rq and it gets the
12992 	 * whole slice. But during force idle, we have siblings acting
12993 	 * like a single runqueue and hence we need to consider runnable
12994 	 * tasks on this CPU and the forced idle CPU. Ideally, we should
12995 	 * go through the forced idle rq, but that would be a perf hit.
12996 	 * We can assume that the forced idle CPU has at least
12997 	 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
12998 	 * if we need to give up the CPU.
12999 	 */
13000 	if (rq->core->core_forceidle_count && rq->cfs.nr_queued == 1 &&
13001 	    __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
13002 		resched_curr(rq);
13003 }
13004 
13005 /*
13006  * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
13007  */
se_fi_update(const struct sched_entity * se,unsigned int fi_seq,bool forceidle)13008 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
13009 			 bool forceidle)
13010 {
13011 	for_each_sched_entity(se) {
13012 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
13013 
13014 		if (forceidle) {
13015 			if (cfs_rq->forceidle_seq == fi_seq)
13016 				break;
13017 			cfs_rq->forceidle_seq = fi_seq;
13018 		}
13019 
13020 		cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
13021 	}
13022 }
13023 
task_vruntime_update(struct rq * rq,struct task_struct * p,bool in_fi)13024 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
13025 {
13026 	struct sched_entity *se = &p->se;
13027 
13028 	if (p->sched_class != &fair_sched_class)
13029 		return;
13030 
13031 	se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
13032 }
13033 
cfs_prio_less(const struct task_struct * a,const struct task_struct * b,bool in_fi)13034 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
13035 			bool in_fi)
13036 {
13037 	struct rq *rq = task_rq(a);
13038 	const struct sched_entity *sea = &a->se;
13039 	const struct sched_entity *seb = &b->se;
13040 	struct cfs_rq *cfs_rqa;
13041 	struct cfs_rq *cfs_rqb;
13042 	s64 delta;
13043 
13044 	WARN_ON_ONCE(task_rq(b)->core != rq->core);
13045 
13046 #ifdef CONFIG_FAIR_GROUP_SCHED
13047 	/*
13048 	 * Find an se in the hierarchy for tasks a and b, such that the se's
13049 	 * are immediate siblings.
13050 	 */
13051 	while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
13052 		int sea_depth = sea->depth;
13053 		int seb_depth = seb->depth;
13054 
13055 		if (sea_depth >= seb_depth)
13056 			sea = parent_entity(sea);
13057 		if (sea_depth <= seb_depth)
13058 			seb = parent_entity(seb);
13059 	}
13060 
13061 	se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
13062 	se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
13063 
13064 	cfs_rqa = sea->cfs_rq;
13065 	cfs_rqb = seb->cfs_rq;
13066 #else
13067 	cfs_rqa = &task_rq(a)->cfs;
13068 	cfs_rqb = &task_rq(b)->cfs;
13069 #endif
13070 
13071 	/*
13072 	 * Find delta after normalizing se's vruntime with its cfs_rq's
13073 	 * min_vruntime_fi, which would have been updated in prior calls
13074 	 * to se_fi_update().
13075 	 */
13076 	delta = (s64)(sea->vruntime - seb->vruntime) +
13077 		(s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
13078 
13079 	return delta > 0;
13080 }
13081 
task_is_throttled_fair(struct task_struct * p,int cpu)13082 static int task_is_throttled_fair(struct task_struct *p, int cpu)
13083 {
13084 	struct cfs_rq *cfs_rq;
13085 
13086 #ifdef CONFIG_FAIR_GROUP_SCHED
13087 	cfs_rq = task_group(p)->cfs_rq[cpu];
13088 #else
13089 	cfs_rq = &cpu_rq(cpu)->cfs;
13090 #endif
13091 	return throttled_hierarchy(cfs_rq);
13092 }
13093 #else
task_tick_core(struct rq * rq,struct task_struct * curr)13094 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
13095 #endif
13096 
13097 /*
13098  * scheduler tick hitting a task of our scheduling class.
13099  *
13100  * NOTE: This function can be called remotely by the tick offload that
13101  * goes along full dynticks. Therefore no local assumption can be made
13102  * and everything must be accessed through the @rq and @curr passed in
13103  * parameters.
13104  */
task_tick_fair(struct rq * rq,struct task_struct * curr,int queued)13105 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
13106 {
13107 	struct cfs_rq *cfs_rq;
13108 	struct sched_entity *se = &curr->se;
13109 
13110 	for_each_sched_entity(se) {
13111 		cfs_rq = cfs_rq_of(se);
13112 		entity_tick(cfs_rq, se, queued);
13113 	}
13114 
13115 	if (static_branch_unlikely(&sched_numa_balancing))
13116 		task_tick_numa(rq, curr);
13117 
13118 	update_misfit_status(curr, rq);
13119 	check_update_overutilized_status(task_rq(curr));
13120 
13121 	task_tick_core(rq, curr);
13122 }
13123 
13124 /*
13125  * called on fork with the child task as argument from the parent's context
13126  *  - child not yet on the tasklist
13127  *  - preemption disabled
13128  */
task_fork_fair(struct task_struct * p)13129 static void task_fork_fair(struct task_struct *p)
13130 {
13131 	set_task_max_allowed_capacity(p);
13132 }
13133 
13134 /*
13135  * Priority of the task has changed. Check to see if we preempt
13136  * the current task.
13137  */
13138 static void
prio_changed_fair(struct rq * rq,struct task_struct * p,int oldprio)13139 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
13140 {
13141 	if (!task_on_rq_queued(p))
13142 		return;
13143 
13144 	if (rq->cfs.nr_queued == 1)
13145 		return;
13146 
13147 	/*
13148 	 * Reschedule if we are currently running on this runqueue and
13149 	 * our priority decreased, or if we are not currently running on
13150 	 * this runqueue and our priority is higher than the current's
13151 	 */
13152 	if (task_current_donor(rq, p)) {
13153 		if (p->prio > oldprio)
13154 			resched_curr(rq);
13155 	} else
13156 		wakeup_preempt(rq, p, 0);
13157 }
13158 
13159 #ifdef CONFIG_FAIR_GROUP_SCHED
13160 /*
13161  * Propagate the changes of the sched_entity across the tg tree to make it
13162  * visible to the root
13163  */
propagate_entity_cfs_rq(struct sched_entity * se)13164 static void propagate_entity_cfs_rq(struct sched_entity *se)
13165 {
13166 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
13167 
13168 	if (cfs_rq_throttled(cfs_rq))
13169 		return;
13170 
13171 	if (!throttled_hierarchy(cfs_rq))
13172 		list_add_leaf_cfs_rq(cfs_rq);
13173 
13174 	/* Start to propagate at parent */
13175 	se = se->parent;
13176 
13177 	for_each_sched_entity(se) {
13178 		cfs_rq = cfs_rq_of(se);
13179 
13180 		update_load_avg(cfs_rq, se, UPDATE_TG);
13181 
13182 		if (cfs_rq_throttled(cfs_rq))
13183 			break;
13184 
13185 		if (!throttled_hierarchy(cfs_rq))
13186 			list_add_leaf_cfs_rq(cfs_rq);
13187 	}
13188 }
13189 #else
propagate_entity_cfs_rq(struct sched_entity * se)13190 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
13191 #endif
13192 
detach_entity_cfs_rq(struct sched_entity * se)13193 static void detach_entity_cfs_rq(struct sched_entity *se)
13194 {
13195 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
13196 
13197 #ifdef CONFIG_SMP
13198 	/*
13199 	 * In case the task sched_avg hasn't been attached:
13200 	 * - A forked task which hasn't been woken up by wake_up_new_task().
13201 	 * - A task which has been woken up by try_to_wake_up() but is
13202 	 *   waiting for actually being woken up by sched_ttwu_pending().
13203 	 */
13204 	if (!se->avg.last_update_time)
13205 		return;
13206 #endif
13207 
13208 	/* Catch up with the cfs_rq and remove our load when we leave */
13209 	update_load_avg(cfs_rq, se, 0);
13210 	detach_entity_load_avg(cfs_rq, se);
13211 	update_tg_load_avg(cfs_rq);
13212 	propagate_entity_cfs_rq(se);
13213 }
13214 
attach_entity_cfs_rq(struct sched_entity * se)13215 static void attach_entity_cfs_rq(struct sched_entity *se)
13216 {
13217 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
13218 
13219 	/* Synchronize entity with its cfs_rq */
13220 	update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
13221 	attach_entity_load_avg(cfs_rq, se);
13222 	update_tg_load_avg(cfs_rq);
13223 	propagate_entity_cfs_rq(se);
13224 }
13225 
detach_task_cfs_rq(struct task_struct * p)13226 static void detach_task_cfs_rq(struct task_struct *p)
13227 {
13228 	struct sched_entity *se = &p->se;
13229 
13230 	detach_entity_cfs_rq(se);
13231 }
13232 
attach_task_cfs_rq(struct task_struct * p)13233 static void attach_task_cfs_rq(struct task_struct *p)
13234 {
13235 	struct sched_entity *se = &p->se;
13236 
13237 	attach_entity_cfs_rq(se);
13238 }
13239 
switched_from_fair(struct rq * rq,struct task_struct * p)13240 static void switched_from_fair(struct rq *rq, struct task_struct *p)
13241 {
13242 	detach_task_cfs_rq(p);
13243 }
13244 
switched_to_fair(struct rq * rq,struct task_struct * p)13245 static void switched_to_fair(struct rq *rq, struct task_struct *p)
13246 {
13247 	WARN_ON_ONCE(p->se.sched_delayed);
13248 
13249 	attach_task_cfs_rq(p);
13250 
13251 	set_task_max_allowed_capacity(p);
13252 
13253 	if (task_on_rq_queued(p)) {
13254 		/*
13255 		 * We were most likely switched from sched_rt, so
13256 		 * kick off the schedule if running, otherwise just see
13257 		 * if we can still preempt the current task.
13258 		 */
13259 		if (task_current_donor(rq, p))
13260 			resched_curr(rq);
13261 		else
13262 			wakeup_preempt(rq, p, 0);
13263 	}
13264 }
13265 
__set_next_task_fair(struct rq * rq,struct task_struct * p,bool first)13266 static void __set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
13267 {
13268 	struct sched_entity *se = &p->se;
13269 
13270 #ifdef CONFIG_SMP
13271 	if (task_on_rq_queued(p)) {
13272 		/*
13273 		 * Move the next running task to the front of the list, so our
13274 		 * cfs_tasks list becomes MRU one.
13275 		 */
13276 		list_move(&se->group_node, &rq->cfs_tasks);
13277 	}
13278 #endif
13279 	if (!first)
13280 		return;
13281 
13282 	WARN_ON_ONCE(se->sched_delayed);
13283 
13284 	if (hrtick_enabled_fair(rq))
13285 		hrtick_start_fair(rq, p);
13286 
13287 	update_misfit_status(p, rq);
13288 	sched_fair_update_stop_tick(rq, p);
13289 }
13290 
13291 /*
13292  * Account for a task changing its policy or group.
13293  *
13294  * This routine is mostly called to set cfs_rq->curr field when a task
13295  * migrates between groups/classes.
13296  */
set_next_task_fair(struct rq * rq,struct task_struct * p,bool first)13297 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
13298 {
13299 	struct sched_entity *se = &p->se;
13300 
13301 	for_each_sched_entity(se) {
13302 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
13303 
13304 		set_next_entity(cfs_rq, se);
13305 		/* ensure bandwidth has been allocated on our new cfs_rq */
13306 		account_cfs_rq_runtime(cfs_rq, 0);
13307 	}
13308 
13309 	__set_next_task_fair(rq, p, first);
13310 }
13311 
init_cfs_rq(struct cfs_rq * cfs_rq)13312 void init_cfs_rq(struct cfs_rq *cfs_rq)
13313 {
13314 	cfs_rq->tasks_timeline = RB_ROOT_CACHED;
13315 	cfs_rq->min_vruntime = (u64)(-(1LL << 20));
13316 #ifdef CONFIG_SMP
13317 	raw_spin_lock_init(&cfs_rq->removed.lock);
13318 #endif
13319 }
13320 
13321 #ifdef CONFIG_FAIR_GROUP_SCHED
task_change_group_fair(struct task_struct * p)13322 static void task_change_group_fair(struct task_struct *p)
13323 {
13324 	/*
13325 	 * We couldn't detach or attach a forked task which
13326 	 * hasn't been woken up by wake_up_new_task().
13327 	 */
13328 	if (READ_ONCE(p->__state) == TASK_NEW)
13329 		return;
13330 
13331 	detach_task_cfs_rq(p);
13332 
13333 #ifdef CONFIG_SMP
13334 	/* Tell se's cfs_rq has been changed -- migrated */
13335 	p->se.avg.last_update_time = 0;
13336 #endif
13337 	set_task_rq(p, task_cpu(p));
13338 	attach_task_cfs_rq(p);
13339 }
13340 
free_fair_sched_group(struct task_group * tg)13341 void free_fair_sched_group(struct task_group *tg)
13342 {
13343 	int i;
13344 
13345 	for_each_possible_cpu(i) {
13346 		if (tg->cfs_rq)
13347 			kfree(tg->cfs_rq[i]);
13348 		if (tg->se)
13349 			kfree(tg->se[i]);
13350 	}
13351 
13352 	kfree(tg->cfs_rq);
13353 	kfree(tg->se);
13354 }
13355 
alloc_fair_sched_group(struct task_group * tg,struct task_group * parent)13356 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
13357 {
13358 	struct sched_entity *se;
13359 	struct cfs_rq *cfs_rq;
13360 	int i;
13361 
13362 	tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
13363 	if (!tg->cfs_rq)
13364 		goto err;
13365 	tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
13366 	if (!tg->se)
13367 		goto err;
13368 
13369 	tg->shares = NICE_0_LOAD;
13370 
13371 	init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent));
13372 
13373 	for_each_possible_cpu(i) {
13374 		cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
13375 				      GFP_KERNEL, cpu_to_node(i));
13376 		if (!cfs_rq)
13377 			goto err;
13378 
13379 		se = kzalloc_node(sizeof(struct sched_entity_stats),
13380 				  GFP_KERNEL, cpu_to_node(i));
13381 		if (!se)
13382 			goto err_free_rq;
13383 
13384 		init_cfs_rq(cfs_rq);
13385 		init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
13386 		init_entity_runnable_average(se);
13387 	}
13388 
13389 	return 1;
13390 
13391 err_free_rq:
13392 	kfree(cfs_rq);
13393 err:
13394 	return 0;
13395 }
13396 
online_fair_sched_group(struct task_group * tg)13397 void online_fair_sched_group(struct task_group *tg)
13398 {
13399 	struct sched_entity *se;
13400 	struct rq_flags rf;
13401 	struct rq *rq;
13402 	int i;
13403 
13404 	for_each_possible_cpu(i) {
13405 		rq = cpu_rq(i);
13406 		se = tg->se[i];
13407 		rq_lock_irq(rq, &rf);
13408 		update_rq_clock(rq);
13409 		attach_entity_cfs_rq(se);
13410 		sync_throttle(tg, i);
13411 		rq_unlock_irq(rq, &rf);
13412 	}
13413 }
13414 
unregister_fair_sched_group(struct task_group * tg)13415 void unregister_fair_sched_group(struct task_group *tg)
13416 {
13417 	int cpu;
13418 
13419 	destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
13420 
13421 	for_each_possible_cpu(cpu) {
13422 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
13423 		struct sched_entity *se = tg->se[cpu];
13424 		struct rq *rq = cpu_rq(cpu);
13425 
13426 		if (se) {
13427 			if (se->sched_delayed) {
13428 				guard(rq_lock_irqsave)(rq);
13429 				if (se->sched_delayed) {
13430 					update_rq_clock(rq);
13431 					dequeue_entities(rq, se, DEQUEUE_SLEEP | DEQUEUE_DELAYED);
13432 				}
13433 				list_del_leaf_cfs_rq(cfs_rq);
13434 			}
13435 			remove_entity_load_avg(se);
13436 		}
13437 
13438 		/*
13439 		 * Only empty task groups can be destroyed; so we can speculatively
13440 		 * check on_list without danger of it being re-added.
13441 		 */
13442 		if (cfs_rq->on_list) {
13443 			guard(rq_lock_irqsave)(rq);
13444 			list_del_leaf_cfs_rq(cfs_rq);
13445 		}
13446 	}
13447 }
13448 
init_tg_cfs_entry(struct task_group * tg,struct cfs_rq * cfs_rq,struct sched_entity * se,int cpu,struct sched_entity * parent)13449 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
13450 			struct sched_entity *se, int cpu,
13451 			struct sched_entity *parent)
13452 {
13453 	struct rq *rq = cpu_rq(cpu);
13454 
13455 	cfs_rq->tg = tg;
13456 	cfs_rq->rq = rq;
13457 	init_cfs_rq_runtime(cfs_rq);
13458 
13459 	tg->cfs_rq[cpu] = cfs_rq;
13460 	tg->se[cpu] = se;
13461 
13462 	/* se could be NULL for root_task_group */
13463 	if (!se)
13464 		return;
13465 
13466 	if (!parent) {
13467 		se->cfs_rq = &rq->cfs;
13468 		se->depth = 0;
13469 	} else {
13470 		se->cfs_rq = parent->my_q;
13471 		se->depth = parent->depth + 1;
13472 	}
13473 
13474 	se->my_q = cfs_rq;
13475 	/* guarantee group entities always have weight */
13476 	update_load_set(&se->load, NICE_0_LOAD);
13477 	se->parent = parent;
13478 }
13479 
13480 static DEFINE_MUTEX(shares_mutex);
13481 
__sched_group_set_shares(struct task_group * tg,unsigned long shares)13482 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
13483 {
13484 	int i;
13485 
13486 	lockdep_assert_held(&shares_mutex);
13487 
13488 	/*
13489 	 * We can't change the weight of the root cgroup.
13490 	 */
13491 	if (!tg->se[0])
13492 		return -EINVAL;
13493 
13494 	shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
13495 
13496 	if (tg->shares == shares)
13497 		return 0;
13498 
13499 	tg->shares = shares;
13500 	for_each_possible_cpu(i) {
13501 		struct rq *rq = cpu_rq(i);
13502 		struct sched_entity *se = tg->se[i];
13503 		struct rq_flags rf;
13504 
13505 		/* Propagate contribution to hierarchy */
13506 		rq_lock_irqsave(rq, &rf);
13507 		update_rq_clock(rq);
13508 		for_each_sched_entity(se) {
13509 			update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
13510 			update_cfs_group(se);
13511 		}
13512 		rq_unlock_irqrestore(rq, &rf);
13513 	}
13514 
13515 	return 0;
13516 }
13517 
sched_group_set_shares(struct task_group * tg,unsigned long shares)13518 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
13519 {
13520 	int ret;
13521 
13522 	mutex_lock(&shares_mutex);
13523 	if (tg_is_idle(tg))
13524 		ret = -EINVAL;
13525 	else
13526 		ret = __sched_group_set_shares(tg, shares);
13527 	mutex_unlock(&shares_mutex);
13528 
13529 	return ret;
13530 }
13531 
sched_group_set_idle(struct task_group * tg,long idle)13532 int sched_group_set_idle(struct task_group *tg, long idle)
13533 {
13534 	int i;
13535 
13536 	if (tg == &root_task_group)
13537 		return -EINVAL;
13538 
13539 	if (idle < 0 || idle > 1)
13540 		return -EINVAL;
13541 
13542 	mutex_lock(&shares_mutex);
13543 
13544 	if (tg->idle == idle) {
13545 		mutex_unlock(&shares_mutex);
13546 		return 0;
13547 	}
13548 
13549 	tg->idle = idle;
13550 
13551 	for_each_possible_cpu(i) {
13552 		struct rq *rq = cpu_rq(i);
13553 		struct sched_entity *se = tg->se[i];
13554 		struct cfs_rq *grp_cfs_rq = tg->cfs_rq[i];
13555 		bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
13556 		long idle_task_delta;
13557 		struct rq_flags rf;
13558 
13559 		rq_lock_irqsave(rq, &rf);
13560 
13561 		grp_cfs_rq->idle = idle;
13562 		if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
13563 			goto next_cpu;
13564 
13565 		idle_task_delta = grp_cfs_rq->h_nr_queued -
13566 				  grp_cfs_rq->h_nr_idle;
13567 		if (!cfs_rq_is_idle(grp_cfs_rq))
13568 			idle_task_delta *= -1;
13569 
13570 		for_each_sched_entity(se) {
13571 			struct cfs_rq *cfs_rq = cfs_rq_of(se);
13572 
13573 			if (!se->on_rq)
13574 				break;
13575 
13576 			cfs_rq->h_nr_idle += idle_task_delta;
13577 
13578 			/* Already accounted at parent level and above. */
13579 			if (cfs_rq_is_idle(cfs_rq))
13580 				break;
13581 		}
13582 
13583 next_cpu:
13584 		rq_unlock_irqrestore(rq, &rf);
13585 	}
13586 
13587 	/* Idle groups have minimum weight. */
13588 	if (tg_is_idle(tg))
13589 		__sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
13590 	else
13591 		__sched_group_set_shares(tg, NICE_0_LOAD);
13592 
13593 	mutex_unlock(&shares_mutex);
13594 	return 0;
13595 }
13596 
13597 #endif /* CONFIG_FAIR_GROUP_SCHED */
13598 
13599 
get_rr_interval_fair(struct rq * rq,struct task_struct * task)13600 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
13601 {
13602 	struct sched_entity *se = &task->se;
13603 	unsigned int rr_interval = 0;
13604 
13605 	/*
13606 	 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
13607 	 * idle runqueue:
13608 	 */
13609 	if (rq->cfs.load.weight)
13610 		rr_interval = NS_TO_JIFFIES(se->slice);
13611 
13612 	return rr_interval;
13613 }
13614 
13615 /*
13616  * All the scheduling class methods:
13617  */
13618 DEFINE_SCHED_CLASS(fair) = {
13619 
13620 	.enqueue_task		= enqueue_task_fair,
13621 	.dequeue_task		= dequeue_task_fair,
13622 	.yield_task		= yield_task_fair,
13623 	.yield_to_task		= yield_to_task_fair,
13624 
13625 	.wakeup_preempt		= check_preempt_wakeup_fair,
13626 
13627 	.pick_task		= pick_task_fair,
13628 	.pick_next_task		= __pick_next_task_fair,
13629 	.put_prev_task		= put_prev_task_fair,
13630 	.set_next_task          = set_next_task_fair,
13631 
13632 #ifdef CONFIG_SMP
13633 	.balance		= balance_fair,
13634 	.select_task_rq		= select_task_rq_fair,
13635 	.migrate_task_rq	= migrate_task_rq_fair,
13636 
13637 	.rq_online		= rq_online_fair,
13638 	.rq_offline		= rq_offline_fair,
13639 
13640 	.task_dead		= task_dead_fair,
13641 	.set_cpus_allowed	= set_cpus_allowed_fair,
13642 #endif
13643 
13644 	.task_tick		= task_tick_fair,
13645 	.task_fork		= task_fork_fair,
13646 
13647 	.reweight_task		= reweight_task_fair,
13648 	.prio_changed		= prio_changed_fair,
13649 	.switched_from		= switched_from_fair,
13650 	.switched_to		= switched_to_fair,
13651 
13652 	.get_rr_interval	= get_rr_interval_fair,
13653 
13654 	.update_curr		= update_curr_fair,
13655 
13656 #ifdef CONFIG_FAIR_GROUP_SCHED
13657 	.task_change_group	= task_change_group_fair,
13658 #endif
13659 
13660 #ifdef CONFIG_SCHED_CORE
13661 	.task_is_throttled	= task_is_throttled_fair,
13662 #endif
13663 
13664 #ifdef CONFIG_UCLAMP_TASK
13665 	.uclamp_enabled		= 1,
13666 #endif
13667 };
13668 
print_cfs_stats(struct seq_file * m,int cpu)13669 void print_cfs_stats(struct seq_file *m, int cpu)
13670 {
13671 	struct cfs_rq *cfs_rq, *pos;
13672 
13673 	rcu_read_lock();
13674 	for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
13675 		print_cfs_rq(m, cpu, cfs_rq);
13676 	rcu_read_unlock();
13677 }
13678 
13679 #ifdef CONFIG_NUMA_BALANCING
show_numa_stats(struct task_struct * p,struct seq_file * m)13680 void show_numa_stats(struct task_struct *p, struct seq_file *m)
13681 {
13682 	int node;
13683 	unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
13684 	struct numa_group *ng;
13685 
13686 	rcu_read_lock();
13687 	ng = rcu_dereference(p->numa_group);
13688 	for_each_online_node(node) {
13689 		if (p->numa_faults) {
13690 			tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
13691 			tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
13692 		}
13693 		if (ng) {
13694 			gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
13695 			gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
13696 		}
13697 		print_numa_stats(m, node, tsf, tpf, gsf, gpf);
13698 	}
13699 	rcu_read_unlock();
13700 }
13701 #endif /* CONFIG_NUMA_BALANCING */
13702 
init_sched_fair_class(void)13703 __init void init_sched_fair_class(void)
13704 {
13705 #ifdef CONFIG_SMP
13706 	int i;
13707 
13708 	for_each_possible_cpu(i) {
13709 		zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
13710 		zalloc_cpumask_var_node(&per_cpu(select_rq_mask,    i), GFP_KERNEL, cpu_to_node(i));
13711 		zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i),
13712 					GFP_KERNEL, cpu_to_node(i));
13713 
13714 #ifdef CONFIG_CFS_BANDWIDTH
13715 		INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
13716 		INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
13717 #endif
13718 	}
13719 
13720 	open_softirq(SCHED_SOFTIRQ, sched_balance_softirq);
13721 
13722 #ifdef CONFIG_NO_HZ_COMMON
13723 	nohz.next_balance = jiffies;
13724 	nohz.next_blocked = jiffies;
13725 	zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
13726 #endif
13727 #endif /* SMP */
13728 
13729 }
13730