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 
41 #include <linux/cpuidle.h>
42 #include <linux/interrupt.h>
43 #include <linux/memory-tiers.h>
44 #include <linux/mempolicy.h>
45 #include <linux/mutex_api.h>
46 #include <linux/profile.h>
47 #include <linux/psi.h>
48 #include <linux/ratelimit.h>
49 #include <linux/task_work.h>
50 #include <linux/rbtree_augmented.h>
51 
52 #include <asm/switch_to.h>
53 
54 #include <linux/sched/cond_resched.h>
55 
56 #include "sched.h"
57 #include "stats.h"
58 #include "autogroup.h"
59 
60 /*
61  * The initial- and re-scaling of tunables is configurable
62  *
63  * Options are:
64  *
65  *   SCHED_TUNABLESCALING_NONE - unscaled, always *1
66  *   SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
67  *   SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
68  *
69  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
70  */
71 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
72 
73 /*
74  * Minimal preemption granularity for CPU-bound tasks:
75  *
76  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
77  */
78 unsigned int sysctl_sched_base_slice			= 750000ULL;
79 static unsigned int normalized_sysctl_sched_base_slice	= 750000ULL;
80 
81 /*
82  * After fork, child runs first. If set to 0 (default) then
83  * parent will (try to) run first.
84  */
85 unsigned int sysctl_sched_child_runs_first __read_mostly;
86 
87 const_debug unsigned int sysctl_sched_migration_cost	= 500000UL;
88 
89 int sched_thermal_decay_shift;
setup_sched_thermal_decay_shift(char * str)90 static int __init setup_sched_thermal_decay_shift(char *str)
91 {
92 	int _shift = 0;
93 
94 	if (kstrtoint(str, 0, &_shift))
95 		pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
96 
97 	sched_thermal_decay_shift = clamp(_shift, 0, 10);
98 	return 1;
99 }
100 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
101 
102 #ifdef CONFIG_SMP
103 /*
104  * For asym packing, by default the lower numbered CPU has higher priority.
105  */
arch_asym_cpu_priority(int cpu)106 int __weak arch_asym_cpu_priority(int cpu)
107 {
108 	return -cpu;
109 }
110 
111 /*
112  * The margin used when comparing utilization with CPU capacity.
113  *
114  * (default: ~20%)
115  */
116 #define fits_capacity(cap, max)	((cap) * 1280 < (max) * 1024)
117 
118 /*
119  * The margin used when comparing CPU capacities.
120  * is 'cap1' noticeably greater than 'cap2'
121  *
122  * (default: ~5%)
123  */
124 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
125 #endif
126 
127 #ifdef CONFIG_CFS_BANDWIDTH
128 /*
129  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
130  * each time a cfs_rq requests quota.
131  *
132  * Note: in the case that the slice exceeds the runtime remaining (either due
133  * to consumption or the quota being specified to be smaller than the slice)
134  * we will always only issue the remaining available time.
135  *
136  * (default: 5 msec, units: microseconds)
137  */
138 static unsigned int sysctl_sched_cfs_bandwidth_slice		= 5000UL;
139 #endif
140 
141 #ifdef CONFIG_NUMA_BALANCING
142 /* Restrict the NUMA promotion throughput (MB/s) for each target node. */
143 static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536;
144 #endif
145 
146 #ifdef CONFIG_SYSCTL
147 static struct ctl_table sched_fair_sysctls[] = {
148 	{
149 		.procname       = "sched_child_runs_first",
150 		.data           = &sysctl_sched_child_runs_first,
151 		.maxlen         = sizeof(unsigned int),
152 		.mode           = 0644,
153 		.proc_handler   = proc_dointvec,
154 	},
155 #ifdef CONFIG_CFS_BANDWIDTH
156 	{
157 		.procname       = "sched_cfs_bandwidth_slice_us",
158 		.data           = &sysctl_sched_cfs_bandwidth_slice,
159 		.maxlen         = sizeof(unsigned int),
160 		.mode           = 0644,
161 		.proc_handler   = proc_dointvec_minmax,
162 		.extra1         = SYSCTL_ONE,
163 	},
164 #endif
165 #ifdef CONFIG_NUMA_BALANCING
166 	{
167 		.procname	= "numa_balancing_promote_rate_limit_MBps",
168 		.data		= &sysctl_numa_balancing_promote_rate_limit,
169 		.maxlen		= sizeof(unsigned int),
170 		.mode		= 0644,
171 		.proc_handler	= proc_dointvec_minmax,
172 		.extra1		= SYSCTL_ZERO,
173 	},
174 #endif /* CONFIG_NUMA_BALANCING */
175 	{}
176 };
177 
sched_fair_sysctl_init(void)178 static int __init sched_fair_sysctl_init(void)
179 {
180 	register_sysctl_init("kernel", sched_fair_sysctls);
181 	return 0;
182 }
183 late_initcall(sched_fair_sysctl_init);
184 #endif
185 
update_load_add(struct load_weight * lw,unsigned long inc)186 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
187 {
188 	lw->weight += inc;
189 	lw->inv_weight = 0;
190 }
191 
update_load_sub(struct load_weight * lw,unsigned long dec)192 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
193 {
194 	lw->weight -= dec;
195 	lw->inv_weight = 0;
196 }
197 
update_load_set(struct load_weight * lw,unsigned long w)198 static inline void update_load_set(struct load_weight *lw, unsigned long w)
199 {
200 	lw->weight = w;
201 	lw->inv_weight = 0;
202 }
203 
204 /*
205  * Increase the granularity value when there are more CPUs,
206  * because with more CPUs the 'effective latency' as visible
207  * to users decreases. But the relationship is not linear,
208  * so pick a second-best guess by going with the log2 of the
209  * number of CPUs.
210  *
211  * This idea comes from the SD scheduler of Con Kolivas:
212  */
get_update_sysctl_factor(void)213 static unsigned int get_update_sysctl_factor(void)
214 {
215 	unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
216 	unsigned int factor;
217 
218 	switch (sysctl_sched_tunable_scaling) {
219 	case SCHED_TUNABLESCALING_NONE:
220 		factor = 1;
221 		break;
222 	case SCHED_TUNABLESCALING_LINEAR:
223 		factor = cpus;
224 		break;
225 	case SCHED_TUNABLESCALING_LOG:
226 	default:
227 		factor = 1 + ilog2(cpus);
228 		break;
229 	}
230 
231 	return factor;
232 }
233 
update_sysctl(void)234 static void update_sysctl(void)
235 {
236 	unsigned int factor = get_update_sysctl_factor();
237 
238 #define SET_SYSCTL(name) \
239 	(sysctl_##name = (factor) * normalized_sysctl_##name)
240 	SET_SYSCTL(sched_base_slice);
241 #undef SET_SYSCTL
242 }
243 
sched_init_granularity(void)244 void __init sched_init_granularity(void)
245 {
246 	update_sysctl();
247 }
248 
249 #define WMULT_CONST	(~0U)
250 #define WMULT_SHIFT	32
251 
__update_inv_weight(struct load_weight * lw)252 static void __update_inv_weight(struct load_weight *lw)
253 {
254 	unsigned long w;
255 
256 	if (likely(lw->inv_weight))
257 		return;
258 
259 	w = scale_load_down(lw->weight);
260 
261 	if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
262 		lw->inv_weight = 1;
263 	else if (unlikely(!w))
264 		lw->inv_weight = WMULT_CONST;
265 	else
266 		lw->inv_weight = WMULT_CONST / w;
267 }
268 
269 /*
270  * delta_exec * weight / lw.weight
271  *   OR
272  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
273  *
274  * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
275  * we're guaranteed shift stays positive because inv_weight is guaranteed to
276  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
277  *
278  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
279  * weight/lw.weight <= 1, and therefore our shift will also be positive.
280  */
__calc_delta(u64 delta_exec,unsigned long weight,struct load_weight * lw)281 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
282 {
283 	u64 fact = scale_load_down(weight);
284 	u32 fact_hi = (u32)(fact >> 32);
285 	int shift = WMULT_SHIFT;
286 	int fs;
287 
288 	__update_inv_weight(lw);
289 
290 	if (unlikely(fact_hi)) {
291 		fs = fls(fact_hi);
292 		shift -= fs;
293 		fact >>= fs;
294 	}
295 
296 	fact = mul_u32_u32(fact, lw->inv_weight);
297 
298 	fact_hi = (u32)(fact >> 32);
299 	if (fact_hi) {
300 		fs = fls(fact_hi);
301 		shift -= fs;
302 		fact >>= fs;
303 	}
304 
305 	return mul_u64_u32_shr(delta_exec, fact, shift);
306 }
307 
308 /*
309  * delta /= w
310  */
calc_delta_fair(u64 delta,struct sched_entity * se)311 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
312 {
313 	if (unlikely(se->load.weight != NICE_0_LOAD))
314 		delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
315 
316 	return delta;
317 }
318 
319 const struct sched_class fair_sched_class;
320 
321 /**************************************************************
322  * CFS operations on generic schedulable entities:
323  */
324 
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 
327 /* Walk up scheduling entities hierarchy */
328 #define for_each_sched_entity(se) \
329 		for (; se; se = se->parent)
330 
list_add_leaf_cfs_rq(struct cfs_rq * cfs_rq)331 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
332 {
333 	struct rq *rq = rq_of(cfs_rq);
334 	int cpu = cpu_of(rq);
335 
336 	if (cfs_rq->on_list)
337 		return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
338 
339 	cfs_rq->on_list = 1;
340 
341 	/*
342 	 * Ensure we either appear before our parent (if already
343 	 * enqueued) or force our parent to appear after us when it is
344 	 * enqueued. The fact that we always enqueue bottom-up
345 	 * reduces this to two cases and a special case for the root
346 	 * cfs_rq. Furthermore, it also means that we will always reset
347 	 * tmp_alone_branch either when the branch is connected
348 	 * to a tree or when we reach the top of the tree
349 	 */
350 	if (cfs_rq->tg->parent &&
351 	    cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
352 		/*
353 		 * If parent is already on the list, we add the child
354 		 * just before. Thanks to circular linked property of
355 		 * the list, this means to put the child at the tail
356 		 * of the list that starts by parent.
357 		 */
358 		list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
359 			&(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
360 		/*
361 		 * The branch is now connected to its tree so we can
362 		 * reset tmp_alone_branch to the beginning of the
363 		 * list.
364 		 */
365 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
366 		return true;
367 	}
368 
369 	if (!cfs_rq->tg->parent) {
370 		/*
371 		 * cfs rq without parent should be put
372 		 * at the tail of the list.
373 		 */
374 		list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
375 			&rq->leaf_cfs_rq_list);
376 		/*
377 		 * We have reach the top of a tree so we can reset
378 		 * tmp_alone_branch to the beginning of the list.
379 		 */
380 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
381 		return true;
382 	}
383 
384 	/*
385 	 * The parent has not already been added so we want to
386 	 * make sure that it will be put after us.
387 	 * tmp_alone_branch points to the begin of the branch
388 	 * where we will add parent.
389 	 */
390 	list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
391 	/*
392 	 * update tmp_alone_branch to points to the new begin
393 	 * of the branch
394 	 */
395 	rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
396 	return false;
397 }
398 
list_del_leaf_cfs_rq(struct cfs_rq * cfs_rq)399 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
400 {
401 	if (cfs_rq->on_list) {
402 		struct rq *rq = rq_of(cfs_rq);
403 
404 		/*
405 		 * With cfs_rq being unthrottled/throttled during an enqueue,
406 		 * it can happen the tmp_alone_branch points the a leaf that
407 		 * we finally want to del. In this case, tmp_alone_branch moves
408 		 * to the prev element but it will point to rq->leaf_cfs_rq_list
409 		 * at the end of the enqueue.
410 		 */
411 		if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
412 			rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
413 
414 		list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
415 		cfs_rq->on_list = 0;
416 	}
417 }
418 
assert_list_leaf_cfs_rq(struct rq * rq)419 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
420 {
421 	SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
422 }
423 
424 /* Iterate thr' all leaf cfs_rq's on a runqueue */
425 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)			\
426 	list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list,	\
427 				 leaf_cfs_rq_list)
428 
429 /* Do the two (enqueued) entities belong to the same group ? */
430 static inline struct cfs_rq *
is_same_group(struct sched_entity * se,struct sched_entity * pse)431 is_same_group(struct sched_entity *se, struct sched_entity *pse)
432 {
433 	if (se->cfs_rq == pse->cfs_rq)
434 		return se->cfs_rq;
435 
436 	return NULL;
437 }
438 
parent_entity(const struct sched_entity * se)439 static inline struct sched_entity *parent_entity(const struct sched_entity *se)
440 {
441 	return se->parent;
442 }
443 
444 static void
find_matching_se(struct sched_entity ** se,struct sched_entity ** pse)445 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
446 {
447 	int se_depth, pse_depth;
448 
449 	/*
450 	 * preemption test can be made between sibling entities who are in the
451 	 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
452 	 * both tasks until we find their ancestors who are siblings of common
453 	 * parent.
454 	 */
455 
456 	/* First walk up until both entities are at same depth */
457 	se_depth = (*se)->depth;
458 	pse_depth = (*pse)->depth;
459 
460 	while (se_depth > pse_depth) {
461 		se_depth--;
462 		*se = parent_entity(*se);
463 	}
464 
465 	while (pse_depth > se_depth) {
466 		pse_depth--;
467 		*pse = parent_entity(*pse);
468 	}
469 
470 	while (!is_same_group(*se, *pse)) {
471 		*se = parent_entity(*se);
472 		*pse = parent_entity(*pse);
473 	}
474 }
475 
tg_is_idle(struct task_group * tg)476 static int tg_is_idle(struct task_group *tg)
477 {
478 	return tg->idle > 0;
479 }
480 
cfs_rq_is_idle(struct cfs_rq * cfs_rq)481 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
482 {
483 	return cfs_rq->idle > 0;
484 }
485 
se_is_idle(struct sched_entity * se)486 static int se_is_idle(struct sched_entity *se)
487 {
488 	if (entity_is_task(se))
489 		return task_has_idle_policy(task_of(se));
490 	return cfs_rq_is_idle(group_cfs_rq(se));
491 }
492 
493 #else	/* !CONFIG_FAIR_GROUP_SCHED */
494 
495 #define for_each_sched_entity(se) \
496 		for (; se; se = NULL)
497 
list_add_leaf_cfs_rq(struct cfs_rq * cfs_rq)498 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
499 {
500 	return true;
501 }
502 
list_del_leaf_cfs_rq(struct cfs_rq * cfs_rq)503 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
504 {
505 }
506 
assert_list_leaf_cfs_rq(struct rq * rq)507 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
508 {
509 }
510 
511 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)	\
512 		for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
513 
parent_entity(struct sched_entity * se)514 static inline struct sched_entity *parent_entity(struct sched_entity *se)
515 {
516 	return NULL;
517 }
518 
519 static inline void
find_matching_se(struct sched_entity ** se,struct sched_entity ** pse)520 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
521 {
522 }
523 
tg_is_idle(struct task_group * tg)524 static inline int tg_is_idle(struct task_group *tg)
525 {
526 	return 0;
527 }
528 
cfs_rq_is_idle(struct cfs_rq * cfs_rq)529 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
530 {
531 	return 0;
532 }
533 
se_is_idle(struct sched_entity * se)534 static int se_is_idle(struct sched_entity *se)
535 {
536 	return 0;
537 }
538 
539 #endif	/* CONFIG_FAIR_GROUP_SCHED */
540 
541 static __always_inline
542 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
543 
544 /**************************************************************
545  * Scheduling class tree data structure manipulation methods:
546  */
547 
max_vruntime(u64 max_vruntime,u64 vruntime)548 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
549 {
550 	s64 delta = (s64)(vruntime - max_vruntime);
551 	if (delta > 0)
552 		max_vruntime = vruntime;
553 
554 	return max_vruntime;
555 }
556 
min_vruntime(u64 min_vruntime,u64 vruntime)557 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
558 {
559 	s64 delta = (s64)(vruntime - min_vruntime);
560 	if (delta < 0)
561 		min_vruntime = vruntime;
562 
563 	return min_vruntime;
564 }
565 
entity_before(const struct sched_entity * a,const struct sched_entity * b)566 static inline bool entity_before(const struct sched_entity *a,
567 				 const struct sched_entity *b)
568 {
569 	return (s64)(a->vruntime - b->vruntime) < 0;
570 }
571 
entity_key(struct cfs_rq * cfs_rq,struct sched_entity * se)572 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
573 {
574 	return (s64)(se->vruntime - cfs_rq->min_vruntime);
575 }
576 
577 #define __node_2_se(node) \
578 	rb_entry((node), struct sched_entity, run_node)
579 
580 /*
581  * Compute virtual time from the per-task service numbers:
582  *
583  * Fair schedulers conserve lag:
584  *
585  *   \Sum lag_i = 0
586  *
587  * Where lag_i is given by:
588  *
589  *   lag_i = S - s_i = w_i * (V - v_i)
590  *
591  * Where S is the ideal service time and V is it's virtual time counterpart.
592  * Therefore:
593  *
594  *   \Sum lag_i = 0
595  *   \Sum w_i * (V - v_i) = 0
596  *   \Sum w_i * V - w_i * v_i = 0
597  *
598  * From which we can solve an expression for V in v_i (which we have in
599  * se->vruntime):
600  *
601  *       \Sum v_i * w_i   \Sum v_i * w_i
602  *   V = -------------- = --------------
603  *          \Sum w_i            W
604  *
605  * Specifically, this is the weighted average of all entity virtual runtimes.
606  *
607  * [[ NOTE: this is only equal to the ideal scheduler under the condition
608  *          that join/leave operations happen at lag_i = 0, otherwise the
609  *          virtual time has non-continguous motion equivalent to:
610  *
611  *	      V +-= lag_i / W
612  *
613  *	    Also see the comment in place_entity() that deals with this. ]]
614  *
615  * However, since v_i is u64, and the multiplcation could easily overflow
616  * transform it into a relative form that uses smaller quantities:
617  *
618  * Substitute: v_i == (v_i - v0) + v0
619  *
620  *     \Sum ((v_i - v0) + v0) * w_i   \Sum (v_i - v0) * w_i
621  * V = ---------------------------- = --------------------- + v0
622  *                  W                            W
623  *
624  * Which we track using:
625  *
626  *                    v0 := cfs_rq->min_vruntime
627  * \Sum (v_i - v0) * w_i := cfs_rq->avg_vruntime
628  *              \Sum w_i := cfs_rq->avg_load
629  *
630  * Since min_vruntime is a monotonic increasing variable that closely tracks
631  * the per-task service, these deltas: (v_i - v), will be in the order of the
632  * maximal (virtual) lag induced in the system due to quantisation.
633  *
634  * Also, we use scale_load_down() to reduce the size.
635  *
636  * As measured, the max (key * weight) value was ~44 bits for a kernel build.
637  */
638 static void
avg_vruntime_add(struct cfs_rq * cfs_rq,struct sched_entity * se)639 avg_vruntime_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
640 {
641 	unsigned long weight = scale_load_down(se->load.weight);
642 	s64 key = entity_key(cfs_rq, se);
643 
644 	cfs_rq->avg_vruntime += key * weight;
645 	cfs_rq->avg_load += weight;
646 }
647 
648 static void
avg_vruntime_sub(struct cfs_rq * cfs_rq,struct sched_entity * se)649 avg_vruntime_sub(struct cfs_rq *cfs_rq, struct sched_entity *se)
650 {
651 	unsigned long weight = scale_load_down(se->load.weight);
652 	s64 key = entity_key(cfs_rq, se);
653 
654 	cfs_rq->avg_vruntime -= key * weight;
655 	cfs_rq->avg_load -= weight;
656 }
657 
658 static inline
avg_vruntime_update(struct cfs_rq * cfs_rq,s64 delta)659 void avg_vruntime_update(struct cfs_rq *cfs_rq, s64 delta)
660 {
661 	/*
662 	 * v' = v + d ==> avg_vruntime' = avg_runtime - d*avg_load
663 	 */
664 	cfs_rq->avg_vruntime -= cfs_rq->avg_load * delta;
665 }
666 
667 /*
668  * Specifically: avg_runtime() + 0 must result in entity_eligible() := true
669  * For this to be so, the result of this function must have a left bias.
670  */
avg_vruntime(struct cfs_rq * cfs_rq)671 u64 avg_vruntime(struct cfs_rq *cfs_rq)
672 {
673 	struct sched_entity *curr = cfs_rq->curr;
674 	s64 avg = cfs_rq->avg_vruntime;
675 	long load = cfs_rq->avg_load;
676 
677 	if (curr && curr->on_rq) {
678 		unsigned long weight = scale_load_down(curr->load.weight);
679 
680 		avg += entity_key(cfs_rq, curr) * weight;
681 		load += weight;
682 	}
683 
684 	if (load) {
685 		/* sign flips effective floor / ceil */
686 		if (avg < 0)
687 			avg -= (load - 1);
688 		avg = div_s64(avg, load);
689 	}
690 
691 	return cfs_rq->min_vruntime + avg;
692 }
693 
694 /*
695  * lag_i = S - s_i = w_i * (V - v_i)
696  *
697  * However, since V is approximated by the weighted average of all entities it
698  * is possible -- by addition/removal/reweight to the tree -- to move V around
699  * and end up with a larger lag than we started with.
700  *
701  * Limit this to either double the slice length with a minimum of TICK_NSEC
702  * since that is the timing granularity.
703  *
704  * EEVDF gives the following limit for a steady state system:
705  *
706  *   -r_max < lag < max(r_max, q)
707  *
708  * XXX could add max_slice to the augmented data to track this.
709  */
update_entity_lag(struct cfs_rq * cfs_rq,struct sched_entity * se)710 static void update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
711 {
712 	s64 lag, limit;
713 
714 	SCHED_WARN_ON(!se->on_rq);
715 	lag = avg_vruntime(cfs_rq) - se->vruntime;
716 
717 	limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se);
718 	se->vlag = clamp(lag, -limit, limit);
719 }
720 
721 /*
722  * Entity is eligible once it received less service than it ought to have,
723  * eg. lag >= 0.
724  *
725  * lag_i = S - s_i = w_i*(V - v_i)
726  *
727  * lag_i >= 0 -> V >= v_i
728  *
729  *     \Sum (v_i - v)*w_i
730  * V = ------------------ + v
731  *          \Sum w_i
732  *
733  * lag_i >= 0 -> \Sum (v_i - v)*w_i >= (v_i - v)*(\Sum w_i)
734  *
735  * Note: using 'avg_vruntime() > se->vruntime' is inacurate due
736  *       to the loss in precision caused by the division.
737  */
entity_eligible(struct cfs_rq * cfs_rq,struct sched_entity * se)738 int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se)
739 {
740 	struct sched_entity *curr = cfs_rq->curr;
741 	s64 avg = cfs_rq->avg_vruntime;
742 	long load = cfs_rq->avg_load;
743 
744 	if (curr && curr->on_rq) {
745 		unsigned long weight = scale_load_down(curr->load.weight);
746 
747 		avg += entity_key(cfs_rq, curr) * weight;
748 		load += weight;
749 	}
750 
751 	return avg >= entity_key(cfs_rq, se) * load;
752 }
753 
__update_min_vruntime(struct cfs_rq * cfs_rq,u64 vruntime)754 static u64 __update_min_vruntime(struct cfs_rq *cfs_rq, u64 vruntime)
755 {
756 	u64 min_vruntime = cfs_rq->min_vruntime;
757 	/*
758 	 * open coded max_vruntime() to allow updating avg_vruntime
759 	 */
760 	s64 delta = (s64)(vruntime - min_vruntime);
761 	if (delta > 0) {
762 		avg_vruntime_update(cfs_rq, delta);
763 		min_vruntime = vruntime;
764 	}
765 	return min_vruntime;
766 }
767 
update_min_vruntime(struct cfs_rq * cfs_rq)768 static void update_min_vruntime(struct cfs_rq *cfs_rq)
769 {
770 	struct sched_entity *se = __pick_first_entity(cfs_rq);
771 	struct sched_entity *curr = cfs_rq->curr;
772 
773 	u64 vruntime = cfs_rq->min_vruntime;
774 
775 	if (curr) {
776 		if (curr->on_rq)
777 			vruntime = curr->vruntime;
778 		else
779 			curr = NULL;
780 	}
781 
782 	if (se) {
783 		if (!curr)
784 			vruntime = se->vruntime;
785 		else
786 			vruntime = min_vruntime(vruntime, se->vruntime);
787 	}
788 
789 	/* ensure we never gain time by being placed backwards. */
790 	u64_u32_store(cfs_rq->min_vruntime,
791 		      __update_min_vruntime(cfs_rq, vruntime));
792 }
793 
__entity_less(struct rb_node * a,const struct rb_node * b)794 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
795 {
796 	return entity_before(__node_2_se(a), __node_2_se(b));
797 }
798 
799 #define deadline_gt(field, lse, rse) ({ (s64)((lse)->field - (rse)->field) > 0; })
800 
__update_min_deadline(struct sched_entity * se,struct rb_node * node)801 static inline void __update_min_deadline(struct sched_entity *se, struct rb_node *node)
802 {
803 	if (node) {
804 		struct sched_entity *rse = __node_2_se(node);
805 		if (deadline_gt(min_deadline, se, rse))
806 			se->min_deadline = rse->min_deadline;
807 	}
808 }
809 
810 /*
811  * se->min_deadline = min(se->deadline, left->min_deadline, right->min_deadline)
812  */
min_deadline_update(struct sched_entity * se,bool exit)813 static inline bool min_deadline_update(struct sched_entity *se, bool exit)
814 {
815 	u64 old_min_deadline = se->min_deadline;
816 	struct rb_node *node = &se->run_node;
817 
818 	se->min_deadline = se->deadline;
819 	__update_min_deadline(se, node->rb_right);
820 	__update_min_deadline(se, node->rb_left);
821 
822 	return se->min_deadline == old_min_deadline;
823 }
824 
825 RB_DECLARE_CALLBACKS(static, min_deadline_cb, struct sched_entity,
826 		     run_node, min_deadline, min_deadline_update);
827 
828 /*
829  * Enqueue an entity into the rb-tree:
830  */
__enqueue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se)831 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
832 {
833 	avg_vruntime_add(cfs_rq, se);
834 	se->min_deadline = se->deadline;
835 	rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
836 				__entity_less, &min_deadline_cb);
837 }
838 
__dequeue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se)839 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
840 {
841 	rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
842 				  &min_deadline_cb);
843 	avg_vruntime_sub(cfs_rq, se);
844 }
845 
__pick_first_entity(struct cfs_rq * cfs_rq)846 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
847 {
848 	struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
849 
850 	if (!left)
851 		return NULL;
852 
853 	return __node_2_se(left);
854 }
855 
856 /*
857  * Earliest Eligible Virtual Deadline First
858  *
859  * In order to provide latency guarantees for different request sizes
860  * EEVDF selects the best runnable task from two criteria:
861  *
862  *  1) the task must be eligible (must be owed service)
863  *
864  *  2) from those tasks that meet 1), we select the one
865  *     with the earliest virtual deadline.
866  *
867  * We can do this in O(log n) time due to an augmented RB-tree. The
868  * tree keeps the entries sorted on service, but also functions as a
869  * heap based on the deadline by keeping:
870  *
871  *  se->min_deadline = min(se->deadline, se->{left,right}->min_deadline)
872  *
873  * Which allows an EDF like search on (sub)trees.
874  */
__pick_eevdf(struct cfs_rq * cfs_rq)875 static struct sched_entity *__pick_eevdf(struct cfs_rq *cfs_rq)
876 {
877 	struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node;
878 	struct sched_entity *curr = cfs_rq->curr;
879 	struct sched_entity *best = NULL;
880 	struct sched_entity *best_left = NULL;
881 
882 	if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))
883 		curr = NULL;
884 	best = curr;
885 
886 	/*
887 	 * Once selected, run a task until it either becomes non-eligible or
888 	 * until it gets a new slice. See the HACK in set_next_entity().
889 	 */
890 	if (sched_feat(RUN_TO_PARITY) && curr && curr->vlag == curr->deadline)
891 		return curr;
892 
893 	while (node) {
894 		struct sched_entity *se = __node_2_se(node);
895 
896 		/*
897 		 * If this entity is not eligible, try the left subtree.
898 		 */
899 		if (!entity_eligible(cfs_rq, se)) {
900 			node = node->rb_left;
901 			continue;
902 		}
903 
904 		/*
905 		 * Now we heap search eligible trees for the best (min_)deadline
906 		 */
907 		if (!best || deadline_gt(deadline, best, se))
908 			best = se;
909 
910 		/*
911 		 * Every se in a left branch is eligible, keep track of the
912 		 * branch with the best min_deadline
913 		 */
914 		if (node->rb_left) {
915 			struct sched_entity *left = __node_2_se(node->rb_left);
916 
917 			if (!best_left || deadline_gt(min_deadline, best_left, left))
918 				best_left = left;
919 
920 			/*
921 			 * min_deadline is in the left branch. rb_left and all
922 			 * descendants are eligible, so immediately switch to the second
923 			 * loop.
924 			 */
925 			if (left->min_deadline == se->min_deadline)
926 				break;
927 		}
928 
929 		/* min_deadline is at this node, no need to look right */
930 		if (se->deadline == se->min_deadline)
931 			break;
932 
933 		/* else min_deadline is in the right branch. */
934 		node = node->rb_right;
935 	}
936 
937 	/*
938 	 * We ran into an eligible node which is itself the best.
939 	 * (Or nr_running == 0 and both are NULL)
940 	 */
941 	if (!best_left || (s64)(best_left->min_deadline - best->deadline) > 0)
942 		return best;
943 
944 	/*
945 	 * Now best_left and all of its children are eligible, and we are just
946 	 * looking for deadline == min_deadline
947 	 */
948 	node = &best_left->run_node;
949 	while (node) {
950 		struct sched_entity *se = __node_2_se(node);
951 
952 		/* min_deadline is the current node */
953 		if (se->deadline == se->min_deadline)
954 			return se;
955 
956 		/* min_deadline is in the left branch */
957 		if (node->rb_left &&
958 		    __node_2_se(node->rb_left)->min_deadline == se->min_deadline) {
959 			node = node->rb_left;
960 			continue;
961 		}
962 
963 		/* else min_deadline is in the right branch */
964 		node = node->rb_right;
965 	}
966 	return NULL;
967 }
968 
pick_eevdf(struct cfs_rq * cfs_rq)969 static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq)
970 {
971 	struct sched_entity *se = __pick_eevdf(cfs_rq);
972 
973 	if (!se) {
974 		struct sched_entity *left = __pick_first_entity(cfs_rq);
975 		if (left) {
976 			pr_err("EEVDF scheduling fail, picking leftmost\n");
977 			return left;
978 		}
979 	}
980 
981 	return se;
982 }
983 
984 #ifdef CONFIG_SCHED_DEBUG
__pick_last_entity(struct cfs_rq * cfs_rq)985 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
986 {
987 	struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
988 
989 	if (!last)
990 		return NULL;
991 
992 	return __node_2_se(last);
993 }
994 
995 /**************************************************************
996  * Scheduling class statistics methods:
997  */
998 #ifdef CONFIG_SMP
sched_update_scaling(void)999 int sched_update_scaling(void)
1000 {
1001 	unsigned int factor = get_update_sysctl_factor();
1002 
1003 #define WRT_SYSCTL(name) \
1004 	(normalized_sysctl_##name = sysctl_##name / (factor))
1005 	WRT_SYSCTL(sched_base_slice);
1006 #undef WRT_SYSCTL
1007 
1008 	return 0;
1009 }
1010 #endif
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 void update_deadline(struct cfs_rq *cfs_rq, struct sched_entity *se)
1020 {
1021 	if ((s64)(se->vruntime - se->deadline) < 0)
1022 		return;
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 	se->slice = sysctl_sched_base_slice;
1030 
1031 	/*
1032 	 * EEVDF: vd_i = ve_i + r_i / w_i
1033 	 */
1034 	se->deadline = se->vruntime + calc_delta_fair(se->slice, se);
1035 
1036 	/*
1037 	 * The task has consumed its request, reschedule.
1038 	 */
1039 	if (cfs_rq->nr_running > 1) {
1040 		resched_curr(rq_of(cfs_rq));
1041 		clear_buddies(cfs_rq, se);
1042 	}
1043 }
1044 
1045 #include "pelt.h"
1046 #ifdef CONFIG_SMP
1047 
1048 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
1049 static unsigned long task_h_load(struct task_struct *p);
1050 static unsigned long capacity_of(int cpu);
1051 
1052 /* Give new sched_entity start runnable values to heavy its load in infant time */
init_entity_runnable_average(struct sched_entity * se)1053 void init_entity_runnable_average(struct sched_entity *se)
1054 {
1055 	struct sched_avg *sa = &se->avg;
1056 
1057 	memset(sa, 0, sizeof(*sa));
1058 
1059 	/*
1060 	 * Tasks are initialized with full load to be seen as heavy tasks until
1061 	 * they get a chance to stabilize to their real load level.
1062 	 * Group entities are initialized with zero load to reflect the fact that
1063 	 * nothing has been attached to the task group yet.
1064 	 */
1065 	if (entity_is_task(se))
1066 		sa->load_avg = scale_load_down(se->load.weight);
1067 
1068 	/* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
1069 }
1070 
1071 /*
1072  * With new tasks being created, their initial util_avgs are extrapolated
1073  * based on the cfs_rq's current util_avg:
1074  *
1075  *   util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
1076  *
1077  * However, in many cases, the above util_avg does not give a desired
1078  * value. Moreover, the sum of the util_avgs may be divergent, such
1079  * as when the series is a harmonic series.
1080  *
1081  * To solve this problem, we also cap the util_avg of successive tasks to
1082  * only 1/2 of the left utilization budget:
1083  *
1084  *   util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
1085  *
1086  * where n denotes the nth task and cpu_scale the CPU capacity.
1087  *
1088  * For example, for a CPU with 1024 of capacity, a simplest series from
1089  * the beginning would be like:
1090  *
1091  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
1092  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
1093  *
1094  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
1095  * if util_avg > util_avg_cap.
1096  */
post_init_entity_util_avg(struct task_struct * p)1097 void post_init_entity_util_avg(struct task_struct *p)
1098 {
1099 	struct sched_entity *se = &p->se;
1100 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
1101 	struct sched_avg *sa = &se->avg;
1102 	long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
1103 	long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
1104 
1105 	if (p->sched_class != &fair_sched_class) {
1106 		/*
1107 		 * For !fair tasks do:
1108 		 *
1109 		update_cfs_rq_load_avg(now, cfs_rq);
1110 		attach_entity_load_avg(cfs_rq, se);
1111 		switched_from_fair(rq, p);
1112 		 *
1113 		 * such that the next switched_to_fair() has the
1114 		 * expected state.
1115 		 */
1116 		se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
1117 		return;
1118 	}
1119 
1120 	if (cap > 0) {
1121 		if (cfs_rq->avg.util_avg != 0) {
1122 			sa->util_avg  = cfs_rq->avg.util_avg * se->load.weight;
1123 			sa->util_avg /= (cfs_rq->avg.load_avg + 1);
1124 
1125 			if (sa->util_avg > cap)
1126 				sa->util_avg = cap;
1127 		} else {
1128 			sa->util_avg = cap;
1129 		}
1130 	}
1131 
1132 	sa->runnable_avg = sa->util_avg;
1133 }
1134 
1135 #else /* !CONFIG_SMP */
init_entity_runnable_average(struct sched_entity * se)1136 void init_entity_runnable_average(struct sched_entity *se)
1137 {
1138 }
post_init_entity_util_avg(struct task_struct * p)1139 void post_init_entity_util_avg(struct task_struct *p)
1140 {
1141 }
update_tg_load_avg(struct cfs_rq * cfs_rq)1142 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
1143 {
1144 }
1145 #endif /* CONFIG_SMP */
1146 
1147 /*
1148  * Update the current task's runtime statistics.
1149  */
update_curr(struct cfs_rq * cfs_rq)1150 static void update_curr(struct cfs_rq *cfs_rq)
1151 {
1152 	struct sched_entity *curr = cfs_rq->curr;
1153 	u64 now = rq_clock_task(rq_of(cfs_rq));
1154 	u64 delta_exec;
1155 
1156 	if (unlikely(!curr))
1157 		return;
1158 
1159 	delta_exec = now - curr->exec_start;
1160 	if (unlikely((s64)delta_exec <= 0))
1161 		return;
1162 
1163 	curr->exec_start = now;
1164 
1165 	if (schedstat_enabled()) {
1166 		struct sched_statistics *stats;
1167 
1168 		stats = __schedstats_from_se(curr);
1169 		__schedstat_set(stats->exec_max,
1170 				max(delta_exec, stats->exec_max));
1171 	}
1172 
1173 	curr->sum_exec_runtime += delta_exec;
1174 	schedstat_add(cfs_rq->exec_clock, delta_exec);
1175 
1176 	curr->vruntime += calc_delta_fair(delta_exec, curr);
1177 	update_deadline(cfs_rq, curr);
1178 	update_min_vruntime(cfs_rq);
1179 
1180 	if (entity_is_task(curr)) {
1181 		struct task_struct *curtask = task_of(curr);
1182 
1183 		trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
1184 		cgroup_account_cputime(curtask, delta_exec);
1185 		account_group_exec_runtime(curtask, delta_exec);
1186 	}
1187 
1188 	account_cfs_rq_runtime(cfs_rq, delta_exec);
1189 }
1190 
update_curr_fair(struct rq * rq)1191 static void update_curr_fair(struct rq *rq)
1192 {
1193 	update_curr(cfs_rq_of(&rq->curr->se));
1194 }
1195 
1196 static inline void
update_stats_wait_start_fair(struct cfs_rq * cfs_rq,struct sched_entity * se)1197 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1198 {
1199 	struct sched_statistics *stats;
1200 	struct task_struct *p = NULL;
1201 
1202 	if (!schedstat_enabled())
1203 		return;
1204 
1205 	stats = __schedstats_from_se(se);
1206 
1207 	if (entity_is_task(se))
1208 		p = task_of(se);
1209 
1210 	__update_stats_wait_start(rq_of(cfs_rq), p, stats);
1211 }
1212 
1213 static inline void
update_stats_wait_end_fair(struct cfs_rq * cfs_rq,struct sched_entity * se)1214 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1215 {
1216 	struct sched_statistics *stats;
1217 	struct task_struct *p = NULL;
1218 
1219 	if (!schedstat_enabled())
1220 		return;
1221 
1222 	stats = __schedstats_from_se(se);
1223 
1224 	/*
1225 	 * When the sched_schedstat changes from 0 to 1, some sched se
1226 	 * maybe already in the runqueue, the se->statistics.wait_start
1227 	 * will be 0.So it will let the delta wrong. We need to avoid this
1228 	 * scenario.
1229 	 */
1230 	if (unlikely(!schedstat_val(stats->wait_start)))
1231 		return;
1232 
1233 	if (entity_is_task(se))
1234 		p = task_of(se);
1235 
1236 	__update_stats_wait_end(rq_of(cfs_rq), p, stats);
1237 }
1238 
1239 static inline void
update_stats_enqueue_sleeper_fair(struct cfs_rq * cfs_rq,struct sched_entity * se)1240 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1241 {
1242 	struct sched_statistics *stats;
1243 	struct task_struct *tsk = NULL;
1244 
1245 	if (!schedstat_enabled())
1246 		return;
1247 
1248 	stats = __schedstats_from_se(se);
1249 
1250 	if (entity_is_task(se))
1251 		tsk = task_of(se);
1252 
1253 	__update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
1254 }
1255 
1256 /*
1257  * Task is being enqueued - update stats:
1258  */
1259 static inline void
update_stats_enqueue_fair(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)1260 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1261 {
1262 	if (!schedstat_enabled())
1263 		return;
1264 
1265 	/*
1266 	 * Are we enqueueing a waiting task? (for current tasks
1267 	 * a dequeue/enqueue event is a NOP)
1268 	 */
1269 	if (se != cfs_rq->curr)
1270 		update_stats_wait_start_fair(cfs_rq, se);
1271 
1272 	if (flags & ENQUEUE_WAKEUP)
1273 		update_stats_enqueue_sleeper_fair(cfs_rq, se);
1274 }
1275 
1276 static inline void
update_stats_dequeue_fair(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)1277 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1278 {
1279 
1280 	if (!schedstat_enabled())
1281 		return;
1282 
1283 	/*
1284 	 * Mark the end of the wait period if dequeueing a
1285 	 * waiting task:
1286 	 */
1287 	if (se != cfs_rq->curr)
1288 		update_stats_wait_end_fair(cfs_rq, se);
1289 
1290 	if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1291 		struct task_struct *tsk = task_of(se);
1292 		unsigned int state;
1293 
1294 		/* XXX racy against TTWU */
1295 		state = READ_ONCE(tsk->__state);
1296 		if (state & TASK_INTERRUPTIBLE)
1297 			__schedstat_set(tsk->stats.sleep_start,
1298 				      rq_clock(rq_of(cfs_rq)));
1299 		if (state & TASK_UNINTERRUPTIBLE)
1300 			__schedstat_set(tsk->stats.block_start,
1301 				      rq_clock(rq_of(cfs_rq)));
1302 	}
1303 }
1304 
1305 /*
1306  * We are picking a new current task - update its stats:
1307  */
1308 static inline void
update_stats_curr_start(struct cfs_rq * cfs_rq,struct sched_entity * se)1309 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1310 {
1311 	/*
1312 	 * We are starting a new run period:
1313 	 */
1314 	se->exec_start = rq_clock_task(rq_of(cfs_rq));
1315 }
1316 
1317 /**************************************************
1318  * Scheduling class queueing methods:
1319  */
1320 
is_core_idle(int cpu)1321 static inline bool is_core_idle(int cpu)
1322 {
1323 #ifdef CONFIG_SCHED_SMT
1324 	int sibling;
1325 
1326 	for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1327 		if (cpu == sibling)
1328 			continue;
1329 
1330 		if (!idle_cpu(sibling))
1331 			return false;
1332 	}
1333 #endif
1334 
1335 	return true;
1336 }
1337 
1338 #ifdef CONFIG_NUMA
1339 #define NUMA_IMBALANCE_MIN 2
1340 
1341 static inline long
adjust_numa_imbalance(int imbalance,int dst_running,int imb_numa_nr)1342 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1343 {
1344 	/*
1345 	 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1346 	 * threshold. Above this threshold, individual tasks may be contending
1347 	 * for both memory bandwidth and any shared HT resources.  This is an
1348 	 * approximation as the number of running tasks may not be related to
1349 	 * the number of busy CPUs due to sched_setaffinity.
1350 	 */
1351 	if (dst_running > imb_numa_nr)
1352 		return imbalance;
1353 
1354 	/*
1355 	 * Allow a small imbalance based on a simple pair of communicating
1356 	 * tasks that remain local when the destination is lightly loaded.
1357 	 */
1358 	if (imbalance <= NUMA_IMBALANCE_MIN)
1359 		return 0;
1360 
1361 	return imbalance;
1362 }
1363 #endif /* CONFIG_NUMA */
1364 
1365 #ifdef CONFIG_NUMA_BALANCING
1366 /*
1367  * Approximate time to scan a full NUMA task in ms. The task scan period is
1368  * calculated based on the tasks virtual memory size and
1369  * numa_balancing_scan_size.
1370  */
1371 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1372 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1373 
1374 /* Portion of address space to scan in MB */
1375 unsigned int sysctl_numa_balancing_scan_size = 256;
1376 
1377 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1378 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1379 
1380 /* The page with hint page fault latency < threshold in ms is considered hot */
1381 unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
1382 
1383 struct numa_group {
1384 	refcount_t refcount;
1385 
1386 	spinlock_t lock; /* nr_tasks, tasks */
1387 	int nr_tasks;
1388 	pid_t gid;
1389 	int active_nodes;
1390 
1391 	struct rcu_head rcu;
1392 	unsigned long total_faults;
1393 	unsigned long max_faults_cpu;
1394 	/*
1395 	 * faults[] array is split into two regions: faults_mem and faults_cpu.
1396 	 *
1397 	 * Faults_cpu is used to decide whether memory should move
1398 	 * towards the CPU. As a consequence, these stats are weighted
1399 	 * more by CPU use than by memory faults.
1400 	 */
1401 	unsigned long faults[];
1402 };
1403 
1404 /*
1405  * For functions that can be called in multiple contexts that permit reading
1406  * ->numa_group (see struct task_struct for locking rules).
1407  */
deref_task_numa_group(struct task_struct * p)1408 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1409 {
1410 	return rcu_dereference_check(p->numa_group, p == current ||
1411 		(lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1412 }
1413 
deref_curr_numa_group(struct task_struct * p)1414 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1415 {
1416 	return rcu_dereference_protected(p->numa_group, p == current);
1417 }
1418 
1419 static inline unsigned long group_faults_priv(struct numa_group *ng);
1420 static inline unsigned long group_faults_shared(struct numa_group *ng);
1421 
task_nr_scan_windows(struct task_struct * p)1422 static unsigned int task_nr_scan_windows(struct task_struct *p)
1423 {
1424 	unsigned long rss = 0;
1425 	unsigned long nr_scan_pages;
1426 
1427 	/*
1428 	 * Calculations based on RSS as non-present and empty pages are skipped
1429 	 * by the PTE scanner and NUMA hinting faults should be trapped based
1430 	 * on resident pages
1431 	 */
1432 	nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1433 	rss = get_mm_rss(p->mm);
1434 	if (!rss)
1435 		rss = nr_scan_pages;
1436 
1437 	rss = round_up(rss, nr_scan_pages);
1438 	return rss / nr_scan_pages;
1439 }
1440 
1441 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1442 #define MAX_SCAN_WINDOW 2560
1443 
task_scan_min(struct task_struct * p)1444 static unsigned int task_scan_min(struct task_struct *p)
1445 {
1446 	unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1447 	unsigned int scan, floor;
1448 	unsigned int windows = 1;
1449 
1450 	if (scan_size < MAX_SCAN_WINDOW)
1451 		windows = MAX_SCAN_WINDOW / scan_size;
1452 	floor = 1000 / windows;
1453 
1454 	scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1455 	return max_t(unsigned int, floor, scan);
1456 }
1457 
task_scan_start(struct task_struct * p)1458 static unsigned int task_scan_start(struct task_struct *p)
1459 {
1460 	unsigned long smin = task_scan_min(p);
1461 	unsigned long period = smin;
1462 	struct numa_group *ng;
1463 
1464 	/* Scale the maximum scan period with the amount of shared memory. */
1465 	rcu_read_lock();
1466 	ng = rcu_dereference(p->numa_group);
1467 	if (ng) {
1468 		unsigned long shared = group_faults_shared(ng);
1469 		unsigned long private = group_faults_priv(ng);
1470 
1471 		period *= refcount_read(&ng->refcount);
1472 		period *= shared + 1;
1473 		period /= private + shared + 1;
1474 	}
1475 	rcu_read_unlock();
1476 
1477 	return max(smin, period);
1478 }
1479 
task_scan_max(struct task_struct * p)1480 static unsigned int task_scan_max(struct task_struct *p)
1481 {
1482 	unsigned long smin = task_scan_min(p);
1483 	unsigned long smax;
1484 	struct numa_group *ng;
1485 
1486 	/* Watch for min being lower than max due to floor calculations */
1487 	smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1488 
1489 	/* Scale the maximum scan period with the amount of shared memory. */
1490 	ng = deref_curr_numa_group(p);
1491 	if (ng) {
1492 		unsigned long shared = group_faults_shared(ng);
1493 		unsigned long private = group_faults_priv(ng);
1494 		unsigned long period = smax;
1495 
1496 		period *= refcount_read(&ng->refcount);
1497 		period *= shared + 1;
1498 		period /= private + shared + 1;
1499 
1500 		smax = max(smax, period);
1501 	}
1502 
1503 	return max(smin, smax);
1504 }
1505 
account_numa_enqueue(struct rq * rq,struct task_struct * p)1506 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1507 {
1508 	rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1509 	rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1510 }
1511 
account_numa_dequeue(struct rq * rq,struct task_struct * p)1512 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1513 {
1514 	rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1515 	rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1516 }
1517 
1518 /* Shared or private faults. */
1519 #define NR_NUMA_HINT_FAULT_TYPES 2
1520 
1521 /* Memory and CPU locality */
1522 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1523 
1524 /* Averaged statistics, and temporary buffers. */
1525 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1526 
task_numa_group_id(struct task_struct * p)1527 pid_t task_numa_group_id(struct task_struct *p)
1528 {
1529 	struct numa_group *ng;
1530 	pid_t gid = 0;
1531 
1532 	rcu_read_lock();
1533 	ng = rcu_dereference(p->numa_group);
1534 	if (ng)
1535 		gid = ng->gid;
1536 	rcu_read_unlock();
1537 
1538 	return gid;
1539 }
1540 
1541 /*
1542  * The averaged statistics, shared & private, memory & CPU,
1543  * occupy the first half of the array. The second half of the
1544  * array is for current counters, which are averaged into the
1545  * first set by task_numa_placement.
1546  */
task_faults_idx(enum numa_faults_stats s,int nid,int priv)1547 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1548 {
1549 	return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1550 }
1551 
task_faults(struct task_struct * p,int nid)1552 static inline unsigned long task_faults(struct task_struct *p, int nid)
1553 {
1554 	if (!p->numa_faults)
1555 		return 0;
1556 
1557 	return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1558 		p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1559 }
1560 
group_faults(struct task_struct * p,int nid)1561 static inline unsigned long group_faults(struct task_struct *p, int nid)
1562 {
1563 	struct numa_group *ng = deref_task_numa_group(p);
1564 
1565 	if (!ng)
1566 		return 0;
1567 
1568 	return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1569 		ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1570 }
1571 
group_faults_cpu(struct numa_group * group,int nid)1572 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1573 {
1574 	return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1575 		group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1576 }
1577 
group_faults_priv(struct numa_group * ng)1578 static inline unsigned long group_faults_priv(struct numa_group *ng)
1579 {
1580 	unsigned long faults = 0;
1581 	int node;
1582 
1583 	for_each_online_node(node) {
1584 		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1585 	}
1586 
1587 	return faults;
1588 }
1589 
group_faults_shared(struct numa_group * ng)1590 static inline unsigned long group_faults_shared(struct numa_group *ng)
1591 {
1592 	unsigned long faults = 0;
1593 	int node;
1594 
1595 	for_each_online_node(node) {
1596 		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1597 	}
1598 
1599 	return faults;
1600 }
1601 
1602 /*
1603  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1604  * considered part of a numa group's pseudo-interleaving set. Migrations
1605  * between these nodes are slowed down, to allow things to settle down.
1606  */
1607 #define ACTIVE_NODE_FRACTION 3
1608 
numa_is_active_node(int nid,struct numa_group * ng)1609 static bool numa_is_active_node(int nid, struct numa_group *ng)
1610 {
1611 	return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1612 }
1613 
1614 /* 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)1615 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1616 					int lim_dist, bool task)
1617 {
1618 	unsigned long score = 0;
1619 	int node, max_dist;
1620 
1621 	/*
1622 	 * All nodes are directly connected, and the same distance
1623 	 * from each other. No need for fancy placement algorithms.
1624 	 */
1625 	if (sched_numa_topology_type == NUMA_DIRECT)
1626 		return 0;
1627 
1628 	/* sched_max_numa_distance may be changed in parallel. */
1629 	max_dist = READ_ONCE(sched_max_numa_distance);
1630 	/*
1631 	 * This code is called for each node, introducing N^2 complexity,
1632 	 * which should be ok given the number of nodes rarely exceeds 8.
1633 	 */
1634 	for_each_online_node(node) {
1635 		unsigned long faults;
1636 		int dist = node_distance(nid, node);
1637 
1638 		/*
1639 		 * The furthest away nodes in the system are not interesting
1640 		 * for placement; nid was already counted.
1641 		 */
1642 		if (dist >= max_dist || node == nid)
1643 			continue;
1644 
1645 		/*
1646 		 * On systems with a backplane NUMA topology, compare groups
1647 		 * of nodes, and move tasks towards the group with the most
1648 		 * memory accesses. When comparing two nodes at distance
1649 		 * "hoplimit", only nodes closer by than "hoplimit" are part
1650 		 * of each group. Skip other nodes.
1651 		 */
1652 		if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1653 			continue;
1654 
1655 		/* Add up the faults from nearby nodes. */
1656 		if (task)
1657 			faults = task_faults(p, node);
1658 		else
1659 			faults = group_faults(p, node);
1660 
1661 		/*
1662 		 * On systems with a glueless mesh NUMA topology, there are
1663 		 * no fixed "groups of nodes". Instead, nodes that are not
1664 		 * directly connected bounce traffic through intermediate
1665 		 * nodes; a numa_group can occupy any set of nodes.
1666 		 * The further away a node is, the less the faults count.
1667 		 * This seems to result in good task placement.
1668 		 */
1669 		if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1670 			faults *= (max_dist - dist);
1671 			faults /= (max_dist - LOCAL_DISTANCE);
1672 		}
1673 
1674 		score += faults;
1675 	}
1676 
1677 	return score;
1678 }
1679 
1680 /*
1681  * These return the fraction of accesses done by a particular task, or
1682  * task group, on a particular numa node.  The group weight is given a
1683  * larger multiplier, in order to group tasks together that are almost
1684  * evenly spread out between numa nodes.
1685  */
task_weight(struct task_struct * p,int nid,int dist)1686 static inline unsigned long task_weight(struct task_struct *p, int nid,
1687 					int dist)
1688 {
1689 	unsigned long faults, total_faults;
1690 
1691 	if (!p->numa_faults)
1692 		return 0;
1693 
1694 	total_faults = p->total_numa_faults;
1695 
1696 	if (!total_faults)
1697 		return 0;
1698 
1699 	faults = task_faults(p, nid);
1700 	faults += score_nearby_nodes(p, nid, dist, true);
1701 
1702 	return 1000 * faults / total_faults;
1703 }
1704 
group_weight(struct task_struct * p,int nid,int dist)1705 static inline unsigned long group_weight(struct task_struct *p, int nid,
1706 					 int dist)
1707 {
1708 	struct numa_group *ng = deref_task_numa_group(p);
1709 	unsigned long faults, total_faults;
1710 
1711 	if (!ng)
1712 		return 0;
1713 
1714 	total_faults = ng->total_faults;
1715 
1716 	if (!total_faults)
1717 		return 0;
1718 
1719 	faults = group_faults(p, nid);
1720 	faults += score_nearby_nodes(p, nid, dist, false);
1721 
1722 	return 1000 * faults / total_faults;
1723 }
1724 
1725 /*
1726  * If memory tiering mode is enabled, cpupid of slow memory page is
1727  * used to record scan time instead of CPU and PID.  When tiering mode
1728  * is disabled at run time, the scan time (in cpupid) will be
1729  * interpreted as CPU and PID.  So CPU needs to be checked to avoid to
1730  * access out of array bound.
1731  */
cpupid_valid(int cpupid)1732 static inline bool cpupid_valid(int cpupid)
1733 {
1734 	return cpupid_to_cpu(cpupid) < nr_cpu_ids;
1735 }
1736 
1737 /*
1738  * For memory tiering mode, if there are enough free pages (more than
1739  * enough watermark defined here) in fast memory node, to take full
1740  * advantage of fast memory capacity, all recently accessed slow
1741  * memory pages will be migrated to fast memory node without
1742  * considering hot threshold.
1743  */
pgdat_free_space_enough(struct pglist_data * pgdat)1744 static bool pgdat_free_space_enough(struct pglist_data *pgdat)
1745 {
1746 	int z;
1747 	unsigned long enough_wmark;
1748 
1749 	enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT,
1750 			   pgdat->node_present_pages >> 4);
1751 	for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1752 		struct zone *zone = pgdat->node_zones + z;
1753 
1754 		if (!populated_zone(zone))
1755 			continue;
1756 
1757 		if (zone_watermark_ok(zone, 0,
1758 				      wmark_pages(zone, WMARK_PROMO) + enough_wmark,
1759 				      ZONE_MOVABLE, 0))
1760 			return true;
1761 	}
1762 	return false;
1763 }
1764 
1765 /*
1766  * For memory tiering mode, when page tables are scanned, the scan
1767  * time will be recorded in struct page in addition to make page
1768  * PROT_NONE for slow memory page.  So when the page is accessed, in
1769  * hint page fault handler, the hint page fault latency is calculated
1770  * via,
1771  *
1772  *	hint page fault latency = hint page fault time - scan time
1773  *
1774  * The smaller the hint page fault latency, the higher the possibility
1775  * for the page to be hot.
1776  */
numa_hint_fault_latency(struct page * page)1777 static int numa_hint_fault_latency(struct page *page)
1778 {
1779 	int last_time, time;
1780 
1781 	time = jiffies_to_msecs(jiffies);
1782 	last_time = xchg_page_access_time(page, time);
1783 
1784 	return (time - last_time) & PAGE_ACCESS_TIME_MASK;
1785 }
1786 
1787 /*
1788  * For memory tiering mode, too high promotion/demotion throughput may
1789  * hurt application latency.  So we provide a mechanism to rate limit
1790  * the number of pages that are tried to be promoted.
1791  */
numa_promotion_rate_limit(struct pglist_data * pgdat,unsigned long rate_limit,int nr)1792 static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
1793 				      unsigned long rate_limit, int nr)
1794 {
1795 	unsigned long nr_cand;
1796 	unsigned int now, start;
1797 
1798 	now = jiffies_to_msecs(jiffies);
1799 	mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr);
1800 	nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1801 	start = pgdat->nbp_rl_start;
1802 	if (now - start > MSEC_PER_SEC &&
1803 	    cmpxchg(&pgdat->nbp_rl_start, start, now) == start)
1804 		pgdat->nbp_rl_nr_cand = nr_cand;
1805 	if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)
1806 		return true;
1807 	return false;
1808 }
1809 
1810 #define NUMA_MIGRATION_ADJUST_STEPS	16
1811 
numa_promotion_adjust_threshold(struct pglist_data * pgdat,unsigned long rate_limit,unsigned int ref_th)1812 static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
1813 					    unsigned long rate_limit,
1814 					    unsigned int ref_th)
1815 {
1816 	unsigned int now, start, th_period, unit_th, th;
1817 	unsigned long nr_cand, ref_cand, diff_cand;
1818 
1819 	now = jiffies_to_msecs(jiffies);
1820 	th_period = sysctl_numa_balancing_scan_period_max;
1821 	start = pgdat->nbp_th_start;
1822 	if (now - start > th_period &&
1823 	    cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
1824 		ref_cand = rate_limit *
1825 			sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
1826 		nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1827 		diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
1828 		unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
1829 		th = pgdat->nbp_threshold ? : ref_th;
1830 		if (diff_cand > ref_cand * 11 / 10)
1831 			th = max(th - unit_th, unit_th);
1832 		else if (diff_cand < ref_cand * 9 / 10)
1833 			th = min(th + unit_th, ref_th * 2);
1834 		pgdat->nbp_th_nr_cand = nr_cand;
1835 		pgdat->nbp_threshold = th;
1836 	}
1837 }
1838 
should_numa_migrate_memory(struct task_struct * p,struct page * page,int src_nid,int dst_cpu)1839 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1840 				int src_nid, int dst_cpu)
1841 {
1842 	struct numa_group *ng = deref_curr_numa_group(p);
1843 	int dst_nid = cpu_to_node(dst_cpu);
1844 	int last_cpupid, this_cpupid;
1845 
1846 	/*
1847 	 * The pages in slow memory node should be migrated according
1848 	 * to hot/cold instead of private/shared.
1849 	 */
1850 	if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1851 	    !node_is_toptier(src_nid)) {
1852 		struct pglist_data *pgdat;
1853 		unsigned long rate_limit;
1854 		unsigned int latency, th, def_th;
1855 
1856 		pgdat = NODE_DATA(dst_nid);
1857 		if (pgdat_free_space_enough(pgdat)) {
1858 			/* workload changed, reset hot threshold */
1859 			pgdat->nbp_threshold = 0;
1860 			return true;
1861 		}
1862 
1863 		def_th = sysctl_numa_balancing_hot_threshold;
1864 		rate_limit = sysctl_numa_balancing_promote_rate_limit << \
1865 			(20 - PAGE_SHIFT);
1866 		numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
1867 
1868 		th = pgdat->nbp_threshold ? : def_th;
1869 		latency = numa_hint_fault_latency(page);
1870 		if (latency >= th)
1871 			return false;
1872 
1873 		return !numa_promotion_rate_limit(pgdat, rate_limit,
1874 						  thp_nr_pages(page));
1875 	}
1876 
1877 	this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1878 	last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1879 
1880 	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1881 	    !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1882 		return false;
1883 
1884 	/*
1885 	 * Allow first faults or private faults to migrate immediately early in
1886 	 * the lifetime of a task. The magic number 4 is based on waiting for
1887 	 * two full passes of the "multi-stage node selection" test that is
1888 	 * executed below.
1889 	 */
1890 	if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1891 	    (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1892 		return true;
1893 
1894 	/*
1895 	 * Multi-stage node selection is used in conjunction with a periodic
1896 	 * migration fault to build a temporal task<->page relation. By using
1897 	 * a two-stage filter we remove short/unlikely relations.
1898 	 *
1899 	 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1900 	 * a task's usage of a particular page (n_p) per total usage of this
1901 	 * page (n_t) (in a given time-span) to a probability.
1902 	 *
1903 	 * Our periodic faults will sample this probability and getting the
1904 	 * same result twice in a row, given these samples are fully
1905 	 * independent, is then given by P(n)^2, provided our sample period
1906 	 * is sufficiently short compared to the usage pattern.
1907 	 *
1908 	 * This quadric squishes small probabilities, making it less likely we
1909 	 * act on an unlikely task<->page relation.
1910 	 */
1911 	if (!cpupid_pid_unset(last_cpupid) &&
1912 				cpupid_to_nid(last_cpupid) != dst_nid)
1913 		return false;
1914 
1915 	/* Always allow migrate on private faults */
1916 	if (cpupid_match_pid(p, last_cpupid))
1917 		return true;
1918 
1919 	/* A shared fault, but p->numa_group has not been set up yet. */
1920 	if (!ng)
1921 		return true;
1922 
1923 	/*
1924 	 * Destination node is much more heavily used than the source
1925 	 * node? Allow migration.
1926 	 */
1927 	if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1928 					ACTIVE_NODE_FRACTION)
1929 		return true;
1930 
1931 	/*
1932 	 * Distribute memory according to CPU & memory use on each node,
1933 	 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1934 	 *
1935 	 * faults_cpu(dst)   3   faults_cpu(src)
1936 	 * --------------- * - > ---------------
1937 	 * faults_mem(dst)   4   faults_mem(src)
1938 	 */
1939 	return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1940 	       group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1941 }
1942 
1943 /*
1944  * 'numa_type' describes the node at the moment of load balancing.
1945  */
1946 enum numa_type {
1947 	/* The node has spare capacity that can be used to run more tasks.  */
1948 	node_has_spare = 0,
1949 	/*
1950 	 * The node is fully used and the tasks don't compete for more CPU
1951 	 * cycles. Nevertheless, some tasks might wait before running.
1952 	 */
1953 	node_fully_busy,
1954 	/*
1955 	 * The node is overloaded and can't provide expected CPU cycles to all
1956 	 * tasks.
1957 	 */
1958 	node_overloaded
1959 };
1960 
1961 /* Cached statistics for all CPUs within a node */
1962 struct numa_stats {
1963 	unsigned long load;
1964 	unsigned long runnable;
1965 	unsigned long util;
1966 	/* Total compute capacity of CPUs on a node */
1967 	unsigned long compute_capacity;
1968 	unsigned int nr_running;
1969 	unsigned int weight;
1970 	enum numa_type node_type;
1971 	int idle_cpu;
1972 };
1973 
1974 struct task_numa_env {
1975 	struct task_struct *p;
1976 
1977 	int src_cpu, src_nid;
1978 	int dst_cpu, dst_nid;
1979 	int imb_numa_nr;
1980 
1981 	struct numa_stats src_stats, dst_stats;
1982 
1983 	int imbalance_pct;
1984 	int dist;
1985 
1986 	struct task_struct *best_task;
1987 	long best_imp;
1988 	int best_cpu;
1989 };
1990 
1991 static unsigned long cpu_load(struct rq *rq);
1992 static unsigned long cpu_runnable(struct rq *rq);
1993 
1994 static inline enum
numa_classify(unsigned int imbalance_pct,struct numa_stats * ns)1995 numa_type numa_classify(unsigned int imbalance_pct,
1996 			 struct numa_stats *ns)
1997 {
1998 	if ((ns->nr_running > ns->weight) &&
1999 	    (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
2000 	     ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
2001 		return node_overloaded;
2002 
2003 	if ((ns->nr_running < ns->weight) ||
2004 	    (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
2005 	     ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
2006 		return node_has_spare;
2007 
2008 	return node_fully_busy;
2009 }
2010 
2011 #ifdef CONFIG_SCHED_SMT
2012 /* Forward declarations of select_idle_sibling helpers */
2013 static inline bool test_idle_cores(int cpu);
numa_idle_core(int idle_core,int cpu)2014 static inline int numa_idle_core(int idle_core, int cpu)
2015 {
2016 	if (!static_branch_likely(&sched_smt_present) ||
2017 	    idle_core >= 0 || !test_idle_cores(cpu))
2018 		return idle_core;
2019 
2020 	/*
2021 	 * Prefer cores instead of packing HT siblings
2022 	 * and triggering future load balancing.
2023 	 */
2024 	if (is_core_idle(cpu))
2025 		idle_core = cpu;
2026 
2027 	return idle_core;
2028 }
2029 #else
numa_idle_core(int idle_core,int cpu)2030 static inline int numa_idle_core(int idle_core, int cpu)
2031 {
2032 	return idle_core;
2033 }
2034 #endif
2035 
2036 /*
2037  * Gather all necessary information to make NUMA balancing placement
2038  * decisions that are compatible with standard load balancer. This
2039  * borrows code and logic from update_sg_lb_stats but sharing a
2040  * common implementation is impractical.
2041  */
update_numa_stats(struct task_numa_env * env,struct numa_stats * ns,int nid,bool find_idle)2042 static void update_numa_stats(struct task_numa_env *env,
2043 			      struct numa_stats *ns, int nid,
2044 			      bool find_idle)
2045 {
2046 	int cpu, idle_core = -1;
2047 
2048 	memset(ns, 0, sizeof(*ns));
2049 	ns->idle_cpu = -1;
2050 
2051 	rcu_read_lock();
2052 	for_each_cpu(cpu, cpumask_of_node(nid)) {
2053 		struct rq *rq = cpu_rq(cpu);
2054 
2055 		ns->load += cpu_load(rq);
2056 		ns->runnable += cpu_runnable(rq);
2057 		ns->util += cpu_util_cfs(cpu);
2058 		ns->nr_running += rq->cfs.h_nr_running;
2059 		ns->compute_capacity += capacity_of(cpu);
2060 
2061 		if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) {
2062 			if (READ_ONCE(rq->numa_migrate_on) ||
2063 			    !cpumask_test_cpu(cpu, env->p->cpus_ptr))
2064 				continue;
2065 
2066 			if (ns->idle_cpu == -1)
2067 				ns->idle_cpu = cpu;
2068 
2069 			idle_core = numa_idle_core(idle_core, cpu);
2070 		}
2071 	}
2072 	rcu_read_unlock();
2073 
2074 	ns->weight = cpumask_weight(cpumask_of_node(nid));
2075 
2076 	ns->node_type = numa_classify(env->imbalance_pct, ns);
2077 
2078 	if (idle_core >= 0)
2079 		ns->idle_cpu = idle_core;
2080 }
2081 
task_numa_assign(struct task_numa_env * env,struct task_struct * p,long imp)2082 static void task_numa_assign(struct task_numa_env *env,
2083 			     struct task_struct *p, long imp)
2084 {
2085 	struct rq *rq = cpu_rq(env->dst_cpu);
2086 
2087 	/* Check if run-queue part of active NUMA balance. */
2088 	if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
2089 		int cpu;
2090 		int start = env->dst_cpu;
2091 
2092 		/* Find alternative idle CPU. */
2093 		for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) {
2094 			if (cpu == env->best_cpu || !idle_cpu(cpu) ||
2095 			    !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
2096 				continue;
2097 			}
2098 
2099 			env->dst_cpu = cpu;
2100 			rq = cpu_rq(env->dst_cpu);
2101 			if (!xchg(&rq->numa_migrate_on, 1))
2102 				goto assign;
2103 		}
2104 
2105 		/* Failed to find an alternative idle CPU */
2106 		return;
2107 	}
2108 
2109 assign:
2110 	/*
2111 	 * Clear previous best_cpu/rq numa-migrate flag, since task now
2112 	 * found a better CPU to move/swap.
2113 	 */
2114 	if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
2115 		rq = cpu_rq(env->best_cpu);
2116 		WRITE_ONCE(rq->numa_migrate_on, 0);
2117 	}
2118 
2119 	if (env->best_task)
2120 		put_task_struct(env->best_task);
2121 	if (p)
2122 		get_task_struct(p);
2123 
2124 	env->best_task = p;
2125 	env->best_imp = imp;
2126 	env->best_cpu = env->dst_cpu;
2127 }
2128 
load_too_imbalanced(long src_load,long dst_load,struct task_numa_env * env)2129 static bool load_too_imbalanced(long src_load, long dst_load,
2130 				struct task_numa_env *env)
2131 {
2132 	long imb, old_imb;
2133 	long orig_src_load, orig_dst_load;
2134 	long src_capacity, dst_capacity;
2135 
2136 	/*
2137 	 * The load is corrected for the CPU capacity available on each node.
2138 	 *
2139 	 * src_load        dst_load
2140 	 * ------------ vs ---------
2141 	 * src_capacity    dst_capacity
2142 	 */
2143 	src_capacity = env->src_stats.compute_capacity;
2144 	dst_capacity = env->dst_stats.compute_capacity;
2145 
2146 	imb = abs(dst_load * src_capacity - src_load * dst_capacity);
2147 
2148 	orig_src_load = env->src_stats.load;
2149 	orig_dst_load = env->dst_stats.load;
2150 
2151 	old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
2152 
2153 	/* Would this change make things worse? */
2154 	return (imb > old_imb);
2155 }
2156 
2157 /*
2158  * Maximum NUMA importance can be 1998 (2*999);
2159  * SMALLIMP @ 30 would be close to 1998/64.
2160  * Used to deter task migration.
2161  */
2162 #define SMALLIMP	30
2163 
2164 /*
2165  * This checks if the overall compute and NUMA accesses of the system would
2166  * be improved if the source tasks was migrated to the target dst_cpu taking
2167  * into account that it might be best if task running on the dst_cpu should
2168  * be exchanged with the source task
2169  */
task_numa_compare(struct task_numa_env * env,long taskimp,long groupimp,bool maymove)2170 static bool task_numa_compare(struct task_numa_env *env,
2171 			      long taskimp, long groupimp, bool maymove)
2172 {
2173 	struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
2174 	struct rq *dst_rq = cpu_rq(env->dst_cpu);
2175 	long imp = p_ng ? groupimp : taskimp;
2176 	struct task_struct *cur;
2177 	long src_load, dst_load;
2178 	int dist = env->dist;
2179 	long moveimp = imp;
2180 	long load;
2181 	bool stopsearch = false;
2182 
2183 	if (READ_ONCE(dst_rq->numa_migrate_on))
2184 		return false;
2185 
2186 	rcu_read_lock();
2187 	cur = rcu_dereference(dst_rq->curr);
2188 	if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
2189 		cur = NULL;
2190 
2191 	/*
2192 	 * Because we have preemption enabled we can get migrated around and
2193 	 * end try selecting ourselves (current == env->p) as a swap candidate.
2194 	 */
2195 	if (cur == env->p) {
2196 		stopsearch = true;
2197 		goto unlock;
2198 	}
2199 
2200 	if (!cur) {
2201 		if (maymove && moveimp >= env->best_imp)
2202 			goto assign;
2203 		else
2204 			goto unlock;
2205 	}
2206 
2207 	/* Skip this swap candidate if cannot move to the source cpu. */
2208 	if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
2209 		goto unlock;
2210 
2211 	/*
2212 	 * Skip this swap candidate if it is not moving to its preferred
2213 	 * node and the best task is.
2214 	 */
2215 	if (env->best_task &&
2216 	    env->best_task->numa_preferred_nid == env->src_nid &&
2217 	    cur->numa_preferred_nid != env->src_nid) {
2218 		goto unlock;
2219 	}
2220 
2221 	/*
2222 	 * "imp" is the fault differential for the source task between the
2223 	 * source and destination node. Calculate the total differential for
2224 	 * the source task and potential destination task. The more negative
2225 	 * the value is, the more remote accesses that would be expected to
2226 	 * be incurred if the tasks were swapped.
2227 	 *
2228 	 * If dst and source tasks are in the same NUMA group, or not
2229 	 * in any group then look only at task weights.
2230 	 */
2231 	cur_ng = rcu_dereference(cur->numa_group);
2232 	if (cur_ng == p_ng) {
2233 		/*
2234 		 * Do not swap within a group or between tasks that have
2235 		 * no group if there is spare capacity. Swapping does
2236 		 * not address the load imbalance and helps one task at
2237 		 * the cost of punishing another.
2238 		 */
2239 		if (env->dst_stats.node_type == node_has_spare)
2240 			goto unlock;
2241 
2242 		imp = taskimp + task_weight(cur, env->src_nid, dist) -
2243 		      task_weight(cur, env->dst_nid, dist);
2244 		/*
2245 		 * Add some hysteresis to prevent swapping the
2246 		 * tasks within a group over tiny differences.
2247 		 */
2248 		if (cur_ng)
2249 			imp -= imp / 16;
2250 	} else {
2251 		/*
2252 		 * Compare the group weights. If a task is all by itself
2253 		 * (not part of a group), use the task weight instead.
2254 		 */
2255 		if (cur_ng && p_ng)
2256 			imp += group_weight(cur, env->src_nid, dist) -
2257 			       group_weight(cur, env->dst_nid, dist);
2258 		else
2259 			imp += task_weight(cur, env->src_nid, dist) -
2260 			       task_weight(cur, env->dst_nid, dist);
2261 	}
2262 
2263 	/* Discourage picking a task already on its preferred node */
2264 	if (cur->numa_preferred_nid == env->dst_nid)
2265 		imp -= imp / 16;
2266 
2267 	/*
2268 	 * Encourage picking a task that moves to its preferred node.
2269 	 * This potentially makes imp larger than it's maximum of
2270 	 * 1998 (see SMALLIMP and task_weight for why) but in this
2271 	 * case, it does not matter.
2272 	 */
2273 	if (cur->numa_preferred_nid == env->src_nid)
2274 		imp += imp / 8;
2275 
2276 	if (maymove && moveimp > imp && moveimp > env->best_imp) {
2277 		imp = moveimp;
2278 		cur = NULL;
2279 		goto assign;
2280 	}
2281 
2282 	/*
2283 	 * Prefer swapping with a task moving to its preferred node over a
2284 	 * task that is not.
2285 	 */
2286 	if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
2287 	    env->best_task->numa_preferred_nid != env->src_nid) {
2288 		goto assign;
2289 	}
2290 
2291 	/*
2292 	 * If the NUMA importance is less than SMALLIMP,
2293 	 * task migration might only result in ping pong
2294 	 * of tasks and also hurt performance due to cache
2295 	 * misses.
2296 	 */
2297 	if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
2298 		goto unlock;
2299 
2300 	/*
2301 	 * In the overloaded case, try and keep the load balanced.
2302 	 */
2303 	load = task_h_load(env->p) - task_h_load(cur);
2304 	if (!load)
2305 		goto assign;
2306 
2307 	dst_load = env->dst_stats.load + load;
2308 	src_load = env->src_stats.load - load;
2309 
2310 	if (load_too_imbalanced(src_load, dst_load, env))
2311 		goto unlock;
2312 
2313 assign:
2314 	/* Evaluate an idle CPU for a task numa move. */
2315 	if (!cur) {
2316 		int cpu = env->dst_stats.idle_cpu;
2317 
2318 		/* Nothing cached so current CPU went idle since the search. */
2319 		if (cpu < 0)
2320 			cpu = env->dst_cpu;
2321 
2322 		/*
2323 		 * If the CPU is no longer truly idle and the previous best CPU
2324 		 * is, keep using it.
2325 		 */
2326 		if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2327 		    idle_cpu(env->best_cpu)) {
2328 			cpu = env->best_cpu;
2329 		}
2330 
2331 		env->dst_cpu = cpu;
2332 	}
2333 
2334 	task_numa_assign(env, cur, imp);
2335 
2336 	/*
2337 	 * If a move to idle is allowed because there is capacity or load
2338 	 * balance improves then stop the search. While a better swap
2339 	 * candidate may exist, a search is not free.
2340 	 */
2341 	if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
2342 		stopsearch = true;
2343 
2344 	/*
2345 	 * If a swap candidate must be identified and the current best task
2346 	 * moves its preferred node then stop the search.
2347 	 */
2348 	if (!maymove && env->best_task &&
2349 	    env->best_task->numa_preferred_nid == env->src_nid) {
2350 		stopsearch = true;
2351 	}
2352 unlock:
2353 	rcu_read_unlock();
2354 
2355 	return stopsearch;
2356 }
2357 
task_numa_find_cpu(struct task_numa_env * env,long taskimp,long groupimp)2358 static void task_numa_find_cpu(struct task_numa_env *env,
2359 				long taskimp, long groupimp)
2360 {
2361 	bool maymove = false;
2362 	int cpu;
2363 
2364 	/*
2365 	 * If dst node has spare capacity, then check if there is an
2366 	 * imbalance that would be overruled by the load balancer.
2367 	 */
2368 	if (env->dst_stats.node_type == node_has_spare) {
2369 		unsigned int imbalance;
2370 		int src_running, dst_running;
2371 
2372 		/*
2373 		 * Would movement cause an imbalance? Note that if src has
2374 		 * more running tasks that the imbalance is ignored as the
2375 		 * move improves the imbalance from the perspective of the
2376 		 * CPU load balancer.
2377 		 * */
2378 		src_running = env->src_stats.nr_running - 1;
2379 		dst_running = env->dst_stats.nr_running + 1;
2380 		imbalance = max(0, dst_running - src_running);
2381 		imbalance = adjust_numa_imbalance(imbalance, dst_running,
2382 						  env->imb_numa_nr);
2383 
2384 		/* Use idle CPU if there is no imbalance */
2385 		if (!imbalance) {
2386 			maymove = true;
2387 			if (env->dst_stats.idle_cpu >= 0) {
2388 				env->dst_cpu = env->dst_stats.idle_cpu;
2389 				task_numa_assign(env, NULL, 0);
2390 				return;
2391 			}
2392 		}
2393 	} else {
2394 		long src_load, dst_load, load;
2395 		/*
2396 		 * If the improvement from just moving env->p direction is better
2397 		 * than swapping tasks around, check if a move is possible.
2398 		 */
2399 		load = task_h_load(env->p);
2400 		dst_load = env->dst_stats.load + load;
2401 		src_load = env->src_stats.load - load;
2402 		maymove = !load_too_imbalanced(src_load, dst_load, env);
2403 	}
2404 
2405 	for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2406 		/* Skip this CPU if the source task cannot migrate */
2407 		if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2408 			continue;
2409 
2410 		env->dst_cpu = cpu;
2411 		if (task_numa_compare(env, taskimp, groupimp, maymove))
2412 			break;
2413 	}
2414 }
2415 
task_numa_migrate(struct task_struct * p)2416 static int task_numa_migrate(struct task_struct *p)
2417 {
2418 	struct task_numa_env env = {
2419 		.p = p,
2420 
2421 		.src_cpu = task_cpu(p),
2422 		.src_nid = task_node(p),
2423 
2424 		.imbalance_pct = 112,
2425 
2426 		.best_task = NULL,
2427 		.best_imp = 0,
2428 		.best_cpu = -1,
2429 	};
2430 	unsigned long taskweight, groupweight;
2431 	struct sched_domain *sd;
2432 	long taskimp, groupimp;
2433 	struct numa_group *ng;
2434 	struct rq *best_rq;
2435 	int nid, ret, dist;
2436 
2437 	/*
2438 	 * Pick the lowest SD_NUMA domain, as that would have the smallest
2439 	 * imbalance and would be the first to start moving tasks about.
2440 	 *
2441 	 * And we want to avoid any moving of tasks about, as that would create
2442 	 * random movement of tasks -- counter the numa conditions we're trying
2443 	 * to satisfy here.
2444 	 */
2445 	rcu_read_lock();
2446 	sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2447 	if (sd) {
2448 		env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2449 		env.imb_numa_nr = sd->imb_numa_nr;
2450 	}
2451 	rcu_read_unlock();
2452 
2453 	/*
2454 	 * Cpusets can break the scheduler domain tree into smaller
2455 	 * balance domains, some of which do not cross NUMA boundaries.
2456 	 * Tasks that are "trapped" in such domains cannot be migrated
2457 	 * elsewhere, so there is no point in (re)trying.
2458 	 */
2459 	if (unlikely(!sd)) {
2460 		sched_setnuma(p, task_node(p));
2461 		return -EINVAL;
2462 	}
2463 
2464 	env.dst_nid = p->numa_preferred_nid;
2465 	dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2466 	taskweight = task_weight(p, env.src_nid, dist);
2467 	groupweight = group_weight(p, env.src_nid, dist);
2468 	update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2469 	taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2470 	groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2471 	update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2472 
2473 	/* Try to find a spot on the preferred nid. */
2474 	task_numa_find_cpu(&env, taskimp, groupimp);
2475 
2476 	/*
2477 	 * Look at other nodes in these cases:
2478 	 * - there is no space available on the preferred_nid
2479 	 * - the task is part of a numa_group that is interleaved across
2480 	 *   multiple NUMA nodes; in order to better consolidate the group,
2481 	 *   we need to check other locations.
2482 	 */
2483 	ng = deref_curr_numa_group(p);
2484 	if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2485 		for_each_node_state(nid, N_CPU) {
2486 			if (nid == env.src_nid || nid == p->numa_preferred_nid)
2487 				continue;
2488 
2489 			dist = node_distance(env.src_nid, env.dst_nid);
2490 			if (sched_numa_topology_type == NUMA_BACKPLANE &&
2491 						dist != env.dist) {
2492 				taskweight = task_weight(p, env.src_nid, dist);
2493 				groupweight = group_weight(p, env.src_nid, dist);
2494 			}
2495 
2496 			/* Only consider nodes where both task and groups benefit */
2497 			taskimp = task_weight(p, nid, dist) - taskweight;
2498 			groupimp = group_weight(p, nid, dist) - groupweight;
2499 			if (taskimp < 0 && groupimp < 0)
2500 				continue;
2501 
2502 			env.dist = dist;
2503 			env.dst_nid = nid;
2504 			update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2505 			task_numa_find_cpu(&env, taskimp, groupimp);
2506 		}
2507 	}
2508 
2509 	/*
2510 	 * If the task is part of a workload that spans multiple NUMA nodes,
2511 	 * and is migrating into one of the workload's active nodes, remember
2512 	 * this node as the task's preferred numa node, so the workload can
2513 	 * settle down.
2514 	 * A task that migrated to a second choice node will be better off
2515 	 * trying for a better one later. Do not set the preferred node here.
2516 	 */
2517 	if (ng) {
2518 		if (env.best_cpu == -1)
2519 			nid = env.src_nid;
2520 		else
2521 			nid = cpu_to_node(env.best_cpu);
2522 
2523 		if (nid != p->numa_preferred_nid)
2524 			sched_setnuma(p, nid);
2525 	}
2526 
2527 	/* No better CPU than the current one was found. */
2528 	if (env.best_cpu == -1) {
2529 		trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2530 		return -EAGAIN;
2531 	}
2532 
2533 	best_rq = cpu_rq(env.best_cpu);
2534 	if (env.best_task == NULL) {
2535 		ret = migrate_task_to(p, env.best_cpu);
2536 		WRITE_ONCE(best_rq->numa_migrate_on, 0);
2537 		if (ret != 0)
2538 			trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2539 		return ret;
2540 	}
2541 
2542 	ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2543 	WRITE_ONCE(best_rq->numa_migrate_on, 0);
2544 
2545 	if (ret != 0)
2546 		trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2547 	put_task_struct(env.best_task);
2548 	return ret;
2549 }
2550 
2551 /* Attempt to migrate a task to a CPU on the preferred node. */
numa_migrate_preferred(struct task_struct * p)2552 static void numa_migrate_preferred(struct task_struct *p)
2553 {
2554 	unsigned long interval = HZ;
2555 
2556 	/* This task has no NUMA fault statistics yet */
2557 	if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2558 		return;
2559 
2560 	/* Periodically retry migrating the task to the preferred node */
2561 	interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2562 	p->numa_migrate_retry = jiffies + interval;
2563 
2564 	/* Success if task is already running on preferred CPU */
2565 	if (task_node(p) == p->numa_preferred_nid)
2566 		return;
2567 
2568 	/* Otherwise, try migrate to a CPU on the preferred node */
2569 	task_numa_migrate(p);
2570 }
2571 
2572 /*
2573  * Find out how many nodes the workload is actively running on. Do this by
2574  * tracking the nodes from which NUMA hinting faults are triggered. This can
2575  * be different from the set of nodes where the workload's memory is currently
2576  * located.
2577  */
numa_group_count_active_nodes(struct numa_group * numa_group)2578 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2579 {
2580 	unsigned long faults, max_faults = 0;
2581 	int nid, active_nodes = 0;
2582 
2583 	for_each_node_state(nid, N_CPU) {
2584 		faults = group_faults_cpu(numa_group, nid);
2585 		if (faults > max_faults)
2586 			max_faults = faults;
2587 	}
2588 
2589 	for_each_node_state(nid, N_CPU) {
2590 		faults = group_faults_cpu(numa_group, nid);
2591 		if (faults * ACTIVE_NODE_FRACTION > max_faults)
2592 			active_nodes++;
2593 	}
2594 
2595 	numa_group->max_faults_cpu = max_faults;
2596 	numa_group->active_nodes = active_nodes;
2597 }
2598 
2599 /*
2600  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2601  * increments. The more local the fault statistics are, the higher the scan
2602  * period will be for the next scan window. If local/(local+remote) ratio is
2603  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2604  * the scan period will decrease. Aim for 70% local accesses.
2605  */
2606 #define NUMA_PERIOD_SLOTS 10
2607 #define NUMA_PERIOD_THRESHOLD 7
2608 
2609 /*
2610  * Increase the scan period (slow down scanning) if the majority of
2611  * our memory is already on our local node, or if the majority of
2612  * the page accesses are shared with other processes.
2613  * Otherwise, decrease the scan period.
2614  */
update_task_scan_period(struct task_struct * p,unsigned long shared,unsigned long private)2615 static void update_task_scan_period(struct task_struct *p,
2616 			unsigned long shared, unsigned long private)
2617 {
2618 	unsigned int period_slot;
2619 	int lr_ratio, ps_ratio;
2620 	int diff;
2621 
2622 	unsigned long remote = p->numa_faults_locality[0];
2623 	unsigned long local = p->numa_faults_locality[1];
2624 
2625 	/*
2626 	 * If there were no record hinting faults then either the task is
2627 	 * completely idle or all activity is in areas that are not of interest
2628 	 * to automatic numa balancing. Related to that, if there were failed
2629 	 * migration then it implies we are migrating too quickly or the local
2630 	 * node is overloaded. In either case, scan slower
2631 	 */
2632 	if (local + shared == 0 || p->numa_faults_locality[2]) {
2633 		p->numa_scan_period = min(p->numa_scan_period_max,
2634 			p->numa_scan_period << 1);
2635 
2636 		p->mm->numa_next_scan = jiffies +
2637 			msecs_to_jiffies(p->numa_scan_period);
2638 
2639 		return;
2640 	}
2641 
2642 	/*
2643 	 * Prepare to scale scan period relative to the current period.
2644 	 *	 == NUMA_PERIOD_THRESHOLD scan period stays the same
2645 	 *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2646 	 *	 >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2647 	 */
2648 	period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2649 	lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2650 	ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2651 
2652 	if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2653 		/*
2654 		 * Most memory accesses are local. There is no need to
2655 		 * do fast NUMA scanning, since memory is already local.
2656 		 */
2657 		int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2658 		if (!slot)
2659 			slot = 1;
2660 		diff = slot * period_slot;
2661 	} else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2662 		/*
2663 		 * Most memory accesses are shared with other tasks.
2664 		 * There is no point in continuing fast NUMA scanning,
2665 		 * since other tasks may just move the memory elsewhere.
2666 		 */
2667 		int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2668 		if (!slot)
2669 			slot = 1;
2670 		diff = slot * period_slot;
2671 	} else {
2672 		/*
2673 		 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2674 		 * yet they are not on the local NUMA node. Speed up
2675 		 * NUMA scanning to get the memory moved over.
2676 		 */
2677 		int ratio = max(lr_ratio, ps_ratio);
2678 		diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2679 	}
2680 
2681 	p->numa_scan_period = clamp(p->numa_scan_period + diff,
2682 			task_scan_min(p), task_scan_max(p));
2683 	memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2684 }
2685 
2686 /*
2687  * Get the fraction of time the task has been running since the last
2688  * NUMA placement cycle. The scheduler keeps similar statistics, but
2689  * decays those on a 32ms period, which is orders of magnitude off
2690  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2691  * stats only if the task is so new there are no NUMA statistics yet.
2692  */
numa_get_avg_runtime(struct task_struct * p,u64 * period)2693 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2694 {
2695 	u64 runtime, delta, now;
2696 	/* Use the start of this time slice to avoid calculations. */
2697 	now = p->se.exec_start;
2698 	runtime = p->se.sum_exec_runtime;
2699 
2700 	if (p->last_task_numa_placement) {
2701 		delta = runtime - p->last_sum_exec_runtime;
2702 		*period = now - p->last_task_numa_placement;
2703 
2704 		/* Avoid time going backwards, prevent potential divide error: */
2705 		if (unlikely((s64)*period < 0))
2706 			*period = 0;
2707 	} else {
2708 		delta = p->se.avg.load_sum;
2709 		*period = LOAD_AVG_MAX;
2710 	}
2711 
2712 	p->last_sum_exec_runtime = runtime;
2713 	p->last_task_numa_placement = now;
2714 
2715 	return delta;
2716 }
2717 
2718 /*
2719  * Determine the preferred nid for a task in a numa_group. This needs to
2720  * be done in a way that produces consistent results with group_weight,
2721  * otherwise workloads might not converge.
2722  */
preferred_group_nid(struct task_struct * p,int nid)2723 static int preferred_group_nid(struct task_struct *p, int nid)
2724 {
2725 	nodemask_t nodes;
2726 	int dist;
2727 
2728 	/* Direct connections between all NUMA nodes. */
2729 	if (sched_numa_topology_type == NUMA_DIRECT)
2730 		return nid;
2731 
2732 	/*
2733 	 * On a system with glueless mesh NUMA topology, group_weight
2734 	 * scores nodes according to the number of NUMA hinting faults on
2735 	 * both the node itself, and on nearby nodes.
2736 	 */
2737 	if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2738 		unsigned long score, max_score = 0;
2739 		int node, max_node = nid;
2740 
2741 		dist = sched_max_numa_distance;
2742 
2743 		for_each_node_state(node, N_CPU) {
2744 			score = group_weight(p, node, dist);
2745 			if (score > max_score) {
2746 				max_score = score;
2747 				max_node = node;
2748 			}
2749 		}
2750 		return max_node;
2751 	}
2752 
2753 	/*
2754 	 * Finding the preferred nid in a system with NUMA backplane
2755 	 * interconnect topology is more involved. The goal is to locate
2756 	 * tasks from numa_groups near each other in the system, and
2757 	 * untangle workloads from different sides of the system. This requires
2758 	 * searching down the hierarchy of node groups, recursively searching
2759 	 * inside the highest scoring group of nodes. The nodemask tricks
2760 	 * keep the complexity of the search down.
2761 	 */
2762 	nodes = node_states[N_CPU];
2763 	for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2764 		unsigned long max_faults = 0;
2765 		nodemask_t max_group = NODE_MASK_NONE;
2766 		int a, b;
2767 
2768 		/* Are there nodes at this distance from each other? */
2769 		if (!find_numa_distance(dist))
2770 			continue;
2771 
2772 		for_each_node_mask(a, nodes) {
2773 			unsigned long faults = 0;
2774 			nodemask_t this_group;
2775 			nodes_clear(this_group);
2776 
2777 			/* Sum group's NUMA faults; includes a==b case. */
2778 			for_each_node_mask(b, nodes) {
2779 				if (node_distance(a, b) < dist) {
2780 					faults += group_faults(p, b);
2781 					node_set(b, this_group);
2782 					node_clear(b, nodes);
2783 				}
2784 			}
2785 
2786 			/* Remember the top group. */
2787 			if (faults > max_faults) {
2788 				max_faults = faults;
2789 				max_group = this_group;
2790 				/*
2791 				 * subtle: at the smallest distance there is
2792 				 * just one node left in each "group", the
2793 				 * winner is the preferred nid.
2794 				 */
2795 				nid = a;
2796 			}
2797 		}
2798 		/* Next round, evaluate the nodes within max_group. */
2799 		if (!max_faults)
2800 			break;
2801 		nodes = max_group;
2802 	}
2803 	return nid;
2804 }
2805 
task_numa_placement(struct task_struct * p)2806 static void task_numa_placement(struct task_struct *p)
2807 {
2808 	int seq, nid, max_nid = NUMA_NO_NODE;
2809 	unsigned long max_faults = 0;
2810 	unsigned long fault_types[2] = { 0, 0 };
2811 	unsigned long total_faults;
2812 	u64 runtime, period;
2813 	spinlock_t *group_lock = NULL;
2814 	struct numa_group *ng;
2815 
2816 	/*
2817 	 * The p->mm->numa_scan_seq field gets updated without
2818 	 * exclusive access. Use READ_ONCE() here to ensure
2819 	 * that the field is read in a single access:
2820 	 */
2821 	seq = READ_ONCE(p->mm->numa_scan_seq);
2822 	if (p->numa_scan_seq == seq)
2823 		return;
2824 	p->numa_scan_seq = seq;
2825 	p->numa_scan_period_max = task_scan_max(p);
2826 
2827 	total_faults = p->numa_faults_locality[0] +
2828 		       p->numa_faults_locality[1];
2829 	runtime = numa_get_avg_runtime(p, &period);
2830 
2831 	/* If the task is part of a group prevent parallel updates to group stats */
2832 	ng = deref_curr_numa_group(p);
2833 	if (ng) {
2834 		group_lock = &ng->lock;
2835 		spin_lock_irq(group_lock);
2836 	}
2837 
2838 	/* Find the node with the highest number of faults */
2839 	for_each_online_node(nid) {
2840 		/* Keep track of the offsets in numa_faults array */
2841 		int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2842 		unsigned long faults = 0, group_faults = 0;
2843 		int priv;
2844 
2845 		for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2846 			long diff, f_diff, f_weight;
2847 
2848 			mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2849 			membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2850 			cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2851 			cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2852 
2853 			/* Decay existing window, copy faults since last scan */
2854 			diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2855 			fault_types[priv] += p->numa_faults[membuf_idx];
2856 			p->numa_faults[membuf_idx] = 0;
2857 
2858 			/*
2859 			 * Normalize the faults_from, so all tasks in a group
2860 			 * count according to CPU use, instead of by the raw
2861 			 * number of faults. Tasks with little runtime have
2862 			 * little over-all impact on throughput, and thus their
2863 			 * faults are less important.
2864 			 */
2865 			f_weight = div64_u64(runtime << 16, period + 1);
2866 			f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2867 				   (total_faults + 1);
2868 			f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2869 			p->numa_faults[cpubuf_idx] = 0;
2870 
2871 			p->numa_faults[mem_idx] += diff;
2872 			p->numa_faults[cpu_idx] += f_diff;
2873 			faults += p->numa_faults[mem_idx];
2874 			p->total_numa_faults += diff;
2875 			if (ng) {
2876 				/*
2877 				 * safe because we can only change our own group
2878 				 *
2879 				 * mem_idx represents the offset for a given
2880 				 * nid and priv in a specific region because it
2881 				 * is at the beginning of the numa_faults array.
2882 				 */
2883 				ng->faults[mem_idx] += diff;
2884 				ng->faults[cpu_idx] += f_diff;
2885 				ng->total_faults += diff;
2886 				group_faults += ng->faults[mem_idx];
2887 			}
2888 		}
2889 
2890 		if (!ng) {
2891 			if (faults > max_faults) {
2892 				max_faults = faults;
2893 				max_nid = nid;
2894 			}
2895 		} else if (group_faults > max_faults) {
2896 			max_faults = group_faults;
2897 			max_nid = nid;
2898 		}
2899 	}
2900 
2901 	/* Cannot migrate task to CPU-less node */
2902 	if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) {
2903 		int near_nid = max_nid;
2904 		int distance, near_distance = INT_MAX;
2905 
2906 		for_each_node_state(nid, N_CPU) {
2907 			distance = node_distance(max_nid, nid);
2908 			if (distance < near_distance) {
2909 				near_nid = nid;
2910 				near_distance = distance;
2911 			}
2912 		}
2913 		max_nid = near_nid;
2914 	}
2915 
2916 	if (ng) {
2917 		numa_group_count_active_nodes(ng);
2918 		spin_unlock_irq(group_lock);
2919 		max_nid = preferred_group_nid(p, max_nid);
2920 	}
2921 
2922 	if (max_faults) {
2923 		/* Set the new preferred node */
2924 		if (max_nid != p->numa_preferred_nid)
2925 			sched_setnuma(p, max_nid);
2926 	}
2927 
2928 	update_task_scan_period(p, fault_types[0], fault_types[1]);
2929 }
2930 
get_numa_group(struct numa_group * grp)2931 static inline int get_numa_group(struct numa_group *grp)
2932 {
2933 	return refcount_inc_not_zero(&grp->refcount);
2934 }
2935 
put_numa_group(struct numa_group * grp)2936 static inline void put_numa_group(struct numa_group *grp)
2937 {
2938 	if (refcount_dec_and_test(&grp->refcount))
2939 		kfree_rcu(grp, rcu);
2940 }
2941 
task_numa_group(struct task_struct * p,int cpupid,int flags,int * priv)2942 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2943 			int *priv)
2944 {
2945 	struct numa_group *grp, *my_grp;
2946 	struct task_struct *tsk;
2947 	bool join = false;
2948 	int cpu = cpupid_to_cpu(cpupid);
2949 	int i;
2950 
2951 	if (unlikely(!deref_curr_numa_group(p))) {
2952 		unsigned int size = sizeof(struct numa_group) +
2953 				    NR_NUMA_HINT_FAULT_STATS *
2954 				    nr_node_ids * sizeof(unsigned long);
2955 
2956 		grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2957 		if (!grp)
2958 			return;
2959 
2960 		refcount_set(&grp->refcount, 1);
2961 		grp->active_nodes = 1;
2962 		grp->max_faults_cpu = 0;
2963 		spin_lock_init(&grp->lock);
2964 		grp->gid = p->pid;
2965 
2966 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2967 			grp->faults[i] = p->numa_faults[i];
2968 
2969 		grp->total_faults = p->total_numa_faults;
2970 
2971 		grp->nr_tasks++;
2972 		rcu_assign_pointer(p->numa_group, grp);
2973 	}
2974 
2975 	rcu_read_lock();
2976 	tsk = READ_ONCE(cpu_rq(cpu)->curr);
2977 
2978 	if (!cpupid_match_pid(tsk, cpupid))
2979 		goto no_join;
2980 
2981 	grp = rcu_dereference(tsk->numa_group);
2982 	if (!grp)
2983 		goto no_join;
2984 
2985 	my_grp = deref_curr_numa_group(p);
2986 	if (grp == my_grp)
2987 		goto no_join;
2988 
2989 	/*
2990 	 * Only join the other group if its bigger; if we're the bigger group,
2991 	 * the other task will join us.
2992 	 */
2993 	if (my_grp->nr_tasks > grp->nr_tasks)
2994 		goto no_join;
2995 
2996 	/*
2997 	 * Tie-break on the grp address.
2998 	 */
2999 	if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
3000 		goto no_join;
3001 
3002 	/* Always join threads in the same process. */
3003 	if (tsk->mm == current->mm)
3004 		join = true;
3005 
3006 	/* Simple filter to avoid false positives due to PID collisions */
3007 	if (flags & TNF_SHARED)
3008 		join = true;
3009 
3010 	/* Update priv based on whether false sharing was detected */
3011 	*priv = !join;
3012 
3013 	if (join && !get_numa_group(grp))
3014 		goto no_join;
3015 
3016 	rcu_read_unlock();
3017 
3018 	if (!join)
3019 		return;
3020 
3021 	WARN_ON_ONCE(irqs_disabled());
3022 	double_lock_irq(&my_grp->lock, &grp->lock);
3023 
3024 	for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
3025 		my_grp->faults[i] -= p->numa_faults[i];
3026 		grp->faults[i] += p->numa_faults[i];
3027 	}
3028 	my_grp->total_faults -= p->total_numa_faults;
3029 	grp->total_faults += p->total_numa_faults;
3030 
3031 	my_grp->nr_tasks--;
3032 	grp->nr_tasks++;
3033 
3034 	spin_unlock(&my_grp->lock);
3035 	spin_unlock_irq(&grp->lock);
3036 
3037 	rcu_assign_pointer(p->numa_group, grp);
3038 
3039 	put_numa_group(my_grp);
3040 	return;
3041 
3042 no_join:
3043 	rcu_read_unlock();
3044 	return;
3045 }
3046 
3047 /*
3048  * Get rid of NUMA statistics associated with a task (either current or dead).
3049  * If @final is set, the task is dead and has reached refcount zero, so we can
3050  * safely free all relevant data structures. Otherwise, there might be
3051  * concurrent reads from places like load balancing and procfs, and we should
3052  * reset the data back to default state without freeing ->numa_faults.
3053  */
task_numa_free(struct task_struct * p,bool final)3054 void task_numa_free(struct task_struct *p, bool final)
3055 {
3056 	/* safe: p either is current or is being freed by current */
3057 	struct numa_group *grp = rcu_dereference_raw(p->numa_group);
3058 	unsigned long *numa_faults = p->numa_faults;
3059 	unsigned long flags;
3060 	int i;
3061 
3062 	if (!numa_faults)
3063 		return;
3064 
3065 	if (grp) {
3066 		spin_lock_irqsave(&grp->lock, flags);
3067 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3068 			grp->faults[i] -= p->numa_faults[i];
3069 		grp->total_faults -= p->total_numa_faults;
3070 
3071 		grp->nr_tasks--;
3072 		spin_unlock_irqrestore(&grp->lock, flags);
3073 		RCU_INIT_POINTER(p->numa_group, NULL);
3074 		put_numa_group(grp);
3075 	}
3076 
3077 	if (final) {
3078 		p->numa_faults = NULL;
3079 		kfree(numa_faults);
3080 	} else {
3081 		p->total_numa_faults = 0;
3082 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3083 			numa_faults[i] = 0;
3084 	}
3085 }
3086 
3087 /*
3088  * Got a PROT_NONE fault for a page on @node.
3089  */
task_numa_fault(int last_cpupid,int mem_node,int pages,int flags)3090 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
3091 {
3092 	struct task_struct *p = current;
3093 	bool migrated = flags & TNF_MIGRATED;
3094 	int cpu_node = task_node(current);
3095 	int local = !!(flags & TNF_FAULT_LOCAL);
3096 	struct numa_group *ng;
3097 	int priv;
3098 
3099 	if (!static_branch_likely(&sched_numa_balancing))
3100 		return;
3101 
3102 	/* for example, ksmd faulting in a user's mm */
3103 	if (!p->mm)
3104 		return;
3105 
3106 	/*
3107 	 * NUMA faults statistics are unnecessary for the slow memory
3108 	 * node for memory tiering mode.
3109 	 */
3110 	if (!node_is_toptier(mem_node) &&
3111 	    (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
3112 	     !cpupid_valid(last_cpupid)))
3113 		return;
3114 
3115 	/* Allocate buffer to track faults on a per-node basis */
3116 	if (unlikely(!p->numa_faults)) {
3117 		int size = sizeof(*p->numa_faults) *
3118 			   NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
3119 
3120 		p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
3121 		if (!p->numa_faults)
3122 			return;
3123 
3124 		p->total_numa_faults = 0;
3125 		memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
3126 	}
3127 
3128 	/*
3129 	 * First accesses are treated as private, otherwise consider accesses
3130 	 * to be private if the accessing pid has not changed
3131 	 */
3132 	if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
3133 		priv = 1;
3134 	} else {
3135 		priv = cpupid_match_pid(p, last_cpupid);
3136 		if (!priv && !(flags & TNF_NO_GROUP))
3137 			task_numa_group(p, last_cpupid, flags, &priv);
3138 	}
3139 
3140 	/*
3141 	 * If a workload spans multiple NUMA nodes, a shared fault that
3142 	 * occurs wholly within the set of nodes that the workload is
3143 	 * actively using should be counted as local. This allows the
3144 	 * scan rate to slow down when a workload has settled down.
3145 	 */
3146 	ng = deref_curr_numa_group(p);
3147 	if (!priv && !local && ng && ng->active_nodes > 1 &&
3148 				numa_is_active_node(cpu_node, ng) &&
3149 				numa_is_active_node(mem_node, ng))
3150 		local = 1;
3151 
3152 	/*
3153 	 * Retry to migrate task to preferred node periodically, in case it
3154 	 * previously failed, or the scheduler moved us.
3155 	 */
3156 	if (time_after(jiffies, p->numa_migrate_retry)) {
3157 		task_numa_placement(p);
3158 		numa_migrate_preferred(p);
3159 	}
3160 
3161 	if (migrated)
3162 		p->numa_pages_migrated += pages;
3163 	if (flags & TNF_MIGRATE_FAIL)
3164 		p->numa_faults_locality[2] += pages;
3165 
3166 	p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
3167 	p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
3168 	p->numa_faults_locality[local] += pages;
3169 }
3170 
reset_ptenuma_scan(struct task_struct * p)3171 static void reset_ptenuma_scan(struct task_struct *p)
3172 {
3173 	/*
3174 	 * We only did a read acquisition of the mmap sem, so
3175 	 * p->mm->numa_scan_seq is written to without exclusive access
3176 	 * and the update is not guaranteed to be atomic. That's not
3177 	 * much of an issue though, since this is just used for
3178 	 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
3179 	 * expensive, to avoid any form of compiler optimizations:
3180 	 */
3181 	WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
3182 	p->mm->numa_scan_offset = 0;
3183 }
3184 
vma_is_accessed(struct vm_area_struct * vma)3185 static bool vma_is_accessed(struct vm_area_struct *vma)
3186 {
3187 	unsigned long pids;
3188 	/*
3189 	 * Allow unconditional access first two times, so that all the (pages)
3190 	 * of VMAs get prot_none fault introduced irrespective of accesses.
3191 	 * This is also done to avoid any side effect of task scanning
3192 	 * amplifying the unfairness of disjoint set of VMAs' access.
3193 	 */
3194 	if (READ_ONCE(current->mm->numa_scan_seq) < 2)
3195 		return true;
3196 
3197 	pids = vma->numab_state->access_pids[0] | vma->numab_state->access_pids[1];
3198 	return test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids);
3199 }
3200 
3201 #define VMA_PID_RESET_PERIOD (4 * sysctl_numa_balancing_scan_delay)
3202 
3203 /*
3204  * The expensive part of numa migration is done from task_work context.
3205  * Triggered from task_tick_numa().
3206  */
task_numa_work(struct callback_head * work)3207 static void task_numa_work(struct callback_head *work)
3208 {
3209 	unsigned long migrate, next_scan, now = jiffies;
3210 	struct task_struct *p = current;
3211 	struct mm_struct *mm = p->mm;
3212 	u64 runtime = p->se.sum_exec_runtime;
3213 	struct vm_area_struct *vma;
3214 	unsigned long start, end;
3215 	unsigned long nr_pte_updates = 0;
3216 	long pages, virtpages;
3217 	struct vma_iterator vmi;
3218 
3219 	SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
3220 
3221 	work->next = work;
3222 	/*
3223 	 * Who cares about NUMA placement when they're dying.
3224 	 *
3225 	 * NOTE: make sure not to dereference p->mm before this check,
3226 	 * exit_task_work() happens _after_ exit_mm() so we could be called
3227 	 * without p->mm even though we still had it when we enqueued this
3228 	 * work.
3229 	 */
3230 	if (p->flags & PF_EXITING)
3231 		return;
3232 
3233 	if (!mm->numa_next_scan) {
3234 		mm->numa_next_scan = now +
3235 			msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3236 	}
3237 
3238 	/*
3239 	 * Enforce maximal scan/migration frequency..
3240 	 */
3241 	migrate = mm->numa_next_scan;
3242 	if (time_before(now, migrate))
3243 		return;
3244 
3245 	if (p->numa_scan_period == 0) {
3246 		p->numa_scan_period_max = task_scan_max(p);
3247 		p->numa_scan_period = task_scan_start(p);
3248 	}
3249 
3250 	next_scan = now + msecs_to_jiffies(p->numa_scan_period);
3251 	if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
3252 		return;
3253 
3254 	/*
3255 	 * Delay this task enough that another task of this mm will likely win
3256 	 * the next time around.
3257 	 */
3258 	p->node_stamp += 2 * TICK_NSEC;
3259 
3260 	start = mm->numa_scan_offset;
3261 	pages = sysctl_numa_balancing_scan_size;
3262 	pages <<= 20 - PAGE_SHIFT; /* MB in pages */
3263 	virtpages = pages * 8;	   /* Scan up to this much virtual space */
3264 	if (!pages)
3265 		return;
3266 
3267 
3268 	if (!mmap_read_trylock(mm))
3269 		return;
3270 	vma_iter_init(&vmi, mm, start);
3271 	vma = vma_next(&vmi);
3272 	if (!vma) {
3273 		reset_ptenuma_scan(p);
3274 		start = 0;
3275 		vma_iter_set(&vmi, start);
3276 		vma = vma_next(&vmi);
3277 	}
3278 
3279 	do {
3280 		if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
3281 			is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3282 			continue;
3283 		}
3284 
3285 		/*
3286 		 * Shared library pages mapped by multiple processes are not
3287 		 * migrated as it is expected they are cache replicated. Avoid
3288 		 * hinting faults in read-only file-backed mappings or the vdso
3289 		 * as migrating the pages will be of marginal benefit.
3290 		 */
3291 		if (!vma->vm_mm ||
3292 		    (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
3293 			continue;
3294 
3295 		/*
3296 		 * Skip inaccessible VMAs to avoid any confusion between
3297 		 * PROT_NONE and NUMA hinting ptes
3298 		 */
3299 		if (!vma_is_accessible(vma))
3300 			continue;
3301 
3302 		/* Initialise new per-VMA NUMAB state. */
3303 		if (!vma->numab_state) {
3304 			vma->numab_state = kzalloc(sizeof(struct vma_numab_state),
3305 				GFP_KERNEL);
3306 			if (!vma->numab_state)
3307 				continue;
3308 
3309 			vma->numab_state->next_scan = now +
3310 				msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3311 
3312 			/* Reset happens after 4 times scan delay of scan start */
3313 			vma->numab_state->next_pid_reset =  vma->numab_state->next_scan +
3314 				msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3315 		}
3316 
3317 		/*
3318 		 * Scanning the VMA's of short lived tasks add more overhead. So
3319 		 * delay the scan for new VMAs.
3320 		 */
3321 		if (mm->numa_scan_seq && time_before(jiffies,
3322 						vma->numab_state->next_scan))
3323 			continue;
3324 
3325 		/* Do not scan the VMA if task has not accessed */
3326 		if (!vma_is_accessed(vma))
3327 			continue;
3328 
3329 		/*
3330 		 * RESET access PIDs regularly for old VMAs. Resetting after checking
3331 		 * vma for recent access to avoid clearing PID info before access..
3332 		 */
3333 		if (mm->numa_scan_seq &&
3334 				time_after(jiffies, vma->numab_state->next_pid_reset)) {
3335 			vma->numab_state->next_pid_reset = vma->numab_state->next_pid_reset +
3336 				msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3337 			vma->numab_state->access_pids[0] = READ_ONCE(vma->numab_state->access_pids[1]);
3338 			vma->numab_state->access_pids[1] = 0;
3339 		}
3340 
3341 		do {
3342 			start = max(start, vma->vm_start);
3343 			end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3344 			end = min(end, vma->vm_end);
3345 			nr_pte_updates = change_prot_numa(vma, start, end);
3346 
3347 			/*
3348 			 * Try to scan sysctl_numa_balancing_size worth of
3349 			 * hpages that have at least one present PTE that
3350 			 * is not already pte-numa. If the VMA contains
3351 			 * areas that are unused or already full of prot_numa
3352 			 * PTEs, scan up to virtpages, to skip through those
3353 			 * areas faster.
3354 			 */
3355 			if (nr_pte_updates)
3356 				pages -= (end - start) >> PAGE_SHIFT;
3357 			virtpages -= (end - start) >> PAGE_SHIFT;
3358 
3359 			start = end;
3360 			if (pages <= 0 || virtpages <= 0)
3361 				goto out;
3362 
3363 			cond_resched();
3364 		} while (end != vma->vm_end);
3365 	} for_each_vma(vmi, vma);
3366 
3367 out:
3368 	/*
3369 	 * It is possible to reach the end of the VMA list but the last few
3370 	 * VMAs are not guaranteed to the vma_migratable. If they are not, we
3371 	 * would find the !migratable VMA on the next scan but not reset the
3372 	 * scanner to the start so check it now.
3373 	 */
3374 	if (vma)
3375 		mm->numa_scan_offset = start;
3376 	else
3377 		reset_ptenuma_scan(p);
3378 	mmap_read_unlock(mm);
3379 
3380 	/*
3381 	 * Make sure tasks use at least 32x as much time to run other code
3382 	 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3383 	 * Usually update_task_scan_period slows down scanning enough; on an
3384 	 * overloaded system we need to limit overhead on a per task basis.
3385 	 */
3386 	if (unlikely(p->se.sum_exec_runtime != runtime)) {
3387 		u64 diff = p->se.sum_exec_runtime - runtime;
3388 		p->node_stamp += 32 * diff;
3389 	}
3390 }
3391 
init_numa_balancing(unsigned long clone_flags,struct task_struct * p)3392 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3393 {
3394 	int mm_users = 0;
3395 	struct mm_struct *mm = p->mm;
3396 
3397 	if (mm) {
3398 		mm_users = atomic_read(&mm->mm_users);
3399 		if (mm_users == 1) {
3400 			mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3401 			mm->numa_scan_seq = 0;
3402 		}
3403 	}
3404 	p->node_stamp			= 0;
3405 	p->numa_scan_seq		= mm ? mm->numa_scan_seq : 0;
3406 	p->numa_scan_period		= sysctl_numa_balancing_scan_delay;
3407 	p->numa_migrate_retry		= 0;
3408 	/* Protect against double add, see task_tick_numa and task_numa_work */
3409 	p->numa_work.next		= &p->numa_work;
3410 	p->numa_faults			= NULL;
3411 	p->numa_pages_migrated		= 0;
3412 	p->total_numa_faults		= 0;
3413 	RCU_INIT_POINTER(p->numa_group, NULL);
3414 	p->last_task_numa_placement	= 0;
3415 	p->last_sum_exec_runtime	= 0;
3416 
3417 	init_task_work(&p->numa_work, task_numa_work);
3418 
3419 	/* New address space, reset the preferred nid */
3420 	if (!(clone_flags & CLONE_VM)) {
3421 		p->numa_preferred_nid = NUMA_NO_NODE;
3422 		return;
3423 	}
3424 
3425 	/*
3426 	 * New thread, keep existing numa_preferred_nid which should be copied
3427 	 * already by arch_dup_task_struct but stagger when scans start.
3428 	 */
3429 	if (mm) {
3430 		unsigned int delay;
3431 
3432 		delay = min_t(unsigned int, task_scan_max(current),
3433 			current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3434 		delay += 2 * TICK_NSEC;
3435 		p->node_stamp = delay;
3436 	}
3437 }
3438 
3439 /*
3440  * Drive the periodic memory faults..
3441  */
task_tick_numa(struct rq * rq,struct task_struct * curr)3442 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3443 {
3444 	struct callback_head *work = &curr->numa_work;
3445 	u64 period, now;
3446 
3447 	/*
3448 	 * We don't care about NUMA placement if we don't have memory.
3449 	 */
3450 	if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3451 		return;
3452 
3453 	/*
3454 	 * Using runtime rather than walltime has the dual advantage that
3455 	 * we (mostly) drive the selection from busy threads and that the
3456 	 * task needs to have done some actual work before we bother with
3457 	 * NUMA placement.
3458 	 */
3459 	now = curr->se.sum_exec_runtime;
3460 	period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3461 
3462 	if (now > curr->node_stamp + period) {
3463 		if (!curr->node_stamp)
3464 			curr->numa_scan_period = task_scan_start(curr);
3465 		curr->node_stamp += period;
3466 
3467 		if (!time_before(jiffies, curr->mm->numa_next_scan))
3468 			task_work_add(curr, work, TWA_RESUME);
3469 	}
3470 }
3471 
update_scan_period(struct task_struct * p,int new_cpu)3472 static void update_scan_period(struct task_struct *p, int new_cpu)
3473 {
3474 	int src_nid = cpu_to_node(task_cpu(p));
3475 	int dst_nid = cpu_to_node(new_cpu);
3476 
3477 	if (!static_branch_likely(&sched_numa_balancing))
3478 		return;
3479 
3480 	if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3481 		return;
3482 
3483 	if (src_nid == dst_nid)
3484 		return;
3485 
3486 	/*
3487 	 * Allow resets if faults have been trapped before one scan
3488 	 * has completed. This is most likely due to a new task that
3489 	 * is pulled cross-node due to wakeups or load balancing.
3490 	 */
3491 	if (p->numa_scan_seq) {
3492 		/*
3493 		 * Avoid scan adjustments if moving to the preferred
3494 		 * node or if the task was not previously running on
3495 		 * the preferred node.
3496 		 */
3497 		if (dst_nid == p->numa_preferred_nid ||
3498 		    (p->numa_preferred_nid != NUMA_NO_NODE &&
3499 			src_nid != p->numa_preferred_nid))
3500 			return;
3501 	}
3502 
3503 	p->numa_scan_period = task_scan_start(p);
3504 }
3505 
3506 #else
task_tick_numa(struct rq * rq,struct task_struct * curr)3507 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3508 {
3509 }
3510 
account_numa_enqueue(struct rq * rq,struct task_struct * p)3511 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3512 {
3513 }
3514 
account_numa_dequeue(struct rq * rq,struct task_struct * p)3515 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3516 {
3517 }
3518 
update_scan_period(struct task_struct * p,int new_cpu)3519 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3520 {
3521 }
3522 
3523 #endif /* CONFIG_NUMA_BALANCING */
3524 
3525 static void
account_entity_enqueue(struct cfs_rq * cfs_rq,struct sched_entity * se)3526 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3527 {
3528 	update_load_add(&cfs_rq->load, se->load.weight);
3529 #ifdef CONFIG_SMP
3530 	if (entity_is_task(se)) {
3531 		struct rq *rq = rq_of(cfs_rq);
3532 
3533 		account_numa_enqueue(rq, task_of(se));
3534 		list_add(&se->group_node, &rq->cfs_tasks);
3535 	}
3536 #endif
3537 	cfs_rq->nr_running++;
3538 	if (se_is_idle(se))
3539 		cfs_rq->idle_nr_running++;
3540 }
3541 
3542 static void
account_entity_dequeue(struct cfs_rq * cfs_rq,struct sched_entity * se)3543 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3544 {
3545 	update_load_sub(&cfs_rq->load, se->load.weight);
3546 #ifdef CONFIG_SMP
3547 	if (entity_is_task(se)) {
3548 		account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3549 		list_del_init(&se->group_node);
3550 	}
3551 #endif
3552 	cfs_rq->nr_running--;
3553 	if (se_is_idle(se))
3554 		cfs_rq->idle_nr_running--;
3555 }
3556 
3557 /*
3558  * Signed add and clamp on underflow.
3559  *
3560  * Explicitly do a load-store to ensure the intermediate value never hits
3561  * memory. This allows lockless observations without ever seeing the negative
3562  * values.
3563  */
3564 #define add_positive(_ptr, _val) do {                           \
3565 	typeof(_ptr) ptr = (_ptr);                              \
3566 	typeof(_val) val = (_val);                              \
3567 	typeof(*ptr) res, var = READ_ONCE(*ptr);                \
3568 								\
3569 	res = var + val;                                        \
3570 								\
3571 	if (val < 0 && res > var)                               \
3572 		res = 0;                                        \
3573 								\
3574 	WRITE_ONCE(*ptr, res);                                  \
3575 } while (0)
3576 
3577 /*
3578  * Unsigned subtract and clamp on underflow.
3579  *
3580  * Explicitly do a load-store to ensure the intermediate value never hits
3581  * memory. This allows lockless observations without ever seeing the negative
3582  * values.
3583  */
3584 #define sub_positive(_ptr, _val) do {				\
3585 	typeof(_ptr) ptr = (_ptr);				\
3586 	typeof(*ptr) val = (_val);				\
3587 	typeof(*ptr) res, var = READ_ONCE(*ptr);		\
3588 	res = var - val;					\
3589 	if (res > var)						\
3590 		res = 0;					\
3591 	WRITE_ONCE(*ptr, res);					\
3592 } while (0)
3593 
3594 /*
3595  * Remove and clamp on negative, from a local variable.
3596  *
3597  * A variant of sub_positive(), which does not use explicit load-store
3598  * and is thus optimized for local variable updates.
3599  */
3600 #define lsub_positive(_ptr, _val) do {				\
3601 	typeof(_ptr) ptr = (_ptr);				\
3602 	*ptr -= min_t(typeof(*ptr), *ptr, _val);		\
3603 } while (0)
3604 
3605 #ifdef CONFIG_SMP
3606 static inline void
enqueue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3607 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3608 {
3609 	cfs_rq->avg.load_avg += se->avg.load_avg;
3610 	cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3611 }
3612 
3613 static inline void
dequeue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3614 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3615 {
3616 	sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3617 	sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3618 	/* See update_cfs_rq_load_avg() */
3619 	cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3620 					  cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3621 }
3622 #else
3623 static inline void
enqueue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3624 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3625 static inline void
dequeue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3626 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3627 #endif
3628 
reweight_eevdf(struct cfs_rq * cfs_rq,struct sched_entity * se,unsigned long weight)3629 static void reweight_eevdf(struct cfs_rq *cfs_rq, struct sched_entity *se,
3630 			   unsigned long weight)
3631 {
3632 	unsigned long old_weight = se->load.weight;
3633 	u64 avruntime = avg_vruntime(cfs_rq);
3634 	s64 vlag, vslice;
3635 
3636 	/*
3637 	 * VRUNTIME
3638 	 * ========
3639 	 *
3640 	 * COROLLARY #1: The virtual runtime of the entity needs to be
3641 	 * adjusted if re-weight at !0-lag point.
3642 	 *
3643 	 * Proof: For contradiction assume this is not true, so we can
3644 	 * re-weight without changing vruntime at !0-lag point.
3645 	 *
3646 	 *             Weight	VRuntime   Avg-VRuntime
3647 	 *     before    w          v            V
3648 	 *      after    w'         v'           V'
3649 	 *
3650 	 * Since lag needs to be preserved through re-weight:
3651 	 *
3652 	 *	lag = (V - v)*w = (V'- v')*w', where v = v'
3653 	 *	==>	V' = (V - v)*w/w' + v		(1)
3654 	 *
3655 	 * Let W be the total weight of the entities before reweight,
3656 	 * since V' is the new weighted average of entities:
3657 	 *
3658 	 *	V' = (WV + w'v - wv) / (W + w' - w)	(2)
3659 	 *
3660 	 * by using (1) & (2) we obtain:
3661 	 *
3662 	 *	(WV + w'v - wv) / (W + w' - w) = (V - v)*w/w' + v
3663 	 *	==> (WV-Wv+Wv+w'v-wv)/(W+w'-w) = (V - v)*w/w' + v
3664 	 *	==> (WV - Wv)/(W + w' - w) + v = (V - v)*w/w' + v
3665 	 *	==>	(V - v)*W/(W + w' - w) = (V - v)*w/w' (3)
3666 	 *
3667 	 * Since we are doing at !0-lag point which means V != v, we
3668 	 * can simplify (3):
3669 	 *
3670 	 *	==>	W / (W + w' - w) = w / w'
3671 	 *	==>	Ww' = Ww + ww' - ww
3672 	 *	==>	W * (w' - w) = w * (w' - w)
3673 	 *	==>	W = w	(re-weight indicates w' != w)
3674 	 *
3675 	 * So the cfs_rq contains only one entity, hence vruntime of
3676 	 * the entity @v should always equal to the cfs_rq's weighted
3677 	 * average vruntime @V, which means we will always re-weight
3678 	 * at 0-lag point, thus breach assumption. Proof completed.
3679 	 *
3680 	 *
3681 	 * COROLLARY #2: Re-weight does NOT affect weighted average
3682 	 * vruntime of all the entities.
3683 	 *
3684 	 * Proof: According to corollary #1, Eq. (1) should be:
3685 	 *
3686 	 *	(V - v)*w = (V' - v')*w'
3687 	 *	==>    v' = V' - (V - v)*w/w'		(4)
3688 	 *
3689 	 * According to the weighted average formula, we have:
3690 	 *
3691 	 *	V' = (WV - wv + w'v') / (W - w + w')
3692 	 *	   = (WV - wv + w'(V' - (V - v)w/w')) / (W - w + w')
3693 	 *	   = (WV - wv + w'V' - Vw + wv) / (W - w + w')
3694 	 *	   = (WV + w'V' - Vw) / (W - w + w')
3695 	 *
3696 	 *	==>  V'*(W - w + w') = WV + w'V' - Vw
3697 	 *	==>	V' * (W - w) = (W - w) * V	(5)
3698 	 *
3699 	 * If the entity is the only one in the cfs_rq, then reweight
3700 	 * always occurs at 0-lag point, so V won't change. Or else
3701 	 * there are other entities, hence W != w, then Eq. (5) turns
3702 	 * into V' = V. So V won't change in either case, proof done.
3703 	 *
3704 	 *
3705 	 * So according to corollary #1 & #2, the effect of re-weight
3706 	 * on vruntime should be:
3707 	 *
3708 	 *	v' = V' - (V - v) * w / w'		(4)
3709 	 *	   = V  - (V - v) * w / w'
3710 	 *	   = V  - vl * w / w'
3711 	 *	   = V  - vl'
3712 	 */
3713 	if (avruntime != se->vruntime) {
3714 		vlag = (s64)(avruntime - se->vruntime);
3715 		vlag = div_s64(vlag * old_weight, weight);
3716 		se->vruntime = avruntime - vlag;
3717 	}
3718 
3719 	/*
3720 	 * DEADLINE
3721 	 * ========
3722 	 *
3723 	 * When the weight changes, the virtual time slope changes and
3724 	 * we should adjust the relative virtual deadline accordingly.
3725 	 *
3726 	 *	d' = v' + (d - v)*w/w'
3727 	 *	   = V' - (V - v)*w/w' + (d - v)*w/w'
3728 	 *	   = V  - (V - v)*w/w' + (d - v)*w/w'
3729 	 *	   = V  + (d - V)*w/w'
3730 	 */
3731 	vslice = (s64)(se->deadline - avruntime);
3732 	vslice = div_s64(vslice * old_weight, weight);
3733 	se->deadline = avruntime + vslice;
3734 }
3735 
reweight_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,unsigned long weight)3736 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3737 			    unsigned long weight)
3738 {
3739 	bool curr = cfs_rq->curr == se;
3740 
3741 	if (se->on_rq) {
3742 		/* commit outstanding execution time */
3743 		if (curr)
3744 			update_curr(cfs_rq);
3745 		else
3746 			__dequeue_entity(cfs_rq, se);
3747 		update_load_sub(&cfs_rq->load, se->load.weight);
3748 	}
3749 	dequeue_load_avg(cfs_rq, se);
3750 
3751 	if (!se->on_rq) {
3752 		/*
3753 		 * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i),
3754 		 * we need to scale se->vlag when w_i changes.
3755 		 */
3756 		se->vlag = div_s64(se->vlag * se->load.weight, weight);
3757 	} else {
3758 		reweight_eevdf(cfs_rq, se, weight);
3759 	}
3760 
3761 	update_load_set(&se->load, weight);
3762 
3763 #ifdef CONFIG_SMP
3764 	do {
3765 		u32 divider = get_pelt_divider(&se->avg);
3766 
3767 		se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3768 	} while (0);
3769 #endif
3770 
3771 	enqueue_load_avg(cfs_rq, se);
3772 	if (se->on_rq) {
3773 		update_load_add(&cfs_rq->load, se->load.weight);
3774 		if (!curr)
3775 			__enqueue_entity(cfs_rq, se);
3776 
3777 		/*
3778 		 * The entity's vruntime has been adjusted, so let's check
3779 		 * whether the rq-wide min_vruntime needs updated too. Since
3780 		 * the calculations above require stable min_vruntime rather
3781 		 * than up-to-date one, we do the update at the end of the
3782 		 * reweight process.
3783 		 */
3784 		update_min_vruntime(cfs_rq);
3785 	}
3786 }
3787 
reweight_task(struct task_struct * p,int prio)3788 void reweight_task(struct task_struct *p, int prio)
3789 {
3790 	struct sched_entity *se = &p->se;
3791 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
3792 	struct load_weight *load = &se->load;
3793 	unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3794 
3795 	reweight_entity(cfs_rq, se, weight);
3796 	load->inv_weight = sched_prio_to_wmult[prio];
3797 }
3798 
3799 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3800 
3801 #ifdef CONFIG_FAIR_GROUP_SCHED
3802 #ifdef CONFIG_SMP
3803 /*
3804  * All this does is approximate the hierarchical proportion which includes that
3805  * global sum we all love to hate.
3806  *
3807  * That is, the weight of a group entity, is the proportional share of the
3808  * group weight based on the group runqueue weights. That is:
3809  *
3810  *                     tg->weight * grq->load.weight
3811  *   ge->load.weight = -----------------------------               (1)
3812  *                       \Sum grq->load.weight
3813  *
3814  * Now, because computing that sum is prohibitively expensive to compute (been
3815  * there, done that) we approximate it with this average stuff. The average
3816  * moves slower and therefore the approximation is cheaper and more stable.
3817  *
3818  * So instead of the above, we substitute:
3819  *
3820  *   grq->load.weight -> grq->avg.load_avg                         (2)
3821  *
3822  * which yields the following:
3823  *
3824  *                     tg->weight * grq->avg.load_avg
3825  *   ge->load.weight = ------------------------------              (3)
3826  *                             tg->load_avg
3827  *
3828  * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3829  *
3830  * That is shares_avg, and it is right (given the approximation (2)).
3831  *
3832  * The problem with it is that because the average is slow -- it was designed
3833  * to be exactly that of course -- this leads to transients in boundary
3834  * conditions. In specific, the case where the group was idle and we start the
3835  * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3836  * yielding bad latency etc..
3837  *
3838  * Now, in that special case (1) reduces to:
3839  *
3840  *                     tg->weight * grq->load.weight
3841  *   ge->load.weight = ----------------------------- = tg->weight   (4)
3842  *                         grp->load.weight
3843  *
3844  * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3845  *
3846  * So what we do is modify our approximation (3) to approach (4) in the (near)
3847  * UP case, like:
3848  *
3849  *   ge->load.weight =
3850  *
3851  *              tg->weight * grq->load.weight
3852  *     ---------------------------------------------------         (5)
3853  *     tg->load_avg - grq->avg.load_avg + grq->load.weight
3854  *
3855  * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3856  * we need to use grq->avg.load_avg as its lower bound, which then gives:
3857  *
3858  *
3859  *                     tg->weight * grq->load.weight
3860  *   ge->load.weight = -----------------------------		   (6)
3861  *                             tg_load_avg'
3862  *
3863  * Where:
3864  *
3865  *   tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3866  *                  max(grq->load.weight, grq->avg.load_avg)
3867  *
3868  * And that is shares_weight and is icky. In the (near) UP case it approaches
3869  * (4) while in the normal case it approaches (3). It consistently
3870  * overestimates the ge->load.weight and therefore:
3871  *
3872  *   \Sum ge->load.weight >= tg->weight
3873  *
3874  * hence icky!
3875  */
calc_group_shares(struct cfs_rq * cfs_rq)3876 static long calc_group_shares(struct cfs_rq *cfs_rq)
3877 {
3878 	long tg_weight, tg_shares, load, shares;
3879 	struct task_group *tg = cfs_rq->tg;
3880 
3881 	tg_shares = READ_ONCE(tg->shares);
3882 
3883 	load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3884 
3885 	tg_weight = atomic_long_read(&tg->load_avg);
3886 
3887 	/* Ensure tg_weight >= load */
3888 	tg_weight -= cfs_rq->tg_load_avg_contrib;
3889 	tg_weight += load;
3890 
3891 	shares = (tg_shares * load);
3892 	if (tg_weight)
3893 		shares /= tg_weight;
3894 
3895 	/*
3896 	 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3897 	 * of a group with small tg->shares value. It is a floor value which is
3898 	 * assigned as a minimum load.weight to the sched_entity representing
3899 	 * the group on a CPU.
3900 	 *
3901 	 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3902 	 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3903 	 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3904 	 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3905 	 * instead of 0.
3906 	 */
3907 	return clamp_t(long, shares, MIN_SHARES, tg_shares);
3908 }
3909 #endif /* CONFIG_SMP */
3910 
3911 /*
3912  * Recomputes the group entity based on the current state of its group
3913  * runqueue.
3914  */
update_cfs_group(struct sched_entity * se)3915 static void update_cfs_group(struct sched_entity *se)
3916 {
3917 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3918 	long shares;
3919 
3920 	if (!gcfs_rq)
3921 		return;
3922 
3923 	if (throttled_hierarchy(gcfs_rq))
3924 		return;
3925 
3926 #ifndef CONFIG_SMP
3927 	shares = READ_ONCE(gcfs_rq->tg->shares);
3928 #else
3929 	shares = calc_group_shares(gcfs_rq);
3930 #endif
3931 	if (unlikely(se->load.weight != shares))
3932 		reweight_entity(cfs_rq_of(se), se, shares);
3933 }
3934 
3935 #else /* CONFIG_FAIR_GROUP_SCHED */
update_cfs_group(struct sched_entity * se)3936 static inline void update_cfs_group(struct sched_entity *se)
3937 {
3938 }
3939 #endif /* CONFIG_FAIR_GROUP_SCHED */
3940 
cfs_rq_util_change(struct cfs_rq * cfs_rq,int flags)3941 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3942 {
3943 	struct rq *rq = rq_of(cfs_rq);
3944 
3945 	if (&rq->cfs == cfs_rq) {
3946 		/*
3947 		 * There are a few boundary cases this might miss but it should
3948 		 * get called often enough that that should (hopefully) not be
3949 		 * a real problem.
3950 		 *
3951 		 * It will not get called when we go idle, because the idle
3952 		 * thread is a different class (!fair), nor will the utilization
3953 		 * number include things like RT tasks.
3954 		 *
3955 		 * As is, the util number is not freq-invariant (we'd have to
3956 		 * implement arch_scale_freq_capacity() for that).
3957 		 *
3958 		 * See cpu_util_cfs().
3959 		 */
3960 		cpufreq_update_util(rq, flags);
3961 	}
3962 }
3963 
3964 #ifdef CONFIG_SMP
load_avg_is_decayed(struct sched_avg * sa)3965 static inline bool load_avg_is_decayed(struct sched_avg *sa)
3966 {
3967 	if (sa->load_sum)
3968 		return false;
3969 
3970 	if (sa->util_sum)
3971 		return false;
3972 
3973 	if (sa->runnable_sum)
3974 		return false;
3975 
3976 	/*
3977 	 * _avg must be null when _sum are null because _avg = _sum / divider
3978 	 * Make sure that rounding and/or propagation of PELT values never
3979 	 * break this.
3980 	 */
3981 	SCHED_WARN_ON(sa->load_avg ||
3982 		      sa->util_avg ||
3983 		      sa->runnable_avg);
3984 
3985 	return true;
3986 }
3987 
cfs_rq_last_update_time(struct cfs_rq * cfs_rq)3988 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3989 {
3990 	return u64_u32_load_copy(cfs_rq->avg.last_update_time,
3991 				 cfs_rq->last_update_time_copy);
3992 }
3993 #ifdef CONFIG_FAIR_GROUP_SCHED
3994 /*
3995  * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3996  * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3997  * bottom-up, we only have to test whether the cfs_rq before us on the list
3998  * is our child.
3999  * If cfs_rq is not on the list, test whether a child needs its to be added to
4000  * connect a branch to the tree  * (see list_add_leaf_cfs_rq() for details).
4001  */
child_cfs_rq_on_list(struct cfs_rq * cfs_rq)4002 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
4003 {
4004 	struct cfs_rq *prev_cfs_rq;
4005 	struct list_head *prev;
4006 
4007 	if (cfs_rq->on_list) {
4008 		prev = cfs_rq->leaf_cfs_rq_list.prev;
4009 	} else {
4010 		struct rq *rq = rq_of(cfs_rq);
4011 
4012 		prev = rq->tmp_alone_branch;
4013 	}
4014 
4015 	prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
4016 
4017 	return (prev_cfs_rq->tg->parent == cfs_rq->tg);
4018 }
4019 
cfs_rq_is_decayed(struct cfs_rq * cfs_rq)4020 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4021 {
4022 	if (cfs_rq->load.weight)
4023 		return false;
4024 
4025 	if (!load_avg_is_decayed(&cfs_rq->avg))
4026 		return false;
4027 
4028 	if (child_cfs_rq_on_list(cfs_rq))
4029 		return false;
4030 
4031 	return true;
4032 }
4033 
4034 /**
4035  * update_tg_load_avg - update the tg's load avg
4036  * @cfs_rq: the cfs_rq whose avg changed
4037  *
4038  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
4039  * However, because tg->load_avg is a global value there are performance
4040  * considerations.
4041  *
4042  * In order to avoid having to look at the other cfs_rq's, we use a
4043  * differential update where we store the last value we propagated. This in
4044  * turn allows skipping updates if the differential is 'small'.
4045  *
4046  * Updating tg's load_avg is necessary before update_cfs_share().
4047  */
update_tg_load_avg(struct cfs_rq * cfs_rq)4048 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
4049 {
4050 	long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
4051 
4052 	/*
4053 	 * No need to update load_avg for root_task_group as it is not used.
4054 	 */
4055 	if (cfs_rq->tg == &root_task_group)
4056 		return;
4057 
4058 	if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
4059 		atomic_long_add(delta, &cfs_rq->tg->load_avg);
4060 		cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
4061 	}
4062 }
4063 
4064 /*
4065  * Called within set_task_rq() right before setting a task's CPU. The
4066  * caller only guarantees p->pi_lock is held; no other assumptions,
4067  * including the state of rq->lock, should be made.
4068  */
set_task_rq_fair(struct sched_entity * se,struct cfs_rq * prev,struct cfs_rq * next)4069 void set_task_rq_fair(struct sched_entity *se,
4070 		      struct cfs_rq *prev, struct cfs_rq *next)
4071 {
4072 	u64 p_last_update_time;
4073 	u64 n_last_update_time;
4074 
4075 	if (!sched_feat(ATTACH_AGE_LOAD))
4076 		return;
4077 
4078 	/*
4079 	 * We are supposed to update the task to "current" time, then its up to
4080 	 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
4081 	 * getting what current time is, so simply throw away the out-of-date
4082 	 * time. This will result in the wakee task is less decayed, but giving
4083 	 * the wakee more load sounds not bad.
4084 	 */
4085 	if (!(se->avg.last_update_time && prev))
4086 		return;
4087 
4088 	p_last_update_time = cfs_rq_last_update_time(prev);
4089 	n_last_update_time = cfs_rq_last_update_time(next);
4090 
4091 	__update_load_avg_blocked_se(p_last_update_time, se);
4092 	se->avg.last_update_time = n_last_update_time;
4093 }
4094 
4095 /*
4096  * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
4097  * propagate its contribution. The key to this propagation is the invariant
4098  * that for each group:
4099  *
4100  *   ge->avg == grq->avg						(1)
4101  *
4102  * _IFF_ we look at the pure running and runnable sums. Because they
4103  * represent the very same entity, just at different points in the hierarchy.
4104  *
4105  * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
4106  * and simply copies the running/runnable sum over (but still wrong, because
4107  * the group entity and group rq do not have their PELT windows aligned).
4108  *
4109  * However, update_tg_cfs_load() is more complex. So we have:
4110  *
4111  *   ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg		(2)
4112  *
4113  * And since, like util, the runnable part should be directly transferable,
4114  * the following would _appear_ to be the straight forward approach:
4115  *
4116  *   grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg	(3)
4117  *
4118  * And per (1) we have:
4119  *
4120  *   ge->avg.runnable_avg == grq->avg.runnable_avg
4121  *
4122  * Which gives:
4123  *
4124  *                      ge->load.weight * grq->avg.load_avg
4125  *   ge->avg.load_avg = -----------------------------------		(4)
4126  *                               grq->load.weight
4127  *
4128  * Except that is wrong!
4129  *
4130  * Because while for entities historical weight is not important and we
4131  * really only care about our future and therefore can consider a pure
4132  * runnable sum, runqueues can NOT do this.
4133  *
4134  * We specifically want runqueues to have a load_avg that includes
4135  * historical weights. Those represent the blocked load, the load we expect
4136  * to (shortly) return to us. This only works by keeping the weights as
4137  * integral part of the sum. We therefore cannot decompose as per (3).
4138  *
4139  * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4140  * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
4141  * rq itself is runnable anywhere between 2/3 and 1 depending on how the
4142  * runnable section of these tasks overlap (or not). If they were to perfectly
4143  * align the rq as a whole would be runnable 2/3 of the time. If however we
4144  * always have at least 1 runnable task, the rq as a whole is always runnable.
4145  *
4146  * So we'll have to approximate.. :/
4147  *
4148  * Given the constraint:
4149  *
4150  *   ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4151  *
4152  * We can construct a rule that adds runnable to a rq by assuming minimal
4153  * overlap.
4154  *
4155  * On removal, we'll assume each task is equally runnable; which yields:
4156  *
4157  *   grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4158  *
4159  * XXX: only do this for the part of runnable > running ?
4160  *
4161  */
4162 static inline void
update_tg_cfs_util(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4163 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4164 {
4165 	long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
4166 	u32 new_sum, divider;
4167 
4168 	/* Nothing to update */
4169 	if (!delta_avg)
4170 		return;
4171 
4172 	/*
4173 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4174 	 * See ___update_load_avg() for details.
4175 	 */
4176 	divider = get_pelt_divider(&cfs_rq->avg);
4177 
4178 
4179 	/* Set new sched_entity's utilization */
4180 	se->avg.util_avg = gcfs_rq->avg.util_avg;
4181 	new_sum = se->avg.util_avg * divider;
4182 	delta_sum = (long)new_sum - (long)se->avg.util_sum;
4183 	se->avg.util_sum = new_sum;
4184 
4185 	/* Update parent cfs_rq utilization */
4186 	add_positive(&cfs_rq->avg.util_avg, delta_avg);
4187 	add_positive(&cfs_rq->avg.util_sum, delta_sum);
4188 
4189 	/* See update_cfs_rq_load_avg() */
4190 	cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4191 					  cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4192 }
4193 
4194 static inline void
update_tg_cfs_runnable(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4195 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4196 {
4197 	long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
4198 	u32 new_sum, divider;
4199 
4200 	/* Nothing to update */
4201 	if (!delta_avg)
4202 		return;
4203 
4204 	/*
4205 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4206 	 * See ___update_load_avg() for details.
4207 	 */
4208 	divider = get_pelt_divider(&cfs_rq->avg);
4209 
4210 	/* Set new sched_entity's runnable */
4211 	se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
4212 	new_sum = se->avg.runnable_avg * divider;
4213 	delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
4214 	se->avg.runnable_sum = new_sum;
4215 
4216 	/* Update parent cfs_rq runnable */
4217 	add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
4218 	add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
4219 	/* See update_cfs_rq_load_avg() */
4220 	cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4221 					      cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4222 }
4223 
4224 static inline void
update_tg_cfs_load(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4225 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4226 {
4227 	long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
4228 	unsigned long load_avg;
4229 	u64 load_sum = 0;
4230 	s64 delta_sum;
4231 	u32 divider;
4232 
4233 	if (!runnable_sum)
4234 		return;
4235 
4236 	gcfs_rq->prop_runnable_sum = 0;
4237 
4238 	/*
4239 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4240 	 * See ___update_load_avg() for details.
4241 	 */
4242 	divider = get_pelt_divider(&cfs_rq->avg);
4243 
4244 	if (runnable_sum >= 0) {
4245 		/*
4246 		 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
4247 		 * the CPU is saturated running == runnable.
4248 		 */
4249 		runnable_sum += se->avg.load_sum;
4250 		runnable_sum = min_t(long, runnable_sum, divider);
4251 	} else {
4252 		/*
4253 		 * Estimate the new unweighted runnable_sum of the gcfs_rq by
4254 		 * assuming all tasks are equally runnable.
4255 		 */
4256 		if (scale_load_down(gcfs_rq->load.weight)) {
4257 			load_sum = div_u64(gcfs_rq->avg.load_sum,
4258 				scale_load_down(gcfs_rq->load.weight));
4259 		}
4260 
4261 		/* But make sure to not inflate se's runnable */
4262 		runnable_sum = min(se->avg.load_sum, load_sum);
4263 	}
4264 
4265 	/*
4266 	 * runnable_sum can't be lower than running_sum
4267 	 * Rescale running sum to be in the same range as runnable sum
4268 	 * running_sum is in [0 : LOAD_AVG_MAX <<  SCHED_CAPACITY_SHIFT]
4269 	 * runnable_sum is in [0 : LOAD_AVG_MAX]
4270 	 */
4271 	running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
4272 	runnable_sum = max(runnable_sum, running_sum);
4273 
4274 	load_sum = se_weight(se) * runnable_sum;
4275 	load_avg = div_u64(load_sum, divider);
4276 
4277 	delta_avg = load_avg - se->avg.load_avg;
4278 	if (!delta_avg)
4279 		return;
4280 
4281 	delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
4282 
4283 	se->avg.load_sum = runnable_sum;
4284 	se->avg.load_avg = load_avg;
4285 	add_positive(&cfs_rq->avg.load_avg, delta_avg);
4286 	add_positive(&cfs_rq->avg.load_sum, delta_sum);
4287 	/* See update_cfs_rq_load_avg() */
4288 	cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
4289 					  cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
4290 }
4291 
add_tg_cfs_propagate(struct cfs_rq * cfs_rq,long runnable_sum)4292 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
4293 {
4294 	cfs_rq->propagate = 1;
4295 	cfs_rq->prop_runnable_sum += runnable_sum;
4296 }
4297 
4298 /* Update task and its cfs_rq load average */
propagate_entity_load_avg(struct sched_entity * se)4299 static inline int propagate_entity_load_avg(struct sched_entity *se)
4300 {
4301 	struct cfs_rq *cfs_rq, *gcfs_rq;
4302 
4303 	if (entity_is_task(se))
4304 		return 0;
4305 
4306 	gcfs_rq = group_cfs_rq(se);
4307 	if (!gcfs_rq->propagate)
4308 		return 0;
4309 
4310 	gcfs_rq->propagate = 0;
4311 
4312 	cfs_rq = cfs_rq_of(se);
4313 
4314 	add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
4315 
4316 	update_tg_cfs_util(cfs_rq, se, gcfs_rq);
4317 	update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
4318 	update_tg_cfs_load(cfs_rq, se, gcfs_rq);
4319 
4320 	trace_pelt_cfs_tp(cfs_rq);
4321 	trace_pelt_se_tp(se);
4322 
4323 	return 1;
4324 }
4325 
4326 /*
4327  * Check if we need to update the load and the utilization of a blocked
4328  * group_entity:
4329  */
skip_blocked_update(struct sched_entity * se)4330 static inline bool skip_blocked_update(struct sched_entity *se)
4331 {
4332 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
4333 
4334 	/*
4335 	 * If sched_entity still have not zero load or utilization, we have to
4336 	 * decay it:
4337 	 */
4338 	if (se->avg.load_avg || se->avg.util_avg)
4339 		return false;
4340 
4341 	/*
4342 	 * If there is a pending propagation, we have to update the load and
4343 	 * the utilization of the sched_entity:
4344 	 */
4345 	if (gcfs_rq->propagate)
4346 		return false;
4347 
4348 	/*
4349 	 * Otherwise, the load and the utilization of the sched_entity is
4350 	 * already zero and there is no pending propagation, so it will be a
4351 	 * waste of time to try to decay it:
4352 	 */
4353 	return true;
4354 }
4355 
4356 #else /* CONFIG_FAIR_GROUP_SCHED */
4357 
update_tg_load_avg(struct cfs_rq * cfs_rq)4358 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
4359 
propagate_entity_load_avg(struct sched_entity * se)4360 static inline int propagate_entity_load_avg(struct sched_entity *se)
4361 {
4362 	return 0;
4363 }
4364 
add_tg_cfs_propagate(struct cfs_rq * cfs_rq,long runnable_sum)4365 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
4366 
4367 #endif /* CONFIG_FAIR_GROUP_SCHED */
4368 
4369 #ifdef CONFIG_NO_HZ_COMMON
migrate_se_pelt_lag(struct sched_entity * se)4370 static inline void migrate_se_pelt_lag(struct sched_entity *se)
4371 {
4372 	u64 throttled = 0, now, lut;
4373 	struct cfs_rq *cfs_rq;
4374 	struct rq *rq;
4375 	bool is_idle;
4376 
4377 	if (load_avg_is_decayed(&se->avg))
4378 		return;
4379 
4380 	cfs_rq = cfs_rq_of(se);
4381 	rq = rq_of(cfs_rq);
4382 
4383 	rcu_read_lock();
4384 	is_idle = is_idle_task(rcu_dereference(rq->curr));
4385 	rcu_read_unlock();
4386 
4387 	/*
4388 	 * The lag estimation comes with a cost we don't want to pay all the
4389 	 * time. Hence, limiting to the case where the source CPU is idle and
4390 	 * we know we are at the greatest risk to have an outdated clock.
4391 	 */
4392 	if (!is_idle)
4393 		return;
4394 
4395 	/*
4396 	 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
4397 	 *
4398 	 *   last_update_time (the cfs_rq's last_update_time)
4399 	 *	= cfs_rq_clock_pelt()@cfs_rq_idle
4400 	 *      = rq_clock_pelt()@cfs_rq_idle
4401 	 *        - cfs->throttled_clock_pelt_time@cfs_rq_idle
4402 	 *
4403 	 *   cfs_idle_lag (delta between rq's update and cfs_rq's update)
4404 	 *      = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
4405 	 *
4406 	 *   rq_idle_lag (delta between now and rq's update)
4407 	 *      = sched_clock_cpu() - rq_clock()@rq_idle
4408 	 *
4409 	 * We can then write:
4410 	 *
4411 	 *    now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
4412 	 *          sched_clock_cpu() - rq_clock()@rq_idle
4413 	 * Where:
4414 	 *      rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
4415 	 *      rq_clock()@rq_idle      is rq->clock_idle
4416 	 *      cfs->throttled_clock_pelt_time@cfs_rq_idle
4417 	 *                              is cfs_rq->throttled_pelt_idle
4418 	 */
4419 
4420 #ifdef CONFIG_CFS_BANDWIDTH
4421 	throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
4422 	/* The clock has been stopped for throttling */
4423 	if (throttled == U64_MAX)
4424 		return;
4425 #endif
4426 	now = u64_u32_load(rq->clock_pelt_idle);
4427 	/*
4428 	 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
4429 	 * is observed the old clock_pelt_idle value and the new clock_idle,
4430 	 * which lead to an underestimation. The opposite would lead to an
4431 	 * overestimation.
4432 	 */
4433 	smp_rmb();
4434 	lut = cfs_rq_last_update_time(cfs_rq);
4435 
4436 	now -= throttled;
4437 	if (now < lut)
4438 		/*
4439 		 * cfs_rq->avg.last_update_time is more recent than our
4440 		 * estimation, let's use it.
4441 		 */
4442 		now = lut;
4443 	else
4444 		now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4445 
4446 	__update_load_avg_blocked_se(now, se);
4447 }
4448 #else
migrate_se_pelt_lag(struct sched_entity * se)4449 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4450 #endif
4451 
4452 /**
4453  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4454  * @now: current time, as per cfs_rq_clock_pelt()
4455  * @cfs_rq: cfs_rq to update
4456  *
4457  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4458  * avg. The immediate corollary is that all (fair) tasks must be attached.
4459  *
4460  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4461  *
4462  * Return: true if the load decayed or we removed load.
4463  *
4464  * Since both these conditions indicate a changed cfs_rq->avg.load we should
4465  * call update_tg_load_avg() when this function returns true.
4466  */
4467 static inline int
update_cfs_rq_load_avg(u64 now,struct cfs_rq * cfs_rq)4468 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4469 {
4470 	unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4471 	struct sched_avg *sa = &cfs_rq->avg;
4472 	int decayed = 0;
4473 
4474 	if (cfs_rq->removed.nr) {
4475 		unsigned long r;
4476 		u32 divider = get_pelt_divider(&cfs_rq->avg);
4477 
4478 		raw_spin_lock(&cfs_rq->removed.lock);
4479 		swap(cfs_rq->removed.util_avg, removed_util);
4480 		swap(cfs_rq->removed.load_avg, removed_load);
4481 		swap(cfs_rq->removed.runnable_avg, removed_runnable);
4482 		cfs_rq->removed.nr = 0;
4483 		raw_spin_unlock(&cfs_rq->removed.lock);
4484 
4485 		r = removed_load;
4486 		sub_positive(&sa->load_avg, r);
4487 		sub_positive(&sa->load_sum, r * divider);
4488 		/* See sa->util_sum below */
4489 		sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4490 
4491 		r = removed_util;
4492 		sub_positive(&sa->util_avg, r);
4493 		sub_positive(&sa->util_sum, r * divider);
4494 		/*
4495 		 * Because of rounding, se->util_sum might ends up being +1 more than
4496 		 * cfs->util_sum. Although this is not a problem by itself, detaching
4497 		 * a lot of tasks with the rounding problem between 2 updates of
4498 		 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4499 		 * cfs_util_avg is not.
4500 		 * Check that util_sum is still above its lower bound for the new
4501 		 * util_avg. Given that period_contrib might have moved since the last
4502 		 * sync, we are only sure that util_sum must be above or equal to
4503 		 *    util_avg * minimum possible divider
4504 		 */
4505 		sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4506 
4507 		r = removed_runnable;
4508 		sub_positive(&sa->runnable_avg, r);
4509 		sub_positive(&sa->runnable_sum, r * divider);
4510 		/* See sa->util_sum above */
4511 		sa->runnable_sum = max_t(u32, sa->runnable_sum,
4512 					      sa->runnable_avg * PELT_MIN_DIVIDER);
4513 
4514 		/*
4515 		 * removed_runnable is the unweighted version of removed_load so we
4516 		 * can use it to estimate removed_load_sum.
4517 		 */
4518 		add_tg_cfs_propagate(cfs_rq,
4519 			-(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4520 
4521 		decayed = 1;
4522 	}
4523 
4524 	decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4525 	u64_u32_store_copy(sa->last_update_time,
4526 			   cfs_rq->last_update_time_copy,
4527 			   sa->last_update_time);
4528 	return decayed;
4529 }
4530 
4531 /**
4532  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4533  * @cfs_rq: cfs_rq to attach to
4534  * @se: sched_entity to attach
4535  *
4536  * Must call update_cfs_rq_load_avg() before this, since we rely on
4537  * cfs_rq->avg.last_update_time being current.
4538  */
attach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)4539 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4540 {
4541 	/*
4542 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4543 	 * See ___update_load_avg() for details.
4544 	 */
4545 	u32 divider = get_pelt_divider(&cfs_rq->avg);
4546 
4547 	/*
4548 	 * When we attach the @se to the @cfs_rq, we must align the decay
4549 	 * window because without that, really weird and wonderful things can
4550 	 * happen.
4551 	 *
4552 	 * XXX illustrate
4553 	 */
4554 	se->avg.last_update_time = cfs_rq->avg.last_update_time;
4555 	se->avg.period_contrib = cfs_rq->avg.period_contrib;
4556 
4557 	/*
4558 	 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4559 	 * period_contrib. This isn't strictly correct, but since we're
4560 	 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4561 	 * _sum a little.
4562 	 */
4563 	se->avg.util_sum = se->avg.util_avg * divider;
4564 
4565 	se->avg.runnable_sum = se->avg.runnable_avg * divider;
4566 
4567 	se->avg.load_sum = se->avg.load_avg * divider;
4568 	if (se_weight(se) < se->avg.load_sum)
4569 		se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4570 	else
4571 		se->avg.load_sum = 1;
4572 
4573 	enqueue_load_avg(cfs_rq, se);
4574 	cfs_rq->avg.util_avg += se->avg.util_avg;
4575 	cfs_rq->avg.util_sum += se->avg.util_sum;
4576 	cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4577 	cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4578 
4579 	add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4580 
4581 	cfs_rq_util_change(cfs_rq, 0);
4582 
4583 	trace_pelt_cfs_tp(cfs_rq);
4584 }
4585 
4586 /**
4587  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4588  * @cfs_rq: cfs_rq to detach from
4589  * @se: sched_entity to detach
4590  *
4591  * Must call update_cfs_rq_load_avg() before this, since we rely on
4592  * cfs_rq->avg.last_update_time being current.
4593  */
detach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)4594 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4595 {
4596 	dequeue_load_avg(cfs_rq, se);
4597 	sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4598 	sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4599 	/* See update_cfs_rq_load_avg() */
4600 	cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4601 					  cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4602 
4603 	sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4604 	sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4605 	/* See update_cfs_rq_load_avg() */
4606 	cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4607 					      cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4608 
4609 	add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4610 
4611 	cfs_rq_util_change(cfs_rq, 0);
4612 
4613 	trace_pelt_cfs_tp(cfs_rq);
4614 }
4615 
4616 /*
4617  * Optional action to be done while updating the load average
4618  */
4619 #define UPDATE_TG	0x1
4620 #define SKIP_AGE_LOAD	0x2
4621 #define DO_ATTACH	0x4
4622 #define DO_DETACH	0x8
4623 
4624 /* Update task and its cfs_rq load average */
update_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)4625 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4626 {
4627 	u64 now = cfs_rq_clock_pelt(cfs_rq);
4628 	int decayed;
4629 
4630 	/*
4631 	 * Track task load average for carrying it to new CPU after migrated, and
4632 	 * track group sched_entity load average for task_h_load calc in migration
4633 	 */
4634 	if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4635 		__update_load_avg_se(now, cfs_rq, se);
4636 
4637 	decayed  = update_cfs_rq_load_avg(now, cfs_rq);
4638 	decayed |= propagate_entity_load_avg(se);
4639 
4640 	if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4641 
4642 		/*
4643 		 * DO_ATTACH means we're here from enqueue_entity().
4644 		 * !last_update_time means we've passed through
4645 		 * migrate_task_rq_fair() indicating we migrated.
4646 		 *
4647 		 * IOW we're enqueueing a task on a new CPU.
4648 		 */
4649 		attach_entity_load_avg(cfs_rq, se);
4650 		update_tg_load_avg(cfs_rq);
4651 
4652 	} else if (flags & DO_DETACH) {
4653 		/*
4654 		 * DO_DETACH means we're here from dequeue_entity()
4655 		 * and we are migrating task out of the CPU.
4656 		 */
4657 		detach_entity_load_avg(cfs_rq, se);
4658 		update_tg_load_avg(cfs_rq);
4659 	} else if (decayed) {
4660 		cfs_rq_util_change(cfs_rq, 0);
4661 
4662 		if (flags & UPDATE_TG)
4663 			update_tg_load_avg(cfs_rq);
4664 	}
4665 }
4666 
4667 /*
4668  * Synchronize entity load avg of dequeued entity without locking
4669  * the previous rq.
4670  */
sync_entity_load_avg(struct sched_entity * se)4671 static void sync_entity_load_avg(struct sched_entity *se)
4672 {
4673 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4674 	u64 last_update_time;
4675 
4676 	last_update_time = cfs_rq_last_update_time(cfs_rq);
4677 	__update_load_avg_blocked_se(last_update_time, se);
4678 }
4679 
4680 /*
4681  * Task first catches up with cfs_rq, and then subtract
4682  * itself from the cfs_rq (task must be off the queue now).
4683  */
remove_entity_load_avg(struct sched_entity * se)4684 static void remove_entity_load_avg(struct sched_entity *se)
4685 {
4686 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4687 	unsigned long flags;
4688 
4689 	/*
4690 	 * tasks cannot exit without having gone through wake_up_new_task() ->
4691 	 * enqueue_task_fair() which will have added things to the cfs_rq,
4692 	 * so we can remove unconditionally.
4693 	 */
4694 
4695 	sync_entity_load_avg(se);
4696 
4697 	raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4698 	++cfs_rq->removed.nr;
4699 	cfs_rq->removed.util_avg	+= se->avg.util_avg;
4700 	cfs_rq->removed.load_avg	+= se->avg.load_avg;
4701 	cfs_rq->removed.runnable_avg	+= se->avg.runnable_avg;
4702 	raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4703 }
4704 
cfs_rq_runnable_avg(struct cfs_rq * cfs_rq)4705 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4706 {
4707 	return cfs_rq->avg.runnable_avg;
4708 }
4709 
cfs_rq_load_avg(struct cfs_rq * cfs_rq)4710 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4711 {
4712 	return cfs_rq->avg.load_avg;
4713 }
4714 
4715 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4716 
task_util(struct task_struct * p)4717 static inline unsigned long task_util(struct task_struct *p)
4718 {
4719 	return READ_ONCE(p->se.avg.util_avg);
4720 }
4721 
_task_util_est(struct task_struct * p)4722 static inline unsigned long _task_util_est(struct task_struct *p)
4723 {
4724 	struct util_est ue = READ_ONCE(p->se.avg.util_est);
4725 
4726 	return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4727 }
4728 
task_util_est(struct task_struct * p)4729 static inline unsigned long task_util_est(struct task_struct *p)
4730 {
4731 	return max(task_util(p), _task_util_est(p));
4732 }
4733 
util_est_enqueue(struct cfs_rq * cfs_rq,struct task_struct * p)4734 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4735 				    struct task_struct *p)
4736 {
4737 	unsigned int enqueued;
4738 
4739 	if (!sched_feat(UTIL_EST))
4740 		return;
4741 
4742 	/* Update root cfs_rq's estimated utilization */
4743 	enqueued  = cfs_rq->avg.util_est.enqueued;
4744 	enqueued += _task_util_est(p);
4745 	WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4746 
4747 	trace_sched_util_est_cfs_tp(cfs_rq);
4748 }
4749 
util_est_dequeue(struct cfs_rq * cfs_rq,struct task_struct * p)4750 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4751 				    struct task_struct *p)
4752 {
4753 	unsigned int enqueued;
4754 
4755 	if (!sched_feat(UTIL_EST))
4756 		return;
4757 
4758 	/* Update root cfs_rq's estimated utilization */
4759 	enqueued  = cfs_rq->avg.util_est.enqueued;
4760 	enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4761 	WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4762 
4763 	trace_sched_util_est_cfs_tp(cfs_rq);
4764 }
4765 
4766 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4767 
4768 /*
4769  * Check if a (signed) value is within a specified (unsigned) margin,
4770  * based on the observation that:
4771  *
4772  *     abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4773  *
4774  * NOTE: this only works when value + margin < INT_MAX.
4775  */
within_margin(int value,int margin)4776 static inline bool within_margin(int value, int margin)
4777 {
4778 	return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4779 }
4780 
util_est_update(struct cfs_rq * cfs_rq,struct task_struct * p,bool task_sleep)4781 static inline void util_est_update(struct cfs_rq *cfs_rq,
4782 				   struct task_struct *p,
4783 				   bool task_sleep)
4784 {
4785 	long last_ewma_diff, last_enqueued_diff;
4786 	struct util_est ue;
4787 
4788 	if (!sched_feat(UTIL_EST))
4789 		return;
4790 
4791 	/*
4792 	 * Skip update of task's estimated utilization when the task has not
4793 	 * yet completed an activation, e.g. being migrated.
4794 	 */
4795 	if (!task_sleep)
4796 		return;
4797 
4798 	/*
4799 	 * If the PELT values haven't changed since enqueue time,
4800 	 * skip the util_est update.
4801 	 */
4802 	ue = p->se.avg.util_est;
4803 	if (ue.enqueued & UTIL_AVG_UNCHANGED)
4804 		return;
4805 
4806 	last_enqueued_diff = ue.enqueued;
4807 
4808 	/*
4809 	 * Reset EWMA on utilization increases, the moving average is used only
4810 	 * to smooth utilization decreases.
4811 	 */
4812 	ue.enqueued = task_util(p);
4813 	if (sched_feat(UTIL_EST_FASTUP)) {
4814 		if (ue.ewma < ue.enqueued) {
4815 			ue.ewma = ue.enqueued;
4816 			goto done;
4817 		}
4818 	}
4819 
4820 	/*
4821 	 * Skip update of task's estimated utilization when its members are
4822 	 * already ~1% close to its last activation value.
4823 	 */
4824 	last_ewma_diff = ue.enqueued - ue.ewma;
4825 	last_enqueued_diff -= ue.enqueued;
4826 	if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4827 		if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4828 			goto done;
4829 
4830 		return;
4831 	}
4832 
4833 	/*
4834 	 * To avoid overestimation of actual task utilization, skip updates if
4835 	 * we cannot grant there is idle time in this CPU.
4836 	 */
4837 	if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4838 		return;
4839 
4840 	/*
4841 	 * Update Task's estimated utilization
4842 	 *
4843 	 * When *p completes an activation we can consolidate another sample
4844 	 * of the task size. This is done by storing the current PELT value
4845 	 * as ue.enqueued and by using this value to update the Exponential
4846 	 * Weighted Moving Average (EWMA):
4847 	 *
4848 	 *  ewma(t) = w *  task_util(p) + (1-w) * ewma(t-1)
4849 	 *          = w *  task_util(p) +         ewma(t-1)  - w * ewma(t-1)
4850 	 *          = w * (task_util(p) -         ewma(t-1)) +     ewma(t-1)
4851 	 *          = w * (      last_ewma_diff            ) +     ewma(t-1)
4852 	 *          = w * (last_ewma_diff  +  ewma(t-1) / w)
4853 	 *
4854 	 * Where 'w' is the weight of new samples, which is configured to be
4855 	 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4856 	 */
4857 	ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4858 	ue.ewma  += last_ewma_diff;
4859 	ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4860 done:
4861 	ue.enqueued |= UTIL_AVG_UNCHANGED;
4862 	WRITE_ONCE(p->se.avg.util_est, ue);
4863 
4864 	trace_sched_util_est_se_tp(&p->se);
4865 }
4866 
util_fits_cpu(unsigned long util,unsigned long uclamp_min,unsigned long uclamp_max,int cpu)4867 static inline int util_fits_cpu(unsigned long util,
4868 				unsigned long uclamp_min,
4869 				unsigned long uclamp_max,
4870 				int cpu)
4871 {
4872 	unsigned long capacity_orig, capacity_orig_thermal;
4873 	unsigned long capacity = capacity_of(cpu);
4874 	bool fits, uclamp_max_fits;
4875 
4876 	/*
4877 	 * Check if the real util fits without any uclamp boost/cap applied.
4878 	 */
4879 	fits = fits_capacity(util, capacity);
4880 
4881 	if (!uclamp_is_used())
4882 		return fits;
4883 
4884 	/*
4885 	 * We must use capacity_orig_of() for comparing against uclamp_min and
4886 	 * uclamp_max. We only care about capacity pressure (by using
4887 	 * capacity_of()) for comparing against the real util.
4888 	 *
4889 	 * If a task is boosted to 1024 for example, we don't want a tiny
4890 	 * pressure to skew the check whether it fits a CPU or not.
4891 	 *
4892 	 * Similarly if a task is capped to capacity_orig_of(little_cpu), it
4893 	 * should fit a little cpu even if there's some pressure.
4894 	 *
4895 	 * Only exception is for thermal pressure since it has a direct impact
4896 	 * on available OPP of the system.
4897 	 *
4898 	 * We honour it for uclamp_min only as a drop in performance level
4899 	 * could result in not getting the requested minimum performance level.
4900 	 *
4901 	 * For uclamp_max, we can tolerate a drop in performance level as the
4902 	 * goal is to cap the task. So it's okay if it's getting less.
4903 	 */
4904 	capacity_orig = capacity_orig_of(cpu);
4905 	capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
4906 
4907 	/*
4908 	 * We want to force a task to fit a cpu as implied by uclamp_max.
4909 	 * But we do have some corner cases to cater for..
4910 	 *
4911 	 *
4912 	 *                                 C=z
4913 	 *   |                             ___
4914 	 *   |                  C=y       |   |
4915 	 *   |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _  uclamp_max
4916 	 *   |      C=x        |   |      |   |
4917 	 *   |      ___        |   |      |   |
4918 	 *   |     |   |       |   |      |   |    (util somewhere in this region)
4919 	 *   |     |   |       |   |      |   |
4920 	 *   |     |   |       |   |      |   |
4921 	 *   +----------------------------------------
4922 	 *         cpu0        cpu1       cpu2
4923 	 *
4924 	 *   In the above example if a task is capped to a specific performance
4925 	 *   point, y, then when:
4926 	 *
4927 	 *   * util = 80% of x then it does not fit on cpu0 and should migrate
4928 	 *     to cpu1
4929 	 *   * util = 80% of y then it is forced to fit on cpu1 to honour
4930 	 *     uclamp_max request.
4931 	 *
4932 	 *   which is what we're enforcing here. A task always fits if
4933 	 *   uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
4934 	 *   the normal upmigration rules should withhold still.
4935 	 *
4936 	 *   Only exception is when we are on max capacity, then we need to be
4937 	 *   careful not to block overutilized state. This is so because:
4938 	 *
4939 	 *     1. There's no concept of capping at max_capacity! We can't go
4940 	 *        beyond this performance level anyway.
4941 	 *     2. The system is being saturated when we're operating near
4942 	 *        max capacity, it doesn't make sense to block overutilized.
4943 	 */
4944 	uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
4945 	uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
4946 	fits = fits || uclamp_max_fits;
4947 
4948 	/*
4949 	 *
4950 	 *                                 C=z
4951 	 *   |                             ___       (region a, capped, util >= uclamp_max)
4952 	 *   |                  C=y       |   |
4953 	 *   |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4954 	 *   |      C=x        |   |      |   |
4955 	 *   |      ___        |   |      |   |      (region b, uclamp_min <= util <= uclamp_max)
4956 	 *   |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
4957 	 *   |     |   |       |   |      |   |
4958 	 *   |     |   |       |   |      |   |      (region c, boosted, util < uclamp_min)
4959 	 *   +----------------------------------------
4960 	 *         cpu0        cpu1       cpu2
4961 	 *
4962 	 * a) If util > uclamp_max, then we're capped, we don't care about
4963 	 *    actual fitness value here. We only care if uclamp_max fits
4964 	 *    capacity without taking margin/pressure into account.
4965 	 *    See comment above.
4966 	 *
4967 	 * b) If uclamp_min <= util <= uclamp_max, then the normal
4968 	 *    fits_capacity() rules apply. Except we need to ensure that we
4969 	 *    enforce we remain within uclamp_max, see comment above.
4970 	 *
4971 	 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
4972 	 *    need to take into account the boosted value fits the CPU without
4973 	 *    taking margin/pressure into account.
4974 	 *
4975 	 * Cases (a) and (b) are handled in the 'fits' variable already. We
4976 	 * just need to consider an extra check for case (c) after ensuring we
4977 	 * handle the case uclamp_min > uclamp_max.
4978 	 */
4979 	uclamp_min = min(uclamp_min, uclamp_max);
4980 	if (fits && (util < uclamp_min) && (uclamp_min > capacity_orig_thermal))
4981 		return -1;
4982 
4983 	return fits;
4984 }
4985 
task_fits_cpu(struct task_struct * p,int cpu)4986 static inline int task_fits_cpu(struct task_struct *p, int cpu)
4987 {
4988 	unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
4989 	unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
4990 	unsigned long util = task_util_est(p);
4991 	/*
4992 	 * Return true only if the cpu fully fits the task requirements, which
4993 	 * include the utilization but also the performance hints.
4994 	 */
4995 	return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
4996 }
4997 
update_misfit_status(struct task_struct * p,struct rq * rq)4998 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4999 {
5000 	if (!sched_asym_cpucap_active())
5001 		return;
5002 
5003 	if (!p || p->nr_cpus_allowed == 1) {
5004 		rq->misfit_task_load = 0;
5005 		return;
5006 	}
5007 
5008 	if (task_fits_cpu(p, cpu_of(rq))) {
5009 		rq->misfit_task_load = 0;
5010 		return;
5011 	}
5012 
5013 	/*
5014 	 * Make sure that misfit_task_load will not be null even if
5015 	 * task_h_load() returns 0.
5016 	 */
5017 	rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
5018 }
5019 
5020 #else /* CONFIG_SMP */
5021 
cfs_rq_is_decayed(struct cfs_rq * cfs_rq)5022 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
5023 {
5024 	return !cfs_rq->nr_running;
5025 }
5026 
5027 #define UPDATE_TG	0x0
5028 #define SKIP_AGE_LOAD	0x0
5029 #define DO_ATTACH	0x0
5030 #define DO_DETACH	0x0
5031 
update_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se,int not_used1)5032 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
5033 {
5034 	cfs_rq_util_change(cfs_rq, 0);
5035 }
5036 
remove_entity_load_avg(struct sched_entity * se)5037 static inline void remove_entity_load_avg(struct sched_entity *se) {}
5038 
5039 static inline void
attach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)5040 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5041 static inline void
detach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)5042 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5043 
newidle_balance(struct rq * rq,struct rq_flags * rf)5044 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
5045 {
5046 	return 0;
5047 }
5048 
5049 static inline void
util_est_enqueue(struct cfs_rq * cfs_rq,struct task_struct * p)5050 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5051 
5052 static inline void
util_est_dequeue(struct cfs_rq * cfs_rq,struct task_struct * p)5053 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5054 
5055 static inline void
util_est_update(struct cfs_rq * cfs_rq,struct task_struct * p,bool task_sleep)5056 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
5057 		bool task_sleep) {}
update_misfit_status(struct task_struct * p,struct rq * rq)5058 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
5059 
5060 #endif /* CONFIG_SMP */
5061 
5062 static void
place_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5063 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5064 {
5065 	u64 vslice, vruntime = avg_vruntime(cfs_rq);
5066 	s64 lag = 0;
5067 
5068 	se->slice = sysctl_sched_base_slice;
5069 	vslice = calc_delta_fair(se->slice, se);
5070 
5071 	/*
5072 	 * Due to how V is constructed as the weighted average of entities,
5073 	 * adding tasks with positive lag, or removing tasks with negative lag
5074 	 * will move 'time' backwards, this can screw around with the lag of
5075 	 * other tasks.
5076 	 *
5077 	 * EEVDF: placement strategy #1 / #2
5078 	 */
5079 	if (sched_feat(PLACE_LAG) && cfs_rq->nr_running) {
5080 		struct sched_entity *curr = cfs_rq->curr;
5081 		unsigned long load;
5082 
5083 		lag = se->vlag;
5084 
5085 		/*
5086 		 * If we want to place a task and preserve lag, we have to
5087 		 * consider the effect of the new entity on the weighted
5088 		 * average and compensate for this, otherwise lag can quickly
5089 		 * evaporate.
5090 		 *
5091 		 * Lag is defined as:
5092 		 *
5093 		 *   lag_i = S - s_i = w_i * (V - v_i)
5094 		 *
5095 		 * To avoid the 'w_i' term all over the place, we only track
5096 		 * the virtual lag:
5097 		 *
5098 		 *   vl_i = V - v_i <=> v_i = V - vl_i
5099 		 *
5100 		 * And we take V to be the weighted average of all v:
5101 		 *
5102 		 *   V = (\Sum w_j*v_j) / W
5103 		 *
5104 		 * Where W is: \Sum w_j
5105 		 *
5106 		 * Then, the weighted average after adding an entity with lag
5107 		 * vl_i is given by:
5108 		 *
5109 		 *   V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i)
5110 		 *      = (W*V + w_i*(V - vl_i)) / (W + w_i)
5111 		 *      = (W*V + w_i*V - w_i*vl_i) / (W + w_i)
5112 		 *      = (V*(W + w_i) - w_i*l) / (W + w_i)
5113 		 *      = V - w_i*vl_i / (W + w_i)
5114 		 *
5115 		 * And the actual lag after adding an entity with vl_i is:
5116 		 *
5117 		 *   vl'_i = V' - v_i
5118 		 *         = V - w_i*vl_i / (W + w_i) - (V - vl_i)
5119 		 *         = vl_i - w_i*vl_i / (W + w_i)
5120 		 *
5121 		 * Which is strictly less than vl_i. So in order to preserve lag
5122 		 * we should inflate the lag before placement such that the
5123 		 * effective lag after placement comes out right.
5124 		 *
5125 		 * As such, invert the above relation for vl'_i to get the vl_i
5126 		 * we need to use such that the lag after placement is the lag
5127 		 * we computed before dequeue.
5128 		 *
5129 		 *   vl'_i = vl_i - w_i*vl_i / (W + w_i)
5130 		 *         = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)
5131 		 *
5132 		 *   (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i
5133 		 *                   = W*vl_i
5134 		 *
5135 		 *   vl_i = (W + w_i)*vl'_i / W
5136 		 */
5137 		load = cfs_rq->avg_load;
5138 		if (curr && curr->on_rq)
5139 			load += scale_load_down(curr->load.weight);
5140 
5141 		lag *= load + scale_load_down(se->load.weight);
5142 		if (WARN_ON_ONCE(!load))
5143 			load = 1;
5144 		lag = div_s64(lag, load);
5145 	}
5146 
5147 	se->vruntime = vruntime - lag;
5148 
5149 	/*
5150 	 * When joining the competition; the exisiting tasks will be,
5151 	 * on average, halfway through their slice, as such start tasks
5152 	 * off with half a slice to ease into the competition.
5153 	 */
5154 	if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL))
5155 		vslice /= 2;
5156 
5157 	/*
5158 	 * EEVDF: vd_i = ve_i + r_i/w_i
5159 	 */
5160 	se->deadline = se->vruntime + vslice;
5161 }
5162 
5163 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
5164 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
5165 
5166 static inline bool cfs_bandwidth_used(void);
5167 
5168 static void
enqueue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5169 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5170 {
5171 	bool curr = cfs_rq->curr == se;
5172 
5173 	/*
5174 	 * If we're the current task, we must renormalise before calling
5175 	 * update_curr().
5176 	 */
5177 	if (curr)
5178 		place_entity(cfs_rq, se, flags);
5179 
5180 	update_curr(cfs_rq);
5181 
5182 	/*
5183 	 * When enqueuing a sched_entity, we must:
5184 	 *   - Update loads to have both entity and cfs_rq synced with now.
5185 	 *   - For group_entity, update its runnable_weight to reflect the new
5186 	 *     h_nr_running of its group cfs_rq.
5187 	 *   - For group_entity, update its weight to reflect the new share of
5188 	 *     its group cfs_rq
5189 	 *   - Add its new weight to cfs_rq->load.weight
5190 	 */
5191 	update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
5192 	se_update_runnable(se);
5193 	/*
5194 	 * XXX update_load_avg() above will have attached us to the pelt sum;
5195 	 * but update_cfs_group() here will re-adjust the weight and have to
5196 	 * undo/redo all that. Seems wasteful.
5197 	 */
5198 	update_cfs_group(se);
5199 
5200 	/*
5201 	 * XXX now that the entity has been re-weighted, and it's lag adjusted,
5202 	 * we can place the entity.
5203 	 */
5204 	if (!curr)
5205 		place_entity(cfs_rq, se, flags);
5206 
5207 	account_entity_enqueue(cfs_rq, se);
5208 
5209 	/* Entity has migrated, no longer consider this task hot */
5210 	if (flags & ENQUEUE_MIGRATED)
5211 		se->exec_start = 0;
5212 
5213 	check_schedstat_required();
5214 	update_stats_enqueue_fair(cfs_rq, se, flags);
5215 	if (!curr)
5216 		__enqueue_entity(cfs_rq, se);
5217 	se->on_rq = 1;
5218 
5219 	if (cfs_rq->nr_running == 1) {
5220 		check_enqueue_throttle(cfs_rq);
5221 		if (!throttled_hierarchy(cfs_rq)) {
5222 			list_add_leaf_cfs_rq(cfs_rq);
5223 		} else {
5224 #ifdef CONFIG_CFS_BANDWIDTH
5225 			struct rq *rq = rq_of(cfs_rq);
5226 
5227 			if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)
5228 				cfs_rq->throttled_clock = rq_clock(rq);
5229 			if (!cfs_rq->throttled_clock_self)
5230 				cfs_rq->throttled_clock_self = rq_clock(rq);
5231 #endif
5232 		}
5233 	}
5234 }
5235 
__clear_buddies_next(struct sched_entity * se)5236 static void __clear_buddies_next(struct sched_entity *se)
5237 {
5238 	for_each_sched_entity(se) {
5239 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
5240 		if (cfs_rq->next != se)
5241 			break;
5242 
5243 		cfs_rq->next = NULL;
5244 	}
5245 }
5246 
clear_buddies(struct cfs_rq * cfs_rq,struct sched_entity * se)5247 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
5248 {
5249 	if (cfs_rq->next == se)
5250 		__clear_buddies_next(se);
5251 }
5252 
5253 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5254 
5255 static void
dequeue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5256 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5257 {
5258 	int action = UPDATE_TG;
5259 
5260 	if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
5261 		action |= DO_DETACH;
5262 
5263 	/*
5264 	 * Update run-time statistics of the 'current'.
5265 	 */
5266 	update_curr(cfs_rq);
5267 
5268 	/*
5269 	 * When dequeuing a sched_entity, we must:
5270 	 *   - Update loads to have both entity and cfs_rq synced with now.
5271 	 *   - For group_entity, update its runnable_weight to reflect the new
5272 	 *     h_nr_running of its group cfs_rq.
5273 	 *   - Subtract its previous weight from cfs_rq->load.weight.
5274 	 *   - For group entity, update its weight to reflect the new share
5275 	 *     of its group cfs_rq.
5276 	 */
5277 	update_load_avg(cfs_rq, se, action);
5278 	se_update_runnable(se);
5279 
5280 	update_stats_dequeue_fair(cfs_rq, se, flags);
5281 
5282 	clear_buddies(cfs_rq, se);
5283 
5284 	update_entity_lag(cfs_rq, se);
5285 	if (se != cfs_rq->curr)
5286 		__dequeue_entity(cfs_rq, se);
5287 	se->on_rq = 0;
5288 	account_entity_dequeue(cfs_rq, se);
5289 
5290 	/* return excess runtime on last dequeue */
5291 	return_cfs_rq_runtime(cfs_rq);
5292 
5293 	update_cfs_group(se);
5294 
5295 	/*
5296 	 * Now advance min_vruntime if @se was the entity holding it back,
5297 	 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
5298 	 * put back on, and if we advance min_vruntime, we'll be placed back
5299 	 * further than we started -- ie. we'll be penalized.
5300 	 */
5301 	if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
5302 		update_min_vruntime(cfs_rq);
5303 
5304 	if (cfs_rq->nr_running == 0)
5305 		update_idle_cfs_rq_clock_pelt(cfs_rq);
5306 }
5307 
5308 static void
set_next_entity(struct cfs_rq * cfs_rq,struct sched_entity * se)5309 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
5310 {
5311 	clear_buddies(cfs_rq, se);
5312 
5313 	/* 'current' is not kept within the tree. */
5314 	if (se->on_rq) {
5315 		/*
5316 		 * Any task has to be enqueued before it get to execute on
5317 		 * a CPU. So account for the time it spent waiting on the
5318 		 * runqueue.
5319 		 */
5320 		update_stats_wait_end_fair(cfs_rq, se);
5321 		__dequeue_entity(cfs_rq, se);
5322 		update_load_avg(cfs_rq, se, UPDATE_TG);
5323 		/*
5324 		 * HACK, stash a copy of deadline at the point of pick in vlag,
5325 		 * which isn't used until dequeue.
5326 		 */
5327 		se->vlag = se->deadline;
5328 	}
5329 
5330 	update_stats_curr_start(cfs_rq, se);
5331 	cfs_rq->curr = se;
5332 
5333 	/*
5334 	 * Track our maximum slice length, if the CPU's load is at
5335 	 * least twice that of our own weight (i.e. dont track it
5336 	 * when there are only lesser-weight tasks around):
5337 	 */
5338 	if (schedstat_enabled() &&
5339 	    rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
5340 		struct sched_statistics *stats;
5341 
5342 		stats = __schedstats_from_se(se);
5343 		__schedstat_set(stats->slice_max,
5344 				max((u64)stats->slice_max,
5345 				    se->sum_exec_runtime - se->prev_sum_exec_runtime));
5346 	}
5347 
5348 	se->prev_sum_exec_runtime = se->sum_exec_runtime;
5349 }
5350 
5351 /*
5352  * Pick the next process, keeping these things in mind, in this order:
5353  * 1) keep things fair between processes/task groups
5354  * 2) pick the "next" process, since someone really wants that to run
5355  * 3) pick the "last" process, for cache locality
5356  * 4) do not run the "skip" process, if something else is available
5357  */
5358 static struct sched_entity *
pick_next_entity(struct cfs_rq * cfs_rq,struct sched_entity * curr)5359 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
5360 {
5361 	/*
5362 	 * Enabling NEXT_BUDDY will affect latency but not fairness.
5363 	 */
5364 	if (sched_feat(NEXT_BUDDY) &&
5365 	    cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next))
5366 		return cfs_rq->next;
5367 
5368 	return pick_eevdf(cfs_rq);
5369 }
5370 
5371 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5372 
put_prev_entity(struct cfs_rq * cfs_rq,struct sched_entity * prev)5373 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5374 {
5375 	/*
5376 	 * If still on the runqueue then deactivate_task()
5377 	 * was not called and update_curr() has to be done:
5378 	 */
5379 	if (prev->on_rq)
5380 		update_curr(cfs_rq);
5381 
5382 	/* throttle cfs_rqs exceeding runtime */
5383 	check_cfs_rq_runtime(cfs_rq);
5384 
5385 	if (prev->on_rq) {
5386 		update_stats_wait_start_fair(cfs_rq, prev);
5387 		/* Put 'current' back into the tree. */
5388 		__enqueue_entity(cfs_rq, prev);
5389 		/* in !on_rq case, update occurred at dequeue */
5390 		update_load_avg(cfs_rq, prev, 0);
5391 	}
5392 	cfs_rq->curr = NULL;
5393 }
5394 
5395 static void
entity_tick(struct cfs_rq * cfs_rq,struct sched_entity * curr,int queued)5396 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5397 {
5398 	/*
5399 	 * Update run-time statistics of the 'current'.
5400 	 */
5401 	update_curr(cfs_rq);
5402 
5403 	/*
5404 	 * Ensure that runnable average is periodically updated.
5405 	 */
5406 	update_load_avg(cfs_rq, curr, UPDATE_TG);
5407 	update_cfs_group(curr);
5408 
5409 #ifdef CONFIG_SCHED_HRTICK
5410 	/*
5411 	 * queued ticks are scheduled to match the slice, so don't bother
5412 	 * validating it and just reschedule.
5413 	 */
5414 	if (queued) {
5415 		resched_curr(rq_of(cfs_rq));
5416 		return;
5417 	}
5418 	/*
5419 	 * don't let the period tick interfere with the hrtick preemption
5420 	 */
5421 	if (!sched_feat(DOUBLE_TICK) &&
5422 			hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5423 		return;
5424 #endif
5425 }
5426 
5427 
5428 /**************************************************
5429  * CFS bandwidth control machinery
5430  */
5431 
5432 #ifdef CONFIG_CFS_BANDWIDTH
5433 
5434 #ifdef CONFIG_JUMP_LABEL
5435 static struct static_key __cfs_bandwidth_used;
5436 
cfs_bandwidth_used(void)5437 static inline bool cfs_bandwidth_used(void)
5438 {
5439 	return static_key_false(&__cfs_bandwidth_used);
5440 }
5441 
cfs_bandwidth_usage_inc(void)5442 void cfs_bandwidth_usage_inc(void)
5443 {
5444 	static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5445 }
5446 
cfs_bandwidth_usage_dec(void)5447 void cfs_bandwidth_usage_dec(void)
5448 {
5449 	static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5450 }
5451 #else /* CONFIG_JUMP_LABEL */
cfs_bandwidth_used(void)5452 static bool cfs_bandwidth_used(void)
5453 {
5454 	return true;
5455 }
5456 
cfs_bandwidth_usage_inc(void)5457 void cfs_bandwidth_usage_inc(void) {}
cfs_bandwidth_usage_dec(void)5458 void cfs_bandwidth_usage_dec(void) {}
5459 #endif /* CONFIG_JUMP_LABEL */
5460 
5461 /*
5462  * default period for cfs group bandwidth.
5463  * default: 0.1s, units: nanoseconds
5464  */
default_cfs_period(void)5465 static inline u64 default_cfs_period(void)
5466 {
5467 	return 100000000ULL;
5468 }
5469 
sched_cfs_bandwidth_slice(void)5470 static inline u64 sched_cfs_bandwidth_slice(void)
5471 {
5472 	return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5473 }
5474 
5475 /*
5476  * Replenish runtime according to assigned quota. We use sched_clock_cpu
5477  * directly instead of rq->clock to avoid adding additional synchronization
5478  * around rq->lock.
5479  *
5480  * requires cfs_b->lock
5481  */
__refill_cfs_bandwidth_runtime(struct cfs_bandwidth * cfs_b)5482 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5483 {
5484 	s64 runtime;
5485 
5486 	if (unlikely(cfs_b->quota == RUNTIME_INF))
5487 		return;
5488 
5489 	cfs_b->runtime += cfs_b->quota;
5490 	runtime = cfs_b->runtime_snap - cfs_b->runtime;
5491 	if (runtime > 0) {
5492 		cfs_b->burst_time += runtime;
5493 		cfs_b->nr_burst++;
5494 	}
5495 
5496 	cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5497 	cfs_b->runtime_snap = cfs_b->runtime;
5498 }
5499 
tg_cfs_bandwidth(struct task_group * tg)5500 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5501 {
5502 	return &tg->cfs_bandwidth;
5503 }
5504 
5505 /* returns 0 on failure to allocate runtime */
__assign_cfs_rq_runtime(struct cfs_bandwidth * cfs_b,struct cfs_rq * cfs_rq,u64 target_runtime)5506 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5507 				   struct cfs_rq *cfs_rq, u64 target_runtime)
5508 {
5509 	u64 min_amount, amount = 0;
5510 
5511 	lockdep_assert_held(&cfs_b->lock);
5512 
5513 	/* note: this is a positive sum as runtime_remaining <= 0 */
5514 	min_amount = target_runtime - cfs_rq->runtime_remaining;
5515 
5516 	if (cfs_b->quota == RUNTIME_INF)
5517 		amount = min_amount;
5518 	else {
5519 		start_cfs_bandwidth(cfs_b);
5520 
5521 		if (cfs_b->runtime > 0) {
5522 			amount = min(cfs_b->runtime, min_amount);
5523 			cfs_b->runtime -= amount;
5524 			cfs_b->idle = 0;
5525 		}
5526 	}
5527 
5528 	cfs_rq->runtime_remaining += amount;
5529 
5530 	return cfs_rq->runtime_remaining > 0;
5531 }
5532 
5533 /* returns 0 on failure to allocate runtime */
assign_cfs_rq_runtime(struct cfs_rq * cfs_rq)5534 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5535 {
5536 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5537 	int ret;
5538 
5539 	raw_spin_lock(&cfs_b->lock);
5540 	ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5541 	raw_spin_unlock(&cfs_b->lock);
5542 
5543 	return ret;
5544 }
5545 
__account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)5546 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5547 {
5548 	/* dock delta_exec before expiring quota (as it could span periods) */
5549 	cfs_rq->runtime_remaining -= delta_exec;
5550 
5551 	if (likely(cfs_rq->runtime_remaining > 0))
5552 		return;
5553 
5554 	if (cfs_rq->throttled)
5555 		return;
5556 	/*
5557 	 * if we're unable to extend our runtime we resched so that the active
5558 	 * hierarchy can be throttled
5559 	 */
5560 	if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5561 		resched_curr(rq_of(cfs_rq));
5562 }
5563 
5564 static __always_inline
account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)5565 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5566 {
5567 	if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5568 		return;
5569 
5570 	__account_cfs_rq_runtime(cfs_rq, delta_exec);
5571 }
5572 
cfs_rq_throttled(struct cfs_rq * cfs_rq)5573 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5574 {
5575 	return cfs_bandwidth_used() && cfs_rq->throttled;
5576 }
5577 
5578 /* check whether cfs_rq, or any parent, is throttled */
throttled_hierarchy(struct cfs_rq * cfs_rq)5579 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5580 {
5581 	return cfs_bandwidth_used() && cfs_rq->throttle_count;
5582 }
5583 
5584 /*
5585  * Ensure that neither of the group entities corresponding to src_cpu or
5586  * dest_cpu are members of a throttled hierarchy when performing group
5587  * load-balance operations.
5588  */
throttled_lb_pair(struct task_group * tg,int src_cpu,int dest_cpu)5589 static inline int throttled_lb_pair(struct task_group *tg,
5590 				    int src_cpu, int dest_cpu)
5591 {
5592 	struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5593 
5594 	src_cfs_rq = tg->cfs_rq[src_cpu];
5595 	dest_cfs_rq = tg->cfs_rq[dest_cpu];
5596 
5597 	return throttled_hierarchy(src_cfs_rq) ||
5598 	       throttled_hierarchy(dest_cfs_rq);
5599 }
5600 
tg_unthrottle_up(struct task_group * tg,void * data)5601 static int tg_unthrottle_up(struct task_group *tg, void *data)
5602 {
5603 	struct rq *rq = data;
5604 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5605 
5606 	cfs_rq->throttle_count--;
5607 	if (!cfs_rq->throttle_count) {
5608 		cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5609 					     cfs_rq->throttled_clock_pelt;
5610 
5611 		/* Add cfs_rq with load or one or more already running entities to the list */
5612 		if (!cfs_rq_is_decayed(cfs_rq))
5613 			list_add_leaf_cfs_rq(cfs_rq);
5614 
5615 		if (cfs_rq->throttled_clock_self) {
5616 			u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;
5617 
5618 			cfs_rq->throttled_clock_self = 0;
5619 
5620 			if (SCHED_WARN_ON((s64)delta < 0))
5621 				delta = 0;
5622 
5623 			cfs_rq->throttled_clock_self_time += delta;
5624 		}
5625 	}
5626 
5627 	return 0;
5628 }
5629 
tg_throttle_down(struct task_group * tg,void * data)5630 static int tg_throttle_down(struct task_group *tg, void *data)
5631 {
5632 	struct rq *rq = data;
5633 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5634 
5635 	/* group is entering throttled state, stop time */
5636 	if (!cfs_rq->throttle_count) {
5637 		cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5638 		list_del_leaf_cfs_rq(cfs_rq);
5639 
5640 		SCHED_WARN_ON(cfs_rq->throttled_clock_self);
5641 		if (cfs_rq->nr_running)
5642 			cfs_rq->throttled_clock_self = rq_clock(rq);
5643 	}
5644 	cfs_rq->throttle_count++;
5645 
5646 	return 0;
5647 }
5648 
throttle_cfs_rq(struct cfs_rq * cfs_rq)5649 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5650 {
5651 	struct rq *rq = rq_of(cfs_rq);
5652 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5653 	struct sched_entity *se;
5654 	long task_delta, idle_task_delta, dequeue = 1;
5655 
5656 	raw_spin_lock(&cfs_b->lock);
5657 	/* This will start the period timer if necessary */
5658 	if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5659 		/*
5660 		 * We have raced with bandwidth becoming available, and if we
5661 		 * actually throttled the timer might not unthrottle us for an
5662 		 * entire period. We additionally needed to make sure that any
5663 		 * subsequent check_cfs_rq_runtime calls agree not to throttle
5664 		 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5665 		 * for 1ns of runtime rather than just check cfs_b.
5666 		 */
5667 		dequeue = 0;
5668 	} else {
5669 		list_add_tail_rcu(&cfs_rq->throttled_list,
5670 				  &cfs_b->throttled_cfs_rq);
5671 	}
5672 	raw_spin_unlock(&cfs_b->lock);
5673 
5674 	if (!dequeue)
5675 		return false;  /* Throttle no longer required. */
5676 
5677 	se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5678 
5679 	/* freeze hierarchy runnable averages while throttled */
5680 	rcu_read_lock();
5681 	walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5682 	rcu_read_unlock();
5683 
5684 	task_delta = cfs_rq->h_nr_running;
5685 	idle_task_delta = cfs_rq->idle_h_nr_running;
5686 	for_each_sched_entity(se) {
5687 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5688 		/* throttled entity or throttle-on-deactivate */
5689 		if (!se->on_rq)
5690 			goto done;
5691 
5692 		dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5693 
5694 		if (cfs_rq_is_idle(group_cfs_rq(se)))
5695 			idle_task_delta = cfs_rq->h_nr_running;
5696 
5697 		qcfs_rq->h_nr_running -= task_delta;
5698 		qcfs_rq->idle_h_nr_running -= idle_task_delta;
5699 
5700 		if (qcfs_rq->load.weight) {
5701 			/* Avoid re-evaluating load for this entity: */
5702 			se = parent_entity(se);
5703 			break;
5704 		}
5705 	}
5706 
5707 	for_each_sched_entity(se) {
5708 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5709 		/* throttled entity or throttle-on-deactivate */
5710 		if (!se->on_rq)
5711 			goto done;
5712 
5713 		update_load_avg(qcfs_rq, se, 0);
5714 		se_update_runnable(se);
5715 
5716 		if (cfs_rq_is_idle(group_cfs_rq(se)))
5717 			idle_task_delta = cfs_rq->h_nr_running;
5718 
5719 		qcfs_rq->h_nr_running -= task_delta;
5720 		qcfs_rq->idle_h_nr_running -= idle_task_delta;
5721 	}
5722 
5723 	/* At this point se is NULL and we are at root level*/
5724 	sub_nr_running(rq, task_delta);
5725 
5726 done:
5727 	/*
5728 	 * Note: distribution will already see us throttled via the
5729 	 * throttled-list.  rq->lock protects completion.
5730 	 */
5731 	cfs_rq->throttled = 1;
5732 	SCHED_WARN_ON(cfs_rq->throttled_clock);
5733 	if (cfs_rq->nr_running)
5734 		cfs_rq->throttled_clock = rq_clock(rq);
5735 	return true;
5736 }
5737 
unthrottle_cfs_rq(struct cfs_rq * cfs_rq)5738 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5739 {
5740 	struct rq *rq = rq_of(cfs_rq);
5741 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5742 	struct sched_entity *se;
5743 	long task_delta, idle_task_delta;
5744 
5745 	se = cfs_rq->tg->se[cpu_of(rq)];
5746 
5747 	cfs_rq->throttled = 0;
5748 
5749 	update_rq_clock(rq);
5750 
5751 	raw_spin_lock(&cfs_b->lock);
5752 	if (cfs_rq->throttled_clock) {
5753 		cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5754 		cfs_rq->throttled_clock = 0;
5755 	}
5756 	list_del_rcu(&cfs_rq->throttled_list);
5757 	raw_spin_unlock(&cfs_b->lock);
5758 
5759 	/* update hierarchical throttle state */
5760 	walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5761 
5762 	if (!cfs_rq->load.weight) {
5763 		if (!cfs_rq->on_list)
5764 			return;
5765 		/*
5766 		 * Nothing to run but something to decay (on_list)?
5767 		 * Complete the branch.
5768 		 */
5769 		for_each_sched_entity(se) {
5770 			if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5771 				break;
5772 		}
5773 		goto unthrottle_throttle;
5774 	}
5775 
5776 	task_delta = cfs_rq->h_nr_running;
5777 	idle_task_delta = cfs_rq->idle_h_nr_running;
5778 	for_each_sched_entity(se) {
5779 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5780 
5781 		if (se->on_rq)
5782 			break;
5783 		enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5784 
5785 		if (cfs_rq_is_idle(group_cfs_rq(se)))
5786 			idle_task_delta = cfs_rq->h_nr_running;
5787 
5788 		qcfs_rq->h_nr_running += task_delta;
5789 		qcfs_rq->idle_h_nr_running += idle_task_delta;
5790 
5791 		/* end evaluation on encountering a throttled cfs_rq */
5792 		if (cfs_rq_throttled(qcfs_rq))
5793 			goto unthrottle_throttle;
5794 	}
5795 
5796 	for_each_sched_entity(se) {
5797 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5798 
5799 		update_load_avg(qcfs_rq, se, UPDATE_TG);
5800 		se_update_runnable(se);
5801 
5802 		if (cfs_rq_is_idle(group_cfs_rq(se)))
5803 			idle_task_delta = cfs_rq->h_nr_running;
5804 
5805 		qcfs_rq->h_nr_running += task_delta;
5806 		qcfs_rq->idle_h_nr_running += idle_task_delta;
5807 
5808 		/* end evaluation on encountering a throttled cfs_rq */
5809 		if (cfs_rq_throttled(qcfs_rq))
5810 			goto unthrottle_throttle;
5811 	}
5812 
5813 	/* At this point se is NULL and we are at root level*/
5814 	add_nr_running(rq, task_delta);
5815 
5816 unthrottle_throttle:
5817 	assert_list_leaf_cfs_rq(rq);
5818 
5819 	/* Determine whether we need to wake up potentially idle CPU: */
5820 	if (rq->curr == rq->idle && rq->cfs.nr_running)
5821 		resched_curr(rq);
5822 }
5823 
5824 #ifdef CONFIG_SMP
__cfsb_csd_unthrottle(void * arg)5825 static void __cfsb_csd_unthrottle(void *arg)
5826 {
5827 	struct cfs_rq *cursor, *tmp;
5828 	struct rq *rq = arg;
5829 	struct rq_flags rf;
5830 
5831 	rq_lock(rq, &rf);
5832 
5833 	/*
5834 	 * Iterating over the list can trigger several call to
5835 	 * update_rq_clock() in unthrottle_cfs_rq().
5836 	 * Do it once and skip the potential next ones.
5837 	 */
5838 	update_rq_clock(rq);
5839 	rq_clock_start_loop_update(rq);
5840 
5841 	/*
5842 	 * Since we hold rq lock we're safe from concurrent manipulation of
5843 	 * the CSD list. However, this RCU critical section annotates the
5844 	 * fact that we pair with sched_free_group_rcu(), so that we cannot
5845 	 * race with group being freed in the window between removing it
5846 	 * from the list and advancing to the next entry in the list.
5847 	 */
5848 	rcu_read_lock();
5849 
5850 	list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
5851 				 throttled_csd_list) {
5852 		list_del_init(&cursor->throttled_csd_list);
5853 
5854 		if (cfs_rq_throttled(cursor))
5855 			unthrottle_cfs_rq(cursor);
5856 	}
5857 
5858 	rcu_read_unlock();
5859 
5860 	rq_clock_stop_loop_update(rq);
5861 	rq_unlock(rq, &rf);
5862 }
5863 
__unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)5864 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5865 {
5866 	struct rq *rq = rq_of(cfs_rq);
5867 	bool first;
5868 
5869 	if (rq == this_rq()) {
5870 		unthrottle_cfs_rq(cfs_rq);
5871 		return;
5872 	}
5873 
5874 	/* Already enqueued */
5875 	if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))
5876 		return;
5877 
5878 	first = list_empty(&rq->cfsb_csd_list);
5879 	list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
5880 	if (first)
5881 		smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
5882 }
5883 #else
__unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)5884 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5885 {
5886 	unthrottle_cfs_rq(cfs_rq);
5887 }
5888 #endif
5889 
unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)5890 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5891 {
5892 	lockdep_assert_rq_held(rq_of(cfs_rq));
5893 
5894 	if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) ||
5895 	    cfs_rq->runtime_remaining <= 0))
5896 		return;
5897 
5898 	__unthrottle_cfs_rq_async(cfs_rq);
5899 }
5900 
distribute_cfs_runtime(struct cfs_bandwidth * cfs_b)5901 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5902 {
5903 	struct cfs_rq *local_unthrottle = NULL;
5904 	int this_cpu = smp_processor_id();
5905 	u64 runtime, remaining = 1;
5906 	bool throttled = false;
5907 	struct cfs_rq *cfs_rq;
5908 	struct rq_flags rf;
5909 	struct rq *rq;
5910 
5911 	rcu_read_lock();
5912 	list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5913 				throttled_list) {
5914 		rq = rq_of(cfs_rq);
5915 
5916 		if (!remaining) {
5917 			throttled = true;
5918 			break;
5919 		}
5920 
5921 		rq_lock_irqsave(rq, &rf);
5922 		if (!cfs_rq_throttled(cfs_rq))
5923 			goto next;
5924 
5925 #ifdef CONFIG_SMP
5926 		/* Already queued for async unthrottle */
5927 		if (!list_empty(&cfs_rq->throttled_csd_list))
5928 			goto next;
5929 #endif
5930 
5931 		/* By the above checks, this should never be true */
5932 		SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5933 
5934 		raw_spin_lock(&cfs_b->lock);
5935 		runtime = -cfs_rq->runtime_remaining + 1;
5936 		if (runtime > cfs_b->runtime)
5937 			runtime = cfs_b->runtime;
5938 		cfs_b->runtime -= runtime;
5939 		remaining = cfs_b->runtime;
5940 		raw_spin_unlock(&cfs_b->lock);
5941 
5942 		cfs_rq->runtime_remaining += runtime;
5943 
5944 		/* we check whether we're throttled above */
5945 		if (cfs_rq->runtime_remaining > 0) {
5946 			if (cpu_of(rq) != this_cpu ||
5947 			    SCHED_WARN_ON(local_unthrottle))
5948 				unthrottle_cfs_rq_async(cfs_rq);
5949 			else
5950 				local_unthrottle = cfs_rq;
5951 		} else {
5952 			throttled = true;
5953 		}
5954 
5955 next:
5956 		rq_unlock_irqrestore(rq, &rf);
5957 	}
5958 	rcu_read_unlock();
5959 
5960 	if (local_unthrottle) {
5961 		rq = cpu_rq(this_cpu);
5962 		rq_lock_irqsave(rq, &rf);
5963 		if (cfs_rq_throttled(local_unthrottle))
5964 			unthrottle_cfs_rq(local_unthrottle);
5965 		rq_unlock_irqrestore(rq, &rf);
5966 	}
5967 
5968 	return throttled;
5969 }
5970 
5971 /*
5972  * Responsible for refilling a task_group's bandwidth and unthrottling its
5973  * cfs_rqs as appropriate. If there has been no activity within the last
5974  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5975  * used to track this state.
5976  */
do_sched_cfs_period_timer(struct cfs_bandwidth * cfs_b,int overrun,unsigned long flags)5977 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5978 {
5979 	int throttled;
5980 
5981 	/* no need to continue the timer with no bandwidth constraint */
5982 	if (cfs_b->quota == RUNTIME_INF)
5983 		goto out_deactivate;
5984 
5985 	throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5986 	cfs_b->nr_periods += overrun;
5987 
5988 	/* Refill extra burst quota even if cfs_b->idle */
5989 	__refill_cfs_bandwidth_runtime(cfs_b);
5990 
5991 	/*
5992 	 * idle depends on !throttled (for the case of a large deficit), and if
5993 	 * we're going inactive then everything else can be deferred
5994 	 */
5995 	if (cfs_b->idle && !throttled)
5996 		goto out_deactivate;
5997 
5998 	if (!throttled) {
5999 		/* mark as potentially idle for the upcoming period */
6000 		cfs_b->idle = 1;
6001 		return 0;
6002 	}
6003 
6004 	/* account preceding periods in which throttling occurred */
6005 	cfs_b->nr_throttled += overrun;
6006 
6007 	/*
6008 	 * This check is repeated as we release cfs_b->lock while we unthrottle.
6009 	 */
6010 	while (throttled && cfs_b->runtime > 0) {
6011 		raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6012 		/* we can't nest cfs_b->lock while distributing bandwidth */
6013 		throttled = distribute_cfs_runtime(cfs_b);
6014 		raw_spin_lock_irqsave(&cfs_b->lock, flags);
6015 	}
6016 
6017 	/*
6018 	 * While we are ensured activity in the period following an
6019 	 * unthrottle, this also covers the case in which the new bandwidth is
6020 	 * insufficient to cover the existing bandwidth deficit.  (Forcing the
6021 	 * timer to remain active while there are any throttled entities.)
6022 	 */
6023 	cfs_b->idle = 0;
6024 
6025 	return 0;
6026 
6027 out_deactivate:
6028 	return 1;
6029 }
6030 
6031 /* a cfs_rq won't donate quota below this amount */
6032 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
6033 /* minimum remaining period time to redistribute slack quota */
6034 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
6035 /* how long we wait to gather additional slack before distributing */
6036 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
6037 
6038 /*
6039  * Are we near the end of the current quota period?
6040  *
6041  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
6042  * hrtimer base being cleared by hrtimer_start. In the case of
6043  * migrate_hrtimers, base is never cleared, so we are fine.
6044  */
runtime_refresh_within(struct cfs_bandwidth * cfs_b,u64 min_expire)6045 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
6046 {
6047 	struct hrtimer *refresh_timer = &cfs_b->period_timer;
6048 	s64 remaining;
6049 
6050 	/* if the call-back is running a quota refresh is already occurring */
6051 	if (hrtimer_callback_running(refresh_timer))
6052 		return 1;
6053 
6054 	/* is a quota refresh about to occur? */
6055 	remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
6056 	if (remaining < (s64)min_expire)
6057 		return 1;
6058 
6059 	return 0;
6060 }
6061 
start_cfs_slack_bandwidth(struct cfs_bandwidth * cfs_b)6062 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
6063 {
6064 	u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
6065 
6066 	/* if there's a quota refresh soon don't bother with slack */
6067 	if (runtime_refresh_within(cfs_b, min_left))
6068 		return;
6069 
6070 	/* don't push forwards an existing deferred unthrottle */
6071 	if (cfs_b->slack_started)
6072 		return;
6073 	cfs_b->slack_started = true;
6074 
6075 	hrtimer_start(&cfs_b->slack_timer,
6076 			ns_to_ktime(cfs_bandwidth_slack_period),
6077 			HRTIMER_MODE_REL);
6078 }
6079 
6080 /* we know any runtime found here is valid as update_curr() precedes return */
__return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6081 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6082 {
6083 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6084 	s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
6085 
6086 	if (slack_runtime <= 0)
6087 		return;
6088 
6089 	raw_spin_lock(&cfs_b->lock);
6090 	if (cfs_b->quota != RUNTIME_INF) {
6091 		cfs_b->runtime += slack_runtime;
6092 
6093 		/* we are under rq->lock, defer unthrottling using a timer */
6094 		if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
6095 		    !list_empty(&cfs_b->throttled_cfs_rq))
6096 			start_cfs_slack_bandwidth(cfs_b);
6097 	}
6098 	raw_spin_unlock(&cfs_b->lock);
6099 
6100 	/* even if it's not valid for return we don't want to try again */
6101 	cfs_rq->runtime_remaining -= slack_runtime;
6102 }
6103 
return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6104 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6105 {
6106 	if (!cfs_bandwidth_used())
6107 		return;
6108 
6109 	if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
6110 		return;
6111 
6112 	__return_cfs_rq_runtime(cfs_rq);
6113 }
6114 
6115 /*
6116  * This is done with a timer (instead of inline with bandwidth return) since
6117  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6118  */
do_sched_cfs_slack_timer(struct cfs_bandwidth * cfs_b)6119 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
6120 {
6121 	u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
6122 	unsigned long flags;
6123 
6124 	/* confirm we're still not at a refresh boundary */
6125 	raw_spin_lock_irqsave(&cfs_b->lock, flags);
6126 	cfs_b->slack_started = false;
6127 
6128 	if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
6129 		raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6130 		return;
6131 	}
6132 
6133 	if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
6134 		runtime = cfs_b->runtime;
6135 
6136 	raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6137 
6138 	if (!runtime)
6139 		return;
6140 
6141 	distribute_cfs_runtime(cfs_b);
6142 }
6143 
6144 /*
6145  * When a group wakes up we want to make sure that its quota is not already
6146  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
6147  * runtime as update_curr() throttling can not trigger until it's on-rq.
6148  */
check_enqueue_throttle(struct cfs_rq * cfs_rq)6149 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
6150 {
6151 	if (!cfs_bandwidth_used())
6152 		return;
6153 
6154 	/* an active group must be handled by the update_curr()->put() path */
6155 	if (!cfs_rq->runtime_enabled || cfs_rq->curr)
6156 		return;
6157 
6158 	/* ensure the group is not already throttled */
6159 	if (cfs_rq_throttled(cfs_rq))
6160 		return;
6161 
6162 	/* update runtime allocation */
6163 	account_cfs_rq_runtime(cfs_rq, 0);
6164 	if (cfs_rq->runtime_remaining <= 0)
6165 		throttle_cfs_rq(cfs_rq);
6166 }
6167 
sync_throttle(struct task_group * tg,int cpu)6168 static void sync_throttle(struct task_group *tg, int cpu)
6169 {
6170 	struct cfs_rq *pcfs_rq, *cfs_rq;
6171 
6172 	if (!cfs_bandwidth_used())
6173 		return;
6174 
6175 	if (!tg->parent)
6176 		return;
6177 
6178 	cfs_rq = tg->cfs_rq[cpu];
6179 	pcfs_rq = tg->parent->cfs_rq[cpu];
6180 
6181 	cfs_rq->throttle_count = pcfs_rq->throttle_count;
6182 	cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
6183 }
6184 
6185 /* conditionally throttle active cfs_rq's from put_prev_entity() */
check_cfs_rq_runtime(struct cfs_rq * cfs_rq)6186 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6187 {
6188 	if (!cfs_bandwidth_used())
6189 		return false;
6190 
6191 	if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
6192 		return false;
6193 
6194 	/*
6195 	 * it's possible for a throttled entity to be forced into a running
6196 	 * state (e.g. set_curr_task), in this case we're finished.
6197 	 */
6198 	if (cfs_rq_throttled(cfs_rq))
6199 		return true;
6200 
6201 	return throttle_cfs_rq(cfs_rq);
6202 }
6203 
sched_cfs_slack_timer(struct hrtimer * timer)6204 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
6205 {
6206 	struct cfs_bandwidth *cfs_b =
6207 		container_of(timer, struct cfs_bandwidth, slack_timer);
6208 
6209 	do_sched_cfs_slack_timer(cfs_b);
6210 
6211 	return HRTIMER_NORESTART;
6212 }
6213 
6214 extern const u64 max_cfs_quota_period;
6215 
sched_cfs_period_timer(struct hrtimer * timer)6216 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
6217 {
6218 	struct cfs_bandwidth *cfs_b =
6219 		container_of(timer, struct cfs_bandwidth, period_timer);
6220 	unsigned long flags;
6221 	int overrun;
6222 	int idle = 0;
6223 	int count = 0;
6224 
6225 	raw_spin_lock_irqsave(&cfs_b->lock, flags);
6226 	for (;;) {
6227 		overrun = hrtimer_forward_now(timer, cfs_b->period);
6228 		if (!overrun)
6229 			break;
6230 
6231 		idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
6232 
6233 		if (++count > 3) {
6234 			u64 new, old = ktime_to_ns(cfs_b->period);
6235 
6236 			/*
6237 			 * Grow period by a factor of 2 to avoid losing precision.
6238 			 * Precision loss in the quota/period ratio can cause __cfs_schedulable
6239 			 * to fail.
6240 			 */
6241 			new = old * 2;
6242 			if (new < max_cfs_quota_period) {
6243 				cfs_b->period = ns_to_ktime(new);
6244 				cfs_b->quota *= 2;
6245 				cfs_b->burst *= 2;
6246 
6247 				pr_warn_ratelimited(
6248 	"cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6249 					smp_processor_id(),
6250 					div_u64(new, NSEC_PER_USEC),
6251 					div_u64(cfs_b->quota, NSEC_PER_USEC));
6252 			} else {
6253 				pr_warn_ratelimited(
6254 	"cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6255 					smp_processor_id(),
6256 					div_u64(old, NSEC_PER_USEC),
6257 					div_u64(cfs_b->quota, NSEC_PER_USEC));
6258 			}
6259 
6260 			/* reset count so we don't come right back in here */
6261 			count = 0;
6262 		}
6263 	}
6264 	if (idle)
6265 		cfs_b->period_active = 0;
6266 	raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6267 
6268 	return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
6269 }
6270 
init_cfs_bandwidth(struct cfs_bandwidth * cfs_b,struct cfs_bandwidth * parent)6271 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent)
6272 {
6273 	raw_spin_lock_init(&cfs_b->lock);
6274 	cfs_b->runtime = 0;
6275 	cfs_b->quota = RUNTIME_INF;
6276 	cfs_b->period = ns_to_ktime(default_cfs_period());
6277 	cfs_b->burst = 0;
6278 	cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;
6279 
6280 	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
6281 	hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
6282 	cfs_b->period_timer.function = sched_cfs_period_timer;
6283 
6284 	/* Add a random offset so that timers interleave */
6285 	hrtimer_set_expires(&cfs_b->period_timer,
6286 			    get_random_u32_below(cfs_b->period));
6287 	hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6288 	cfs_b->slack_timer.function = sched_cfs_slack_timer;
6289 	cfs_b->slack_started = false;
6290 }
6291 
init_cfs_rq_runtime(struct cfs_rq * cfs_rq)6292 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6293 {
6294 	cfs_rq->runtime_enabled = 0;
6295 	INIT_LIST_HEAD(&cfs_rq->throttled_list);
6296 #ifdef CONFIG_SMP
6297 	INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
6298 #endif
6299 }
6300 
start_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6301 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6302 {
6303 	lockdep_assert_held(&cfs_b->lock);
6304 
6305 	if (cfs_b->period_active)
6306 		return;
6307 
6308 	cfs_b->period_active = 1;
6309 	hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
6310 	hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
6311 }
6312 
destroy_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6313 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6314 {
6315 	int __maybe_unused i;
6316 
6317 	/* init_cfs_bandwidth() was not called */
6318 	if (!cfs_b->throttled_cfs_rq.next)
6319 		return;
6320 
6321 	hrtimer_cancel(&cfs_b->period_timer);
6322 	hrtimer_cancel(&cfs_b->slack_timer);
6323 
6324 	/*
6325 	 * It is possible that we still have some cfs_rq's pending on a CSD
6326 	 * list, though this race is very rare. In order for this to occur, we
6327 	 * must have raced with the last task leaving the group while there
6328 	 * exist throttled cfs_rq(s), and the period_timer must have queued the
6329 	 * CSD item but the remote cpu has not yet processed it. To handle this,
6330 	 * we can simply flush all pending CSD work inline here. We're
6331 	 * guaranteed at this point that no additional cfs_rq of this group can
6332 	 * join a CSD list.
6333 	 */
6334 #ifdef CONFIG_SMP
6335 	for_each_possible_cpu(i) {
6336 		struct rq *rq = cpu_rq(i);
6337 		unsigned long flags;
6338 
6339 		if (list_empty(&rq->cfsb_csd_list))
6340 			continue;
6341 
6342 		local_irq_save(flags);
6343 		__cfsb_csd_unthrottle(rq);
6344 		local_irq_restore(flags);
6345 	}
6346 #endif
6347 }
6348 
6349 /*
6350  * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
6351  *
6352  * The race is harmless, since modifying bandwidth settings of unhooked group
6353  * bits doesn't do much.
6354  */
6355 
6356 /* cpu online callback */
update_runtime_enabled(struct rq * rq)6357 static void __maybe_unused update_runtime_enabled(struct rq *rq)
6358 {
6359 	struct task_group *tg;
6360 
6361 	lockdep_assert_rq_held(rq);
6362 
6363 	rcu_read_lock();
6364 	list_for_each_entry_rcu(tg, &task_groups, list) {
6365 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6366 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6367 
6368 		raw_spin_lock(&cfs_b->lock);
6369 		cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6370 		raw_spin_unlock(&cfs_b->lock);
6371 	}
6372 	rcu_read_unlock();
6373 }
6374 
6375 /* cpu offline callback */
unthrottle_offline_cfs_rqs(struct rq * rq)6376 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6377 {
6378 	struct task_group *tg;
6379 
6380 	lockdep_assert_rq_held(rq);
6381 
6382 	/*
6383 	 * The rq clock has already been updated in the
6384 	 * set_rq_offline(), so we should skip updating
6385 	 * the rq clock again in unthrottle_cfs_rq().
6386 	 */
6387 	rq_clock_start_loop_update(rq);
6388 
6389 	rcu_read_lock();
6390 	list_for_each_entry_rcu(tg, &task_groups, list) {
6391 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6392 
6393 		if (!cfs_rq->runtime_enabled)
6394 			continue;
6395 
6396 		/*
6397 		 * clock_task is not advancing so we just need to make sure
6398 		 * there's some valid quota amount
6399 		 */
6400 		cfs_rq->runtime_remaining = 1;
6401 		/*
6402 		 * Offline rq is schedulable till CPU is completely disabled
6403 		 * in take_cpu_down(), so we prevent new cfs throttling here.
6404 		 */
6405 		cfs_rq->runtime_enabled = 0;
6406 
6407 		if (cfs_rq_throttled(cfs_rq))
6408 			unthrottle_cfs_rq(cfs_rq);
6409 	}
6410 	rcu_read_unlock();
6411 
6412 	rq_clock_stop_loop_update(rq);
6413 }
6414 
cfs_task_bw_constrained(struct task_struct * p)6415 bool cfs_task_bw_constrained(struct task_struct *p)
6416 {
6417 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
6418 
6419 	if (!cfs_bandwidth_used())
6420 		return false;
6421 
6422 	if (cfs_rq->runtime_enabled ||
6423 	    tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)
6424 		return true;
6425 
6426 	return false;
6427 }
6428 
6429 #ifdef CONFIG_NO_HZ_FULL
6430 /* called from pick_next_task_fair() */
sched_fair_update_stop_tick(struct rq * rq,struct task_struct * p)6431 static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
6432 {
6433 	int cpu = cpu_of(rq);
6434 
6435 	if (!sched_feat(HZ_BW) || !cfs_bandwidth_used())
6436 		return;
6437 
6438 	if (!tick_nohz_full_cpu(cpu))
6439 		return;
6440 
6441 	if (rq->nr_running != 1)
6442 		return;
6443 
6444 	/*
6445 	 *  We know there is only one task runnable and we've just picked it. The
6446 	 *  normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
6447 	 *  be otherwise able to stop the tick. Just need to check if we are using
6448 	 *  bandwidth control.
6449 	 */
6450 	if (cfs_task_bw_constrained(p))
6451 		tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
6452 }
6453 #endif
6454 
6455 #else /* CONFIG_CFS_BANDWIDTH */
6456 
cfs_bandwidth_used(void)6457 static inline bool cfs_bandwidth_used(void)
6458 {
6459 	return false;
6460 }
6461 
account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)6462 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
check_cfs_rq_runtime(struct cfs_rq * cfs_rq)6463 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
check_enqueue_throttle(struct cfs_rq * cfs_rq)6464 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
sync_throttle(struct task_group * tg,int cpu)6465 static inline void sync_throttle(struct task_group *tg, int cpu) {}
return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6466 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6467 
cfs_rq_throttled(struct cfs_rq * cfs_rq)6468 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6469 {
6470 	return 0;
6471 }
6472 
throttled_hierarchy(struct cfs_rq * cfs_rq)6473 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6474 {
6475 	return 0;
6476 }
6477 
throttled_lb_pair(struct task_group * tg,int src_cpu,int dest_cpu)6478 static inline int throttled_lb_pair(struct task_group *tg,
6479 				    int src_cpu, int dest_cpu)
6480 {
6481 	return 0;
6482 }
6483 
6484 #ifdef CONFIG_FAIR_GROUP_SCHED
init_cfs_bandwidth(struct cfs_bandwidth * cfs_b,struct cfs_bandwidth * parent)6485 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {}
init_cfs_rq_runtime(struct cfs_rq * cfs_rq)6486 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6487 #endif
6488 
tg_cfs_bandwidth(struct task_group * tg)6489 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6490 {
6491 	return NULL;
6492 }
destroy_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6493 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
update_runtime_enabled(struct rq * rq)6494 static inline void update_runtime_enabled(struct rq *rq) {}
unthrottle_offline_cfs_rqs(struct rq * rq)6495 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6496 #ifdef CONFIG_CGROUP_SCHED
cfs_task_bw_constrained(struct task_struct * p)6497 bool cfs_task_bw_constrained(struct task_struct *p)
6498 {
6499 	return false;
6500 }
6501 #endif
6502 #endif /* CONFIG_CFS_BANDWIDTH */
6503 
6504 #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL)
sched_fair_update_stop_tick(struct rq * rq,struct task_struct * p)6505 static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {}
6506 #endif
6507 
6508 /**************************************************
6509  * CFS operations on tasks:
6510  */
6511 
6512 #ifdef CONFIG_SCHED_HRTICK
hrtick_start_fair(struct rq * rq,struct task_struct * p)6513 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6514 {
6515 	struct sched_entity *se = &p->se;
6516 
6517 	SCHED_WARN_ON(task_rq(p) != rq);
6518 
6519 	if (rq->cfs.h_nr_running > 1) {
6520 		u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6521 		u64 slice = se->slice;
6522 		s64 delta = slice - ran;
6523 
6524 		if (delta < 0) {
6525 			if (task_current(rq, p))
6526 				resched_curr(rq);
6527 			return;
6528 		}
6529 		hrtick_start(rq, delta);
6530 	}
6531 }
6532 
6533 /*
6534  * called from enqueue/dequeue and updates the hrtick when the
6535  * current task is from our class and nr_running is low enough
6536  * to matter.
6537  */
hrtick_update(struct rq * rq)6538 static void hrtick_update(struct rq *rq)
6539 {
6540 	struct task_struct *curr = rq->curr;
6541 
6542 	if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
6543 		return;
6544 
6545 	hrtick_start_fair(rq, curr);
6546 }
6547 #else /* !CONFIG_SCHED_HRTICK */
6548 static inline void
hrtick_start_fair(struct rq * rq,struct task_struct * p)6549 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6550 {
6551 }
6552 
hrtick_update(struct rq * rq)6553 static inline void hrtick_update(struct rq *rq)
6554 {
6555 }
6556 #endif
6557 
6558 #ifdef CONFIG_SMP
cpu_overutilized(int cpu)6559 static inline bool cpu_overutilized(int cpu)
6560 {
6561 	unsigned long rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6562 	unsigned long rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6563 
6564 	/* Return true only if the utilization doesn't fit CPU's capacity */
6565 	return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6566 }
6567 
update_overutilized_status(struct rq * rq)6568 static inline void update_overutilized_status(struct rq *rq)
6569 {
6570 	if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
6571 		WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
6572 		trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
6573 	}
6574 }
6575 #else
update_overutilized_status(struct rq * rq)6576 static inline void update_overutilized_status(struct rq *rq) { }
6577 #endif
6578 
6579 /* Runqueue only has SCHED_IDLE tasks enqueued */
sched_idle_rq(struct rq * rq)6580 static int sched_idle_rq(struct rq *rq)
6581 {
6582 	return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6583 			rq->nr_running);
6584 }
6585 
6586 #ifdef CONFIG_SMP
sched_idle_cpu(int cpu)6587 static int sched_idle_cpu(int cpu)
6588 {
6589 	return sched_idle_rq(cpu_rq(cpu));
6590 }
6591 #endif
6592 
6593 /*
6594  * The enqueue_task method is called before nr_running is
6595  * increased. Here we update the fair scheduling stats and
6596  * then put the task into the rbtree:
6597  */
6598 static void
enqueue_task_fair(struct rq * rq,struct task_struct * p,int flags)6599 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6600 {
6601 	struct cfs_rq *cfs_rq;
6602 	struct sched_entity *se = &p->se;
6603 	int idle_h_nr_running = task_has_idle_policy(p);
6604 	int task_new = !(flags & ENQUEUE_WAKEUP);
6605 
6606 	/*
6607 	 * The code below (indirectly) updates schedutil which looks at
6608 	 * the cfs_rq utilization to select a frequency.
6609 	 * Let's add the task's estimated utilization to the cfs_rq's
6610 	 * estimated utilization, before we update schedutil.
6611 	 */
6612 	util_est_enqueue(&rq->cfs, p);
6613 
6614 	/*
6615 	 * If in_iowait is set, the code below may not trigger any cpufreq
6616 	 * utilization updates, so do it here explicitly with the IOWAIT flag
6617 	 * passed.
6618 	 */
6619 	if (p->in_iowait)
6620 		cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6621 
6622 	for_each_sched_entity(se) {
6623 		if (se->on_rq)
6624 			break;
6625 		cfs_rq = cfs_rq_of(se);
6626 		enqueue_entity(cfs_rq, se, flags);
6627 
6628 		cfs_rq->h_nr_running++;
6629 		cfs_rq->idle_h_nr_running += idle_h_nr_running;
6630 
6631 		if (cfs_rq_is_idle(cfs_rq))
6632 			idle_h_nr_running = 1;
6633 
6634 		/* end evaluation on encountering a throttled cfs_rq */
6635 		if (cfs_rq_throttled(cfs_rq))
6636 			goto enqueue_throttle;
6637 
6638 		flags = ENQUEUE_WAKEUP;
6639 	}
6640 
6641 	for_each_sched_entity(se) {
6642 		cfs_rq = cfs_rq_of(se);
6643 
6644 		update_load_avg(cfs_rq, se, UPDATE_TG);
6645 		se_update_runnable(se);
6646 		update_cfs_group(se);
6647 
6648 		cfs_rq->h_nr_running++;
6649 		cfs_rq->idle_h_nr_running += idle_h_nr_running;
6650 
6651 		if (cfs_rq_is_idle(cfs_rq))
6652 			idle_h_nr_running = 1;
6653 
6654 		/* end evaluation on encountering a throttled cfs_rq */
6655 		if (cfs_rq_throttled(cfs_rq))
6656 			goto enqueue_throttle;
6657 	}
6658 
6659 	/* At this point se is NULL and we are at root level*/
6660 	add_nr_running(rq, 1);
6661 
6662 	/*
6663 	 * Since new tasks are assigned an initial util_avg equal to
6664 	 * half of the spare capacity of their CPU, tiny tasks have the
6665 	 * ability to cross the overutilized threshold, which will
6666 	 * result in the load balancer ruining all the task placement
6667 	 * done by EAS. As a way to mitigate that effect, do not account
6668 	 * for the first enqueue operation of new tasks during the
6669 	 * overutilized flag detection.
6670 	 *
6671 	 * A better way of solving this problem would be to wait for
6672 	 * the PELT signals of tasks to converge before taking them
6673 	 * into account, but that is not straightforward to implement,
6674 	 * and the following generally works well enough in practice.
6675 	 */
6676 	if (!task_new)
6677 		update_overutilized_status(rq);
6678 
6679 enqueue_throttle:
6680 	assert_list_leaf_cfs_rq(rq);
6681 
6682 	hrtick_update(rq);
6683 }
6684 
6685 static void set_next_buddy(struct sched_entity *se);
6686 
6687 /*
6688  * The dequeue_task method is called before nr_running is
6689  * decreased. We remove the task from the rbtree and
6690  * update the fair scheduling stats:
6691  */
dequeue_task_fair(struct rq * rq,struct task_struct * p,int flags)6692 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6693 {
6694 	struct cfs_rq *cfs_rq;
6695 	struct sched_entity *se = &p->se;
6696 	int task_sleep = flags & DEQUEUE_SLEEP;
6697 	int idle_h_nr_running = task_has_idle_policy(p);
6698 	bool was_sched_idle = sched_idle_rq(rq);
6699 
6700 	util_est_dequeue(&rq->cfs, p);
6701 
6702 	for_each_sched_entity(se) {
6703 		cfs_rq = cfs_rq_of(se);
6704 		dequeue_entity(cfs_rq, se, flags);
6705 
6706 		cfs_rq->h_nr_running--;
6707 		cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6708 
6709 		if (cfs_rq_is_idle(cfs_rq))
6710 			idle_h_nr_running = 1;
6711 
6712 		/* end evaluation on encountering a throttled cfs_rq */
6713 		if (cfs_rq_throttled(cfs_rq))
6714 			goto dequeue_throttle;
6715 
6716 		/* Don't dequeue parent if it has other entities besides us */
6717 		if (cfs_rq->load.weight) {
6718 			/* Avoid re-evaluating load for this entity: */
6719 			se = parent_entity(se);
6720 			/*
6721 			 * Bias pick_next to pick a task from this cfs_rq, as
6722 			 * p is sleeping when it is within its sched_slice.
6723 			 */
6724 			if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6725 				set_next_buddy(se);
6726 			break;
6727 		}
6728 		flags |= DEQUEUE_SLEEP;
6729 	}
6730 
6731 	for_each_sched_entity(se) {
6732 		cfs_rq = cfs_rq_of(se);
6733 
6734 		update_load_avg(cfs_rq, se, UPDATE_TG);
6735 		se_update_runnable(se);
6736 		update_cfs_group(se);
6737 
6738 		cfs_rq->h_nr_running--;
6739 		cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6740 
6741 		if (cfs_rq_is_idle(cfs_rq))
6742 			idle_h_nr_running = 1;
6743 
6744 		/* end evaluation on encountering a throttled cfs_rq */
6745 		if (cfs_rq_throttled(cfs_rq))
6746 			goto dequeue_throttle;
6747 
6748 	}
6749 
6750 	/* At this point se is NULL and we are at root level*/
6751 	sub_nr_running(rq, 1);
6752 
6753 	/* balance early to pull high priority tasks */
6754 	if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6755 		rq->next_balance = jiffies;
6756 
6757 dequeue_throttle:
6758 	util_est_update(&rq->cfs, p, task_sleep);
6759 	hrtick_update(rq);
6760 }
6761 
6762 #ifdef CONFIG_SMP
6763 
6764 /* Working cpumask for: load_balance, load_balance_newidle. */
6765 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6766 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6767 static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask);
6768 
6769 #ifdef CONFIG_NO_HZ_COMMON
6770 
6771 static struct {
6772 	cpumask_var_t idle_cpus_mask;
6773 	atomic_t nr_cpus;
6774 	int has_blocked;		/* Idle CPUS has blocked load */
6775 	int needs_update;		/* Newly idle CPUs need their next_balance collated */
6776 	unsigned long next_balance;     /* in jiffy units */
6777 	unsigned long next_blocked;	/* Next update of blocked load in jiffies */
6778 } nohz ____cacheline_aligned;
6779 
6780 #endif /* CONFIG_NO_HZ_COMMON */
6781 
cpu_load(struct rq * rq)6782 static unsigned long cpu_load(struct rq *rq)
6783 {
6784 	return cfs_rq_load_avg(&rq->cfs);
6785 }
6786 
6787 /*
6788  * cpu_load_without - compute CPU load without any contributions from *p
6789  * @cpu: the CPU which load is requested
6790  * @p: the task which load should be discounted
6791  *
6792  * The load of a CPU is defined by the load of tasks currently enqueued on that
6793  * CPU as well as tasks which are currently sleeping after an execution on that
6794  * CPU.
6795  *
6796  * This method returns the load of the specified CPU by discounting the load of
6797  * the specified task, whenever the task is currently contributing to the CPU
6798  * load.
6799  */
cpu_load_without(struct rq * rq,struct task_struct * p)6800 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6801 {
6802 	struct cfs_rq *cfs_rq;
6803 	unsigned int load;
6804 
6805 	/* Task has no contribution or is new */
6806 	if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6807 		return cpu_load(rq);
6808 
6809 	cfs_rq = &rq->cfs;
6810 	load = READ_ONCE(cfs_rq->avg.load_avg);
6811 
6812 	/* Discount task's util from CPU's util */
6813 	lsub_positive(&load, task_h_load(p));
6814 
6815 	return load;
6816 }
6817 
cpu_runnable(struct rq * rq)6818 static unsigned long cpu_runnable(struct rq *rq)
6819 {
6820 	return cfs_rq_runnable_avg(&rq->cfs);
6821 }
6822 
cpu_runnable_without(struct rq * rq,struct task_struct * p)6823 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6824 {
6825 	struct cfs_rq *cfs_rq;
6826 	unsigned int runnable;
6827 
6828 	/* Task has no contribution or is new */
6829 	if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6830 		return cpu_runnable(rq);
6831 
6832 	cfs_rq = &rq->cfs;
6833 	runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6834 
6835 	/* Discount task's runnable from CPU's runnable */
6836 	lsub_positive(&runnable, p->se.avg.runnable_avg);
6837 
6838 	return runnable;
6839 }
6840 
capacity_of(int cpu)6841 static unsigned long capacity_of(int cpu)
6842 {
6843 	return cpu_rq(cpu)->cpu_capacity;
6844 }
6845 
record_wakee(struct task_struct * p)6846 static void record_wakee(struct task_struct *p)
6847 {
6848 	/*
6849 	 * Only decay a single time; tasks that have less then 1 wakeup per
6850 	 * jiffy will not have built up many flips.
6851 	 */
6852 	if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6853 		current->wakee_flips >>= 1;
6854 		current->wakee_flip_decay_ts = jiffies;
6855 	}
6856 
6857 	if (current->last_wakee != p) {
6858 		current->last_wakee = p;
6859 		current->wakee_flips++;
6860 	}
6861 }
6862 
6863 /*
6864  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6865  *
6866  * A waker of many should wake a different task than the one last awakened
6867  * at a frequency roughly N times higher than one of its wakees.
6868  *
6869  * In order to determine whether we should let the load spread vs consolidating
6870  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6871  * partner, and a factor of lls_size higher frequency in the other.
6872  *
6873  * With both conditions met, we can be relatively sure that the relationship is
6874  * non-monogamous, with partner count exceeding socket size.
6875  *
6876  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6877  * whatever is irrelevant, spread criteria is apparent partner count exceeds
6878  * socket size.
6879  */
wake_wide(struct task_struct * p)6880 static int wake_wide(struct task_struct *p)
6881 {
6882 	unsigned int master = current->wakee_flips;
6883 	unsigned int slave = p->wakee_flips;
6884 	int factor = __this_cpu_read(sd_llc_size);
6885 
6886 	if (master < slave)
6887 		swap(master, slave);
6888 	if (slave < factor || master < slave * factor)
6889 		return 0;
6890 	return 1;
6891 }
6892 
6893 /*
6894  * The purpose of wake_affine() is to quickly determine on which CPU we can run
6895  * soonest. For the purpose of speed we only consider the waking and previous
6896  * CPU.
6897  *
6898  * wake_affine_idle() - only considers 'now', it check if the waking CPU is
6899  *			cache-affine and is (or	will be) idle.
6900  *
6901  * wake_affine_weight() - considers the weight to reflect the average
6902  *			  scheduling latency of the CPUs. This seems to work
6903  *			  for the overloaded case.
6904  */
6905 static int
wake_affine_idle(int this_cpu,int prev_cpu,int sync)6906 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
6907 {
6908 	/*
6909 	 * If this_cpu is idle, it implies the wakeup is from interrupt
6910 	 * context. Only allow the move if cache is shared. Otherwise an
6911 	 * interrupt intensive workload could force all tasks onto one
6912 	 * node depending on the IO topology or IRQ affinity settings.
6913 	 *
6914 	 * If the prev_cpu is idle and cache affine then avoid a migration.
6915 	 * There is no guarantee that the cache hot data from an interrupt
6916 	 * is more important than cache hot data on the prev_cpu and from
6917 	 * a cpufreq perspective, it's better to have higher utilisation
6918 	 * on one CPU.
6919 	 */
6920 	if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
6921 		return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
6922 
6923 	if (sync && cpu_rq(this_cpu)->nr_running == 1)
6924 		return this_cpu;
6925 
6926 	if (available_idle_cpu(prev_cpu))
6927 		return prev_cpu;
6928 
6929 	return nr_cpumask_bits;
6930 }
6931 
6932 static int
wake_affine_weight(struct sched_domain * sd,struct task_struct * p,int this_cpu,int prev_cpu,int sync)6933 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
6934 		   int this_cpu, int prev_cpu, int sync)
6935 {
6936 	s64 this_eff_load, prev_eff_load;
6937 	unsigned long task_load;
6938 
6939 	this_eff_load = cpu_load(cpu_rq(this_cpu));
6940 
6941 	if (sync) {
6942 		unsigned long current_load = task_h_load(current);
6943 
6944 		if (current_load > this_eff_load)
6945 			return this_cpu;
6946 
6947 		this_eff_load -= current_load;
6948 	}
6949 
6950 	task_load = task_h_load(p);
6951 
6952 	this_eff_load += task_load;
6953 	if (sched_feat(WA_BIAS))
6954 		this_eff_load *= 100;
6955 	this_eff_load *= capacity_of(prev_cpu);
6956 
6957 	prev_eff_load = cpu_load(cpu_rq(prev_cpu));
6958 	prev_eff_load -= task_load;
6959 	if (sched_feat(WA_BIAS))
6960 		prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
6961 	prev_eff_load *= capacity_of(this_cpu);
6962 
6963 	/*
6964 	 * If sync, adjust the weight of prev_eff_load such that if
6965 	 * prev_eff == this_eff that select_idle_sibling() will consider
6966 	 * stacking the wakee on top of the waker if no other CPU is
6967 	 * idle.
6968 	 */
6969 	if (sync)
6970 		prev_eff_load += 1;
6971 
6972 	return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
6973 }
6974 
wake_affine(struct sched_domain * sd,struct task_struct * p,int this_cpu,int prev_cpu,int sync)6975 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
6976 		       int this_cpu, int prev_cpu, int sync)
6977 {
6978 	int target = nr_cpumask_bits;
6979 
6980 	if (sched_feat(WA_IDLE))
6981 		target = wake_affine_idle(this_cpu, prev_cpu, sync);
6982 
6983 	if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
6984 		target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
6985 
6986 	schedstat_inc(p->stats.nr_wakeups_affine_attempts);
6987 	if (target != this_cpu)
6988 		return prev_cpu;
6989 
6990 	schedstat_inc(sd->ttwu_move_affine);
6991 	schedstat_inc(p->stats.nr_wakeups_affine);
6992 	return target;
6993 }
6994 
6995 static struct sched_group *
6996 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
6997 
6998 /*
6999  * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
7000  */
7001 static int
find_idlest_group_cpu(struct sched_group * group,struct task_struct * p,int this_cpu)7002 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
7003 {
7004 	unsigned long load, min_load = ULONG_MAX;
7005 	unsigned int min_exit_latency = UINT_MAX;
7006 	u64 latest_idle_timestamp = 0;
7007 	int least_loaded_cpu = this_cpu;
7008 	int shallowest_idle_cpu = -1;
7009 	int i;
7010 
7011 	/* Check if we have any choice: */
7012 	if (group->group_weight == 1)
7013 		return cpumask_first(sched_group_span(group));
7014 
7015 	/* Traverse only the allowed CPUs */
7016 	for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
7017 		struct rq *rq = cpu_rq(i);
7018 
7019 		if (!sched_core_cookie_match(rq, p))
7020 			continue;
7021 
7022 		if (sched_idle_cpu(i))
7023 			return i;
7024 
7025 		if (available_idle_cpu(i)) {
7026 			struct cpuidle_state *idle = idle_get_state(rq);
7027 			if (idle && idle->exit_latency < min_exit_latency) {
7028 				/*
7029 				 * We give priority to a CPU whose idle state
7030 				 * has the smallest exit latency irrespective
7031 				 * of any idle timestamp.
7032 				 */
7033 				min_exit_latency = idle->exit_latency;
7034 				latest_idle_timestamp = rq->idle_stamp;
7035 				shallowest_idle_cpu = i;
7036 			} else if ((!idle || idle->exit_latency == min_exit_latency) &&
7037 				   rq->idle_stamp > latest_idle_timestamp) {
7038 				/*
7039 				 * If equal or no active idle state, then
7040 				 * the most recently idled CPU might have
7041 				 * a warmer cache.
7042 				 */
7043 				latest_idle_timestamp = rq->idle_stamp;
7044 				shallowest_idle_cpu = i;
7045 			}
7046 		} else if (shallowest_idle_cpu == -1) {
7047 			load = cpu_load(cpu_rq(i));
7048 			if (load < min_load) {
7049 				min_load = load;
7050 				least_loaded_cpu = i;
7051 			}
7052 		}
7053 	}
7054 
7055 	return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
7056 }
7057 
find_idlest_cpu(struct sched_domain * sd,struct task_struct * p,int cpu,int prev_cpu,int sd_flag)7058 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
7059 				  int cpu, int prev_cpu, int sd_flag)
7060 {
7061 	int new_cpu = cpu;
7062 
7063 	if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
7064 		return prev_cpu;
7065 
7066 	/*
7067 	 * We need task's util for cpu_util_without, sync it up to
7068 	 * prev_cpu's last_update_time.
7069 	 */
7070 	if (!(sd_flag & SD_BALANCE_FORK))
7071 		sync_entity_load_avg(&p->se);
7072 
7073 	while (sd) {
7074 		struct sched_group *group;
7075 		struct sched_domain *tmp;
7076 		int weight;
7077 
7078 		if (!(sd->flags & sd_flag)) {
7079 			sd = sd->child;
7080 			continue;
7081 		}
7082 
7083 		group = find_idlest_group(sd, p, cpu);
7084 		if (!group) {
7085 			sd = sd->child;
7086 			continue;
7087 		}
7088 
7089 		new_cpu = find_idlest_group_cpu(group, p, cpu);
7090 		if (new_cpu == cpu) {
7091 			/* Now try balancing at a lower domain level of 'cpu': */
7092 			sd = sd->child;
7093 			continue;
7094 		}
7095 
7096 		/* Now try balancing at a lower domain level of 'new_cpu': */
7097 		cpu = new_cpu;
7098 		weight = sd->span_weight;
7099 		sd = NULL;
7100 		for_each_domain(cpu, tmp) {
7101 			if (weight <= tmp->span_weight)
7102 				break;
7103 			if (tmp->flags & sd_flag)
7104 				sd = tmp;
7105 		}
7106 	}
7107 
7108 	return new_cpu;
7109 }
7110 
__select_idle_cpu(int cpu,struct task_struct * p)7111 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
7112 {
7113 	if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
7114 	    sched_cpu_cookie_match(cpu_rq(cpu), p))
7115 		return cpu;
7116 
7117 	return -1;
7118 }
7119 
7120 #ifdef CONFIG_SCHED_SMT
7121 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7122 EXPORT_SYMBOL_GPL(sched_smt_present);
7123 
set_idle_cores(int cpu,int val)7124 static inline void set_idle_cores(int cpu, int val)
7125 {
7126 	struct sched_domain_shared *sds;
7127 
7128 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7129 	if (sds)
7130 		WRITE_ONCE(sds->has_idle_cores, val);
7131 }
7132 
test_idle_cores(int cpu)7133 static inline bool test_idle_cores(int cpu)
7134 {
7135 	struct sched_domain_shared *sds;
7136 
7137 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7138 	if (sds)
7139 		return READ_ONCE(sds->has_idle_cores);
7140 
7141 	return false;
7142 }
7143 
7144 /*
7145  * Scans the local SMT mask to see if the entire core is idle, and records this
7146  * information in sd_llc_shared->has_idle_cores.
7147  *
7148  * Since SMT siblings share all cache levels, inspecting this limited remote
7149  * state should be fairly cheap.
7150  */
__update_idle_core(struct rq * rq)7151 void __update_idle_core(struct rq *rq)
7152 {
7153 	int core = cpu_of(rq);
7154 	int cpu;
7155 
7156 	rcu_read_lock();
7157 	if (test_idle_cores(core))
7158 		goto unlock;
7159 
7160 	for_each_cpu(cpu, cpu_smt_mask(core)) {
7161 		if (cpu == core)
7162 			continue;
7163 
7164 		if (!available_idle_cpu(cpu))
7165 			goto unlock;
7166 	}
7167 
7168 	set_idle_cores(core, 1);
7169 unlock:
7170 	rcu_read_unlock();
7171 }
7172 
7173 /*
7174  * Scan the entire LLC domain for idle cores; this dynamically switches off if
7175  * there are no idle cores left in the system; tracked through
7176  * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7177  */
select_idle_core(struct task_struct * p,int core,struct cpumask * cpus,int * idle_cpu)7178 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7179 {
7180 	bool idle = true;
7181 	int cpu;
7182 
7183 	for_each_cpu(cpu, cpu_smt_mask(core)) {
7184 		if (!available_idle_cpu(cpu)) {
7185 			idle = false;
7186 			if (*idle_cpu == -1) {
7187 				if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
7188 					*idle_cpu = cpu;
7189 					break;
7190 				}
7191 				continue;
7192 			}
7193 			break;
7194 		}
7195 		if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
7196 			*idle_cpu = cpu;
7197 	}
7198 
7199 	if (idle)
7200 		return core;
7201 
7202 	cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
7203 	return -1;
7204 }
7205 
7206 /*
7207  * Scan the local SMT mask for idle CPUs.
7208  */
select_idle_smt(struct task_struct * p,int target)7209 static int select_idle_smt(struct task_struct *p, int target)
7210 {
7211 	int cpu;
7212 
7213 	for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
7214 		if (cpu == target)
7215 			continue;
7216 		if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
7217 			return cpu;
7218 	}
7219 
7220 	return -1;
7221 }
7222 
7223 #else /* CONFIG_SCHED_SMT */
7224 
set_idle_cores(int cpu,int val)7225 static inline void set_idle_cores(int cpu, int val)
7226 {
7227 }
7228 
test_idle_cores(int cpu)7229 static inline bool test_idle_cores(int cpu)
7230 {
7231 	return false;
7232 }
7233 
select_idle_core(struct task_struct * p,int core,struct cpumask * cpus,int * idle_cpu)7234 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7235 {
7236 	return __select_idle_cpu(core, p);
7237 }
7238 
select_idle_smt(struct task_struct * p,int target)7239 static inline int select_idle_smt(struct task_struct *p, int target)
7240 {
7241 	return -1;
7242 }
7243 
7244 #endif /* CONFIG_SCHED_SMT */
7245 
7246 /*
7247  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
7248  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7249  * average idle time for this rq (as found in rq->avg_idle).
7250  */
select_idle_cpu(struct task_struct * p,struct sched_domain * sd,bool has_idle_core,int target)7251 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
7252 {
7253 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7254 	int i, cpu, idle_cpu = -1, nr = INT_MAX;
7255 	struct sched_domain_shared *sd_share;
7256 	struct rq *this_rq = this_rq();
7257 	int this = smp_processor_id();
7258 	struct sched_domain *this_sd = NULL;
7259 	u64 time = 0;
7260 
7261 	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7262 
7263 	if (sched_feat(SIS_PROP) && !has_idle_core) {
7264 		u64 avg_cost, avg_idle, span_avg;
7265 		unsigned long now = jiffies;
7266 
7267 		this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
7268 		if (!this_sd)
7269 			return -1;
7270 
7271 		/*
7272 		 * If we're busy, the assumption that the last idle period
7273 		 * predicts the future is flawed; age away the remaining
7274 		 * predicted idle time.
7275 		 */
7276 		if (unlikely(this_rq->wake_stamp < now)) {
7277 			while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
7278 				this_rq->wake_stamp++;
7279 				this_rq->wake_avg_idle >>= 1;
7280 			}
7281 		}
7282 
7283 		avg_idle = this_rq->wake_avg_idle;
7284 		avg_cost = this_sd->avg_scan_cost + 1;
7285 
7286 		span_avg = sd->span_weight * avg_idle;
7287 		if (span_avg > 4*avg_cost)
7288 			nr = div_u64(span_avg, avg_cost);
7289 		else
7290 			nr = 4;
7291 
7292 		time = cpu_clock(this);
7293 	}
7294 
7295 	if (sched_feat(SIS_UTIL)) {
7296 		sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
7297 		if (sd_share) {
7298 			/* because !--nr is the condition to stop scan */
7299 			nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
7300 			/* overloaded LLC is unlikely to have idle cpu/core */
7301 			if (nr == 1)
7302 				return -1;
7303 		}
7304 	}
7305 
7306 	for_each_cpu_wrap(cpu, cpus, target + 1) {
7307 		if (has_idle_core) {
7308 			i = select_idle_core(p, cpu, cpus, &idle_cpu);
7309 			if ((unsigned int)i < nr_cpumask_bits)
7310 				return i;
7311 
7312 		} else {
7313 			if (!--nr)
7314 				return -1;
7315 			idle_cpu = __select_idle_cpu(cpu, p);
7316 			if ((unsigned int)idle_cpu < nr_cpumask_bits)
7317 				break;
7318 		}
7319 	}
7320 
7321 	if (has_idle_core)
7322 		set_idle_cores(target, false);
7323 
7324 	if (sched_feat(SIS_PROP) && this_sd && !has_idle_core) {
7325 		time = cpu_clock(this) - time;
7326 
7327 		/*
7328 		 * Account for the scan cost of wakeups against the average
7329 		 * idle time.
7330 		 */
7331 		this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
7332 
7333 		update_avg(&this_sd->avg_scan_cost, time);
7334 	}
7335 
7336 	return idle_cpu;
7337 }
7338 
7339 /*
7340  * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
7341  * the task fits. If no CPU is big enough, but there are idle ones, try to
7342  * maximize capacity.
7343  */
7344 static int
select_idle_capacity(struct task_struct * p,struct sched_domain * sd,int target)7345 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
7346 {
7347 	unsigned long task_util, util_min, util_max, best_cap = 0;
7348 	int fits, best_fits = 0;
7349 	int cpu, best_cpu = -1;
7350 	struct cpumask *cpus;
7351 
7352 	cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7353 	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7354 
7355 	task_util = task_util_est(p);
7356 	util_min = uclamp_eff_value(p, UCLAMP_MIN);
7357 	util_max = uclamp_eff_value(p, UCLAMP_MAX);
7358 
7359 	for_each_cpu_wrap(cpu, cpus, target) {
7360 		unsigned long cpu_cap = capacity_of(cpu);
7361 
7362 		if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
7363 			continue;
7364 
7365 		fits = util_fits_cpu(task_util, util_min, util_max, cpu);
7366 
7367 		/* This CPU fits with all requirements */
7368 		if (fits > 0)
7369 			return cpu;
7370 		/*
7371 		 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
7372 		 * Look for the CPU with best capacity.
7373 		 */
7374 		else if (fits < 0)
7375 			cpu_cap = capacity_orig_of(cpu) - thermal_load_avg(cpu_rq(cpu));
7376 
7377 		/*
7378 		 * First, select CPU which fits better (-1 being better than 0).
7379 		 * Then, select the one with best capacity at same level.
7380 		 */
7381 		if ((fits < best_fits) ||
7382 		    ((fits == best_fits) && (cpu_cap > best_cap))) {
7383 			best_cap = cpu_cap;
7384 			best_cpu = cpu;
7385 			best_fits = fits;
7386 		}
7387 	}
7388 
7389 	return best_cpu;
7390 }
7391 
asym_fits_cpu(unsigned long util,unsigned long util_min,unsigned long util_max,int cpu)7392 static inline bool asym_fits_cpu(unsigned long util,
7393 				 unsigned long util_min,
7394 				 unsigned long util_max,
7395 				 int cpu)
7396 {
7397 	if (sched_asym_cpucap_active())
7398 		/*
7399 		 * Return true only if the cpu fully fits the task requirements
7400 		 * which include the utilization and the performance hints.
7401 		 */
7402 		return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7403 
7404 	return true;
7405 }
7406 
7407 /*
7408  * Try and locate an idle core/thread in the LLC cache domain.
7409  */
select_idle_sibling(struct task_struct * p,int prev,int target)7410 static int select_idle_sibling(struct task_struct *p, int prev, int target)
7411 {
7412 	bool has_idle_core = false;
7413 	struct sched_domain *sd;
7414 	unsigned long task_util, util_min, util_max;
7415 	int i, recent_used_cpu;
7416 
7417 	/*
7418 	 * On asymmetric system, update task utilization because we will check
7419 	 * that the task fits with cpu's capacity.
7420 	 */
7421 	if (sched_asym_cpucap_active()) {
7422 		sync_entity_load_avg(&p->se);
7423 		task_util = task_util_est(p);
7424 		util_min = uclamp_eff_value(p, UCLAMP_MIN);
7425 		util_max = uclamp_eff_value(p, UCLAMP_MAX);
7426 	}
7427 
7428 	/*
7429 	 * per-cpu select_rq_mask usage
7430 	 */
7431 	lockdep_assert_irqs_disabled();
7432 
7433 	if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7434 	    asym_fits_cpu(task_util, util_min, util_max, target))
7435 		return target;
7436 
7437 	/*
7438 	 * If the previous CPU is cache affine and idle, don't be stupid:
7439 	 */
7440 	if (prev != target && cpus_share_cache(prev, target) &&
7441 	    (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7442 	    asym_fits_cpu(task_util, util_min, util_max, prev))
7443 		return prev;
7444 
7445 	/*
7446 	 * Allow a per-cpu kthread to stack with the wakee if the
7447 	 * kworker thread and the tasks previous CPUs are the same.
7448 	 * The assumption is that the wakee queued work for the
7449 	 * per-cpu kthread that is now complete and the wakeup is
7450 	 * essentially a sync wakeup. An obvious example of this
7451 	 * pattern is IO completions.
7452 	 */
7453 	if (is_per_cpu_kthread(current) &&
7454 	    in_task() &&
7455 	    prev == smp_processor_id() &&
7456 	    this_rq()->nr_running <= 1 &&
7457 	    asym_fits_cpu(task_util, util_min, util_max, prev)) {
7458 		return prev;
7459 	}
7460 
7461 	/* Check a recently used CPU as a potential idle candidate: */
7462 	recent_used_cpu = p->recent_used_cpu;
7463 	p->recent_used_cpu = prev;
7464 	if (recent_used_cpu != prev &&
7465 	    recent_used_cpu != target &&
7466 	    cpus_share_cache(recent_used_cpu, target) &&
7467 	    (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7468 	    cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&
7469 	    asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7470 		return recent_used_cpu;
7471 	}
7472 
7473 	/*
7474 	 * For asymmetric CPU capacity systems, our domain of interest is
7475 	 * sd_asym_cpucapacity rather than sd_llc.
7476 	 */
7477 	if (sched_asym_cpucap_active()) {
7478 		sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7479 		/*
7480 		 * On an asymmetric CPU capacity system where an exclusive
7481 		 * cpuset defines a symmetric island (i.e. one unique
7482 		 * capacity_orig value through the cpuset), the key will be set
7483 		 * but the CPUs within that cpuset will not have a domain with
7484 		 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7485 		 * capacity path.
7486 		 */
7487 		if (sd) {
7488 			i = select_idle_capacity(p, sd, target);
7489 			return ((unsigned)i < nr_cpumask_bits) ? i : target;
7490 		}
7491 	}
7492 
7493 	sd = rcu_dereference(per_cpu(sd_llc, target));
7494 	if (!sd)
7495 		return target;
7496 
7497 	if (sched_smt_active()) {
7498 		has_idle_core = test_idle_cores(target);
7499 
7500 		if (!has_idle_core && cpus_share_cache(prev, target)) {
7501 			i = select_idle_smt(p, prev);
7502 			if ((unsigned int)i < nr_cpumask_bits)
7503 				return i;
7504 		}
7505 	}
7506 
7507 	i = select_idle_cpu(p, sd, has_idle_core, target);
7508 	if ((unsigned)i < nr_cpumask_bits)
7509 		return i;
7510 
7511 	return target;
7512 }
7513 
7514 /**
7515  * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7516  * @cpu: the CPU to get the utilization for
7517  * @p: task for which the CPU utilization should be predicted or NULL
7518  * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7519  * @boost: 1 to enable boosting, otherwise 0
7520  *
7521  * The unit of the return value must be the same as the one of CPU capacity
7522  * so that CPU utilization can be compared with CPU capacity.
7523  *
7524  * CPU utilization is the sum of running time of runnable tasks plus the
7525  * recent utilization of currently non-runnable tasks on that CPU.
7526  * It represents the amount of CPU capacity currently used by CFS tasks in
7527  * the range [0..max CPU capacity] with max CPU capacity being the CPU
7528  * capacity at f_max.
7529  *
7530  * The estimated CPU utilization is defined as the maximum between CPU
7531  * utilization and sum of the estimated utilization of the currently
7532  * runnable tasks on that CPU. It preserves a utilization "snapshot" of
7533  * previously-executed tasks, which helps better deduce how busy a CPU will
7534  * be when a long-sleeping task wakes up. The contribution to CPU utilization
7535  * of such a task would be significantly decayed at this point of time.
7536  *
7537  * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization).
7538  * CPU contention for CFS tasks can be detected by CPU runnable > CPU
7539  * utilization. Boosting is implemented in cpu_util() so that internal
7540  * users (e.g. EAS) can use it next to external users (e.g. schedutil),
7541  * latter via cpu_util_cfs_boost().
7542  *
7543  * CPU utilization can be higher than the current CPU capacity
7544  * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because
7545  * of rounding errors as well as task migrations or wakeups of new tasks.
7546  * CPU utilization has to be capped to fit into the [0..max CPU capacity]
7547  * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%)
7548  * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7549  * capacity. CPU utilization is allowed to overshoot current CPU capacity
7550  * though since this is useful for predicting the CPU capacity required
7551  * after task migrations (scheduler-driven DVFS).
7552  *
7553  * Return: (Boosted) (estimated) utilization for the specified CPU.
7554  */
7555 static unsigned long
cpu_util(int cpu,struct task_struct * p,int dst_cpu,int boost)7556 cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
7557 {
7558 	struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
7559 	unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
7560 	unsigned long runnable;
7561 
7562 	if (boost) {
7563 		runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
7564 		util = max(util, runnable);
7565 	}
7566 
7567 	/*
7568 	 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
7569 	 * contribution. If @p migrates from another CPU to @cpu add its
7570 	 * contribution. In all the other cases @cpu is not impacted by the
7571 	 * migration so its util_avg is already correct.
7572 	 */
7573 	if (p && task_cpu(p) == cpu && dst_cpu != cpu)
7574 		lsub_positive(&util, task_util(p));
7575 	else if (p && task_cpu(p) != cpu && dst_cpu == cpu)
7576 		util += task_util(p);
7577 
7578 	if (sched_feat(UTIL_EST)) {
7579 		unsigned long util_est;
7580 
7581 		util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
7582 
7583 		/*
7584 		 * During wake-up @p isn't enqueued yet and doesn't contribute
7585 		 * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
7586 		 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
7587 		 * has been enqueued.
7588 		 *
7589 		 * During exec (@dst_cpu = -1) @p is enqueued and does
7590 		 * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
7591 		 * Remove it to "simulate" cpu_util without @p's contribution.
7592 		 *
7593 		 * Despite the task_on_rq_queued(@p) check there is still a
7594 		 * small window for a possible race when an exec
7595 		 * select_task_rq_fair() races with LB's detach_task().
7596 		 *
7597 		 *   detach_task()
7598 		 *     deactivate_task()
7599 		 *       p->on_rq = TASK_ON_RQ_MIGRATING;
7600 		 *       -------------------------------- A
7601 		 *       dequeue_task()                    \
7602 		 *         dequeue_task_fair()              + Race Time
7603 		 *           util_est_dequeue()            /
7604 		 *       -------------------------------- B
7605 		 *
7606 		 * The additional check "current == p" is required to further
7607 		 * reduce the race window.
7608 		 */
7609 		if (dst_cpu == cpu)
7610 			util_est += _task_util_est(p);
7611 		else if (p && unlikely(task_on_rq_queued(p) || current == p))
7612 			lsub_positive(&util_est, _task_util_est(p));
7613 
7614 		util = max(util, util_est);
7615 	}
7616 
7617 	return min(util, capacity_orig_of(cpu));
7618 }
7619 
cpu_util_cfs(int cpu)7620 unsigned long cpu_util_cfs(int cpu)
7621 {
7622 	return cpu_util(cpu, NULL, -1, 0);
7623 }
7624 
cpu_util_cfs_boost(int cpu)7625 unsigned long cpu_util_cfs_boost(int cpu)
7626 {
7627 	return cpu_util(cpu, NULL, -1, 1);
7628 }
7629 
7630 /*
7631  * cpu_util_without: compute cpu utilization without any contributions from *p
7632  * @cpu: the CPU which utilization is requested
7633  * @p: the task which utilization should be discounted
7634  *
7635  * The utilization of a CPU is defined by the utilization of tasks currently
7636  * enqueued on that CPU as well as tasks which are currently sleeping after an
7637  * execution on that CPU.
7638  *
7639  * This method returns the utilization of the specified CPU by discounting the
7640  * utilization of the specified task, whenever the task is currently
7641  * contributing to the CPU utilization.
7642  */
cpu_util_without(int cpu,struct task_struct * p)7643 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7644 {
7645 	/* Task has no contribution or is new */
7646 	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7647 		p = NULL;
7648 
7649 	return cpu_util(cpu, p, -1, 0);
7650 }
7651 
7652 /*
7653  * energy_env - Utilization landscape for energy estimation.
7654  * @task_busy_time: Utilization contribution by the task for which we test the
7655  *                  placement. Given by eenv_task_busy_time().
7656  * @pd_busy_time:   Utilization of the whole perf domain without the task
7657  *                  contribution. Given by eenv_pd_busy_time().
7658  * @cpu_cap:        Maximum CPU capacity for the perf domain.
7659  * @pd_cap:         Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7660  */
7661 struct energy_env {
7662 	unsigned long task_busy_time;
7663 	unsigned long pd_busy_time;
7664 	unsigned long cpu_cap;
7665 	unsigned long pd_cap;
7666 };
7667 
7668 /*
7669  * Compute the task busy time for compute_energy(). This time cannot be
7670  * injected directly into effective_cpu_util() because of the IRQ scaling.
7671  * The latter only makes sense with the most recent CPUs where the task has
7672  * run.
7673  */
eenv_task_busy_time(struct energy_env * eenv,struct task_struct * p,int prev_cpu)7674 static inline void eenv_task_busy_time(struct energy_env *eenv,
7675 				       struct task_struct *p, int prev_cpu)
7676 {
7677 	unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7678 	unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7679 
7680 	if (unlikely(irq >= max_cap))
7681 		busy_time = max_cap;
7682 	else
7683 		busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7684 
7685 	eenv->task_busy_time = busy_time;
7686 }
7687 
7688 /*
7689  * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7690  * utilization for each @pd_cpus, it however doesn't take into account
7691  * clamping since the ratio (utilization / cpu_capacity) is already enough to
7692  * scale the EM reported power consumption at the (eventually clamped)
7693  * cpu_capacity.
7694  *
7695  * The contribution of the task @p for which we want to estimate the
7696  * energy cost is removed (by cpu_util()) and must be calculated
7697  * separately (see eenv_task_busy_time). This ensures:
7698  *
7699  *   - A stable PD utilization, no matter which CPU of that PD we want to place
7700  *     the task on.
7701  *
7702  *   - A fair comparison between CPUs as the task contribution (task_util())
7703  *     will always be the same no matter which CPU utilization we rely on
7704  *     (util_avg or util_est).
7705  *
7706  * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7707  * exceed @eenv->pd_cap.
7708  */
eenv_pd_busy_time(struct energy_env * eenv,struct cpumask * pd_cpus,struct task_struct * p)7709 static inline void eenv_pd_busy_time(struct energy_env *eenv,
7710 				     struct cpumask *pd_cpus,
7711 				     struct task_struct *p)
7712 {
7713 	unsigned long busy_time = 0;
7714 	int cpu;
7715 
7716 	for_each_cpu(cpu, pd_cpus) {
7717 		unsigned long util = cpu_util(cpu, p, -1, 0);
7718 
7719 		busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
7720 	}
7721 
7722 	eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7723 }
7724 
7725 /*
7726  * Compute the maximum utilization for compute_energy() when the task @p
7727  * is placed on the cpu @dst_cpu.
7728  *
7729  * Returns the maximum utilization among @eenv->cpus. This utilization can't
7730  * exceed @eenv->cpu_cap.
7731  */
7732 static inline unsigned long
eenv_pd_max_util(struct energy_env * eenv,struct cpumask * pd_cpus,struct task_struct * p,int dst_cpu)7733 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7734 		 struct task_struct *p, int dst_cpu)
7735 {
7736 	unsigned long max_util = 0;
7737 	int cpu;
7738 
7739 	for_each_cpu(cpu, pd_cpus) {
7740 		struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7741 		unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
7742 		unsigned long eff_util;
7743 
7744 		/*
7745 		 * Performance domain frequency: utilization clamping
7746 		 * must be considered since it affects the selection
7747 		 * of the performance domain frequency.
7748 		 * NOTE: in case RT tasks are running, by default the
7749 		 * FREQUENCY_UTIL's utilization can be max OPP.
7750 		 */
7751 		eff_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
7752 		max_util = max(max_util, eff_util);
7753 	}
7754 
7755 	return min(max_util, eenv->cpu_cap);
7756 }
7757 
7758 /*
7759  * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7760  * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7761  * contribution is ignored.
7762  */
7763 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)7764 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7765 	       struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7766 {
7767 	unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7768 	unsigned long busy_time = eenv->pd_busy_time;
7769 
7770 	if (dst_cpu >= 0)
7771 		busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7772 
7773 	return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7774 }
7775 
7776 /*
7777  * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7778  * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7779  * spare capacity in each performance domain and uses it as a potential
7780  * candidate to execute the task. Then, it uses the Energy Model to figure
7781  * out which of the CPU candidates is the most energy-efficient.
7782  *
7783  * The rationale for this heuristic is as follows. In a performance domain,
7784  * all the most energy efficient CPU candidates (according to the Energy
7785  * Model) are those for which we'll request a low frequency. When there are
7786  * several CPUs for which the frequency request will be the same, we don't
7787  * have enough data to break the tie between them, because the Energy Model
7788  * only includes active power costs. With this model, if we assume that
7789  * frequency requests follow utilization (e.g. using schedutil), the CPU with
7790  * the maximum spare capacity in a performance domain is guaranteed to be among
7791  * the best candidates of the performance domain.
7792  *
7793  * In practice, it could be preferable from an energy standpoint to pack
7794  * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7795  * but that could also hurt our chances to go cluster idle, and we have no
7796  * ways to tell with the current Energy Model if this is actually a good
7797  * idea or not. So, find_energy_efficient_cpu() basically favors
7798  * cluster-packing, and spreading inside a cluster. That should at least be
7799  * a good thing for latency, and this is consistent with the idea that most
7800  * of the energy savings of EAS come from the asymmetry of the system, and
7801  * not so much from breaking the tie between identical CPUs. That's also the
7802  * reason why EAS is enabled in the topology code only for systems where
7803  * SD_ASYM_CPUCAPACITY is set.
7804  *
7805  * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7806  * they don't have any useful utilization data yet and it's not possible to
7807  * forecast their impact on energy consumption. Consequently, they will be
7808  * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7809  * to be energy-inefficient in some use-cases. The alternative would be to
7810  * bias new tasks towards specific types of CPUs first, or to try to infer
7811  * their util_avg from the parent task, but those heuristics could hurt
7812  * other use-cases too. So, until someone finds a better way to solve this,
7813  * let's keep things simple by re-using the existing slow path.
7814  */
find_energy_efficient_cpu(struct task_struct * p,int prev_cpu)7815 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7816 {
7817 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7818 	unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7819 	unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7820 	unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7821 	struct root_domain *rd = this_rq()->rd;
7822 	int cpu, best_energy_cpu, target = -1;
7823 	int prev_fits = -1, best_fits = -1;
7824 	unsigned long best_thermal_cap = 0;
7825 	unsigned long prev_thermal_cap = 0;
7826 	struct sched_domain *sd;
7827 	struct perf_domain *pd;
7828 	struct energy_env eenv;
7829 
7830 	rcu_read_lock();
7831 	pd = rcu_dereference(rd->pd);
7832 	if (!pd || READ_ONCE(rd->overutilized))
7833 		goto unlock;
7834 
7835 	/*
7836 	 * Energy-aware wake-up happens on the lowest sched_domain starting
7837 	 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7838 	 */
7839 	sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7840 	while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7841 		sd = sd->parent;
7842 	if (!sd)
7843 		goto unlock;
7844 
7845 	target = prev_cpu;
7846 
7847 	sync_entity_load_avg(&p->se);
7848 	if (!task_util_est(p) && p_util_min == 0)
7849 		goto unlock;
7850 
7851 	eenv_task_busy_time(&eenv, p, prev_cpu);
7852 
7853 	for (; pd; pd = pd->next) {
7854 		unsigned long util_min = p_util_min, util_max = p_util_max;
7855 		unsigned long cpu_cap, cpu_thermal_cap, util;
7856 		long prev_spare_cap = -1, max_spare_cap = -1;
7857 		unsigned long rq_util_min, rq_util_max;
7858 		unsigned long cur_delta, base_energy;
7859 		int max_spare_cap_cpu = -1;
7860 		int fits, max_fits = -1;
7861 
7862 		cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
7863 
7864 		if (cpumask_empty(cpus))
7865 			continue;
7866 
7867 		/* Account thermal pressure for the energy estimation */
7868 		cpu = cpumask_first(cpus);
7869 		cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
7870 		cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
7871 
7872 		eenv.cpu_cap = cpu_thermal_cap;
7873 		eenv.pd_cap = 0;
7874 
7875 		for_each_cpu(cpu, cpus) {
7876 			struct rq *rq = cpu_rq(cpu);
7877 
7878 			eenv.pd_cap += cpu_thermal_cap;
7879 
7880 			if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7881 				continue;
7882 
7883 			if (!cpumask_test_cpu(cpu, p->cpus_ptr))
7884 				continue;
7885 
7886 			util = cpu_util(cpu, p, cpu, 0);
7887 			cpu_cap = capacity_of(cpu);
7888 
7889 			/*
7890 			 * Skip CPUs that cannot satisfy the capacity request.
7891 			 * IOW, placing the task there would make the CPU
7892 			 * overutilized. Take uclamp into account to see how
7893 			 * much capacity we can get out of the CPU; this is
7894 			 * aligned with sched_cpu_util().
7895 			 */
7896 			if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
7897 				/*
7898 				 * Open code uclamp_rq_util_with() except for
7899 				 * the clamp() part. Ie: apply max aggregation
7900 				 * only. util_fits_cpu() logic requires to
7901 				 * operate on non clamped util but must use the
7902 				 * max-aggregated uclamp_{min, max}.
7903 				 */
7904 				rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
7905 				rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
7906 
7907 				util_min = max(rq_util_min, p_util_min);
7908 				util_max = max(rq_util_max, p_util_max);
7909 			}
7910 
7911 			fits = util_fits_cpu(util, util_min, util_max, cpu);
7912 			if (!fits)
7913 				continue;
7914 
7915 			lsub_positive(&cpu_cap, util);
7916 
7917 			if (cpu == prev_cpu) {
7918 				/* Always use prev_cpu as a candidate. */
7919 				prev_spare_cap = cpu_cap;
7920 				prev_fits = fits;
7921 			} else if ((fits > max_fits) ||
7922 				   ((fits == max_fits) && ((long)cpu_cap > max_spare_cap))) {
7923 				/*
7924 				 * Find the CPU with the maximum spare capacity
7925 				 * among the remaining CPUs in the performance
7926 				 * domain.
7927 				 */
7928 				max_spare_cap = cpu_cap;
7929 				max_spare_cap_cpu = cpu;
7930 				max_fits = fits;
7931 			}
7932 		}
7933 
7934 		if (max_spare_cap_cpu < 0 && prev_spare_cap < 0)
7935 			continue;
7936 
7937 		eenv_pd_busy_time(&eenv, cpus, p);
7938 		/* Compute the 'base' energy of the pd, without @p */
7939 		base_energy = compute_energy(&eenv, pd, cpus, p, -1);
7940 
7941 		/* Evaluate the energy impact of using prev_cpu. */
7942 		if (prev_spare_cap > -1) {
7943 			prev_delta = compute_energy(&eenv, pd, cpus, p,
7944 						    prev_cpu);
7945 			/* CPU utilization has changed */
7946 			if (prev_delta < base_energy)
7947 				goto unlock;
7948 			prev_delta -= base_energy;
7949 			prev_thermal_cap = cpu_thermal_cap;
7950 			best_delta = min(best_delta, prev_delta);
7951 		}
7952 
7953 		/* Evaluate the energy impact of using max_spare_cap_cpu. */
7954 		if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
7955 			/* Current best energy cpu fits better */
7956 			if (max_fits < best_fits)
7957 				continue;
7958 
7959 			/*
7960 			 * Both don't fit performance hint (i.e. uclamp_min)
7961 			 * but best energy cpu has better capacity.
7962 			 */
7963 			if ((max_fits < 0) &&
7964 			    (cpu_thermal_cap <= best_thermal_cap))
7965 				continue;
7966 
7967 			cur_delta = compute_energy(&eenv, pd, cpus, p,
7968 						   max_spare_cap_cpu);
7969 			/* CPU utilization has changed */
7970 			if (cur_delta < base_energy)
7971 				goto unlock;
7972 			cur_delta -= base_energy;
7973 
7974 			/*
7975 			 * Both fit for the task but best energy cpu has lower
7976 			 * energy impact.
7977 			 */
7978 			if ((max_fits > 0) && (best_fits > 0) &&
7979 			    (cur_delta >= best_delta))
7980 				continue;
7981 
7982 			best_delta = cur_delta;
7983 			best_energy_cpu = max_spare_cap_cpu;
7984 			best_fits = max_fits;
7985 			best_thermal_cap = cpu_thermal_cap;
7986 		}
7987 	}
7988 	rcu_read_unlock();
7989 
7990 	if ((best_fits > prev_fits) ||
7991 	    ((best_fits > 0) && (best_delta < prev_delta)) ||
7992 	    ((best_fits < 0) && (best_thermal_cap > prev_thermal_cap)))
7993 		target = best_energy_cpu;
7994 
7995 	return target;
7996 
7997 unlock:
7998 	rcu_read_unlock();
7999 
8000 	return target;
8001 }
8002 
8003 /*
8004  * select_task_rq_fair: Select target runqueue for the waking task in domains
8005  * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
8006  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
8007  *
8008  * Balances load by selecting the idlest CPU in the idlest group, or under
8009  * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
8010  *
8011  * Returns the target CPU number.
8012  */
8013 static int
select_task_rq_fair(struct task_struct * p,int prev_cpu,int wake_flags)8014 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
8015 {
8016 	int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
8017 	struct sched_domain *tmp, *sd = NULL;
8018 	int cpu = smp_processor_id();
8019 	int new_cpu = prev_cpu;
8020 	int want_affine = 0;
8021 	/* SD_flags and WF_flags share the first nibble */
8022 	int sd_flag = wake_flags & 0xF;
8023 
8024 	/*
8025 	 * required for stable ->cpus_allowed
8026 	 */
8027 	lockdep_assert_held(&p->pi_lock);
8028 	if (wake_flags & WF_TTWU) {
8029 		record_wakee(p);
8030 
8031 		if ((wake_flags & WF_CURRENT_CPU) &&
8032 		    cpumask_test_cpu(cpu, p->cpus_ptr))
8033 			return cpu;
8034 
8035 		if (sched_energy_enabled()) {
8036 			new_cpu = find_energy_efficient_cpu(p, prev_cpu);
8037 			if (new_cpu >= 0)
8038 				return new_cpu;
8039 			new_cpu = prev_cpu;
8040 		}
8041 
8042 		want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
8043 	}
8044 
8045 	rcu_read_lock();
8046 	for_each_domain(cpu, tmp) {
8047 		/*
8048 		 * If both 'cpu' and 'prev_cpu' are part of this domain,
8049 		 * cpu is a valid SD_WAKE_AFFINE target.
8050 		 */
8051 		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
8052 		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
8053 			if (cpu != prev_cpu)
8054 				new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
8055 
8056 			sd = NULL; /* Prefer wake_affine over balance flags */
8057 			break;
8058 		}
8059 
8060 		/*
8061 		 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
8062 		 * usually do not have SD_BALANCE_WAKE set. That means wakeup
8063 		 * will usually go to the fast path.
8064 		 */
8065 		if (tmp->flags & sd_flag)
8066 			sd = tmp;
8067 		else if (!want_affine)
8068 			break;
8069 	}
8070 
8071 	if (unlikely(sd)) {
8072 		/* Slow path */
8073 		new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
8074 	} else if (wake_flags & WF_TTWU) { /* XXX always ? */
8075 		/* Fast path */
8076 		new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
8077 	}
8078 	rcu_read_unlock();
8079 
8080 	return new_cpu;
8081 }
8082 
8083 /*
8084  * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
8085  * cfs_rq_of(p) references at time of call are still valid and identify the
8086  * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
8087  */
migrate_task_rq_fair(struct task_struct * p,int new_cpu)8088 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
8089 {
8090 	struct sched_entity *se = &p->se;
8091 
8092 	if (!task_on_rq_migrating(p)) {
8093 		remove_entity_load_avg(se);
8094 
8095 		/*
8096 		 * Here, the task's PELT values have been updated according to
8097 		 * the current rq's clock. But if that clock hasn't been
8098 		 * updated in a while, a substantial idle time will be missed,
8099 		 * leading to an inflation after wake-up on the new rq.
8100 		 *
8101 		 * Estimate the missing time from the cfs_rq last_update_time
8102 		 * and update sched_avg to improve the PELT continuity after
8103 		 * migration.
8104 		 */
8105 		migrate_se_pelt_lag(se);
8106 	}
8107 
8108 	/* Tell new CPU we are migrated */
8109 	se->avg.last_update_time = 0;
8110 
8111 	update_scan_period(p, new_cpu);
8112 }
8113 
task_dead_fair(struct task_struct * p)8114 static void task_dead_fair(struct task_struct *p)
8115 {
8116 	remove_entity_load_avg(&p->se);
8117 }
8118 
8119 static int
balance_fair(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)8120 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8121 {
8122 	if (rq->nr_running)
8123 		return 1;
8124 
8125 	return newidle_balance(rq, rf) != 0;
8126 }
8127 #endif /* CONFIG_SMP */
8128 
set_next_buddy(struct sched_entity * se)8129 static void set_next_buddy(struct sched_entity *se)
8130 {
8131 	for_each_sched_entity(se) {
8132 		if (SCHED_WARN_ON(!se->on_rq))
8133 			return;
8134 		if (se_is_idle(se))
8135 			return;
8136 		cfs_rq_of(se)->next = se;
8137 	}
8138 }
8139 
8140 /*
8141  * Preempt the current task with a newly woken task if needed:
8142  */
check_preempt_wakeup(struct rq * rq,struct task_struct * p,int wake_flags)8143 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
8144 {
8145 	struct task_struct *curr = rq->curr;
8146 	struct sched_entity *se = &curr->se, *pse = &p->se;
8147 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8148 	int next_buddy_marked = 0;
8149 	int cse_is_idle, pse_is_idle;
8150 
8151 	if (unlikely(se == pse))
8152 		return;
8153 
8154 	/*
8155 	 * This is possible from callers such as attach_tasks(), in which we
8156 	 * unconditionally check_preempt_curr() after an enqueue (which may have
8157 	 * lead to a throttle).  This both saves work and prevents false
8158 	 * next-buddy nomination below.
8159 	 */
8160 	if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
8161 		return;
8162 
8163 	if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) {
8164 		set_next_buddy(pse);
8165 		next_buddy_marked = 1;
8166 	}
8167 
8168 	/*
8169 	 * We can come here with TIF_NEED_RESCHED already set from new task
8170 	 * wake up path.
8171 	 *
8172 	 * Note: this also catches the edge-case of curr being in a throttled
8173 	 * group (e.g. via set_curr_task), since update_curr() (in the
8174 	 * enqueue of curr) will have resulted in resched being set.  This
8175 	 * prevents us from potentially nominating it as a false LAST_BUDDY
8176 	 * below.
8177 	 */
8178 	if (test_tsk_need_resched(curr))
8179 		return;
8180 
8181 	/* Idle tasks are by definition preempted by non-idle tasks. */
8182 	if (unlikely(task_has_idle_policy(curr)) &&
8183 	    likely(!task_has_idle_policy(p)))
8184 		goto preempt;
8185 
8186 	/*
8187 	 * Batch and idle tasks do not preempt non-idle tasks (their preemption
8188 	 * is driven by the tick):
8189 	 */
8190 	if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
8191 		return;
8192 
8193 	find_matching_se(&se, &pse);
8194 	WARN_ON_ONCE(!pse);
8195 
8196 	cse_is_idle = se_is_idle(se);
8197 	pse_is_idle = se_is_idle(pse);
8198 
8199 	/*
8200 	 * Preempt an idle group in favor of a non-idle group (and don't preempt
8201 	 * in the inverse case).
8202 	 */
8203 	if (cse_is_idle && !pse_is_idle)
8204 		goto preempt;
8205 	if (cse_is_idle != pse_is_idle)
8206 		return;
8207 
8208 	cfs_rq = cfs_rq_of(se);
8209 	update_curr(cfs_rq);
8210 
8211 	/*
8212 	 * XXX pick_eevdf(cfs_rq) != se ?
8213 	 */
8214 	if (pick_eevdf(cfs_rq) == pse)
8215 		goto preempt;
8216 
8217 	return;
8218 
8219 preempt:
8220 	resched_curr(rq);
8221 }
8222 
8223 #ifdef CONFIG_SMP
pick_task_fair(struct rq * rq)8224 static struct task_struct *pick_task_fair(struct rq *rq)
8225 {
8226 	struct sched_entity *se;
8227 	struct cfs_rq *cfs_rq;
8228 
8229 again:
8230 	cfs_rq = &rq->cfs;
8231 	if (!cfs_rq->nr_running)
8232 		return NULL;
8233 
8234 	do {
8235 		struct sched_entity *curr = cfs_rq->curr;
8236 
8237 		/* When we pick for a remote RQ, we'll not have done put_prev_entity() */
8238 		if (curr) {
8239 			if (curr->on_rq)
8240 				update_curr(cfs_rq);
8241 			else
8242 				curr = NULL;
8243 
8244 			if (unlikely(check_cfs_rq_runtime(cfs_rq)))
8245 				goto again;
8246 		}
8247 
8248 		se = pick_next_entity(cfs_rq, curr);
8249 		cfs_rq = group_cfs_rq(se);
8250 	} while (cfs_rq);
8251 
8252 	return task_of(se);
8253 }
8254 #endif
8255 
8256 struct task_struct *
pick_next_task_fair(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)8257 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8258 {
8259 	struct cfs_rq *cfs_rq = &rq->cfs;
8260 	struct sched_entity *se;
8261 	struct task_struct *p;
8262 	int new_tasks;
8263 
8264 again:
8265 	if (!sched_fair_runnable(rq))
8266 		goto idle;
8267 
8268 #ifdef CONFIG_FAIR_GROUP_SCHED
8269 	if (!prev || prev->sched_class != &fair_sched_class)
8270 		goto simple;
8271 
8272 	/*
8273 	 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
8274 	 * likely that a next task is from the same cgroup as the current.
8275 	 *
8276 	 * Therefore attempt to avoid putting and setting the entire cgroup
8277 	 * hierarchy, only change the part that actually changes.
8278 	 */
8279 
8280 	do {
8281 		struct sched_entity *curr = cfs_rq->curr;
8282 
8283 		/*
8284 		 * Since we got here without doing put_prev_entity() we also
8285 		 * have to consider cfs_rq->curr. If it is still a runnable
8286 		 * entity, update_curr() will update its vruntime, otherwise
8287 		 * forget we've ever seen it.
8288 		 */
8289 		if (curr) {
8290 			if (curr->on_rq)
8291 				update_curr(cfs_rq);
8292 			else
8293 				curr = NULL;
8294 
8295 			/*
8296 			 * This call to check_cfs_rq_runtime() will do the
8297 			 * throttle and dequeue its entity in the parent(s).
8298 			 * Therefore the nr_running test will indeed
8299 			 * be correct.
8300 			 */
8301 			if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
8302 				cfs_rq = &rq->cfs;
8303 
8304 				if (!cfs_rq->nr_running)
8305 					goto idle;
8306 
8307 				goto simple;
8308 			}
8309 		}
8310 
8311 		se = pick_next_entity(cfs_rq, curr);
8312 		cfs_rq = group_cfs_rq(se);
8313 	} while (cfs_rq);
8314 
8315 	p = task_of(se);
8316 
8317 	/*
8318 	 * Since we haven't yet done put_prev_entity and if the selected task
8319 	 * is a different task than we started out with, try and touch the
8320 	 * least amount of cfs_rqs.
8321 	 */
8322 	if (prev != p) {
8323 		struct sched_entity *pse = &prev->se;
8324 
8325 		while (!(cfs_rq = is_same_group(se, pse))) {
8326 			int se_depth = se->depth;
8327 			int pse_depth = pse->depth;
8328 
8329 			if (se_depth <= pse_depth) {
8330 				put_prev_entity(cfs_rq_of(pse), pse);
8331 				pse = parent_entity(pse);
8332 			}
8333 			if (se_depth >= pse_depth) {
8334 				set_next_entity(cfs_rq_of(se), se);
8335 				se = parent_entity(se);
8336 			}
8337 		}
8338 
8339 		put_prev_entity(cfs_rq, pse);
8340 		set_next_entity(cfs_rq, se);
8341 	}
8342 
8343 	goto done;
8344 simple:
8345 #endif
8346 	if (prev)
8347 		put_prev_task(rq, prev);
8348 
8349 	do {
8350 		se = pick_next_entity(cfs_rq, NULL);
8351 		set_next_entity(cfs_rq, se);
8352 		cfs_rq = group_cfs_rq(se);
8353 	} while (cfs_rq);
8354 
8355 	p = task_of(se);
8356 
8357 done: __maybe_unused;
8358 #ifdef CONFIG_SMP
8359 	/*
8360 	 * Move the next running task to the front of
8361 	 * the list, so our cfs_tasks list becomes MRU
8362 	 * one.
8363 	 */
8364 	list_move(&p->se.group_node, &rq->cfs_tasks);
8365 #endif
8366 
8367 	if (hrtick_enabled_fair(rq))
8368 		hrtick_start_fair(rq, p);
8369 
8370 	update_misfit_status(p, rq);
8371 	sched_fair_update_stop_tick(rq, p);
8372 
8373 	return p;
8374 
8375 idle:
8376 	if (!rf)
8377 		return NULL;
8378 
8379 	new_tasks = newidle_balance(rq, rf);
8380 
8381 	/*
8382 	 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
8383 	 * possible for any higher priority task to appear. In that case we
8384 	 * must re-start the pick_next_entity() loop.
8385 	 */
8386 	if (new_tasks < 0)
8387 		return RETRY_TASK;
8388 
8389 	if (new_tasks > 0)
8390 		goto again;
8391 
8392 	/*
8393 	 * rq is about to be idle, check if we need to update the
8394 	 * lost_idle_time of clock_pelt
8395 	 */
8396 	update_idle_rq_clock_pelt(rq);
8397 
8398 	return NULL;
8399 }
8400 
__pick_next_task_fair(struct rq * rq)8401 static struct task_struct *__pick_next_task_fair(struct rq *rq)
8402 {
8403 	return pick_next_task_fair(rq, NULL, NULL);
8404 }
8405 
8406 /*
8407  * Account for a descheduled task:
8408  */
put_prev_task_fair(struct rq * rq,struct task_struct * prev)8409 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
8410 {
8411 	struct sched_entity *se = &prev->se;
8412 	struct cfs_rq *cfs_rq;
8413 
8414 	for_each_sched_entity(se) {
8415 		cfs_rq = cfs_rq_of(se);
8416 		put_prev_entity(cfs_rq, se);
8417 	}
8418 }
8419 
8420 /*
8421  * sched_yield() is very simple
8422  */
yield_task_fair(struct rq * rq)8423 static void yield_task_fair(struct rq *rq)
8424 {
8425 	struct task_struct *curr = rq->curr;
8426 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8427 	struct sched_entity *se = &curr->se;
8428 
8429 	/*
8430 	 * Are we the only task in the tree?
8431 	 */
8432 	if (unlikely(rq->nr_running == 1))
8433 		return;
8434 
8435 	clear_buddies(cfs_rq, se);
8436 
8437 	update_rq_clock(rq);
8438 	/*
8439 	 * Update run-time statistics of the 'current'.
8440 	 */
8441 	update_curr(cfs_rq);
8442 	/*
8443 	 * Tell update_rq_clock() that we've just updated,
8444 	 * so we don't do microscopic update in schedule()
8445 	 * and double the fastpath cost.
8446 	 */
8447 	rq_clock_skip_update(rq);
8448 
8449 	se->deadline += calc_delta_fair(se->slice, se);
8450 }
8451 
yield_to_task_fair(struct rq * rq,struct task_struct * p)8452 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
8453 {
8454 	struct sched_entity *se = &p->se;
8455 
8456 	/* throttled hierarchies are not runnable */
8457 	if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
8458 		return false;
8459 
8460 	/* Tell the scheduler that we'd really like pse to run next. */
8461 	set_next_buddy(se);
8462 
8463 	yield_task_fair(rq);
8464 
8465 	return true;
8466 }
8467 
8468 #ifdef CONFIG_SMP
8469 /**************************************************
8470  * Fair scheduling class load-balancing methods.
8471  *
8472  * BASICS
8473  *
8474  * The purpose of load-balancing is to achieve the same basic fairness the
8475  * per-CPU scheduler provides, namely provide a proportional amount of compute
8476  * time to each task. This is expressed in the following equation:
8477  *
8478  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
8479  *
8480  * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
8481  * W_i,0 is defined as:
8482  *
8483  *   W_i,0 = \Sum_j w_i,j                                             (2)
8484  *
8485  * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
8486  * is derived from the nice value as per sched_prio_to_weight[].
8487  *
8488  * The weight average is an exponential decay average of the instantaneous
8489  * weight:
8490  *
8491  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
8492  *
8493  * C_i is the compute capacity of CPU i, typically it is the
8494  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
8495  * can also include other factors [XXX].
8496  *
8497  * To achieve this balance we define a measure of imbalance which follows
8498  * directly from (1):
8499  *
8500  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
8501  *
8502  * We them move tasks around to minimize the imbalance. In the continuous
8503  * function space it is obvious this converges, in the discrete case we get
8504  * a few fun cases generally called infeasible weight scenarios.
8505  *
8506  * [XXX expand on:
8507  *     - infeasible weights;
8508  *     - local vs global optima in the discrete case. ]
8509  *
8510  *
8511  * SCHED DOMAINS
8512  *
8513  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
8514  * for all i,j solution, we create a tree of CPUs that follows the hardware
8515  * topology where each level pairs two lower groups (or better). This results
8516  * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
8517  * tree to only the first of the previous level and we decrease the frequency
8518  * of load-balance at each level inv. proportional to the number of CPUs in
8519  * the groups.
8520  *
8521  * This yields:
8522  *
8523  *     log_2 n     1     n
8524  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
8525  *     i = 0      2^i   2^i
8526  *                               `- size of each group
8527  *         |         |     `- number of CPUs doing load-balance
8528  *         |         `- freq
8529  *         `- sum over all levels
8530  *
8531  * Coupled with a limit on how many tasks we can migrate every balance pass,
8532  * this makes (5) the runtime complexity of the balancer.
8533  *
8534  * An important property here is that each CPU is still (indirectly) connected
8535  * to every other CPU in at most O(log n) steps:
8536  *
8537  * The adjacency matrix of the resulting graph is given by:
8538  *
8539  *             log_2 n
8540  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
8541  *             k = 0
8542  *
8543  * And you'll find that:
8544  *
8545  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
8546  *
8547  * Showing there's indeed a path between every CPU in at most O(log n) steps.
8548  * The task movement gives a factor of O(m), giving a convergence complexity
8549  * of:
8550  *
8551  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
8552  *
8553  *
8554  * WORK CONSERVING
8555  *
8556  * In order to avoid CPUs going idle while there's still work to do, new idle
8557  * balancing is more aggressive and has the newly idle CPU iterate up the domain
8558  * tree itself instead of relying on other CPUs to bring it work.
8559  *
8560  * This adds some complexity to both (5) and (8) but it reduces the total idle
8561  * time.
8562  *
8563  * [XXX more?]
8564  *
8565  *
8566  * CGROUPS
8567  *
8568  * Cgroups make a horror show out of (2), instead of a simple sum we get:
8569  *
8570  *                                s_k,i
8571  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
8572  *                                 S_k
8573  *
8574  * Where
8575  *
8576  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
8577  *
8578  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8579  *
8580  * The big problem is S_k, its a global sum needed to compute a local (W_i)
8581  * property.
8582  *
8583  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8584  *      rewrite all of this once again.]
8585  */
8586 
8587 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8588 
8589 enum fbq_type { regular, remote, all };
8590 
8591 /*
8592  * 'group_type' describes the group of CPUs at the moment of load balancing.
8593  *
8594  * The enum is ordered by pulling priority, with the group with lowest priority
8595  * first so the group_type can simply be compared when selecting the busiest
8596  * group. See update_sd_pick_busiest().
8597  */
8598 enum group_type {
8599 	/* The group has spare capacity that can be used to run more tasks.  */
8600 	group_has_spare = 0,
8601 	/*
8602 	 * The group is fully used and the tasks don't compete for more CPU
8603 	 * cycles. Nevertheless, some tasks might wait before running.
8604 	 */
8605 	group_fully_busy,
8606 	/*
8607 	 * One task doesn't fit with CPU's capacity and must be migrated to a
8608 	 * more powerful CPU.
8609 	 */
8610 	group_misfit_task,
8611 	/*
8612 	 * Balance SMT group that's fully busy. Can benefit from migration
8613 	 * a task on SMT with busy sibling to another CPU on idle core.
8614 	 */
8615 	group_smt_balance,
8616 	/*
8617 	 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8618 	 * and the task should be migrated to it instead of running on the
8619 	 * current CPU.
8620 	 */
8621 	group_asym_packing,
8622 	/*
8623 	 * The tasks' affinity constraints previously prevented the scheduler
8624 	 * from balancing the load across the system.
8625 	 */
8626 	group_imbalanced,
8627 	/*
8628 	 * The CPU is overloaded and can't provide expected CPU cycles to all
8629 	 * tasks.
8630 	 */
8631 	group_overloaded
8632 };
8633 
8634 enum migration_type {
8635 	migrate_load = 0,
8636 	migrate_util,
8637 	migrate_task,
8638 	migrate_misfit
8639 };
8640 
8641 #define LBF_ALL_PINNED	0x01
8642 #define LBF_NEED_BREAK	0x02
8643 #define LBF_DST_PINNED  0x04
8644 #define LBF_SOME_PINNED	0x08
8645 #define LBF_ACTIVE_LB	0x10
8646 
8647 struct lb_env {
8648 	struct sched_domain	*sd;
8649 
8650 	struct rq		*src_rq;
8651 	int			src_cpu;
8652 
8653 	int			dst_cpu;
8654 	struct rq		*dst_rq;
8655 
8656 	struct cpumask		*dst_grpmask;
8657 	int			new_dst_cpu;
8658 	enum cpu_idle_type	idle;
8659 	long			imbalance;
8660 	/* The set of CPUs under consideration for load-balancing */
8661 	struct cpumask		*cpus;
8662 
8663 	unsigned int		flags;
8664 
8665 	unsigned int		loop;
8666 	unsigned int		loop_break;
8667 	unsigned int		loop_max;
8668 
8669 	enum fbq_type		fbq_type;
8670 	enum migration_type	migration_type;
8671 	struct list_head	tasks;
8672 };
8673 
8674 /*
8675  * Is this task likely cache-hot:
8676  */
task_hot(struct task_struct * p,struct lb_env * env)8677 static int task_hot(struct task_struct *p, struct lb_env *env)
8678 {
8679 	s64 delta;
8680 
8681 	lockdep_assert_rq_held(env->src_rq);
8682 
8683 	if (p->sched_class != &fair_sched_class)
8684 		return 0;
8685 
8686 	if (unlikely(task_has_idle_policy(p)))
8687 		return 0;
8688 
8689 	/* SMT siblings share cache */
8690 	if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8691 		return 0;
8692 
8693 	/*
8694 	 * Buddy candidates are cache hot:
8695 	 */
8696 	if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8697 	    (&p->se == cfs_rq_of(&p->se)->next))
8698 		return 1;
8699 
8700 	if (sysctl_sched_migration_cost == -1)
8701 		return 1;
8702 
8703 	/*
8704 	 * Don't migrate task if the task's cookie does not match
8705 	 * with the destination CPU's core cookie.
8706 	 */
8707 	if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8708 		return 1;
8709 
8710 	if (sysctl_sched_migration_cost == 0)
8711 		return 0;
8712 
8713 	delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8714 
8715 	return delta < (s64)sysctl_sched_migration_cost;
8716 }
8717 
8718 #ifdef CONFIG_NUMA_BALANCING
8719 /*
8720  * Returns 1, if task migration degrades locality
8721  * Returns 0, if task migration improves locality i.e migration preferred.
8722  * Returns -1, if task migration is not affected by locality.
8723  */
migrate_degrades_locality(struct task_struct * p,struct lb_env * env)8724 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8725 {
8726 	struct numa_group *numa_group = rcu_dereference(p->numa_group);
8727 	unsigned long src_weight, dst_weight;
8728 	int src_nid, dst_nid, dist;
8729 
8730 	if (!static_branch_likely(&sched_numa_balancing))
8731 		return -1;
8732 
8733 	if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8734 		return -1;
8735 
8736 	src_nid = cpu_to_node(env->src_cpu);
8737 	dst_nid = cpu_to_node(env->dst_cpu);
8738 
8739 	if (src_nid == dst_nid)
8740 		return -1;
8741 
8742 	/* Migrating away from the preferred node is always bad. */
8743 	if (src_nid == p->numa_preferred_nid) {
8744 		if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8745 			return 1;
8746 		else
8747 			return -1;
8748 	}
8749 
8750 	/* Encourage migration to the preferred node. */
8751 	if (dst_nid == p->numa_preferred_nid)
8752 		return 0;
8753 
8754 	/* Leaving a core idle is often worse than degrading locality. */
8755 	if (env->idle == CPU_IDLE)
8756 		return -1;
8757 
8758 	dist = node_distance(src_nid, dst_nid);
8759 	if (numa_group) {
8760 		src_weight = group_weight(p, src_nid, dist);
8761 		dst_weight = group_weight(p, dst_nid, dist);
8762 	} else {
8763 		src_weight = task_weight(p, src_nid, dist);
8764 		dst_weight = task_weight(p, dst_nid, dist);
8765 	}
8766 
8767 	return dst_weight < src_weight;
8768 }
8769 
8770 #else
migrate_degrades_locality(struct task_struct * p,struct lb_env * env)8771 static inline int migrate_degrades_locality(struct task_struct *p,
8772 					     struct lb_env *env)
8773 {
8774 	return -1;
8775 }
8776 #endif
8777 
8778 /*
8779  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8780  */
8781 static
can_migrate_task(struct task_struct * p,struct lb_env * env)8782 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8783 {
8784 	int tsk_cache_hot;
8785 
8786 	lockdep_assert_rq_held(env->src_rq);
8787 
8788 	/*
8789 	 * We do not migrate tasks that are:
8790 	 * 1) throttled_lb_pair, or
8791 	 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8792 	 * 3) running (obviously), or
8793 	 * 4) are cache-hot on their current CPU.
8794 	 */
8795 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8796 		return 0;
8797 
8798 	/* Disregard pcpu kthreads; they are where they need to be. */
8799 	if (kthread_is_per_cpu(p))
8800 		return 0;
8801 
8802 	if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8803 		int cpu;
8804 
8805 		schedstat_inc(p->stats.nr_failed_migrations_affine);
8806 
8807 		env->flags |= LBF_SOME_PINNED;
8808 
8809 		/*
8810 		 * Remember if this task can be migrated to any other CPU in
8811 		 * our sched_group. We may want to revisit it if we couldn't
8812 		 * meet load balance goals by pulling other tasks on src_cpu.
8813 		 *
8814 		 * Avoid computing new_dst_cpu
8815 		 * - for NEWLY_IDLE
8816 		 * - if we have already computed one in current iteration
8817 		 * - if it's an active balance
8818 		 */
8819 		if (env->idle == CPU_NEWLY_IDLE ||
8820 		    env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8821 			return 0;
8822 
8823 		/* Prevent to re-select dst_cpu via env's CPUs: */
8824 		for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8825 			if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8826 				env->flags |= LBF_DST_PINNED;
8827 				env->new_dst_cpu = cpu;
8828 				break;
8829 			}
8830 		}
8831 
8832 		return 0;
8833 	}
8834 
8835 	/* Record that we found at least one task that could run on dst_cpu */
8836 	env->flags &= ~LBF_ALL_PINNED;
8837 
8838 	if (task_on_cpu(env->src_rq, p)) {
8839 		schedstat_inc(p->stats.nr_failed_migrations_running);
8840 		return 0;
8841 	}
8842 
8843 	/*
8844 	 * Aggressive migration if:
8845 	 * 1) active balance
8846 	 * 2) destination numa is preferred
8847 	 * 3) task is cache cold, or
8848 	 * 4) too many balance attempts have failed.
8849 	 */
8850 	if (env->flags & LBF_ACTIVE_LB)
8851 		return 1;
8852 
8853 	tsk_cache_hot = migrate_degrades_locality(p, env);
8854 	if (tsk_cache_hot == -1)
8855 		tsk_cache_hot = task_hot(p, env);
8856 
8857 	if (tsk_cache_hot <= 0 ||
8858 	    env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
8859 		if (tsk_cache_hot == 1) {
8860 			schedstat_inc(env->sd->lb_hot_gained[env->idle]);
8861 			schedstat_inc(p->stats.nr_forced_migrations);
8862 		}
8863 		return 1;
8864 	}
8865 
8866 	schedstat_inc(p->stats.nr_failed_migrations_hot);
8867 	return 0;
8868 }
8869 
8870 /*
8871  * detach_task() -- detach the task for the migration specified in env
8872  */
detach_task(struct task_struct * p,struct lb_env * env)8873 static void detach_task(struct task_struct *p, struct lb_env *env)
8874 {
8875 	lockdep_assert_rq_held(env->src_rq);
8876 
8877 	deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
8878 	set_task_cpu(p, env->dst_cpu);
8879 }
8880 
8881 /*
8882  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
8883  * part of active balancing operations within "domain".
8884  *
8885  * Returns a task if successful and NULL otherwise.
8886  */
detach_one_task(struct lb_env * env)8887 static struct task_struct *detach_one_task(struct lb_env *env)
8888 {
8889 	struct task_struct *p;
8890 
8891 	lockdep_assert_rq_held(env->src_rq);
8892 
8893 	list_for_each_entry_reverse(p,
8894 			&env->src_rq->cfs_tasks, se.group_node) {
8895 		if (!can_migrate_task(p, env))
8896 			continue;
8897 
8898 		detach_task(p, env);
8899 
8900 		/*
8901 		 * Right now, this is only the second place where
8902 		 * lb_gained[env->idle] is updated (other is detach_tasks)
8903 		 * so we can safely collect stats here rather than
8904 		 * inside detach_tasks().
8905 		 */
8906 		schedstat_inc(env->sd->lb_gained[env->idle]);
8907 		return p;
8908 	}
8909 	return NULL;
8910 }
8911 
8912 /*
8913  * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
8914  * busiest_rq, as part of a balancing operation within domain "sd".
8915  *
8916  * Returns number of detached tasks if successful and 0 otherwise.
8917  */
detach_tasks(struct lb_env * env)8918 static int detach_tasks(struct lb_env *env)
8919 {
8920 	struct list_head *tasks = &env->src_rq->cfs_tasks;
8921 	unsigned long util, load;
8922 	struct task_struct *p;
8923 	int detached = 0;
8924 
8925 	lockdep_assert_rq_held(env->src_rq);
8926 
8927 	/*
8928 	 * Source run queue has been emptied by another CPU, clear
8929 	 * LBF_ALL_PINNED flag as we will not test any task.
8930 	 */
8931 	if (env->src_rq->nr_running <= 1) {
8932 		env->flags &= ~LBF_ALL_PINNED;
8933 		return 0;
8934 	}
8935 
8936 	if (env->imbalance <= 0)
8937 		return 0;
8938 
8939 	while (!list_empty(tasks)) {
8940 		/*
8941 		 * We don't want to steal all, otherwise we may be treated likewise,
8942 		 * which could at worst lead to a livelock crash.
8943 		 */
8944 		if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
8945 			break;
8946 
8947 		env->loop++;
8948 		/*
8949 		 * We've more or less seen every task there is, call it quits
8950 		 * unless we haven't found any movable task yet.
8951 		 */
8952 		if (env->loop > env->loop_max &&
8953 		    !(env->flags & LBF_ALL_PINNED))
8954 			break;
8955 
8956 		/* take a breather every nr_migrate tasks */
8957 		if (env->loop > env->loop_break) {
8958 			env->loop_break += SCHED_NR_MIGRATE_BREAK;
8959 			env->flags |= LBF_NEED_BREAK;
8960 			break;
8961 		}
8962 
8963 		p = list_last_entry(tasks, struct task_struct, se.group_node);
8964 
8965 		if (!can_migrate_task(p, env))
8966 			goto next;
8967 
8968 		switch (env->migration_type) {
8969 		case migrate_load:
8970 			/*
8971 			 * Depending of the number of CPUs and tasks and the
8972 			 * cgroup hierarchy, task_h_load() can return a null
8973 			 * value. Make sure that env->imbalance decreases
8974 			 * otherwise detach_tasks() will stop only after
8975 			 * detaching up to loop_max tasks.
8976 			 */
8977 			load = max_t(unsigned long, task_h_load(p), 1);
8978 
8979 			if (sched_feat(LB_MIN) &&
8980 			    load < 16 && !env->sd->nr_balance_failed)
8981 				goto next;
8982 
8983 			/*
8984 			 * Make sure that we don't migrate too much load.
8985 			 * Nevertheless, let relax the constraint if
8986 			 * scheduler fails to find a good waiting task to
8987 			 * migrate.
8988 			 */
8989 			if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
8990 				goto next;
8991 
8992 			env->imbalance -= load;
8993 			break;
8994 
8995 		case migrate_util:
8996 			util = task_util_est(p);
8997 
8998 			if (util > env->imbalance)
8999 				goto next;
9000 
9001 			env->imbalance -= util;
9002 			break;
9003 
9004 		case migrate_task:
9005 			env->imbalance--;
9006 			break;
9007 
9008 		case migrate_misfit:
9009 			/* This is not a misfit task */
9010 			if (task_fits_cpu(p, env->src_cpu))
9011 				goto next;
9012 
9013 			env->imbalance = 0;
9014 			break;
9015 		}
9016 
9017 		detach_task(p, env);
9018 		list_add(&p->se.group_node, &env->tasks);
9019 
9020 		detached++;
9021 
9022 #ifdef CONFIG_PREEMPTION
9023 		/*
9024 		 * NEWIDLE balancing is a source of latency, so preemptible
9025 		 * kernels will stop after the first task is detached to minimize
9026 		 * the critical section.
9027 		 */
9028 		if (env->idle == CPU_NEWLY_IDLE)
9029 			break;
9030 #endif
9031 
9032 		/*
9033 		 * We only want to steal up to the prescribed amount of
9034 		 * load/util/tasks.
9035 		 */
9036 		if (env->imbalance <= 0)
9037 			break;
9038 
9039 		continue;
9040 next:
9041 		list_move(&p->se.group_node, tasks);
9042 	}
9043 
9044 	/*
9045 	 * Right now, this is one of only two places we collect this stat
9046 	 * so we can safely collect detach_one_task() stats here rather
9047 	 * than inside detach_one_task().
9048 	 */
9049 	schedstat_add(env->sd->lb_gained[env->idle], detached);
9050 
9051 	return detached;
9052 }
9053 
9054 /*
9055  * attach_task() -- attach the task detached by detach_task() to its new rq.
9056  */
attach_task(struct rq * rq,struct task_struct * p)9057 static void attach_task(struct rq *rq, struct task_struct *p)
9058 {
9059 	lockdep_assert_rq_held(rq);
9060 
9061 	WARN_ON_ONCE(task_rq(p) != rq);
9062 	activate_task(rq, p, ENQUEUE_NOCLOCK);
9063 	check_preempt_curr(rq, p, 0);
9064 }
9065 
9066 /*
9067  * attach_one_task() -- attaches the task returned from detach_one_task() to
9068  * its new rq.
9069  */
attach_one_task(struct rq * rq,struct task_struct * p)9070 static void attach_one_task(struct rq *rq, struct task_struct *p)
9071 {
9072 	struct rq_flags rf;
9073 
9074 	rq_lock(rq, &rf);
9075 	update_rq_clock(rq);
9076 	attach_task(rq, p);
9077 	rq_unlock(rq, &rf);
9078 }
9079 
9080 /*
9081  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
9082  * new rq.
9083  */
attach_tasks(struct lb_env * env)9084 static void attach_tasks(struct lb_env *env)
9085 {
9086 	struct list_head *tasks = &env->tasks;
9087 	struct task_struct *p;
9088 	struct rq_flags rf;
9089 
9090 	rq_lock(env->dst_rq, &rf);
9091 	update_rq_clock(env->dst_rq);
9092 
9093 	while (!list_empty(tasks)) {
9094 		p = list_first_entry(tasks, struct task_struct, se.group_node);
9095 		list_del_init(&p->se.group_node);
9096 
9097 		attach_task(env->dst_rq, p);
9098 	}
9099 
9100 	rq_unlock(env->dst_rq, &rf);
9101 }
9102 
9103 #ifdef CONFIG_NO_HZ_COMMON
cfs_rq_has_blocked(struct cfs_rq * cfs_rq)9104 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
9105 {
9106 	if (cfs_rq->avg.load_avg)
9107 		return true;
9108 
9109 	if (cfs_rq->avg.util_avg)
9110 		return true;
9111 
9112 	return false;
9113 }
9114 
others_have_blocked(struct rq * rq)9115 static inline bool others_have_blocked(struct rq *rq)
9116 {
9117 	if (READ_ONCE(rq->avg_rt.util_avg))
9118 		return true;
9119 
9120 	if (READ_ONCE(rq->avg_dl.util_avg))
9121 		return true;
9122 
9123 	if (thermal_load_avg(rq))
9124 		return true;
9125 
9126 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
9127 	if (READ_ONCE(rq->avg_irq.util_avg))
9128 		return true;
9129 #endif
9130 
9131 	return false;
9132 }
9133 
update_blocked_load_tick(struct rq * rq)9134 static inline void update_blocked_load_tick(struct rq *rq)
9135 {
9136 	WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
9137 }
9138 
update_blocked_load_status(struct rq * rq,bool has_blocked)9139 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
9140 {
9141 	if (!has_blocked)
9142 		rq->has_blocked_load = 0;
9143 }
9144 #else
cfs_rq_has_blocked(struct cfs_rq * cfs_rq)9145 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
others_have_blocked(struct rq * rq)9146 static inline bool others_have_blocked(struct rq *rq) { return false; }
update_blocked_load_tick(struct rq * rq)9147 static inline void update_blocked_load_tick(struct rq *rq) {}
update_blocked_load_status(struct rq * rq,bool has_blocked)9148 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
9149 #endif
9150 
__update_blocked_others(struct rq * rq,bool * done)9151 static bool __update_blocked_others(struct rq *rq, bool *done)
9152 {
9153 	const struct sched_class *curr_class;
9154 	u64 now = rq_clock_pelt(rq);
9155 	unsigned long thermal_pressure;
9156 	bool decayed;
9157 
9158 	/*
9159 	 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
9160 	 * DL and IRQ signals have been updated before updating CFS.
9161 	 */
9162 	curr_class = rq->curr->sched_class;
9163 
9164 	thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
9165 
9166 	decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
9167 		  update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
9168 		  update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
9169 		  update_irq_load_avg(rq, 0);
9170 
9171 	if (others_have_blocked(rq))
9172 		*done = false;
9173 
9174 	return decayed;
9175 }
9176 
9177 #ifdef CONFIG_FAIR_GROUP_SCHED
9178 
__update_blocked_fair(struct rq * rq,bool * done)9179 static bool __update_blocked_fair(struct rq *rq, bool *done)
9180 {
9181 	struct cfs_rq *cfs_rq, *pos;
9182 	bool decayed = false;
9183 	int cpu = cpu_of(rq);
9184 
9185 	/*
9186 	 * Iterates the task_group tree in a bottom up fashion, see
9187 	 * list_add_leaf_cfs_rq() for details.
9188 	 */
9189 	for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
9190 		struct sched_entity *se;
9191 
9192 		if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
9193 			update_tg_load_avg(cfs_rq);
9194 
9195 			if (cfs_rq->nr_running == 0)
9196 				update_idle_cfs_rq_clock_pelt(cfs_rq);
9197 
9198 			if (cfs_rq == &rq->cfs)
9199 				decayed = true;
9200 		}
9201 
9202 		/* Propagate pending load changes to the parent, if any: */
9203 		se = cfs_rq->tg->se[cpu];
9204 		if (se && !skip_blocked_update(se))
9205 			update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9206 
9207 		/*
9208 		 * There can be a lot of idle CPU cgroups.  Don't let fully
9209 		 * decayed cfs_rqs linger on the list.
9210 		 */
9211 		if (cfs_rq_is_decayed(cfs_rq))
9212 			list_del_leaf_cfs_rq(cfs_rq);
9213 
9214 		/* Don't need periodic decay once load/util_avg are null */
9215 		if (cfs_rq_has_blocked(cfs_rq))
9216 			*done = false;
9217 	}
9218 
9219 	return decayed;
9220 }
9221 
9222 /*
9223  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
9224  * This needs to be done in a top-down fashion because the load of a child
9225  * group is a fraction of its parents load.
9226  */
update_cfs_rq_h_load(struct cfs_rq * cfs_rq)9227 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
9228 {
9229 	struct rq *rq = rq_of(cfs_rq);
9230 	struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
9231 	unsigned long now = jiffies;
9232 	unsigned long load;
9233 
9234 	if (cfs_rq->last_h_load_update == now)
9235 		return;
9236 
9237 	WRITE_ONCE(cfs_rq->h_load_next, NULL);
9238 	for_each_sched_entity(se) {
9239 		cfs_rq = cfs_rq_of(se);
9240 		WRITE_ONCE(cfs_rq->h_load_next, se);
9241 		if (cfs_rq->last_h_load_update == now)
9242 			break;
9243 	}
9244 
9245 	if (!se) {
9246 		cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
9247 		cfs_rq->last_h_load_update = now;
9248 	}
9249 
9250 	while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
9251 		load = cfs_rq->h_load;
9252 		load = div64_ul(load * se->avg.load_avg,
9253 			cfs_rq_load_avg(cfs_rq) + 1);
9254 		cfs_rq = group_cfs_rq(se);
9255 		cfs_rq->h_load = load;
9256 		cfs_rq->last_h_load_update = now;
9257 	}
9258 }
9259 
task_h_load(struct task_struct * p)9260 static unsigned long task_h_load(struct task_struct *p)
9261 {
9262 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
9263 
9264 	update_cfs_rq_h_load(cfs_rq);
9265 	return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
9266 			cfs_rq_load_avg(cfs_rq) + 1);
9267 }
9268 #else
__update_blocked_fair(struct rq * rq,bool * done)9269 static bool __update_blocked_fair(struct rq *rq, bool *done)
9270 {
9271 	struct cfs_rq *cfs_rq = &rq->cfs;
9272 	bool decayed;
9273 
9274 	decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
9275 	if (cfs_rq_has_blocked(cfs_rq))
9276 		*done = false;
9277 
9278 	return decayed;
9279 }
9280 
task_h_load(struct task_struct * p)9281 static unsigned long task_h_load(struct task_struct *p)
9282 {
9283 	return p->se.avg.load_avg;
9284 }
9285 #endif
9286 
update_blocked_averages(int cpu)9287 static void update_blocked_averages(int cpu)
9288 {
9289 	bool decayed = false, done = true;
9290 	struct rq *rq = cpu_rq(cpu);
9291 	struct rq_flags rf;
9292 
9293 	rq_lock_irqsave(rq, &rf);
9294 	update_blocked_load_tick(rq);
9295 	update_rq_clock(rq);
9296 
9297 	decayed |= __update_blocked_others(rq, &done);
9298 	decayed |= __update_blocked_fair(rq, &done);
9299 
9300 	update_blocked_load_status(rq, !done);
9301 	if (decayed)
9302 		cpufreq_update_util(rq, 0);
9303 	rq_unlock_irqrestore(rq, &rf);
9304 }
9305 
9306 /********** Helpers for find_busiest_group ************************/
9307 
9308 /*
9309  * sg_lb_stats - stats of a sched_group required for load_balancing
9310  */
9311 struct sg_lb_stats {
9312 	unsigned long avg_load; /*Avg load across the CPUs of the group */
9313 	unsigned long group_load; /* Total load over the CPUs of the group */
9314 	unsigned long group_capacity;
9315 	unsigned long group_util; /* Total utilization over the CPUs of the group */
9316 	unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
9317 	unsigned int sum_nr_running; /* Nr of tasks running in the group */
9318 	unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
9319 	unsigned int idle_cpus;
9320 	unsigned int group_weight;
9321 	enum group_type group_type;
9322 	unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
9323 	unsigned int group_smt_balance;  /* Task on busy SMT be moved */
9324 	unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
9325 #ifdef CONFIG_NUMA_BALANCING
9326 	unsigned int nr_numa_running;
9327 	unsigned int nr_preferred_running;
9328 #endif
9329 };
9330 
9331 /*
9332  * sd_lb_stats - Structure to store the statistics of a sched_domain
9333  *		 during load balancing.
9334  */
9335 struct sd_lb_stats {
9336 	struct sched_group *busiest;	/* Busiest group in this sd */
9337 	struct sched_group *local;	/* Local group in this sd */
9338 	unsigned long total_load;	/* Total load of all groups in sd */
9339 	unsigned long total_capacity;	/* Total capacity of all groups in sd */
9340 	unsigned long avg_load;	/* Average load across all groups in sd */
9341 	unsigned int prefer_sibling; /* tasks should go to sibling first */
9342 
9343 	struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
9344 	struct sg_lb_stats local_stat;	/* Statistics of the local group */
9345 };
9346 
init_sd_lb_stats(struct sd_lb_stats * sds)9347 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
9348 {
9349 	/*
9350 	 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
9351 	 * local_stat because update_sg_lb_stats() does a full clear/assignment.
9352 	 * We must however set busiest_stat::group_type and
9353 	 * busiest_stat::idle_cpus to the worst busiest group because
9354 	 * update_sd_pick_busiest() reads these before assignment.
9355 	 */
9356 	*sds = (struct sd_lb_stats){
9357 		.busiest = NULL,
9358 		.local = NULL,
9359 		.total_load = 0UL,
9360 		.total_capacity = 0UL,
9361 		.busiest_stat = {
9362 			.idle_cpus = UINT_MAX,
9363 			.group_type = group_has_spare,
9364 		},
9365 	};
9366 }
9367 
scale_rt_capacity(int cpu)9368 static unsigned long scale_rt_capacity(int cpu)
9369 {
9370 	struct rq *rq = cpu_rq(cpu);
9371 	unsigned long max = arch_scale_cpu_capacity(cpu);
9372 	unsigned long used, free;
9373 	unsigned long irq;
9374 
9375 	irq = cpu_util_irq(rq);
9376 
9377 	if (unlikely(irq >= max))
9378 		return 1;
9379 
9380 	/*
9381 	 * avg_rt.util_avg and avg_dl.util_avg track binary signals
9382 	 * (running and not running) with weights 0 and 1024 respectively.
9383 	 * avg_thermal.load_avg tracks thermal pressure and the weighted
9384 	 * average uses the actual delta max capacity(load).
9385 	 */
9386 	used = READ_ONCE(rq->avg_rt.util_avg);
9387 	used += READ_ONCE(rq->avg_dl.util_avg);
9388 	used += thermal_load_avg(rq);
9389 
9390 	if (unlikely(used >= max))
9391 		return 1;
9392 
9393 	free = max - used;
9394 
9395 	return scale_irq_capacity(free, irq, max);
9396 }
9397 
update_cpu_capacity(struct sched_domain * sd,int cpu)9398 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
9399 {
9400 	unsigned long capacity = scale_rt_capacity(cpu);
9401 	struct sched_group *sdg = sd->groups;
9402 
9403 	cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
9404 
9405 	if (!capacity)
9406 		capacity = 1;
9407 
9408 	cpu_rq(cpu)->cpu_capacity = capacity;
9409 	trace_sched_cpu_capacity_tp(cpu_rq(cpu));
9410 
9411 	sdg->sgc->capacity = capacity;
9412 	sdg->sgc->min_capacity = capacity;
9413 	sdg->sgc->max_capacity = capacity;
9414 }
9415 
update_group_capacity(struct sched_domain * sd,int cpu)9416 void update_group_capacity(struct sched_domain *sd, int cpu)
9417 {
9418 	struct sched_domain *child = sd->child;
9419 	struct sched_group *group, *sdg = sd->groups;
9420 	unsigned long capacity, min_capacity, max_capacity;
9421 	unsigned long interval;
9422 
9423 	interval = msecs_to_jiffies(sd->balance_interval);
9424 	interval = clamp(interval, 1UL, max_load_balance_interval);
9425 	sdg->sgc->next_update = jiffies + interval;
9426 
9427 	if (!child) {
9428 		update_cpu_capacity(sd, cpu);
9429 		return;
9430 	}
9431 
9432 	capacity = 0;
9433 	min_capacity = ULONG_MAX;
9434 	max_capacity = 0;
9435 
9436 	if (child->flags & SD_OVERLAP) {
9437 		/*
9438 		 * SD_OVERLAP domains cannot assume that child groups
9439 		 * span the current group.
9440 		 */
9441 
9442 		for_each_cpu(cpu, sched_group_span(sdg)) {
9443 			unsigned long cpu_cap = capacity_of(cpu);
9444 
9445 			capacity += cpu_cap;
9446 			min_capacity = min(cpu_cap, min_capacity);
9447 			max_capacity = max(cpu_cap, max_capacity);
9448 		}
9449 	} else  {
9450 		/*
9451 		 * !SD_OVERLAP domains can assume that child groups
9452 		 * span the current group.
9453 		 */
9454 
9455 		group = child->groups;
9456 		do {
9457 			struct sched_group_capacity *sgc = group->sgc;
9458 
9459 			capacity += sgc->capacity;
9460 			min_capacity = min(sgc->min_capacity, min_capacity);
9461 			max_capacity = max(sgc->max_capacity, max_capacity);
9462 			group = group->next;
9463 		} while (group != child->groups);
9464 	}
9465 
9466 	sdg->sgc->capacity = capacity;
9467 	sdg->sgc->min_capacity = min_capacity;
9468 	sdg->sgc->max_capacity = max_capacity;
9469 }
9470 
9471 /*
9472  * Check whether the capacity of the rq has been noticeably reduced by side
9473  * activity. The imbalance_pct is used for the threshold.
9474  * Return true is the capacity is reduced
9475  */
9476 static inline int
check_cpu_capacity(struct rq * rq,struct sched_domain * sd)9477 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
9478 {
9479 	return ((rq->cpu_capacity * sd->imbalance_pct) <
9480 				(rq->cpu_capacity_orig * 100));
9481 }
9482 
9483 /*
9484  * Check whether a rq has a misfit task and if it looks like we can actually
9485  * help that task: we can migrate the task to a CPU of higher capacity, or
9486  * the task's current CPU is heavily pressured.
9487  */
check_misfit_status(struct rq * rq,struct sched_domain * sd)9488 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
9489 {
9490 	return rq->misfit_task_load &&
9491 		(rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
9492 		 check_cpu_capacity(rq, sd));
9493 }
9494 
9495 /*
9496  * Group imbalance indicates (and tries to solve) the problem where balancing
9497  * groups is inadequate due to ->cpus_ptr constraints.
9498  *
9499  * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9500  * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9501  * Something like:
9502  *
9503  *	{ 0 1 2 3 } { 4 5 6 7 }
9504  *	        *     * * *
9505  *
9506  * If we were to balance group-wise we'd place two tasks in the first group and
9507  * two tasks in the second group. Clearly this is undesired as it will overload
9508  * cpu 3 and leave one of the CPUs in the second group unused.
9509  *
9510  * The current solution to this issue is detecting the skew in the first group
9511  * by noticing the lower domain failed to reach balance and had difficulty
9512  * moving tasks due to affinity constraints.
9513  *
9514  * When this is so detected; this group becomes a candidate for busiest; see
9515  * update_sd_pick_busiest(). And calculate_imbalance() and
9516  * find_busiest_group() avoid some of the usual balance conditions to allow it
9517  * to create an effective group imbalance.
9518  *
9519  * This is a somewhat tricky proposition since the next run might not find the
9520  * group imbalance and decide the groups need to be balanced again. A most
9521  * subtle and fragile situation.
9522  */
9523 
sg_imbalanced(struct sched_group * group)9524 static inline int sg_imbalanced(struct sched_group *group)
9525 {
9526 	return group->sgc->imbalance;
9527 }
9528 
9529 /*
9530  * group_has_capacity returns true if the group has spare capacity that could
9531  * be used by some tasks.
9532  * We consider that a group has spare capacity if the number of task is
9533  * smaller than the number of CPUs or if the utilization is lower than the
9534  * available capacity for CFS tasks.
9535  * For the latter, we use a threshold to stabilize the state, to take into
9536  * account the variance of the tasks' load and to return true if the available
9537  * capacity in meaningful for the load balancer.
9538  * As an example, an available capacity of 1% can appear but it doesn't make
9539  * any benefit for the load balance.
9540  */
9541 static inline bool
group_has_capacity(unsigned int imbalance_pct,struct sg_lb_stats * sgs)9542 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9543 {
9544 	if (sgs->sum_nr_running < sgs->group_weight)
9545 		return true;
9546 
9547 	if ((sgs->group_capacity * imbalance_pct) <
9548 			(sgs->group_runnable * 100))
9549 		return false;
9550 
9551 	if ((sgs->group_capacity * 100) >
9552 			(sgs->group_util * imbalance_pct))
9553 		return true;
9554 
9555 	return false;
9556 }
9557 
9558 /*
9559  *  group_is_overloaded returns true if the group has more tasks than it can
9560  *  handle.
9561  *  group_is_overloaded is not equals to !group_has_capacity because a group
9562  *  with the exact right number of tasks, has no more spare capacity but is not
9563  *  overloaded so both group_has_capacity and group_is_overloaded return
9564  *  false.
9565  */
9566 static inline bool
group_is_overloaded(unsigned int imbalance_pct,struct sg_lb_stats * sgs)9567 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9568 {
9569 	if (sgs->sum_nr_running <= sgs->group_weight)
9570 		return false;
9571 
9572 	if ((sgs->group_capacity * 100) <
9573 			(sgs->group_util * imbalance_pct))
9574 		return true;
9575 
9576 	if ((sgs->group_capacity * imbalance_pct) <
9577 			(sgs->group_runnable * 100))
9578 		return true;
9579 
9580 	return false;
9581 }
9582 
9583 static inline enum
group_classify(unsigned int imbalance_pct,struct sched_group * group,struct sg_lb_stats * sgs)9584 group_type group_classify(unsigned int imbalance_pct,
9585 			  struct sched_group *group,
9586 			  struct sg_lb_stats *sgs)
9587 {
9588 	if (group_is_overloaded(imbalance_pct, sgs))
9589 		return group_overloaded;
9590 
9591 	if (sg_imbalanced(group))
9592 		return group_imbalanced;
9593 
9594 	if (sgs->group_asym_packing)
9595 		return group_asym_packing;
9596 
9597 	if (sgs->group_smt_balance)
9598 		return group_smt_balance;
9599 
9600 	if (sgs->group_misfit_task_load)
9601 		return group_misfit_task;
9602 
9603 	if (!group_has_capacity(imbalance_pct, sgs))
9604 		return group_fully_busy;
9605 
9606 	return group_has_spare;
9607 }
9608 
9609 /**
9610  * sched_use_asym_prio - Check whether asym_packing priority must be used
9611  * @sd:		The scheduling domain of the load balancing
9612  * @cpu:	A CPU
9613  *
9614  * Always use CPU priority when balancing load between SMT siblings. When
9615  * balancing load between cores, it is not sufficient that @cpu is idle. Only
9616  * use CPU priority if the whole core is idle.
9617  *
9618  * Returns: True if the priority of @cpu must be followed. False otherwise.
9619  */
sched_use_asym_prio(struct sched_domain * sd,int cpu)9620 static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
9621 {
9622 	if (!sched_smt_active())
9623 		return true;
9624 
9625 	return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);
9626 }
9627 
9628 /**
9629  * sched_asym - Check if the destination CPU can do asym_packing load balance
9630  * @env:	The load balancing environment
9631  * @sds:	Load-balancing data with statistics of the local group
9632  * @sgs:	Load-balancing statistics of the candidate busiest group
9633  * @group:	The candidate busiest group
9634  *
9635  * @env::dst_cpu can do asym_packing if it has higher priority than the
9636  * preferred CPU of @group.
9637  *
9638  * SMT is a special case. If we are balancing load between cores, @env::dst_cpu
9639  * can do asym_packing balance only if all its SMT siblings are idle. Also, it
9640  * can only do it if @group is an SMT group and has exactly on busy CPU. Larger
9641  * imbalances in the number of CPUS are dealt with in find_busiest_group().
9642  *
9643  * If we are balancing load within an SMT core, or at DIE domain level, always
9644  * proceed.
9645  *
9646  * Return: true if @env::dst_cpu can do with asym_packing load balance. False
9647  * otherwise.
9648  */
9649 static inline bool
sched_asym(struct lb_env * env,struct sd_lb_stats * sds,struct sg_lb_stats * sgs,struct sched_group * group)9650 sched_asym(struct lb_env *env, struct sd_lb_stats *sds,  struct sg_lb_stats *sgs,
9651 	   struct sched_group *group)
9652 {
9653 	/* Ensure that the whole local core is idle, if applicable. */
9654 	if (!sched_use_asym_prio(env->sd, env->dst_cpu))
9655 		return false;
9656 
9657 	/*
9658 	 * CPU priorities does not make sense for SMT cores with more than one
9659 	 * busy sibling.
9660 	 */
9661 	if (group->flags & SD_SHARE_CPUCAPACITY) {
9662 		if (sgs->group_weight - sgs->idle_cpus != 1)
9663 			return false;
9664 	}
9665 
9666 	return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
9667 }
9668 
9669 /* 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)9670 static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1,
9671 				    struct sched_group *sg2)
9672 {
9673 	if (!sg1 || !sg2)
9674 		return false;
9675 
9676 	return (sg1->flags & SD_SHARE_CPUCAPACITY) !=
9677 		(sg2->flags & SD_SHARE_CPUCAPACITY);
9678 }
9679 
smt_balance(struct lb_env * env,struct sg_lb_stats * sgs,struct sched_group * group)9680 static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs,
9681 			       struct sched_group *group)
9682 {
9683 	if (env->idle == CPU_NOT_IDLE)
9684 		return false;
9685 
9686 	/*
9687 	 * For SMT source group, it is better to move a task
9688 	 * to a CPU that doesn't have multiple tasks sharing its CPU capacity.
9689 	 * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY
9690 	 * will not be on.
9691 	 */
9692 	if (group->flags & SD_SHARE_CPUCAPACITY &&
9693 	    sgs->sum_h_nr_running > 1)
9694 		return true;
9695 
9696 	return false;
9697 }
9698 
sibling_imbalance(struct lb_env * env,struct sd_lb_stats * sds,struct sg_lb_stats * busiest,struct sg_lb_stats * local)9699 static inline long sibling_imbalance(struct lb_env *env,
9700 				    struct sd_lb_stats *sds,
9701 				    struct sg_lb_stats *busiest,
9702 				    struct sg_lb_stats *local)
9703 {
9704 	int ncores_busiest, ncores_local;
9705 	long imbalance;
9706 
9707 	if (env->idle == CPU_NOT_IDLE || !busiest->sum_nr_running)
9708 		return 0;
9709 
9710 	ncores_busiest = sds->busiest->cores;
9711 	ncores_local = sds->local->cores;
9712 
9713 	if (ncores_busiest == ncores_local) {
9714 		imbalance = busiest->sum_nr_running;
9715 		lsub_positive(&imbalance, local->sum_nr_running);
9716 		return imbalance;
9717 	}
9718 
9719 	/* Balance such that nr_running/ncores ratio are same on both groups */
9720 	imbalance = ncores_local * busiest->sum_nr_running;
9721 	lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);
9722 	/* Normalize imbalance and do rounding on normalization */
9723 	imbalance = 2 * imbalance + ncores_local + ncores_busiest;
9724 	imbalance /= ncores_local + ncores_busiest;
9725 
9726 	/* Take advantage of resource in an empty sched group */
9727 	if (imbalance <= 1 && local->sum_nr_running == 0 &&
9728 	    busiest->sum_nr_running > 1)
9729 		imbalance = 2;
9730 
9731 	return imbalance;
9732 }
9733 
9734 static inline bool
sched_reduced_capacity(struct rq * rq,struct sched_domain * sd)9735 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9736 {
9737 	/*
9738 	 * When there is more than 1 task, the group_overloaded case already
9739 	 * takes care of cpu with reduced capacity
9740 	 */
9741 	if (rq->cfs.h_nr_running != 1)
9742 		return false;
9743 
9744 	return check_cpu_capacity(rq, sd);
9745 }
9746 
9747 /**
9748  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9749  * @env: The load balancing environment.
9750  * @sds: Load-balancing data with statistics of the local group.
9751  * @group: sched_group whose statistics are to be updated.
9752  * @sgs: variable to hold the statistics for this group.
9753  * @sg_status: Holds flag indicating the status of the sched_group
9754  */
update_sg_lb_stats(struct lb_env * env,struct sd_lb_stats * sds,struct sched_group * group,struct sg_lb_stats * sgs,int * sg_status)9755 static inline void update_sg_lb_stats(struct lb_env *env,
9756 				      struct sd_lb_stats *sds,
9757 				      struct sched_group *group,
9758 				      struct sg_lb_stats *sgs,
9759 				      int *sg_status)
9760 {
9761 	int i, nr_running, local_group;
9762 
9763 	memset(sgs, 0, sizeof(*sgs));
9764 
9765 	local_group = group == sds->local;
9766 
9767 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9768 		struct rq *rq = cpu_rq(i);
9769 		unsigned long load = cpu_load(rq);
9770 
9771 		sgs->group_load += load;
9772 		sgs->group_util += cpu_util_cfs(i);
9773 		sgs->group_runnable += cpu_runnable(rq);
9774 		sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9775 
9776 		nr_running = rq->nr_running;
9777 		sgs->sum_nr_running += nr_running;
9778 
9779 		if (nr_running > 1)
9780 			*sg_status |= SG_OVERLOAD;
9781 
9782 		if (cpu_overutilized(i))
9783 			*sg_status |= SG_OVERUTILIZED;
9784 
9785 #ifdef CONFIG_NUMA_BALANCING
9786 		sgs->nr_numa_running += rq->nr_numa_running;
9787 		sgs->nr_preferred_running += rq->nr_preferred_running;
9788 #endif
9789 		/*
9790 		 * No need to call idle_cpu() if nr_running is not 0
9791 		 */
9792 		if (!nr_running && idle_cpu(i)) {
9793 			sgs->idle_cpus++;
9794 			/* Idle cpu can't have misfit task */
9795 			continue;
9796 		}
9797 
9798 		if (local_group)
9799 			continue;
9800 
9801 		if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9802 			/* Check for a misfit task on the cpu */
9803 			if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9804 				sgs->group_misfit_task_load = rq->misfit_task_load;
9805 				*sg_status |= SG_OVERLOAD;
9806 			}
9807 		} else if ((env->idle != CPU_NOT_IDLE) &&
9808 			   sched_reduced_capacity(rq, env->sd)) {
9809 			/* Check for a task running on a CPU with reduced capacity */
9810 			if (sgs->group_misfit_task_load < load)
9811 				sgs->group_misfit_task_load = load;
9812 		}
9813 	}
9814 
9815 	sgs->group_capacity = group->sgc->capacity;
9816 
9817 	sgs->group_weight = group->group_weight;
9818 
9819 	/* Check if dst CPU is idle and preferred to this group */
9820 	if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
9821 	    env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9822 	    sched_asym(env, sds, sgs, group)) {
9823 		sgs->group_asym_packing = 1;
9824 	}
9825 
9826 	/* Check for loaded SMT group to be balanced to dst CPU */
9827 	if (!local_group && smt_balance(env, sgs, group))
9828 		sgs->group_smt_balance = 1;
9829 
9830 	sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9831 
9832 	/* Computing avg_load makes sense only when group is overloaded */
9833 	if (sgs->group_type == group_overloaded)
9834 		sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9835 				sgs->group_capacity;
9836 }
9837 
9838 /**
9839  * update_sd_pick_busiest - return 1 on busiest group
9840  * @env: The load balancing environment.
9841  * @sds: sched_domain statistics
9842  * @sg: sched_group candidate to be checked for being the busiest
9843  * @sgs: sched_group statistics
9844  *
9845  * Determine if @sg is a busier group than the previously selected
9846  * busiest group.
9847  *
9848  * Return: %true if @sg is a busier group than the previously selected
9849  * busiest group. %false otherwise.
9850  */
update_sd_pick_busiest(struct lb_env * env,struct sd_lb_stats * sds,struct sched_group * sg,struct sg_lb_stats * sgs)9851 static bool update_sd_pick_busiest(struct lb_env *env,
9852 				   struct sd_lb_stats *sds,
9853 				   struct sched_group *sg,
9854 				   struct sg_lb_stats *sgs)
9855 {
9856 	struct sg_lb_stats *busiest = &sds->busiest_stat;
9857 
9858 	/* Make sure that there is at least one task to pull */
9859 	if (!sgs->sum_h_nr_running)
9860 		return false;
9861 
9862 	/*
9863 	 * Don't try to pull misfit tasks we can't help.
9864 	 * We can use max_capacity here as reduction in capacity on some
9865 	 * CPUs in the group should either be possible to resolve
9866 	 * internally or be covered by avg_load imbalance (eventually).
9867 	 */
9868 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9869 	    (sgs->group_type == group_misfit_task) &&
9870 	    (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
9871 	     sds->local_stat.group_type != group_has_spare))
9872 		return false;
9873 
9874 	if (sgs->group_type > busiest->group_type)
9875 		return true;
9876 
9877 	if (sgs->group_type < busiest->group_type)
9878 		return false;
9879 
9880 	/*
9881 	 * The candidate and the current busiest group are the same type of
9882 	 * group. Let check which one is the busiest according to the type.
9883 	 */
9884 
9885 	switch (sgs->group_type) {
9886 	case group_overloaded:
9887 		/* Select the overloaded group with highest avg_load. */
9888 		if (sgs->avg_load <= busiest->avg_load)
9889 			return false;
9890 		break;
9891 
9892 	case group_imbalanced:
9893 		/*
9894 		 * Select the 1st imbalanced group as we don't have any way to
9895 		 * choose one more than another.
9896 		 */
9897 		return false;
9898 
9899 	case group_asym_packing:
9900 		/* Prefer to move from lowest priority CPU's work */
9901 		if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
9902 			return false;
9903 		break;
9904 
9905 	case group_misfit_task:
9906 		/*
9907 		 * If we have more than one misfit sg go with the biggest
9908 		 * misfit.
9909 		 */
9910 		if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
9911 			return false;
9912 		break;
9913 
9914 	case group_smt_balance:
9915 		/*
9916 		 * Check if we have spare CPUs on either SMT group to
9917 		 * choose has spare or fully busy handling.
9918 		 */
9919 		if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)
9920 			goto has_spare;
9921 
9922 		fallthrough;
9923 
9924 	case group_fully_busy:
9925 		/*
9926 		 * Select the fully busy group with highest avg_load. In
9927 		 * theory, there is no need to pull task from such kind of
9928 		 * group because tasks have all compute capacity that they need
9929 		 * but we can still improve the overall throughput by reducing
9930 		 * contention when accessing shared HW resources.
9931 		 *
9932 		 * XXX for now avg_load is not computed and always 0 so we
9933 		 * select the 1st one, except if @sg is composed of SMT
9934 		 * siblings.
9935 		 */
9936 
9937 		if (sgs->avg_load < busiest->avg_load)
9938 			return false;
9939 
9940 		if (sgs->avg_load == busiest->avg_load) {
9941 			/*
9942 			 * SMT sched groups need more help than non-SMT groups.
9943 			 * If @sg happens to also be SMT, either choice is good.
9944 			 */
9945 			if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)
9946 				return false;
9947 		}
9948 
9949 		break;
9950 
9951 	case group_has_spare:
9952 		/*
9953 		 * Do not pick sg with SMT CPUs over sg with pure CPUs,
9954 		 * as we do not want to pull task off SMT core with one task
9955 		 * and make the core idle.
9956 		 */
9957 		if (smt_vs_nonsmt_groups(sds->busiest, sg)) {
9958 			if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)
9959 				return false;
9960 			else
9961 				return true;
9962 		}
9963 has_spare:
9964 
9965 		/*
9966 		 * Select not overloaded group with lowest number of idle cpus
9967 		 * and highest number of running tasks. We could also compare
9968 		 * the spare capacity which is more stable but it can end up
9969 		 * that the group has less spare capacity but finally more idle
9970 		 * CPUs which means less opportunity to pull tasks.
9971 		 */
9972 		if (sgs->idle_cpus > busiest->idle_cpus)
9973 			return false;
9974 		else if ((sgs->idle_cpus == busiest->idle_cpus) &&
9975 			 (sgs->sum_nr_running <= busiest->sum_nr_running))
9976 			return false;
9977 
9978 		break;
9979 	}
9980 
9981 	/*
9982 	 * Candidate sg has no more than one task per CPU and has higher
9983 	 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
9984 	 * throughput. Maximize throughput, power/energy consequences are not
9985 	 * considered.
9986 	 */
9987 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9988 	    (sgs->group_type <= group_fully_busy) &&
9989 	    (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
9990 		return false;
9991 
9992 	return true;
9993 }
9994 
9995 #ifdef CONFIG_NUMA_BALANCING
fbq_classify_group(struct sg_lb_stats * sgs)9996 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9997 {
9998 	if (sgs->sum_h_nr_running > sgs->nr_numa_running)
9999 		return regular;
10000 	if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
10001 		return remote;
10002 	return all;
10003 }
10004 
fbq_classify_rq(struct rq * rq)10005 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10006 {
10007 	if (rq->nr_running > rq->nr_numa_running)
10008 		return regular;
10009 	if (rq->nr_running > rq->nr_preferred_running)
10010 		return remote;
10011 	return all;
10012 }
10013 #else
fbq_classify_group(struct sg_lb_stats * sgs)10014 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10015 {
10016 	return all;
10017 }
10018 
fbq_classify_rq(struct rq * rq)10019 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10020 {
10021 	return regular;
10022 }
10023 #endif /* CONFIG_NUMA_BALANCING */
10024 
10025 
10026 struct sg_lb_stats;
10027 
10028 /*
10029  * task_running_on_cpu - return 1 if @p is running on @cpu.
10030  */
10031 
task_running_on_cpu(int cpu,struct task_struct * p)10032 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
10033 {
10034 	/* Task has no contribution or is new */
10035 	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
10036 		return 0;
10037 
10038 	if (task_on_rq_queued(p))
10039 		return 1;
10040 
10041 	return 0;
10042 }
10043 
10044 /**
10045  * idle_cpu_without - would a given CPU be idle without p ?
10046  * @cpu: the processor on which idleness is tested.
10047  * @p: task which should be ignored.
10048  *
10049  * Return: 1 if the CPU would be idle. 0 otherwise.
10050  */
idle_cpu_without(int cpu,struct task_struct * p)10051 static int idle_cpu_without(int cpu, struct task_struct *p)
10052 {
10053 	struct rq *rq = cpu_rq(cpu);
10054 
10055 	if (rq->curr != rq->idle && rq->curr != p)
10056 		return 0;
10057 
10058 	/*
10059 	 * rq->nr_running can't be used but an updated version without the
10060 	 * impact of p on cpu must be used instead. The updated nr_running
10061 	 * be computed and tested before calling idle_cpu_without().
10062 	 */
10063 
10064 #ifdef CONFIG_SMP
10065 	if (rq->ttwu_pending)
10066 		return 0;
10067 #endif
10068 
10069 	return 1;
10070 }
10071 
10072 /*
10073  * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
10074  * @sd: The sched_domain level to look for idlest group.
10075  * @group: sched_group whose statistics are to be updated.
10076  * @sgs: variable to hold the statistics for this group.
10077  * @p: The task for which we look for the idlest group/CPU.
10078  */
update_sg_wakeup_stats(struct sched_domain * sd,struct sched_group * group,struct sg_lb_stats * sgs,struct task_struct * p)10079 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
10080 					  struct sched_group *group,
10081 					  struct sg_lb_stats *sgs,
10082 					  struct task_struct *p)
10083 {
10084 	int i, nr_running;
10085 
10086 	memset(sgs, 0, sizeof(*sgs));
10087 
10088 	/* Assume that task can't fit any CPU of the group */
10089 	if (sd->flags & SD_ASYM_CPUCAPACITY)
10090 		sgs->group_misfit_task_load = 1;
10091 
10092 	for_each_cpu(i, sched_group_span(group)) {
10093 		struct rq *rq = cpu_rq(i);
10094 		unsigned int local;
10095 
10096 		sgs->group_load += cpu_load_without(rq, p);
10097 		sgs->group_util += cpu_util_without(i, p);
10098 		sgs->group_runnable += cpu_runnable_without(rq, p);
10099 		local = task_running_on_cpu(i, p);
10100 		sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
10101 
10102 		nr_running = rq->nr_running - local;
10103 		sgs->sum_nr_running += nr_running;
10104 
10105 		/*
10106 		 * No need to call idle_cpu_without() if nr_running is not 0
10107 		 */
10108 		if (!nr_running && idle_cpu_without(i, p))
10109 			sgs->idle_cpus++;
10110 
10111 		/* Check if task fits in the CPU */
10112 		if (sd->flags & SD_ASYM_CPUCAPACITY &&
10113 		    sgs->group_misfit_task_load &&
10114 		    task_fits_cpu(p, i))
10115 			sgs->group_misfit_task_load = 0;
10116 
10117 	}
10118 
10119 	sgs->group_capacity = group->sgc->capacity;
10120 
10121 	sgs->group_weight = group->group_weight;
10122 
10123 	sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
10124 
10125 	/*
10126 	 * Computing avg_load makes sense only when group is fully busy or
10127 	 * overloaded
10128 	 */
10129 	if (sgs->group_type == group_fully_busy ||
10130 		sgs->group_type == group_overloaded)
10131 		sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10132 				sgs->group_capacity;
10133 }
10134 
update_pick_idlest(struct sched_group * idlest,struct sg_lb_stats * idlest_sgs,struct sched_group * group,struct sg_lb_stats * sgs)10135 static bool update_pick_idlest(struct sched_group *idlest,
10136 			       struct sg_lb_stats *idlest_sgs,
10137 			       struct sched_group *group,
10138 			       struct sg_lb_stats *sgs)
10139 {
10140 	if (sgs->group_type < idlest_sgs->group_type)
10141 		return true;
10142 
10143 	if (sgs->group_type > idlest_sgs->group_type)
10144 		return false;
10145 
10146 	/*
10147 	 * The candidate and the current idlest group are the same type of
10148 	 * group. Let check which one is the idlest according to the type.
10149 	 */
10150 
10151 	switch (sgs->group_type) {
10152 	case group_overloaded:
10153 	case group_fully_busy:
10154 		/* Select the group with lowest avg_load. */
10155 		if (idlest_sgs->avg_load <= sgs->avg_load)
10156 			return false;
10157 		break;
10158 
10159 	case group_imbalanced:
10160 	case group_asym_packing:
10161 	case group_smt_balance:
10162 		/* Those types are not used in the slow wakeup path */
10163 		return false;
10164 
10165 	case group_misfit_task:
10166 		/* Select group with the highest max capacity */
10167 		if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
10168 			return false;
10169 		break;
10170 
10171 	case group_has_spare:
10172 		/* Select group with most idle CPUs */
10173 		if (idlest_sgs->idle_cpus > sgs->idle_cpus)
10174 			return false;
10175 
10176 		/* Select group with lowest group_util */
10177 		if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
10178 			idlest_sgs->group_util <= sgs->group_util)
10179 			return false;
10180 
10181 		break;
10182 	}
10183 
10184 	return true;
10185 }
10186 
10187 /*
10188  * find_idlest_group() finds and returns the least busy CPU group within the
10189  * domain.
10190  *
10191  * Assumes p is allowed on at least one CPU in sd.
10192  */
10193 static struct sched_group *
find_idlest_group(struct sched_domain * sd,struct task_struct * p,int this_cpu)10194 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
10195 {
10196 	struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
10197 	struct sg_lb_stats local_sgs, tmp_sgs;
10198 	struct sg_lb_stats *sgs;
10199 	unsigned long imbalance;
10200 	struct sg_lb_stats idlest_sgs = {
10201 			.avg_load = UINT_MAX,
10202 			.group_type = group_overloaded,
10203 	};
10204 
10205 	do {
10206 		int local_group;
10207 
10208 		/* Skip over this group if it has no CPUs allowed */
10209 		if (!cpumask_intersects(sched_group_span(group),
10210 					p->cpus_ptr))
10211 			continue;
10212 
10213 		/* Skip over this group if no cookie matched */
10214 		if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
10215 			continue;
10216 
10217 		local_group = cpumask_test_cpu(this_cpu,
10218 					       sched_group_span(group));
10219 
10220 		if (local_group) {
10221 			sgs = &local_sgs;
10222 			local = group;
10223 		} else {
10224 			sgs = &tmp_sgs;
10225 		}
10226 
10227 		update_sg_wakeup_stats(sd, group, sgs, p);
10228 
10229 		if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
10230 			idlest = group;
10231 			idlest_sgs = *sgs;
10232 		}
10233 
10234 	} while (group = group->next, group != sd->groups);
10235 
10236 
10237 	/* There is no idlest group to push tasks to */
10238 	if (!idlest)
10239 		return NULL;
10240 
10241 	/* The local group has been skipped because of CPU affinity */
10242 	if (!local)
10243 		return idlest;
10244 
10245 	/*
10246 	 * If the local group is idler than the selected idlest group
10247 	 * don't try and push the task.
10248 	 */
10249 	if (local_sgs.group_type < idlest_sgs.group_type)
10250 		return NULL;
10251 
10252 	/*
10253 	 * If the local group is busier than the selected idlest group
10254 	 * try and push the task.
10255 	 */
10256 	if (local_sgs.group_type > idlest_sgs.group_type)
10257 		return idlest;
10258 
10259 	switch (local_sgs.group_type) {
10260 	case group_overloaded:
10261 	case group_fully_busy:
10262 
10263 		/* Calculate allowed imbalance based on load */
10264 		imbalance = scale_load_down(NICE_0_LOAD) *
10265 				(sd->imbalance_pct-100) / 100;
10266 
10267 		/*
10268 		 * When comparing groups across NUMA domains, it's possible for
10269 		 * the local domain to be very lightly loaded relative to the
10270 		 * remote domains but "imbalance" skews the comparison making
10271 		 * remote CPUs look much more favourable. When considering
10272 		 * cross-domain, add imbalance to the load on the remote node
10273 		 * and consider staying local.
10274 		 */
10275 
10276 		if ((sd->flags & SD_NUMA) &&
10277 		    ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
10278 			return NULL;
10279 
10280 		/*
10281 		 * If the local group is less loaded than the selected
10282 		 * idlest group don't try and push any tasks.
10283 		 */
10284 		if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
10285 			return NULL;
10286 
10287 		if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
10288 			return NULL;
10289 		break;
10290 
10291 	case group_imbalanced:
10292 	case group_asym_packing:
10293 	case group_smt_balance:
10294 		/* Those type are not used in the slow wakeup path */
10295 		return NULL;
10296 
10297 	case group_misfit_task:
10298 		/* Select group with the highest max capacity */
10299 		if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
10300 			return NULL;
10301 		break;
10302 
10303 	case group_has_spare:
10304 #ifdef CONFIG_NUMA
10305 		if (sd->flags & SD_NUMA) {
10306 			int imb_numa_nr = sd->imb_numa_nr;
10307 #ifdef CONFIG_NUMA_BALANCING
10308 			int idlest_cpu;
10309 			/*
10310 			 * If there is spare capacity at NUMA, try to select
10311 			 * the preferred node
10312 			 */
10313 			if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
10314 				return NULL;
10315 
10316 			idlest_cpu = cpumask_first(sched_group_span(idlest));
10317 			if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
10318 				return idlest;
10319 #endif /* CONFIG_NUMA_BALANCING */
10320 			/*
10321 			 * Otherwise, keep the task close to the wakeup source
10322 			 * and improve locality if the number of running tasks
10323 			 * would remain below threshold where an imbalance is
10324 			 * allowed while accounting for the possibility the
10325 			 * task is pinned to a subset of CPUs. If there is a
10326 			 * real need of migration, periodic load balance will
10327 			 * take care of it.
10328 			 */
10329 			if (p->nr_cpus_allowed != NR_CPUS) {
10330 				struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
10331 
10332 				cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
10333 				imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
10334 			}
10335 
10336 			imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
10337 			if (!adjust_numa_imbalance(imbalance,
10338 						   local_sgs.sum_nr_running + 1,
10339 						   imb_numa_nr)) {
10340 				return NULL;
10341 			}
10342 		}
10343 #endif /* CONFIG_NUMA */
10344 
10345 		/*
10346 		 * Select group with highest number of idle CPUs. We could also
10347 		 * compare the utilization which is more stable but it can end
10348 		 * up that the group has less spare capacity but finally more
10349 		 * idle CPUs which means more opportunity to run task.
10350 		 */
10351 		if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
10352 			return NULL;
10353 		break;
10354 	}
10355 
10356 	return idlest;
10357 }
10358 
update_idle_cpu_scan(struct lb_env * env,unsigned long sum_util)10359 static void update_idle_cpu_scan(struct lb_env *env,
10360 				 unsigned long sum_util)
10361 {
10362 	struct sched_domain_shared *sd_share;
10363 	int llc_weight, pct;
10364 	u64 x, y, tmp;
10365 	/*
10366 	 * Update the number of CPUs to scan in LLC domain, which could
10367 	 * be used as a hint in select_idle_cpu(). The update of sd_share
10368 	 * could be expensive because it is within a shared cache line.
10369 	 * So the write of this hint only occurs during periodic load
10370 	 * balancing, rather than CPU_NEWLY_IDLE, because the latter
10371 	 * can fire way more frequently than the former.
10372 	 */
10373 	if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
10374 		return;
10375 
10376 	llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
10377 	if (env->sd->span_weight != llc_weight)
10378 		return;
10379 
10380 	sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
10381 	if (!sd_share)
10382 		return;
10383 
10384 	/*
10385 	 * The number of CPUs to search drops as sum_util increases, when
10386 	 * sum_util hits 85% or above, the scan stops.
10387 	 * The reason to choose 85% as the threshold is because this is the
10388 	 * imbalance_pct(117) when a LLC sched group is overloaded.
10389 	 *
10390 	 * let y = SCHED_CAPACITY_SCALE - p * x^2                       [1]
10391 	 * and y'= y / SCHED_CAPACITY_SCALE
10392 	 *
10393 	 * x is the ratio of sum_util compared to the CPU capacity:
10394 	 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
10395 	 * y' is the ratio of CPUs to be scanned in the LLC domain,
10396 	 * and the number of CPUs to scan is calculated by:
10397 	 *
10398 	 * nr_scan = llc_weight * y'                                    [2]
10399 	 *
10400 	 * When x hits the threshold of overloaded, AKA, when
10401 	 * x = 100 / pct, y drops to 0. According to [1],
10402 	 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
10403 	 *
10404 	 * Scale x by SCHED_CAPACITY_SCALE:
10405 	 * x' = sum_util / llc_weight;                                  [3]
10406 	 *
10407 	 * and finally [1] becomes:
10408 	 * y = SCHED_CAPACITY_SCALE -
10409 	 *     x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE)            [4]
10410 	 *
10411 	 */
10412 	/* equation [3] */
10413 	x = sum_util;
10414 	do_div(x, llc_weight);
10415 
10416 	/* equation [4] */
10417 	pct = env->sd->imbalance_pct;
10418 	tmp = x * x * pct * pct;
10419 	do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
10420 	tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
10421 	y = SCHED_CAPACITY_SCALE - tmp;
10422 
10423 	/* equation [2] */
10424 	y *= llc_weight;
10425 	do_div(y, SCHED_CAPACITY_SCALE);
10426 	if ((int)y != sd_share->nr_idle_scan)
10427 		WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
10428 }
10429 
10430 /**
10431  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
10432  * @env: The load balancing environment.
10433  * @sds: variable to hold the statistics for this sched_domain.
10434  */
10435 
update_sd_lb_stats(struct lb_env * env,struct sd_lb_stats * sds)10436 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
10437 {
10438 	struct sched_group *sg = env->sd->groups;
10439 	struct sg_lb_stats *local = &sds->local_stat;
10440 	struct sg_lb_stats tmp_sgs;
10441 	unsigned long sum_util = 0;
10442 	int sg_status = 0;
10443 
10444 	do {
10445 		struct sg_lb_stats *sgs = &tmp_sgs;
10446 		int local_group;
10447 
10448 		local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
10449 		if (local_group) {
10450 			sds->local = sg;
10451 			sgs = local;
10452 
10453 			if (env->idle != CPU_NEWLY_IDLE ||
10454 			    time_after_eq(jiffies, sg->sgc->next_update))
10455 				update_group_capacity(env->sd, env->dst_cpu);
10456 		}
10457 
10458 		update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
10459 
10460 		if (local_group)
10461 			goto next_group;
10462 
10463 
10464 		if (update_sd_pick_busiest(env, sds, sg, sgs)) {
10465 			sds->busiest = sg;
10466 			sds->busiest_stat = *sgs;
10467 		}
10468 
10469 next_group:
10470 		/* Now, start updating sd_lb_stats */
10471 		sds->total_load += sgs->group_load;
10472 		sds->total_capacity += sgs->group_capacity;
10473 
10474 		sum_util += sgs->group_util;
10475 		sg = sg->next;
10476 	} while (sg != env->sd->groups);
10477 
10478 	/*
10479 	 * Indicate that the child domain of the busiest group prefers tasks
10480 	 * go to a child's sibling domains first. NB the flags of a sched group
10481 	 * are those of the child domain.
10482 	 */
10483 	if (sds->busiest)
10484 		sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);
10485 
10486 
10487 	if (env->sd->flags & SD_NUMA)
10488 		env->fbq_type = fbq_classify_group(&sds->busiest_stat);
10489 
10490 	if (!env->sd->parent) {
10491 		struct root_domain *rd = env->dst_rq->rd;
10492 
10493 		/* update overload indicator if we are at root domain */
10494 		WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
10495 
10496 		/* Update over-utilization (tipping point, U >= 0) indicator */
10497 		WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
10498 		trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
10499 	} else if (sg_status & SG_OVERUTILIZED) {
10500 		struct root_domain *rd = env->dst_rq->rd;
10501 
10502 		WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
10503 		trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
10504 	}
10505 
10506 	update_idle_cpu_scan(env, sum_util);
10507 }
10508 
10509 /**
10510  * calculate_imbalance - Calculate the amount of imbalance present within the
10511  *			 groups of a given sched_domain during load balance.
10512  * @env: load balance environment
10513  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
10514  */
calculate_imbalance(struct lb_env * env,struct sd_lb_stats * sds)10515 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
10516 {
10517 	struct sg_lb_stats *local, *busiest;
10518 
10519 	local = &sds->local_stat;
10520 	busiest = &sds->busiest_stat;
10521 
10522 	if (busiest->group_type == group_misfit_task) {
10523 		if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
10524 			/* Set imbalance to allow misfit tasks to be balanced. */
10525 			env->migration_type = migrate_misfit;
10526 			env->imbalance = 1;
10527 		} else {
10528 			/*
10529 			 * Set load imbalance to allow moving task from cpu
10530 			 * with reduced capacity.
10531 			 */
10532 			env->migration_type = migrate_load;
10533 			env->imbalance = busiest->group_misfit_task_load;
10534 		}
10535 		return;
10536 	}
10537 
10538 	if (busiest->group_type == group_asym_packing) {
10539 		/*
10540 		 * In case of asym capacity, we will try to migrate all load to
10541 		 * the preferred CPU.
10542 		 */
10543 		env->migration_type = migrate_task;
10544 		env->imbalance = busiest->sum_h_nr_running;
10545 		return;
10546 	}
10547 
10548 	if (busiest->group_type == group_smt_balance) {
10549 		/* Reduce number of tasks sharing CPU capacity */
10550 		env->migration_type = migrate_task;
10551 		env->imbalance = 1;
10552 		return;
10553 	}
10554 
10555 	if (busiest->group_type == group_imbalanced) {
10556 		/*
10557 		 * In the group_imb case we cannot rely on group-wide averages
10558 		 * to ensure CPU-load equilibrium, try to move any task to fix
10559 		 * the imbalance. The next load balance will take care of
10560 		 * balancing back the system.
10561 		 */
10562 		env->migration_type = migrate_task;
10563 		env->imbalance = 1;
10564 		return;
10565 	}
10566 
10567 	/*
10568 	 * Try to use spare capacity of local group without overloading it or
10569 	 * emptying busiest.
10570 	 */
10571 	if (local->group_type == group_has_spare) {
10572 		if ((busiest->group_type > group_fully_busy) &&
10573 		    !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
10574 			/*
10575 			 * If busiest is overloaded, try to fill spare
10576 			 * capacity. This might end up creating spare capacity
10577 			 * in busiest or busiest still being overloaded but
10578 			 * there is no simple way to directly compute the
10579 			 * amount of load to migrate in order to balance the
10580 			 * system.
10581 			 */
10582 			env->migration_type = migrate_util;
10583 			env->imbalance = max(local->group_capacity, local->group_util) -
10584 					 local->group_util;
10585 
10586 			/*
10587 			 * In some cases, the group's utilization is max or even
10588 			 * higher than capacity because of migrations but the
10589 			 * local CPU is (newly) idle. There is at least one
10590 			 * waiting task in this overloaded busiest group. Let's
10591 			 * try to pull it.
10592 			 */
10593 			if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
10594 				env->migration_type = migrate_task;
10595 				env->imbalance = 1;
10596 			}
10597 
10598 			return;
10599 		}
10600 
10601 		if (busiest->group_weight == 1 || sds->prefer_sibling) {
10602 			/*
10603 			 * When prefer sibling, evenly spread running tasks on
10604 			 * groups.
10605 			 */
10606 			env->migration_type = migrate_task;
10607 			env->imbalance = sibling_imbalance(env, sds, busiest, local);
10608 		} else {
10609 
10610 			/*
10611 			 * If there is no overload, we just want to even the number of
10612 			 * idle cpus.
10613 			 */
10614 			env->migration_type = migrate_task;
10615 			env->imbalance = max_t(long, 0,
10616 					       (local->idle_cpus - busiest->idle_cpus));
10617 		}
10618 
10619 #ifdef CONFIG_NUMA
10620 		/* Consider allowing a small imbalance between NUMA groups */
10621 		if (env->sd->flags & SD_NUMA) {
10622 			env->imbalance = adjust_numa_imbalance(env->imbalance,
10623 							       local->sum_nr_running + 1,
10624 							       env->sd->imb_numa_nr);
10625 		}
10626 #endif
10627 
10628 		/* Number of tasks to move to restore balance */
10629 		env->imbalance >>= 1;
10630 
10631 		return;
10632 	}
10633 
10634 	/*
10635 	 * Local is fully busy but has to take more load to relieve the
10636 	 * busiest group
10637 	 */
10638 	if (local->group_type < group_overloaded) {
10639 		/*
10640 		 * Local will become overloaded so the avg_load metrics are
10641 		 * finally needed.
10642 		 */
10643 
10644 		local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10645 				  local->group_capacity;
10646 
10647 		/*
10648 		 * If the local group is more loaded than the selected
10649 		 * busiest group don't try to pull any tasks.
10650 		 */
10651 		if (local->avg_load >= busiest->avg_load) {
10652 			env->imbalance = 0;
10653 			return;
10654 		}
10655 
10656 		sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10657 				sds->total_capacity;
10658 
10659 		/*
10660 		 * If the local group is more loaded than the average system
10661 		 * load, don't try to pull any tasks.
10662 		 */
10663 		if (local->avg_load >= sds->avg_load) {
10664 			env->imbalance = 0;
10665 			return;
10666 		}
10667 
10668 	}
10669 
10670 	/*
10671 	 * Both group are or will become overloaded and we're trying to get all
10672 	 * the CPUs to the average_load, so we don't want to push ourselves
10673 	 * above the average load, nor do we wish to reduce the max loaded CPU
10674 	 * below the average load. At the same time, we also don't want to
10675 	 * reduce the group load below the group capacity. Thus we look for
10676 	 * the minimum possible imbalance.
10677 	 */
10678 	env->migration_type = migrate_load;
10679 	env->imbalance = min(
10680 		(busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10681 		(sds->avg_load - local->avg_load) * local->group_capacity
10682 	) / SCHED_CAPACITY_SCALE;
10683 }
10684 
10685 /******* find_busiest_group() helpers end here *********************/
10686 
10687 /*
10688  * Decision matrix according to the local and busiest group type:
10689  *
10690  * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10691  * has_spare        nr_idle   balanced   N/A    N/A  balanced   balanced
10692  * fully_busy       nr_idle   nr_idle    N/A    N/A  balanced   balanced
10693  * misfit_task      force     N/A        N/A    N/A  N/A        N/A
10694  * asym_packing     force     force      N/A    N/A  force      force
10695  * imbalanced       force     force      N/A    N/A  force      force
10696  * overloaded       force     force      N/A    N/A  force      avg_load
10697  *
10698  * N/A :      Not Applicable because already filtered while updating
10699  *            statistics.
10700  * balanced : The system is balanced for these 2 groups.
10701  * force :    Calculate the imbalance as load migration is probably needed.
10702  * avg_load : Only if imbalance is significant enough.
10703  * nr_idle :  dst_cpu is not busy and the number of idle CPUs is quite
10704  *            different in groups.
10705  */
10706 
10707 /**
10708  * find_busiest_group - Returns the busiest group within the sched_domain
10709  * if there is an imbalance.
10710  * @env: The load balancing environment.
10711  *
10712  * Also calculates the amount of runnable load which should be moved
10713  * to restore balance.
10714  *
10715  * Return:	- The busiest group if imbalance exists.
10716  */
find_busiest_group(struct lb_env * env)10717 static struct sched_group *find_busiest_group(struct lb_env *env)
10718 {
10719 	struct sg_lb_stats *local, *busiest;
10720 	struct sd_lb_stats sds;
10721 
10722 	init_sd_lb_stats(&sds);
10723 
10724 	/*
10725 	 * Compute the various statistics relevant for load balancing at
10726 	 * this level.
10727 	 */
10728 	update_sd_lb_stats(env, &sds);
10729 
10730 	/* There is no busy sibling group to pull tasks from */
10731 	if (!sds.busiest)
10732 		goto out_balanced;
10733 
10734 	busiest = &sds.busiest_stat;
10735 
10736 	/* Misfit tasks should be dealt with regardless of the avg load */
10737 	if (busiest->group_type == group_misfit_task)
10738 		goto force_balance;
10739 
10740 	if (sched_energy_enabled()) {
10741 		struct root_domain *rd = env->dst_rq->rd;
10742 
10743 		if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
10744 			goto out_balanced;
10745 	}
10746 
10747 	/* ASYM feature bypasses nice load balance check */
10748 	if (busiest->group_type == group_asym_packing)
10749 		goto force_balance;
10750 
10751 	/*
10752 	 * If the busiest group is imbalanced the below checks don't
10753 	 * work because they assume all things are equal, which typically
10754 	 * isn't true due to cpus_ptr constraints and the like.
10755 	 */
10756 	if (busiest->group_type == group_imbalanced)
10757 		goto force_balance;
10758 
10759 	local = &sds.local_stat;
10760 	/*
10761 	 * If the local group is busier than the selected busiest group
10762 	 * don't try and pull any tasks.
10763 	 */
10764 	if (local->group_type > busiest->group_type)
10765 		goto out_balanced;
10766 
10767 	/*
10768 	 * When groups are overloaded, use the avg_load to ensure fairness
10769 	 * between tasks.
10770 	 */
10771 	if (local->group_type == group_overloaded) {
10772 		/*
10773 		 * If the local group is more loaded than the selected
10774 		 * busiest group don't try to pull any tasks.
10775 		 */
10776 		if (local->avg_load >= busiest->avg_load)
10777 			goto out_balanced;
10778 
10779 		/* XXX broken for overlapping NUMA groups */
10780 		sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10781 				sds.total_capacity;
10782 
10783 		/*
10784 		 * Don't pull any tasks if this group is already above the
10785 		 * domain average load.
10786 		 */
10787 		if (local->avg_load >= sds.avg_load)
10788 			goto out_balanced;
10789 
10790 		/*
10791 		 * If the busiest group is more loaded, use imbalance_pct to be
10792 		 * conservative.
10793 		 */
10794 		if (100 * busiest->avg_load <=
10795 				env->sd->imbalance_pct * local->avg_load)
10796 			goto out_balanced;
10797 	}
10798 
10799 	/*
10800 	 * Try to move all excess tasks to a sibling domain of the busiest
10801 	 * group's child domain.
10802 	 */
10803 	if (sds.prefer_sibling && local->group_type == group_has_spare &&
10804 	    sibling_imbalance(env, &sds, busiest, local) > 1)
10805 		goto force_balance;
10806 
10807 	if (busiest->group_type != group_overloaded) {
10808 		if (env->idle == CPU_NOT_IDLE) {
10809 			/*
10810 			 * If the busiest group is not overloaded (and as a
10811 			 * result the local one too) but this CPU is already
10812 			 * busy, let another idle CPU try to pull task.
10813 			 */
10814 			goto out_balanced;
10815 		}
10816 
10817 		if (busiest->group_type == group_smt_balance &&
10818 		    smt_vs_nonsmt_groups(sds.local, sds.busiest)) {
10819 			/* Let non SMT CPU pull from SMT CPU sharing with sibling */
10820 			goto force_balance;
10821 		}
10822 
10823 		if (busiest->group_weight > 1 &&
10824 		    local->idle_cpus <= (busiest->idle_cpus + 1)) {
10825 			/*
10826 			 * If the busiest group is not overloaded
10827 			 * and there is no imbalance between this and busiest
10828 			 * group wrt idle CPUs, it is balanced. The imbalance
10829 			 * becomes significant if the diff is greater than 1
10830 			 * otherwise we might end up to just move the imbalance
10831 			 * on another group. Of course this applies only if
10832 			 * there is more than 1 CPU per group.
10833 			 */
10834 			goto out_balanced;
10835 		}
10836 
10837 		if (busiest->sum_h_nr_running == 1) {
10838 			/*
10839 			 * busiest doesn't have any tasks waiting to run
10840 			 */
10841 			goto out_balanced;
10842 		}
10843 	}
10844 
10845 force_balance:
10846 	/* Looks like there is an imbalance. Compute it */
10847 	calculate_imbalance(env, &sds);
10848 	return env->imbalance ? sds.busiest : NULL;
10849 
10850 out_balanced:
10851 	env->imbalance = 0;
10852 	return NULL;
10853 }
10854 
10855 /*
10856  * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10857  */
find_busiest_queue(struct lb_env * env,struct sched_group * group)10858 static struct rq *find_busiest_queue(struct lb_env *env,
10859 				     struct sched_group *group)
10860 {
10861 	struct rq *busiest = NULL, *rq;
10862 	unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
10863 	unsigned int busiest_nr = 0;
10864 	int i;
10865 
10866 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10867 		unsigned long capacity, load, util;
10868 		unsigned int nr_running;
10869 		enum fbq_type rt;
10870 
10871 		rq = cpu_rq(i);
10872 		rt = fbq_classify_rq(rq);
10873 
10874 		/*
10875 		 * We classify groups/runqueues into three groups:
10876 		 *  - regular: there are !numa tasks
10877 		 *  - remote:  there are numa tasks that run on the 'wrong' node
10878 		 *  - all:     there is no distinction
10879 		 *
10880 		 * In order to avoid migrating ideally placed numa tasks,
10881 		 * ignore those when there's better options.
10882 		 *
10883 		 * If we ignore the actual busiest queue to migrate another
10884 		 * task, the next balance pass can still reduce the busiest
10885 		 * queue by moving tasks around inside the node.
10886 		 *
10887 		 * If we cannot move enough load due to this classification
10888 		 * the next pass will adjust the group classification and
10889 		 * allow migration of more tasks.
10890 		 *
10891 		 * Both cases only affect the total convergence complexity.
10892 		 */
10893 		if (rt > env->fbq_type)
10894 			continue;
10895 
10896 		nr_running = rq->cfs.h_nr_running;
10897 		if (!nr_running)
10898 			continue;
10899 
10900 		capacity = capacity_of(i);
10901 
10902 		/*
10903 		 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
10904 		 * eventually lead to active_balancing high->low capacity.
10905 		 * Higher per-CPU capacity is considered better than balancing
10906 		 * average load.
10907 		 */
10908 		if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
10909 		    !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
10910 		    nr_running == 1)
10911 			continue;
10912 
10913 		/*
10914 		 * Make sure we only pull tasks from a CPU of lower priority
10915 		 * when balancing between SMT siblings.
10916 		 *
10917 		 * If balancing between cores, let lower priority CPUs help
10918 		 * SMT cores with more than one busy sibling.
10919 		 */
10920 		if ((env->sd->flags & SD_ASYM_PACKING) &&
10921 		    sched_use_asym_prio(env->sd, i) &&
10922 		    sched_asym_prefer(i, env->dst_cpu) &&
10923 		    nr_running == 1)
10924 			continue;
10925 
10926 		switch (env->migration_type) {
10927 		case migrate_load:
10928 			/*
10929 			 * When comparing with load imbalance, use cpu_load()
10930 			 * which is not scaled with the CPU capacity.
10931 			 */
10932 			load = cpu_load(rq);
10933 
10934 			if (nr_running == 1 && load > env->imbalance &&
10935 			    !check_cpu_capacity(rq, env->sd))
10936 				break;
10937 
10938 			/*
10939 			 * For the load comparisons with the other CPUs,
10940 			 * consider the cpu_load() scaled with the CPU
10941 			 * capacity, so that the load can be moved away
10942 			 * from the CPU that is potentially running at a
10943 			 * lower capacity.
10944 			 *
10945 			 * Thus we're looking for max(load_i / capacity_i),
10946 			 * crosswise multiplication to rid ourselves of the
10947 			 * division works out to:
10948 			 * load_i * capacity_j > load_j * capacity_i;
10949 			 * where j is our previous maximum.
10950 			 */
10951 			if (load * busiest_capacity > busiest_load * capacity) {
10952 				busiest_load = load;
10953 				busiest_capacity = capacity;
10954 				busiest = rq;
10955 			}
10956 			break;
10957 
10958 		case migrate_util:
10959 			util = cpu_util_cfs_boost(i);
10960 
10961 			/*
10962 			 * Don't try to pull utilization from a CPU with one
10963 			 * running task. Whatever its utilization, we will fail
10964 			 * detach the task.
10965 			 */
10966 			if (nr_running <= 1)
10967 				continue;
10968 
10969 			if (busiest_util < util) {
10970 				busiest_util = util;
10971 				busiest = rq;
10972 			}
10973 			break;
10974 
10975 		case migrate_task:
10976 			if (busiest_nr < nr_running) {
10977 				busiest_nr = nr_running;
10978 				busiest = rq;
10979 			}
10980 			break;
10981 
10982 		case migrate_misfit:
10983 			/*
10984 			 * For ASYM_CPUCAPACITY domains with misfit tasks we
10985 			 * simply seek the "biggest" misfit task.
10986 			 */
10987 			if (rq->misfit_task_load > busiest_load) {
10988 				busiest_load = rq->misfit_task_load;
10989 				busiest = rq;
10990 			}
10991 
10992 			break;
10993 
10994 		}
10995 	}
10996 
10997 	return busiest;
10998 }
10999 
11000 /*
11001  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
11002  * so long as it is large enough.
11003  */
11004 #define MAX_PINNED_INTERVAL	512
11005 
11006 static inline bool
asym_active_balance(struct lb_env * env)11007 asym_active_balance(struct lb_env *env)
11008 {
11009 	/*
11010 	 * ASYM_PACKING needs to force migrate tasks from busy but lower
11011 	 * priority CPUs in order to pack all tasks in the highest priority
11012 	 * CPUs. When done between cores, do it only if the whole core if the
11013 	 * whole core is idle.
11014 	 *
11015 	 * If @env::src_cpu is an SMT core with busy siblings, let
11016 	 * the lower priority @env::dst_cpu help it. Do not follow
11017 	 * CPU priority.
11018 	 */
11019 	return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
11020 	       sched_use_asym_prio(env->sd, env->dst_cpu) &&
11021 	       (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
11022 		!sched_use_asym_prio(env->sd, env->src_cpu));
11023 }
11024 
11025 static inline bool
imbalanced_active_balance(struct lb_env * env)11026 imbalanced_active_balance(struct lb_env *env)
11027 {
11028 	struct sched_domain *sd = env->sd;
11029 
11030 	/*
11031 	 * The imbalanced case includes the case of pinned tasks preventing a fair
11032 	 * distribution of the load on the system but also the even distribution of the
11033 	 * threads on a system with spare capacity
11034 	 */
11035 	if ((env->migration_type == migrate_task) &&
11036 	    (sd->nr_balance_failed > sd->cache_nice_tries+2))
11037 		return 1;
11038 
11039 	return 0;
11040 }
11041 
need_active_balance(struct lb_env * env)11042 static int need_active_balance(struct lb_env *env)
11043 {
11044 	struct sched_domain *sd = env->sd;
11045 
11046 	if (asym_active_balance(env))
11047 		return 1;
11048 
11049 	if (imbalanced_active_balance(env))
11050 		return 1;
11051 
11052 	/*
11053 	 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
11054 	 * It's worth migrating the task if the src_cpu's capacity is reduced
11055 	 * because of other sched_class or IRQs if more capacity stays
11056 	 * available on dst_cpu.
11057 	 */
11058 	if ((env->idle != CPU_NOT_IDLE) &&
11059 	    (env->src_rq->cfs.h_nr_running == 1)) {
11060 		if ((check_cpu_capacity(env->src_rq, sd)) &&
11061 		    (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
11062 			return 1;
11063 	}
11064 
11065 	if (env->migration_type == migrate_misfit)
11066 		return 1;
11067 
11068 	return 0;
11069 }
11070 
11071 static int active_load_balance_cpu_stop(void *data);
11072 
should_we_balance(struct lb_env * env)11073 static int should_we_balance(struct lb_env *env)
11074 {
11075 	struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask);
11076 	struct sched_group *sg = env->sd->groups;
11077 	int cpu, idle_smt = -1;
11078 
11079 	/*
11080 	 * Ensure the balancing environment is consistent; can happen
11081 	 * when the softirq triggers 'during' hotplug.
11082 	 */
11083 	if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
11084 		return 0;
11085 
11086 	/*
11087 	 * In the newly idle case, we will allow all the CPUs
11088 	 * to do the newly idle load balance.
11089 	 *
11090 	 * However, we bail out if we already have tasks or a wakeup pending,
11091 	 * to optimize wakeup latency.
11092 	 */
11093 	if (env->idle == CPU_NEWLY_IDLE) {
11094 		if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
11095 			return 0;
11096 		return 1;
11097 	}
11098 
11099 	cpumask_copy(swb_cpus, group_balance_mask(sg));
11100 	/* Try to find first idle CPU */
11101 	for_each_cpu_and(cpu, swb_cpus, env->cpus) {
11102 		if (!idle_cpu(cpu))
11103 			continue;
11104 
11105 		/*
11106 		 * Don't balance to idle SMT in busy core right away when
11107 		 * balancing cores, but remember the first idle SMT CPU for
11108 		 * later consideration.  Find CPU on an idle core first.
11109 		 */
11110 		if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) {
11111 			if (idle_smt == -1)
11112 				idle_smt = cpu;
11113 			/*
11114 			 * If the core is not idle, and first SMT sibling which is
11115 			 * idle has been found, then its not needed to check other
11116 			 * SMT siblings for idleness:
11117 			 */
11118 #ifdef CONFIG_SCHED_SMT
11119 			cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu));
11120 #endif
11121 			continue;
11122 		}
11123 
11124 		/*
11125 		 * Are we the first idle core in a non-SMT domain or higher,
11126 		 * or the first idle CPU in a SMT domain?
11127 		 */
11128 		return cpu == env->dst_cpu;
11129 	}
11130 
11131 	/* Are we the first idle CPU with busy siblings? */
11132 	if (idle_smt != -1)
11133 		return idle_smt == env->dst_cpu;
11134 
11135 	/* Are we the first CPU of this group ? */
11136 	return group_balance_cpu(sg) == env->dst_cpu;
11137 }
11138 
11139 /*
11140  * Check this_cpu to ensure it is balanced within domain. Attempt to move
11141  * tasks if there is an imbalance.
11142  */
load_balance(int this_cpu,struct rq * this_rq,struct sched_domain * sd,enum cpu_idle_type idle,int * continue_balancing)11143 static int load_balance(int this_cpu, struct rq *this_rq,
11144 			struct sched_domain *sd, enum cpu_idle_type idle,
11145 			int *continue_balancing)
11146 {
11147 	int ld_moved, cur_ld_moved, active_balance = 0;
11148 	struct sched_domain *sd_parent = sd->parent;
11149 	struct sched_group *group;
11150 	struct rq *busiest;
11151 	struct rq_flags rf;
11152 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
11153 	struct lb_env env = {
11154 		.sd		= sd,
11155 		.dst_cpu	= this_cpu,
11156 		.dst_rq		= this_rq,
11157 		.dst_grpmask    = group_balance_mask(sd->groups),
11158 		.idle		= idle,
11159 		.loop_break	= SCHED_NR_MIGRATE_BREAK,
11160 		.cpus		= cpus,
11161 		.fbq_type	= all,
11162 		.tasks		= LIST_HEAD_INIT(env.tasks),
11163 	};
11164 
11165 	cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
11166 
11167 	schedstat_inc(sd->lb_count[idle]);
11168 
11169 redo:
11170 	if (!should_we_balance(&env)) {
11171 		*continue_balancing = 0;
11172 		goto out_balanced;
11173 	}
11174 
11175 	group = find_busiest_group(&env);
11176 	if (!group) {
11177 		schedstat_inc(sd->lb_nobusyg[idle]);
11178 		goto out_balanced;
11179 	}
11180 
11181 	busiest = find_busiest_queue(&env, group);
11182 	if (!busiest) {
11183 		schedstat_inc(sd->lb_nobusyq[idle]);
11184 		goto out_balanced;
11185 	}
11186 
11187 	WARN_ON_ONCE(busiest == env.dst_rq);
11188 
11189 	schedstat_add(sd->lb_imbalance[idle], env.imbalance);
11190 
11191 	env.src_cpu = busiest->cpu;
11192 	env.src_rq = busiest;
11193 
11194 	ld_moved = 0;
11195 	/* Clear this flag as soon as we find a pullable task */
11196 	env.flags |= LBF_ALL_PINNED;
11197 	if (busiest->nr_running > 1) {
11198 		/*
11199 		 * Attempt to move tasks. If find_busiest_group has found
11200 		 * an imbalance but busiest->nr_running <= 1, the group is
11201 		 * still unbalanced. ld_moved simply stays zero, so it is
11202 		 * correctly treated as an imbalance.
11203 		 */
11204 		env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
11205 
11206 more_balance:
11207 		rq_lock_irqsave(busiest, &rf);
11208 		update_rq_clock(busiest);
11209 
11210 		/*
11211 		 * cur_ld_moved - load moved in current iteration
11212 		 * ld_moved     - cumulative load moved across iterations
11213 		 */
11214 		cur_ld_moved = detach_tasks(&env);
11215 
11216 		/*
11217 		 * We've detached some tasks from busiest_rq. Every
11218 		 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
11219 		 * unlock busiest->lock, and we are able to be sure
11220 		 * that nobody can manipulate the tasks in parallel.
11221 		 * See task_rq_lock() family for the details.
11222 		 */
11223 
11224 		rq_unlock(busiest, &rf);
11225 
11226 		if (cur_ld_moved) {
11227 			attach_tasks(&env);
11228 			ld_moved += cur_ld_moved;
11229 		}
11230 
11231 		local_irq_restore(rf.flags);
11232 
11233 		if (env.flags & LBF_NEED_BREAK) {
11234 			env.flags &= ~LBF_NEED_BREAK;
11235 			/* Stop if we tried all running tasks */
11236 			if (env.loop < busiest->nr_running)
11237 				goto more_balance;
11238 		}
11239 
11240 		/*
11241 		 * Revisit (affine) tasks on src_cpu that couldn't be moved to
11242 		 * us and move them to an alternate dst_cpu in our sched_group
11243 		 * where they can run. The upper limit on how many times we
11244 		 * iterate on same src_cpu is dependent on number of CPUs in our
11245 		 * sched_group.
11246 		 *
11247 		 * This changes load balance semantics a bit on who can move
11248 		 * load to a given_cpu. In addition to the given_cpu itself
11249 		 * (or a ilb_cpu acting on its behalf where given_cpu is
11250 		 * nohz-idle), we now have balance_cpu in a position to move
11251 		 * load to given_cpu. In rare situations, this may cause
11252 		 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
11253 		 * _independently_ and at _same_ time to move some load to
11254 		 * given_cpu) causing excess load to be moved to given_cpu.
11255 		 * This however should not happen so much in practice and
11256 		 * moreover subsequent load balance cycles should correct the
11257 		 * excess load moved.
11258 		 */
11259 		if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
11260 
11261 			/* Prevent to re-select dst_cpu via env's CPUs */
11262 			__cpumask_clear_cpu(env.dst_cpu, env.cpus);
11263 
11264 			env.dst_rq	 = cpu_rq(env.new_dst_cpu);
11265 			env.dst_cpu	 = env.new_dst_cpu;
11266 			env.flags	&= ~LBF_DST_PINNED;
11267 			env.loop	 = 0;
11268 			env.loop_break	 = SCHED_NR_MIGRATE_BREAK;
11269 
11270 			/*
11271 			 * Go back to "more_balance" rather than "redo" since we
11272 			 * need to continue with same src_cpu.
11273 			 */
11274 			goto more_balance;
11275 		}
11276 
11277 		/*
11278 		 * We failed to reach balance because of affinity.
11279 		 */
11280 		if (sd_parent) {
11281 			int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11282 
11283 			if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
11284 				*group_imbalance = 1;
11285 		}
11286 
11287 		/* All tasks on this runqueue were pinned by CPU affinity */
11288 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
11289 			__cpumask_clear_cpu(cpu_of(busiest), cpus);
11290 			/*
11291 			 * Attempting to continue load balancing at the current
11292 			 * sched_domain level only makes sense if there are
11293 			 * active CPUs remaining as possible busiest CPUs to
11294 			 * pull load from which are not contained within the
11295 			 * destination group that is receiving any migrated
11296 			 * load.
11297 			 */
11298 			if (!cpumask_subset(cpus, env.dst_grpmask)) {
11299 				env.loop = 0;
11300 				env.loop_break = SCHED_NR_MIGRATE_BREAK;
11301 				goto redo;
11302 			}
11303 			goto out_all_pinned;
11304 		}
11305 	}
11306 
11307 	if (!ld_moved) {
11308 		schedstat_inc(sd->lb_failed[idle]);
11309 		/*
11310 		 * Increment the failure counter only on periodic balance.
11311 		 * We do not want newidle balance, which can be very
11312 		 * frequent, pollute the failure counter causing
11313 		 * excessive cache_hot migrations and active balances.
11314 		 */
11315 		if (idle != CPU_NEWLY_IDLE)
11316 			sd->nr_balance_failed++;
11317 
11318 		if (need_active_balance(&env)) {
11319 			unsigned long flags;
11320 
11321 			raw_spin_rq_lock_irqsave(busiest, flags);
11322 
11323 			/*
11324 			 * Don't kick the active_load_balance_cpu_stop,
11325 			 * if the curr task on busiest CPU can't be
11326 			 * moved to this_cpu:
11327 			 */
11328 			if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
11329 				raw_spin_rq_unlock_irqrestore(busiest, flags);
11330 				goto out_one_pinned;
11331 			}
11332 
11333 			/* Record that we found at least one task that could run on this_cpu */
11334 			env.flags &= ~LBF_ALL_PINNED;
11335 
11336 			/*
11337 			 * ->active_balance synchronizes accesses to
11338 			 * ->active_balance_work.  Once set, it's cleared
11339 			 * only after active load balance is finished.
11340 			 */
11341 			if (!busiest->active_balance) {
11342 				busiest->active_balance = 1;
11343 				busiest->push_cpu = this_cpu;
11344 				active_balance = 1;
11345 			}
11346 
11347 			preempt_disable();
11348 			raw_spin_rq_unlock_irqrestore(busiest, flags);
11349 			if (active_balance) {
11350 				stop_one_cpu_nowait(cpu_of(busiest),
11351 					active_load_balance_cpu_stop, busiest,
11352 					&busiest->active_balance_work);
11353 			}
11354 			preempt_enable();
11355 		}
11356 	} else {
11357 		sd->nr_balance_failed = 0;
11358 	}
11359 
11360 	if (likely(!active_balance) || need_active_balance(&env)) {
11361 		/* We were unbalanced, so reset the balancing interval */
11362 		sd->balance_interval = sd->min_interval;
11363 	}
11364 
11365 	goto out;
11366 
11367 out_balanced:
11368 	/*
11369 	 * We reach balance although we may have faced some affinity
11370 	 * constraints. Clear the imbalance flag only if other tasks got
11371 	 * a chance to move and fix the imbalance.
11372 	 */
11373 	if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
11374 		int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11375 
11376 		if (*group_imbalance)
11377 			*group_imbalance = 0;
11378 	}
11379 
11380 out_all_pinned:
11381 	/*
11382 	 * We reach balance because all tasks are pinned at this level so
11383 	 * we can't migrate them. Let the imbalance flag set so parent level
11384 	 * can try to migrate them.
11385 	 */
11386 	schedstat_inc(sd->lb_balanced[idle]);
11387 
11388 	sd->nr_balance_failed = 0;
11389 
11390 out_one_pinned:
11391 	ld_moved = 0;
11392 
11393 	/*
11394 	 * newidle_balance() disregards balance intervals, so we could
11395 	 * repeatedly reach this code, which would lead to balance_interval
11396 	 * skyrocketing in a short amount of time. Skip the balance_interval
11397 	 * increase logic to avoid that.
11398 	 */
11399 	if (env.idle == CPU_NEWLY_IDLE)
11400 		goto out;
11401 
11402 	/* tune up the balancing interval */
11403 	if ((env.flags & LBF_ALL_PINNED &&
11404 	     sd->balance_interval < MAX_PINNED_INTERVAL) ||
11405 	    sd->balance_interval < sd->max_interval)
11406 		sd->balance_interval *= 2;
11407 out:
11408 	return ld_moved;
11409 }
11410 
11411 static inline unsigned long
get_sd_balance_interval(struct sched_domain * sd,int cpu_busy)11412 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
11413 {
11414 	unsigned long interval = sd->balance_interval;
11415 
11416 	if (cpu_busy)
11417 		interval *= sd->busy_factor;
11418 
11419 	/* scale ms to jiffies */
11420 	interval = msecs_to_jiffies(interval);
11421 
11422 	/*
11423 	 * Reduce likelihood of busy balancing at higher domains racing with
11424 	 * balancing at lower domains by preventing their balancing periods
11425 	 * from being multiples of each other.
11426 	 */
11427 	if (cpu_busy)
11428 		interval -= 1;
11429 
11430 	interval = clamp(interval, 1UL, max_load_balance_interval);
11431 
11432 	return interval;
11433 }
11434 
11435 static inline void
update_next_balance(struct sched_domain * sd,unsigned long * next_balance)11436 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
11437 {
11438 	unsigned long interval, next;
11439 
11440 	/* used by idle balance, so cpu_busy = 0 */
11441 	interval = get_sd_balance_interval(sd, 0);
11442 	next = sd->last_balance + interval;
11443 
11444 	if (time_after(*next_balance, next))
11445 		*next_balance = next;
11446 }
11447 
11448 /*
11449  * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
11450  * running tasks off the busiest CPU onto idle CPUs. It requires at
11451  * least 1 task to be running on each physical CPU where possible, and
11452  * avoids physical / logical imbalances.
11453  */
active_load_balance_cpu_stop(void * data)11454 static int active_load_balance_cpu_stop(void *data)
11455 {
11456 	struct rq *busiest_rq = data;
11457 	int busiest_cpu = cpu_of(busiest_rq);
11458 	int target_cpu = busiest_rq->push_cpu;
11459 	struct rq *target_rq = cpu_rq(target_cpu);
11460 	struct sched_domain *sd;
11461 	struct task_struct *p = NULL;
11462 	struct rq_flags rf;
11463 
11464 	rq_lock_irq(busiest_rq, &rf);
11465 	/*
11466 	 * Between queueing the stop-work and running it is a hole in which
11467 	 * CPUs can become inactive. We should not move tasks from or to
11468 	 * inactive CPUs.
11469 	 */
11470 	if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
11471 		goto out_unlock;
11472 
11473 	/* Make sure the requested CPU hasn't gone down in the meantime: */
11474 	if (unlikely(busiest_cpu != smp_processor_id() ||
11475 		     !busiest_rq->active_balance))
11476 		goto out_unlock;
11477 
11478 	/* Is there any task to move? */
11479 	if (busiest_rq->nr_running <= 1)
11480 		goto out_unlock;
11481 
11482 	/*
11483 	 * This condition is "impossible", if it occurs
11484 	 * we need to fix it. Originally reported by
11485 	 * Bjorn Helgaas on a 128-CPU setup.
11486 	 */
11487 	WARN_ON_ONCE(busiest_rq == target_rq);
11488 
11489 	/* Search for an sd spanning us and the target CPU. */
11490 	rcu_read_lock();
11491 	for_each_domain(target_cpu, sd) {
11492 		if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
11493 			break;
11494 	}
11495 
11496 	if (likely(sd)) {
11497 		struct lb_env env = {
11498 			.sd		= sd,
11499 			.dst_cpu	= target_cpu,
11500 			.dst_rq		= target_rq,
11501 			.src_cpu	= busiest_rq->cpu,
11502 			.src_rq		= busiest_rq,
11503 			.idle		= CPU_IDLE,
11504 			.flags		= LBF_ACTIVE_LB,
11505 		};
11506 
11507 		schedstat_inc(sd->alb_count);
11508 		update_rq_clock(busiest_rq);
11509 
11510 		p = detach_one_task(&env);
11511 		if (p) {
11512 			schedstat_inc(sd->alb_pushed);
11513 			/* Active balancing done, reset the failure counter. */
11514 			sd->nr_balance_failed = 0;
11515 		} else {
11516 			schedstat_inc(sd->alb_failed);
11517 		}
11518 	}
11519 	rcu_read_unlock();
11520 out_unlock:
11521 	busiest_rq->active_balance = 0;
11522 	rq_unlock(busiest_rq, &rf);
11523 
11524 	if (p)
11525 		attach_one_task(target_rq, p);
11526 
11527 	local_irq_enable();
11528 
11529 	return 0;
11530 }
11531 
11532 static DEFINE_SPINLOCK(balancing);
11533 
11534 /*
11535  * Scale the max load_balance interval with the number of CPUs in the system.
11536  * This trades load-balance latency on larger machines for less cross talk.
11537  */
update_max_interval(void)11538 void update_max_interval(void)
11539 {
11540 	max_load_balance_interval = HZ*num_online_cpus()/10;
11541 }
11542 
update_newidle_cost(struct sched_domain * sd,u64 cost)11543 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
11544 {
11545 	if (cost > sd->max_newidle_lb_cost) {
11546 		/*
11547 		 * Track max cost of a domain to make sure to not delay the
11548 		 * next wakeup on the CPU.
11549 		 */
11550 		sd->max_newidle_lb_cost = cost;
11551 		sd->last_decay_max_lb_cost = jiffies;
11552 	} else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
11553 		/*
11554 		 * Decay the newidle max times by ~1% per second to ensure that
11555 		 * it is not outdated and the current max cost is actually
11556 		 * shorter.
11557 		 */
11558 		sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
11559 		sd->last_decay_max_lb_cost = jiffies;
11560 
11561 		return true;
11562 	}
11563 
11564 	return false;
11565 }
11566 
11567 /*
11568  * It checks each scheduling domain to see if it is due to be balanced,
11569  * and initiates a balancing operation if so.
11570  *
11571  * Balancing parameters are set up in init_sched_domains.
11572  */
rebalance_domains(struct rq * rq,enum cpu_idle_type idle)11573 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
11574 {
11575 	int continue_balancing = 1;
11576 	int cpu = rq->cpu;
11577 	int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11578 	unsigned long interval;
11579 	struct sched_domain *sd;
11580 	/* Earliest time when we have to do rebalance again */
11581 	unsigned long next_balance = jiffies + 60*HZ;
11582 	int update_next_balance = 0;
11583 	int need_serialize, need_decay = 0;
11584 	u64 max_cost = 0;
11585 
11586 	rcu_read_lock();
11587 	for_each_domain(cpu, sd) {
11588 		/*
11589 		 * Decay the newidle max times here because this is a regular
11590 		 * visit to all the domains.
11591 		 */
11592 		need_decay = update_newidle_cost(sd, 0);
11593 		max_cost += sd->max_newidle_lb_cost;
11594 
11595 		/*
11596 		 * Stop the load balance at this level. There is another
11597 		 * CPU in our sched group which is doing load balancing more
11598 		 * actively.
11599 		 */
11600 		if (!continue_balancing) {
11601 			if (need_decay)
11602 				continue;
11603 			break;
11604 		}
11605 
11606 		interval = get_sd_balance_interval(sd, busy);
11607 
11608 		need_serialize = sd->flags & SD_SERIALIZE;
11609 		if (need_serialize) {
11610 			if (!spin_trylock(&balancing))
11611 				goto out;
11612 		}
11613 
11614 		if (time_after_eq(jiffies, sd->last_balance + interval)) {
11615 			if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
11616 				/*
11617 				 * The LBF_DST_PINNED logic could have changed
11618 				 * env->dst_cpu, so we can't know our idle
11619 				 * state even if we migrated tasks. Update it.
11620 				 */
11621 				idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
11622 				busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11623 			}
11624 			sd->last_balance = jiffies;
11625 			interval = get_sd_balance_interval(sd, busy);
11626 		}
11627 		if (need_serialize)
11628 			spin_unlock(&balancing);
11629 out:
11630 		if (time_after(next_balance, sd->last_balance + interval)) {
11631 			next_balance = sd->last_balance + interval;
11632 			update_next_balance = 1;
11633 		}
11634 	}
11635 	if (need_decay) {
11636 		/*
11637 		 * Ensure the rq-wide value also decays but keep it at a
11638 		 * reasonable floor to avoid funnies with rq->avg_idle.
11639 		 */
11640 		rq->max_idle_balance_cost =
11641 			max((u64)sysctl_sched_migration_cost, max_cost);
11642 	}
11643 	rcu_read_unlock();
11644 
11645 	/*
11646 	 * next_balance will be updated only when there is a need.
11647 	 * When the cpu is attached to null domain for ex, it will not be
11648 	 * updated.
11649 	 */
11650 	if (likely(update_next_balance))
11651 		rq->next_balance = next_balance;
11652 
11653 }
11654 
on_null_domain(struct rq * rq)11655 static inline int on_null_domain(struct rq *rq)
11656 {
11657 	return unlikely(!rcu_dereference_sched(rq->sd));
11658 }
11659 
11660 #ifdef CONFIG_NO_HZ_COMMON
11661 /*
11662  * idle load balancing details
11663  * - When one of the busy CPUs notice that there may be an idle rebalancing
11664  *   needed, they will kick the idle load balancer, which then does idle
11665  *   load balancing for all the idle CPUs.
11666  * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
11667  *   anywhere yet.
11668  */
11669 
find_new_ilb(void)11670 static inline int find_new_ilb(void)
11671 {
11672 	int ilb;
11673 	const struct cpumask *hk_mask;
11674 
11675 	hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11676 
11677 	for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
11678 
11679 		if (ilb == smp_processor_id())
11680 			continue;
11681 
11682 		if (idle_cpu(ilb))
11683 			return ilb;
11684 	}
11685 
11686 	return nr_cpu_ids;
11687 }
11688 
11689 /*
11690  * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
11691  * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11692  */
kick_ilb(unsigned int flags)11693 static void kick_ilb(unsigned int flags)
11694 {
11695 	int ilb_cpu;
11696 
11697 	/*
11698 	 * Increase nohz.next_balance only when if full ilb is triggered but
11699 	 * not if we only update stats.
11700 	 */
11701 	if (flags & NOHZ_BALANCE_KICK)
11702 		nohz.next_balance = jiffies+1;
11703 
11704 	ilb_cpu = find_new_ilb();
11705 
11706 	if (ilb_cpu >= nr_cpu_ids)
11707 		return;
11708 
11709 	/*
11710 	 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11711 	 * the first flag owns it; cleared by nohz_csd_func().
11712 	 */
11713 	flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11714 	if (flags & NOHZ_KICK_MASK)
11715 		return;
11716 
11717 	/*
11718 	 * This way we generate an IPI on the target CPU which
11719 	 * is idle. And the softirq performing nohz idle load balance
11720 	 * will be run before returning from the IPI.
11721 	 */
11722 	smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11723 }
11724 
11725 /*
11726  * Current decision point for kicking the idle load balancer in the presence
11727  * of idle CPUs in the system.
11728  */
nohz_balancer_kick(struct rq * rq)11729 static void nohz_balancer_kick(struct rq *rq)
11730 {
11731 	unsigned long now = jiffies;
11732 	struct sched_domain_shared *sds;
11733 	struct sched_domain *sd;
11734 	int nr_busy, i, cpu = rq->cpu;
11735 	unsigned int flags = 0;
11736 
11737 	if (unlikely(rq->idle_balance))
11738 		return;
11739 
11740 	/*
11741 	 * We may be recently in ticked or tickless idle mode. At the first
11742 	 * busy tick after returning from idle, we will update the busy stats.
11743 	 */
11744 	nohz_balance_exit_idle(rq);
11745 
11746 	/*
11747 	 * None are in tickless mode and hence no need for NOHZ idle load
11748 	 * balancing.
11749 	 */
11750 	if (likely(!atomic_read(&nohz.nr_cpus)))
11751 		return;
11752 
11753 	if (READ_ONCE(nohz.has_blocked) &&
11754 	    time_after(now, READ_ONCE(nohz.next_blocked)))
11755 		flags = NOHZ_STATS_KICK;
11756 
11757 	if (time_before(now, nohz.next_balance))
11758 		goto out;
11759 
11760 	if (rq->nr_running >= 2) {
11761 		flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11762 		goto out;
11763 	}
11764 
11765 	rcu_read_lock();
11766 
11767 	sd = rcu_dereference(rq->sd);
11768 	if (sd) {
11769 		/*
11770 		 * If there's a CFS task and the current CPU has reduced
11771 		 * capacity; kick the ILB to see if there's a better CPU to run
11772 		 * on.
11773 		 */
11774 		if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11775 			flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11776 			goto unlock;
11777 		}
11778 	}
11779 
11780 	sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11781 	if (sd) {
11782 		/*
11783 		 * When ASYM_PACKING; see if there's a more preferred CPU
11784 		 * currently idle; in which case, kick the ILB to move tasks
11785 		 * around.
11786 		 *
11787 		 * When balancing betwen cores, all the SMT siblings of the
11788 		 * preferred CPU must be idle.
11789 		 */
11790 		for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11791 			if (sched_use_asym_prio(sd, i) &&
11792 			    sched_asym_prefer(i, cpu)) {
11793 				flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11794 				goto unlock;
11795 			}
11796 		}
11797 	}
11798 
11799 	sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11800 	if (sd) {
11801 		/*
11802 		 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11803 		 * to run the misfit task on.
11804 		 */
11805 		if (check_misfit_status(rq, sd)) {
11806 			flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11807 			goto unlock;
11808 		}
11809 
11810 		/*
11811 		 * For asymmetric systems, we do not want to nicely balance
11812 		 * cache use, instead we want to embrace asymmetry and only
11813 		 * ensure tasks have enough CPU capacity.
11814 		 *
11815 		 * Skip the LLC logic because it's not relevant in that case.
11816 		 */
11817 		goto unlock;
11818 	}
11819 
11820 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
11821 	if (sds) {
11822 		/*
11823 		 * If there is an imbalance between LLC domains (IOW we could
11824 		 * increase the overall cache use), we need some less-loaded LLC
11825 		 * domain to pull some load. Likewise, we may need to spread
11826 		 * load within the current LLC domain (e.g. packed SMT cores but
11827 		 * other CPUs are idle). We can't really know from here how busy
11828 		 * the others are - so just get a nohz balance going if it looks
11829 		 * like this LLC domain has tasks we could move.
11830 		 */
11831 		nr_busy = atomic_read(&sds->nr_busy_cpus);
11832 		if (nr_busy > 1) {
11833 			flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11834 			goto unlock;
11835 		}
11836 	}
11837 unlock:
11838 	rcu_read_unlock();
11839 out:
11840 	if (READ_ONCE(nohz.needs_update))
11841 		flags |= NOHZ_NEXT_KICK;
11842 
11843 	if (flags)
11844 		kick_ilb(flags);
11845 }
11846 
set_cpu_sd_state_busy(int cpu)11847 static void set_cpu_sd_state_busy(int cpu)
11848 {
11849 	struct sched_domain *sd;
11850 
11851 	rcu_read_lock();
11852 	sd = rcu_dereference(per_cpu(sd_llc, cpu));
11853 
11854 	if (!sd || !sd->nohz_idle)
11855 		goto unlock;
11856 	sd->nohz_idle = 0;
11857 
11858 	atomic_inc(&sd->shared->nr_busy_cpus);
11859 unlock:
11860 	rcu_read_unlock();
11861 }
11862 
nohz_balance_exit_idle(struct rq * rq)11863 void nohz_balance_exit_idle(struct rq *rq)
11864 {
11865 	SCHED_WARN_ON(rq != this_rq());
11866 
11867 	if (likely(!rq->nohz_tick_stopped))
11868 		return;
11869 
11870 	rq->nohz_tick_stopped = 0;
11871 	cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
11872 	atomic_dec(&nohz.nr_cpus);
11873 
11874 	set_cpu_sd_state_busy(rq->cpu);
11875 }
11876 
set_cpu_sd_state_idle(int cpu)11877 static void set_cpu_sd_state_idle(int cpu)
11878 {
11879 	struct sched_domain *sd;
11880 
11881 	rcu_read_lock();
11882 	sd = rcu_dereference(per_cpu(sd_llc, cpu));
11883 
11884 	if (!sd || sd->nohz_idle)
11885 		goto unlock;
11886 	sd->nohz_idle = 1;
11887 
11888 	atomic_dec(&sd->shared->nr_busy_cpus);
11889 unlock:
11890 	rcu_read_unlock();
11891 }
11892 
11893 /*
11894  * This routine will record that the CPU is going idle with tick stopped.
11895  * This info will be used in performing idle load balancing in the future.
11896  */
nohz_balance_enter_idle(int cpu)11897 void nohz_balance_enter_idle(int cpu)
11898 {
11899 	struct rq *rq = cpu_rq(cpu);
11900 
11901 	SCHED_WARN_ON(cpu != smp_processor_id());
11902 
11903 	/* If this CPU is going down, then nothing needs to be done: */
11904 	if (!cpu_active(cpu))
11905 		return;
11906 
11907 	/* Spare idle load balancing on CPUs that don't want to be disturbed: */
11908 	if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
11909 		return;
11910 
11911 	/*
11912 	 * Can be set safely without rq->lock held
11913 	 * If a clear happens, it will have evaluated last additions because
11914 	 * rq->lock is held during the check and the clear
11915 	 */
11916 	rq->has_blocked_load = 1;
11917 
11918 	/*
11919 	 * The tick is still stopped but load could have been added in the
11920 	 * meantime. We set the nohz.has_blocked flag to trig a check of the
11921 	 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
11922 	 * of nohz.has_blocked can only happen after checking the new load
11923 	 */
11924 	if (rq->nohz_tick_stopped)
11925 		goto out;
11926 
11927 	/* If we're a completely isolated CPU, we don't play: */
11928 	if (on_null_domain(rq))
11929 		return;
11930 
11931 	rq->nohz_tick_stopped = 1;
11932 
11933 	cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
11934 	atomic_inc(&nohz.nr_cpus);
11935 
11936 	/*
11937 	 * Ensures that if nohz_idle_balance() fails to observe our
11938 	 * @idle_cpus_mask store, it must observe the @has_blocked
11939 	 * and @needs_update stores.
11940 	 */
11941 	smp_mb__after_atomic();
11942 
11943 	set_cpu_sd_state_idle(cpu);
11944 
11945 	WRITE_ONCE(nohz.needs_update, 1);
11946 out:
11947 	/*
11948 	 * Each time a cpu enter idle, we assume that it has blocked load and
11949 	 * enable the periodic update of the load of idle cpus
11950 	 */
11951 	WRITE_ONCE(nohz.has_blocked, 1);
11952 }
11953 
update_nohz_stats(struct rq * rq)11954 static bool update_nohz_stats(struct rq *rq)
11955 {
11956 	unsigned int cpu = rq->cpu;
11957 
11958 	if (!rq->has_blocked_load)
11959 		return false;
11960 
11961 	if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
11962 		return false;
11963 
11964 	if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
11965 		return true;
11966 
11967 	update_blocked_averages(cpu);
11968 
11969 	return rq->has_blocked_load;
11970 }
11971 
11972 /*
11973  * Internal function that runs load balance for all idle cpus. The load balance
11974  * can be a simple update of blocked load or a complete load balance with
11975  * tasks movement depending of flags.
11976  */
_nohz_idle_balance(struct rq * this_rq,unsigned int flags)11977 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
11978 {
11979 	/* Earliest time when we have to do rebalance again */
11980 	unsigned long now = jiffies;
11981 	unsigned long next_balance = now + 60*HZ;
11982 	bool has_blocked_load = false;
11983 	int update_next_balance = 0;
11984 	int this_cpu = this_rq->cpu;
11985 	int balance_cpu;
11986 	struct rq *rq;
11987 
11988 	SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
11989 
11990 	/*
11991 	 * We assume there will be no idle load after this update and clear
11992 	 * the has_blocked flag. If a cpu enters idle in the mean time, it will
11993 	 * set the has_blocked flag and trigger another update of idle load.
11994 	 * Because a cpu that becomes idle, is added to idle_cpus_mask before
11995 	 * setting the flag, we are sure to not clear the state and not
11996 	 * check the load of an idle cpu.
11997 	 *
11998 	 * Same applies to idle_cpus_mask vs needs_update.
11999 	 */
12000 	if (flags & NOHZ_STATS_KICK)
12001 		WRITE_ONCE(nohz.has_blocked, 0);
12002 	if (flags & NOHZ_NEXT_KICK)
12003 		WRITE_ONCE(nohz.needs_update, 0);
12004 
12005 	/*
12006 	 * Ensures that if we miss the CPU, we must see the has_blocked
12007 	 * store from nohz_balance_enter_idle().
12008 	 */
12009 	smp_mb();
12010 
12011 	/*
12012 	 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
12013 	 * chance for other idle cpu to pull load.
12014 	 */
12015 	for_each_cpu_wrap(balance_cpu,  nohz.idle_cpus_mask, this_cpu+1) {
12016 		if (!idle_cpu(balance_cpu))
12017 			continue;
12018 
12019 		/*
12020 		 * If this CPU gets work to do, stop the load balancing
12021 		 * work being done for other CPUs. Next load
12022 		 * balancing owner will pick it up.
12023 		 */
12024 		if (need_resched()) {
12025 			if (flags & NOHZ_STATS_KICK)
12026 				has_blocked_load = true;
12027 			if (flags & NOHZ_NEXT_KICK)
12028 				WRITE_ONCE(nohz.needs_update, 1);
12029 			goto abort;
12030 		}
12031 
12032 		rq = cpu_rq(balance_cpu);
12033 
12034 		if (flags & NOHZ_STATS_KICK)
12035 			has_blocked_load |= update_nohz_stats(rq);
12036 
12037 		/*
12038 		 * If time for next balance is due,
12039 		 * do the balance.
12040 		 */
12041 		if (time_after_eq(jiffies, rq->next_balance)) {
12042 			struct rq_flags rf;
12043 
12044 			rq_lock_irqsave(rq, &rf);
12045 			update_rq_clock(rq);
12046 			rq_unlock_irqrestore(rq, &rf);
12047 
12048 			if (flags & NOHZ_BALANCE_KICK)
12049 				rebalance_domains(rq, CPU_IDLE);
12050 		}
12051 
12052 		if (time_after(next_balance, rq->next_balance)) {
12053 			next_balance = rq->next_balance;
12054 			update_next_balance = 1;
12055 		}
12056 	}
12057 
12058 	/*
12059 	 * next_balance will be updated only when there is a need.
12060 	 * When the CPU is attached to null domain for ex, it will not be
12061 	 * updated.
12062 	 */
12063 	if (likely(update_next_balance))
12064 		nohz.next_balance = next_balance;
12065 
12066 	if (flags & NOHZ_STATS_KICK)
12067 		WRITE_ONCE(nohz.next_blocked,
12068 			   now + msecs_to_jiffies(LOAD_AVG_PERIOD));
12069 
12070 abort:
12071 	/* There is still blocked load, enable periodic update */
12072 	if (has_blocked_load)
12073 		WRITE_ONCE(nohz.has_blocked, 1);
12074 }
12075 
12076 /*
12077  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
12078  * rebalancing for all the cpus for whom scheduler ticks are stopped.
12079  */
nohz_idle_balance(struct rq * this_rq,enum cpu_idle_type idle)12080 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12081 {
12082 	unsigned int flags = this_rq->nohz_idle_balance;
12083 
12084 	if (!flags)
12085 		return false;
12086 
12087 	this_rq->nohz_idle_balance = 0;
12088 
12089 	if (idle != CPU_IDLE)
12090 		return false;
12091 
12092 	_nohz_idle_balance(this_rq, flags);
12093 
12094 	return true;
12095 }
12096 
12097 /*
12098  * Check if we need to run the ILB for updating blocked load before entering
12099  * idle state.
12100  */
nohz_run_idle_balance(int cpu)12101 void nohz_run_idle_balance(int cpu)
12102 {
12103 	unsigned int flags;
12104 
12105 	flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
12106 
12107 	/*
12108 	 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
12109 	 * (ie NOHZ_STATS_KICK set) and will do the same.
12110 	 */
12111 	if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
12112 		_nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
12113 }
12114 
nohz_newidle_balance(struct rq * this_rq)12115 static void nohz_newidle_balance(struct rq *this_rq)
12116 {
12117 	int this_cpu = this_rq->cpu;
12118 
12119 	/*
12120 	 * This CPU doesn't want to be disturbed by scheduler
12121 	 * housekeeping
12122 	 */
12123 	if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
12124 		return;
12125 
12126 	/* Will wake up very soon. No time for doing anything else*/
12127 	if (this_rq->avg_idle < sysctl_sched_migration_cost)
12128 		return;
12129 
12130 	/* Don't need to update blocked load of idle CPUs*/
12131 	if (!READ_ONCE(nohz.has_blocked) ||
12132 	    time_before(jiffies, READ_ONCE(nohz.next_blocked)))
12133 		return;
12134 
12135 	/*
12136 	 * Set the need to trigger ILB in order to update blocked load
12137 	 * before entering idle state.
12138 	 */
12139 	atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
12140 }
12141 
12142 #else /* !CONFIG_NO_HZ_COMMON */
nohz_balancer_kick(struct rq * rq)12143 static inline void nohz_balancer_kick(struct rq *rq) { }
12144 
nohz_idle_balance(struct rq * this_rq,enum cpu_idle_type idle)12145 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12146 {
12147 	return false;
12148 }
12149 
nohz_newidle_balance(struct rq * this_rq)12150 static inline void nohz_newidle_balance(struct rq *this_rq) { }
12151 #endif /* CONFIG_NO_HZ_COMMON */
12152 
12153 /*
12154  * newidle_balance is called by schedule() if this_cpu is about to become
12155  * idle. Attempts to pull tasks from other CPUs.
12156  *
12157  * Returns:
12158  *   < 0 - we released the lock and there are !fair tasks present
12159  *     0 - failed, no new tasks
12160  *   > 0 - success, new (fair) tasks present
12161  */
newidle_balance(struct rq * this_rq,struct rq_flags * rf)12162 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
12163 {
12164 	unsigned long next_balance = jiffies + HZ;
12165 	int this_cpu = this_rq->cpu;
12166 	u64 t0, t1, curr_cost = 0;
12167 	struct sched_domain *sd;
12168 	int pulled_task = 0;
12169 
12170 	update_misfit_status(NULL, this_rq);
12171 
12172 	/*
12173 	 * There is a task waiting to run. No need to search for one.
12174 	 * Return 0; the task will be enqueued when switching to idle.
12175 	 */
12176 	if (this_rq->ttwu_pending)
12177 		return 0;
12178 
12179 	/*
12180 	 * We must set idle_stamp _before_ calling idle_balance(), such that we
12181 	 * measure the duration of idle_balance() as idle time.
12182 	 */
12183 	this_rq->idle_stamp = rq_clock(this_rq);
12184 
12185 	/*
12186 	 * Do not pull tasks towards !active CPUs...
12187 	 */
12188 	if (!cpu_active(this_cpu))
12189 		return 0;
12190 
12191 	/*
12192 	 * This is OK, because current is on_cpu, which avoids it being picked
12193 	 * for load-balance and preemption/IRQs are still disabled avoiding
12194 	 * further scheduler activity on it and we're being very careful to
12195 	 * re-start the picking loop.
12196 	 */
12197 	rq_unpin_lock(this_rq, rf);
12198 
12199 	rcu_read_lock();
12200 	sd = rcu_dereference_check_sched_domain(this_rq->sd);
12201 
12202 	if (!READ_ONCE(this_rq->rd->overload) ||
12203 	    (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
12204 
12205 		if (sd)
12206 			update_next_balance(sd, &next_balance);
12207 		rcu_read_unlock();
12208 
12209 		goto out;
12210 	}
12211 	rcu_read_unlock();
12212 
12213 	raw_spin_rq_unlock(this_rq);
12214 
12215 	t0 = sched_clock_cpu(this_cpu);
12216 	update_blocked_averages(this_cpu);
12217 
12218 	rcu_read_lock();
12219 	for_each_domain(this_cpu, sd) {
12220 		int continue_balancing = 1;
12221 		u64 domain_cost;
12222 
12223 		update_next_balance(sd, &next_balance);
12224 
12225 		if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
12226 			break;
12227 
12228 		if (sd->flags & SD_BALANCE_NEWIDLE) {
12229 
12230 			pulled_task = load_balance(this_cpu, this_rq,
12231 						   sd, CPU_NEWLY_IDLE,
12232 						   &continue_balancing);
12233 
12234 			t1 = sched_clock_cpu(this_cpu);
12235 			domain_cost = t1 - t0;
12236 			update_newidle_cost(sd, domain_cost);
12237 
12238 			curr_cost += domain_cost;
12239 			t0 = t1;
12240 		}
12241 
12242 		/*
12243 		 * Stop searching for tasks to pull if there are
12244 		 * now runnable tasks on this rq.
12245 		 */
12246 		if (pulled_task || this_rq->nr_running > 0 ||
12247 		    this_rq->ttwu_pending)
12248 			break;
12249 	}
12250 	rcu_read_unlock();
12251 
12252 	raw_spin_rq_lock(this_rq);
12253 
12254 	if (curr_cost > this_rq->max_idle_balance_cost)
12255 		this_rq->max_idle_balance_cost = curr_cost;
12256 
12257 	/*
12258 	 * While browsing the domains, we released the rq lock, a task could
12259 	 * have been enqueued in the meantime. Since we're not going idle,
12260 	 * pretend we pulled a task.
12261 	 */
12262 	if (this_rq->cfs.h_nr_running && !pulled_task)
12263 		pulled_task = 1;
12264 
12265 	/* Is there a task of a high priority class? */
12266 	if (this_rq->nr_running != this_rq->cfs.h_nr_running)
12267 		pulled_task = -1;
12268 
12269 out:
12270 	/* Move the next balance forward */
12271 	if (time_after(this_rq->next_balance, next_balance))
12272 		this_rq->next_balance = next_balance;
12273 
12274 	if (pulled_task)
12275 		this_rq->idle_stamp = 0;
12276 	else
12277 		nohz_newidle_balance(this_rq);
12278 
12279 	rq_repin_lock(this_rq, rf);
12280 
12281 	return pulled_task;
12282 }
12283 
12284 /*
12285  * run_rebalance_domains is triggered when needed from the scheduler tick.
12286  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
12287  */
run_rebalance_domains(struct softirq_action * h)12288 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
12289 {
12290 	struct rq *this_rq = this_rq();
12291 	enum cpu_idle_type idle = this_rq->idle_balance ?
12292 						CPU_IDLE : CPU_NOT_IDLE;
12293 
12294 	/*
12295 	 * If this CPU has a pending nohz_balance_kick, then do the
12296 	 * balancing on behalf of the other idle CPUs whose ticks are
12297 	 * stopped. Do nohz_idle_balance *before* rebalance_domains to
12298 	 * give the idle CPUs a chance to load balance. Else we may
12299 	 * load balance only within the local sched_domain hierarchy
12300 	 * and abort nohz_idle_balance altogether if we pull some load.
12301 	 */
12302 	if (nohz_idle_balance(this_rq, idle))
12303 		return;
12304 
12305 	/* normal load balance */
12306 	update_blocked_averages(this_rq->cpu);
12307 	rebalance_domains(this_rq, idle);
12308 }
12309 
12310 /*
12311  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
12312  */
trigger_load_balance(struct rq * rq)12313 void trigger_load_balance(struct rq *rq)
12314 {
12315 	/*
12316 	 * Don't need to rebalance while attached to NULL domain or
12317 	 * runqueue CPU is not active
12318 	 */
12319 	if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
12320 		return;
12321 
12322 	if (time_after_eq(jiffies, rq->next_balance))
12323 		raise_softirq(SCHED_SOFTIRQ);
12324 
12325 	nohz_balancer_kick(rq);
12326 }
12327 
rq_online_fair(struct rq * rq)12328 static void rq_online_fair(struct rq *rq)
12329 {
12330 	update_sysctl();
12331 
12332 	update_runtime_enabled(rq);
12333 }
12334 
rq_offline_fair(struct rq * rq)12335 static void rq_offline_fair(struct rq *rq)
12336 {
12337 	update_sysctl();
12338 
12339 	/* Ensure any throttled groups are reachable by pick_next_task */
12340 	unthrottle_offline_cfs_rqs(rq);
12341 }
12342 
12343 #endif /* CONFIG_SMP */
12344 
12345 #ifdef CONFIG_SCHED_CORE
12346 static inline bool
__entity_slice_used(struct sched_entity * se,int min_nr_tasks)12347 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
12348 {
12349 	u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
12350 	u64 slice = se->slice;
12351 
12352 	return (rtime * min_nr_tasks > slice);
12353 }
12354 
12355 #define MIN_NR_TASKS_DURING_FORCEIDLE	2
task_tick_core(struct rq * rq,struct task_struct * curr)12356 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
12357 {
12358 	if (!sched_core_enabled(rq))
12359 		return;
12360 
12361 	/*
12362 	 * If runqueue has only one task which used up its slice and
12363 	 * if the sibling is forced idle, then trigger schedule to
12364 	 * give forced idle task a chance.
12365 	 *
12366 	 * sched_slice() considers only this active rq and it gets the
12367 	 * whole slice. But during force idle, we have siblings acting
12368 	 * like a single runqueue and hence we need to consider runnable
12369 	 * tasks on this CPU and the forced idle CPU. Ideally, we should
12370 	 * go through the forced idle rq, but that would be a perf hit.
12371 	 * We can assume that the forced idle CPU has at least
12372 	 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
12373 	 * if we need to give up the CPU.
12374 	 */
12375 	if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
12376 	    __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
12377 		resched_curr(rq);
12378 }
12379 
12380 /*
12381  * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
12382  */
se_fi_update(const struct sched_entity * se,unsigned int fi_seq,bool forceidle)12383 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
12384 			 bool forceidle)
12385 {
12386 	for_each_sched_entity(se) {
12387 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
12388 
12389 		if (forceidle) {
12390 			if (cfs_rq->forceidle_seq == fi_seq)
12391 				break;
12392 			cfs_rq->forceidle_seq = fi_seq;
12393 		}
12394 
12395 		cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
12396 	}
12397 }
12398 
task_vruntime_update(struct rq * rq,struct task_struct * p,bool in_fi)12399 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
12400 {
12401 	struct sched_entity *se = &p->se;
12402 
12403 	if (p->sched_class != &fair_sched_class)
12404 		return;
12405 
12406 	se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
12407 }
12408 
cfs_prio_less(const struct task_struct * a,const struct task_struct * b,bool in_fi)12409 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
12410 			bool in_fi)
12411 {
12412 	struct rq *rq = task_rq(a);
12413 	const struct sched_entity *sea = &a->se;
12414 	const struct sched_entity *seb = &b->se;
12415 	struct cfs_rq *cfs_rqa;
12416 	struct cfs_rq *cfs_rqb;
12417 	s64 delta;
12418 
12419 	SCHED_WARN_ON(task_rq(b)->core != rq->core);
12420 
12421 #ifdef CONFIG_FAIR_GROUP_SCHED
12422 	/*
12423 	 * Find an se in the hierarchy for tasks a and b, such that the se's
12424 	 * are immediate siblings.
12425 	 */
12426 	while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
12427 		int sea_depth = sea->depth;
12428 		int seb_depth = seb->depth;
12429 
12430 		if (sea_depth >= seb_depth)
12431 			sea = parent_entity(sea);
12432 		if (sea_depth <= seb_depth)
12433 			seb = parent_entity(seb);
12434 	}
12435 
12436 	se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
12437 	se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
12438 
12439 	cfs_rqa = sea->cfs_rq;
12440 	cfs_rqb = seb->cfs_rq;
12441 #else
12442 	cfs_rqa = &task_rq(a)->cfs;
12443 	cfs_rqb = &task_rq(b)->cfs;
12444 #endif
12445 
12446 	/*
12447 	 * Find delta after normalizing se's vruntime with its cfs_rq's
12448 	 * min_vruntime_fi, which would have been updated in prior calls
12449 	 * to se_fi_update().
12450 	 */
12451 	delta = (s64)(sea->vruntime - seb->vruntime) +
12452 		(s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
12453 
12454 	return delta > 0;
12455 }
12456 
task_is_throttled_fair(struct task_struct * p,int cpu)12457 static int task_is_throttled_fair(struct task_struct *p, int cpu)
12458 {
12459 	struct cfs_rq *cfs_rq;
12460 
12461 #ifdef CONFIG_FAIR_GROUP_SCHED
12462 	cfs_rq = task_group(p)->cfs_rq[cpu];
12463 #else
12464 	cfs_rq = &cpu_rq(cpu)->cfs;
12465 #endif
12466 	return throttled_hierarchy(cfs_rq);
12467 }
12468 #else
task_tick_core(struct rq * rq,struct task_struct * curr)12469 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
12470 #endif
12471 
12472 /*
12473  * scheduler tick hitting a task of our scheduling class.
12474  *
12475  * NOTE: This function can be called remotely by the tick offload that
12476  * goes along full dynticks. Therefore no local assumption can be made
12477  * and everything must be accessed through the @rq and @curr passed in
12478  * parameters.
12479  */
task_tick_fair(struct rq * rq,struct task_struct * curr,int queued)12480 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
12481 {
12482 	struct cfs_rq *cfs_rq;
12483 	struct sched_entity *se = &curr->se;
12484 
12485 	for_each_sched_entity(se) {
12486 		cfs_rq = cfs_rq_of(se);
12487 		entity_tick(cfs_rq, se, queued);
12488 	}
12489 
12490 	if (static_branch_unlikely(&sched_numa_balancing))
12491 		task_tick_numa(rq, curr);
12492 
12493 	update_misfit_status(curr, rq);
12494 	update_overutilized_status(task_rq(curr));
12495 
12496 	task_tick_core(rq, curr);
12497 }
12498 
12499 /*
12500  * called on fork with the child task as argument from the parent's context
12501  *  - child not yet on the tasklist
12502  *  - preemption disabled
12503  */
task_fork_fair(struct task_struct * p)12504 static void task_fork_fair(struct task_struct *p)
12505 {
12506 	struct sched_entity *se = &p->se, *curr;
12507 	struct cfs_rq *cfs_rq;
12508 	struct rq *rq = this_rq();
12509 	struct rq_flags rf;
12510 
12511 	rq_lock(rq, &rf);
12512 	update_rq_clock(rq);
12513 
12514 	cfs_rq = task_cfs_rq(current);
12515 	curr = cfs_rq->curr;
12516 	if (curr)
12517 		update_curr(cfs_rq);
12518 	place_entity(cfs_rq, se, ENQUEUE_INITIAL);
12519 	rq_unlock(rq, &rf);
12520 }
12521 
12522 /*
12523  * Priority of the task has changed. Check to see if we preempt
12524  * the current task.
12525  */
12526 static void
prio_changed_fair(struct rq * rq,struct task_struct * p,int oldprio)12527 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
12528 {
12529 	if (!task_on_rq_queued(p))
12530 		return;
12531 
12532 	if (rq->cfs.nr_running == 1)
12533 		return;
12534 
12535 	/*
12536 	 * Reschedule if we are currently running on this runqueue and
12537 	 * our priority decreased, or if we are not currently running on
12538 	 * this runqueue and our priority is higher than the current's
12539 	 */
12540 	if (task_current(rq, p)) {
12541 		if (p->prio > oldprio)
12542 			resched_curr(rq);
12543 	} else
12544 		check_preempt_curr(rq, p, 0);
12545 }
12546 
12547 #ifdef CONFIG_FAIR_GROUP_SCHED
12548 /*
12549  * Propagate the changes of the sched_entity across the tg tree to make it
12550  * visible to the root
12551  */
propagate_entity_cfs_rq(struct sched_entity * se)12552 static void propagate_entity_cfs_rq(struct sched_entity *se)
12553 {
12554 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
12555 
12556 	if (cfs_rq_throttled(cfs_rq))
12557 		return;
12558 
12559 	if (!throttled_hierarchy(cfs_rq))
12560 		list_add_leaf_cfs_rq(cfs_rq);
12561 
12562 	/* Start to propagate at parent */
12563 	se = se->parent;
12564 
12565 	for_each_sched_entity(se) {
12566 		cfs_rq = cfs_rq_of(se);
12567 
12568 		update_load_avg(cfs_rq, se, UPDATE_TG);
12569 
12570 		if (cfs_rq_throttled(cfs_rq))
12571 			break;
12572 
12573 		if (!throttled_hierarchy(cfs_rq))
12574 			list_add_leaf_cfs_rq(cfs_rq);
12575 	}
12576 }
12577 #else
propagate_entity_cfs_rq(struct sched_entity * se)12578 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
12579 #endif
12580 
detach_entity_cfs_rq(struct sched_entity * se)12581 static void detach_entity_cfs_rq(struct sched_entity *se)
12582 {
12583 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
12584 
12585 #ifdef CONFIG_SMP
12586 	/*
12587 	 * In case the task sched_avg hasn't been attached:
12588 	 * - A forked task which hasn't been woken up by wake_up_new_task().
12589 	 * - A task which has been woken up by try_to_wake_up() but is
12590 	 *   waiting for actually being woken up by sched_ttwu_pending().
12591 	 */
12592 	if (!se->avg.last_update_time)
12593 		return;
12594 #endif
12595 
12596 	/* Catch up with the cfs_rq and remove our load when we leave */
12597 	update_load_avg(cfs_rq, se, 0);
12598 	detach_entity_load_avg(cfs_rq, se);
12599 	update_tg_load_avg(cfs_rq);
12600 	propagate_entity_cfs_rq(se);
12601 }
12602 
attach_entity_cfs_rq(struct sched_entity * se)12603 static void attach_entity_cfs_rq(struct sched_entity *se)
12604 {
12605 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
12606 
12607 	/* Synchronize entity with its cfs_rq */
12608 	update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
12609 	attach_entity_load_avg(cfs_rq, se);
12610 	update_tg_load_avg(cfs_rq);
12611 	propagate_entity_cfs_rq(se);
12612 }
12613 
detach_task_cfs_rq(struct task_struct * p)12614 static void detach_task_cfs_rq(struct task_struct *p)
12615 {
12616 	struct sched_entity *se = &p->se;
12617 
12618 	detach_entity_cfs_rq(se);
12619 }
12620 
attach_task_cfs_rq(struct task_struct * p)12621 static void attach_task_cfs_rq(struct task_struct *p)
12622 {
12623 	struct sched_entity *se = &p->se;
12624 
12625 	attach_entity_cfs_rq(se);
12626 }
12627 
switched_from_fair(struct rq * rq,struct task_struct * p)12628 static void switched_from_fair(struct rq *rq, struct task_struct *p)
12629 {
12630 	detach_task_cfs_rq(p);
12631 }
12632 
switched_to_fair(struct rq * rq,struct task_struct * p)12633 static void switched_to_fair(struct rq *rq, struct task_struct *p)
12634 {
12635 	attach_task_cfs_rq(p);
12636 
12637 	if (task_on_rq_queued(p)) {
12638 		/*
12639 		 * We were most likely switched from sched_rt, so
12640 		 * kick off the schedule if running, otherwise just see
12641 		 * if we can still preempt the current task.
12642 		 */
12643 		if (task_current(rq, p))
12644 			resched_curr(rq);
12645 		else
12646 			check_preempt_curr(rq, p, 0);
12647 	}
12648 }
12649 
12650 /* Account for a task changing its policy or group.
12651  *
12652  * This routine is mostly called to set cfs_rq->curr field when a task
12653  * migrates between groups/classes.
12654  */
set_next_task_fair(struct rq * rq,struct task_struct * p,bool first)12655 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12656 {
12657 	struct sched_entity *se = &p->se;
12658 
12659 #ifdef CONFIG_SMP
12660 	if (task_on_rq_queued(p)) {
12661 		/*
12662 		 * Move the next running task to the front of the list, so our
12663 		 * cfs_tasks list becomes MRU one.
12664 		 */
12665 		list_move(&se->group_node, &rq->cfs_tasks);
12666 	}
12667 #endif
12668 
12669 	for_each_sched_entity(se) {
12670 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
12671 
12672 		set_next_entity(cfs_rq, se);
12673 		/* ensure bandwidth has been allocated on our new cfs_rq */
12674 		account_cfs_rq_runtime(cfs_rq, 0);
12675 	}
12676 }
12677 
init_cfs_rq(struct cfs_rq * cfs_rq)12678 void init_cfs_rq(struct cfs_rq *cfs_rq)
12679 {
12680 	cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12681 	u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12682 #ifdef CONFIG_SMP
12683 	raw_spin_lock_init(&cfs_rq->removed.lock);
12684 #endif
12685 }
12686 
12687 #ifdef CONFIG_FAIR_GROUP_SCHED
task_change_group_fair(struct task_struct * p)12688 static void task_change_group_fair(struct task_struct *p)
12689 {
12690 	/*
12691 	 * We couldn't detach or attach a forked task which
12692 	 * hasn't been woken up by wake_up_new_task().
12693 	 */
12694 	if (READ_ONCE(p->__state) == TASK_NEW)
12695 		return;
12696 
12697 	detach_task_cfs_rq(p);
12698 
12699 #ifdef CONFIG_SMP
12700 	/* Tell se's cfs_rq has been changed -- migrated */
12701 	p->se.avg.last_update_time = 0;
12702 #endif
12703 	set_task_rq(p, task_cpu(p));
12704 	attach_task_cfs_rq(p);
12705 }
12706 
free_fair_sched_group(struct task_group * tg)12707 void free_fair_sched_group(struct task_group *tg)
12708 {
12709 	int i;
12710 
12711 	for_each_possible_cpu(i) {
12712 		if (tg->cfs_rq)
12713 			kfree(tg->cfs_rq[i]);
12714 		if (tg->se)
12715 			kfree(tg->se[i]);
12716 	}
12717 
12718 	kfree(tg->cfs_rq);
12719 	kfree(tg->se);
12720 }
12721 
alloc_fair_sched_group(struct task_group * tg,struct task_group * parent)12722 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12723 {
12724 	struct sched_entity *se;
12725 	struct cfs_rq *cfs_rq;
12726 	int i;
12727 
12728 	tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12729 	if (!tg->cfs_rq)
12730 		goto err;
12731 	tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12732 	if (!tg->se)
12733 		goto err;
12734 
12735 	tg->shares = NICE_0_LOAD;
12736 
12737 	init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent));
12738 
12739 	for_each_possible_cpu(i) {
12740 		cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12741 				      GFP_KERNEL, cpu_to_node(i));
12742 		if (!cfs_rq)
12743 			goto err;
12744 
12745 		se = kzalloc_node(sizeof(struct sched_entity_stats),
12746 				  GFP_KERNEL, cpu_to_node(i));
12747 		if (!se)
12748 			goto err_free_rq;
12749 
12750 		init_cfs_rq(cfs_rq);
12751 		init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12752 		init_entity_runnable_average(se);
12753 	}
12754 
12755 	return 1;
12756 
12757 err_free_rq:
12758 	kfree(cfs_rq);
12759 err:
12760 	return 0;
12761 }
12762 
online_fair_sched_group(struct task_group * tg)12763 void online_fair_sched_group(struct task_group *tg)
12764 {
12765 	struct sched_entity *se;
12766 	struct rq_flags rf;
12767 	struct rq *rq;
12768 	int i;
12769 
12770 	for_each_possible_cpu(i) {
12771 		rq = cpu_rq(i);
12772 		se = tg->se[i];
12773 		rq_lock_irq(rq, &rf);
12774 		update_rq_clock(rq);
12775 		attach_entity_cfs_rq(se);
12776 		sync_throttle(tg, i);
12777 		rq_unlock_irq(rq, &rf);
12778 	}
12779 }
12780 
unregister_fair_sched_group(struct task_group * tg)12781 void unregister_fair_sched_group(struct task_group *tg)
12782 {
12783 	unsigned long flags;
12784 	struct rq *rq;
12785 	int cpu;
12786 
12787 	destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12788 
12789 	for_each_possible_cpu(cpu) {
12790 		if (tg->se[cpu])
12791 			remove_entity_load_avg(tg->se[cpu]);
12792 
12793 		/*
12794 		 * Only empty task groups can be destroyed; so we can speculatively
12795 		 * check on_list without danger of it being re-added.
12796 		 */
12797 		if (!tg->cfs_rq[cpu]->on_list)
12798 			continue;
12799 
12800 		rq = cpu_rq(cpu);
12801 
12802 		raw_spin_rq_lock_irqsave(rq, flags);
12803 		list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
12804 		raw_spin_rq_unlock_irqrestore(rq, flags);
12805 	}
12806 }
12807 
init_tg_cfs_entry(struct task_group * tg,struct cfs_rq * cfs_rq,struct sched_entity * se,int cpu,struct sched_entity * parent)12808 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
12809 			struct sched_entity *se, int cpu,
12810 			struct sched_entity *parent)
12811 {
12812 	struct rq *rq = cpu_rq(cpu);
12813 
12814 	cfs_rq->tg = tg;
12815 	cfs_rq->rq = rq;
12816 	init_cfs_rq_runtime(cfs_rq);
12817 
12818 	tg->cfs_rq[cpu] = cfs_rq;
12819 	tg->se[cpu] = se;
12820 
12821 	/* se could be NULL for root_task_group */
12822 	if (!se)
12823 		return;
12824 
12825 	if (!parent) {
12826 		se->cfs_rq = &rq->cfs;
12827 		se->depth = 0;
12828 	} else {
12829 		se->cfs_rq = parent->my_q;
12830 		se->depth = parent->depth + 1;
12831 	}
12832 
12833 	se->my_q = cfs_rq;
12834 	/* guarantee group entities always have weight */
12835 	update_load_set(&se->load, NICE_0_LOAD);
12836 	se->parent = parent;
12837 }
12838 
12839 static DEFINE_MUTEX(shares_mutex);
12840 
__sched_group_set_shares(struct task_group * tg,unsigned long shares)12841 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
12842 {
12843 	int i;
12844 
12845 	lockdep_assert_held(&shares_mutex);
12846 
12847 	/*
12848 	 * We can't change the weight of the root cgroup.
12849 	 */
12850 	if (!tg->se[0])
12851 		return -EINVAL;
12852 
12853 	shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12854 
12855 	if (tg->shares == shares)
12856 		return 0;
12857 
12858 	tg->shares = shares;
12859 	for_each_possible_cpu(i) {
12860 		struct rq *rq = cpu_rq(i);
12861 		struct sched_entity *se = tg->se[i];
12862 		struct rq_flags rf;
12863 
12864 		/* Propagate contribution to hierarchy */
12865 		rq_lock_irqsave(rq, &rf);
12866 		update_rq_clock(rq);
12867 		for_each_sched_entity(se) {
12868 			update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
12869 			update_cfs_group(se);
12870 		}
12871 		rq_unlock_irqrestore(rq, &rf);
12872 	}
12873 
12874 	return 0;
12875 }
12876 
sched_group_set_shares(struct task_group * tg,unsigned long shares)12877 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
12878 {
12879 	int ret;
12880 
12881 	mutex_lock(&shares_mutex);
12882 	if (tg_is_idle(tg))
12883 		ret = -EINVAL;
12884 	else
12885 		ret = __sched_group_set_shares(tg, shares);
12886 	mutex_unlock(&shares_mutex);
12887 
12888 	return ret;
12889 }
12890 
sched_group_set_idle(struct task_group * tg,long idle)12891 int sched_group_set_idle(struct task_group *tg, long idle)
12892 {
12893 	int i;
12894 
12895 	if (tg == &root_task_group)
12896 		return -EINVAL;
12897 
12898 	if (idle < 0 || idle > 1)
12899 		return -EINVAL;
12900 
12901 	mutex_lock(&shares_mutex);
12902 
12903 	if (tg->idle == idle) {
12904 		mutex_unlock(&shares_mutex);
12905 		return 0;
12906 	}
12907 
12908 	tg->idle = idle;
12909 
12910 	for_each_possible_cpu(i) {
12911 		struct rq *rq = cpu_rq(i);
12912 		struct sched_entity *se = tg->se[i];
12913 		struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
12914 		bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
12915 		long idle_task_delta;
12916 		struct rq_flags rf;
12917 
12918 		rq_lock_irqsave(rq, &rf);
12919 
12920 		grp_cfs_rq->idle = idle;
12921 		if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
12922 			goto next_cpu;
12923 
12924 		if (se->on_rq) {
12925 			parent_cfs_rq = cfs_rq_of(se);
12926 			if (cfs_rq_is_idle(grp_cfs_rq))
12927 				parent_cfs_rq->idle_nr_running++;
12928 			else
12929 				parent_cfs_rq->idle_nr_running--;
12930 		}
12931 
12932 		idle_task_delta = grp_cfs_rq->h_nr_running -
12933 				  grp_cfs_rq->idle_h_nr_running;
12934 		if (!cfs_rq_is_idle(grp_cfs_rq))
12935 			idle_task_delta *= -1;
12936 
12937 		for_each_sched_entity(se) {
12938 			struct cfs_rq *cfs_rq = cfs_rq_of(se);
12939 
12940 			if (!se->on_rq)
12941 				break;
12942 
12943 			cfs_rq->idle_h_nr_running += idle_task_delta;
12944 
12945 			/* Already accounted at parent level and above. */
12946 			if (cfs_rq_is_idle(cfs_rq))
12947 				break;
12948 		}
12949 
12950 next_cpu:
12951 		rq_unlock_irqrestore(rq, &rf);
12952 	}
12953 
12954 	/* Idle groups have minimum weight. */
12955 	if (tg_is_idle(tg))
12956 		__sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
12957 	else
12958 		__sched_group_set_shares(tg, NICE_0_LOAD);
12959 
12960 	mutex_unlock(&shares_mutex);
12961 	return 0;
12962 }
12963 
12964 #else /* CONFIG_FAIR_GROUP_SCHED */
12965 
free_fair_sched_group(struct task_group * tg)12966 void free_fair_sched_group(struct task_group *tg) { }
12967 
alloc_fair_sched_group(struct task_group * tg,struct task_group * parent)12968 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12969 {
12970 	return 1;
12971 }
12972 
online_fair_sched_group(struct task_group * tg)12973 void online_fair_sched_group(struct task_group *tg) { }
12974 
unregister_fair_sched_group(struct task_group * tg)12975 void unregister_fair_sched_group(struct task_group *tg) { }
12976 
12977 #endif /* CONFIG_FAIR_GROUP_SCHED */
12978 
12979 
get_rr_interval_fair(struct rq * rq,struct task_struct * task)12980 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
12981 {
12982 	struct sched_entity *se = &task->se;
12983 	unsigned int rr_interval = 0;
12984 
12985 	/*
12986 	 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
12987 	 * idle runqueue:
12988 	 */
12989 	if (rq->cfs.load.weight)
12990 		rr_interval = NS_TO_JIFFIES(se->slice);
12991 
12992 	return rr_interval;
12993 }
12994 
12995 /*
12996  * All the scheduling class methods:
12997  */
12998 DEFINE_SCHED_CLASS(fair) = {
12999 
13000 	.enqueue_task		= enqueue_task_fair,
13001 	.dequeue_task		= dequeue_task_fair,
13002 	.yield_task		= yield_task_fair,
13003 	.yield_to_task		= yield_to_task_fair,
13004 
13005 	.check_preempt_curr	= check_preempt_wakeup,
13006 
13007 	.pick_next_task		= __pick_next_task_fair,
13008 	.put_prev_task		= put_prev_task_fair,
13009 	.set_next_task          = set_next_task_fair,
13010 
13011 #ifdef CONFIG_SMP
13012 	.balance		= balance_fair,
13013 	.pick_task		= pick_task_fair,
13014 	.select_task_rq		= select_task_rq_fair,
13015 	.migrate_task_rq	= migrate_task_rq_fair,
13016 
13017 	.rq_online		= rq_online_fair,
13018 	.rq_offline		= rq_offline_fair,
13019 
13020 	.task_dead		= task_dead_fair,
13021 	.set_cpus_allowed	= set_cpus_allowed_common,
13022 #endif
13023 
13024 	.task_tick		= task_tick_fair,
13025 	.task_fork		= task_fork_fair,
13026 
13027 	.prio_changed		= prio_changed_fair,
13028 	.switched_from		= switched_from_fair,
13029 	.switched_to		= switched_to_fair,
13030 
13031 	.get_rr_interval	= get_rr_interval_fair,
13032 
13033 	.update_curr		= update_curr_fair,
13034 
13035 #ifdef CONFIG_FAIR_GROUP_SCHED
13036 	.task_change_group	= task_change_group_fair,
13037 #endif
13038 
13039 #ifdef CONFIG_SCHED_CORE
13040 	.task_is_throttled	= task_is_throttled_fair,
13041 #endif
13042 
13043 #ifdef CONFIG_UCLAMP_TASK
13044 	.uclamp_enabled		= 1,
13045 #endif
13046 };
13047 
13048 #ifdef CONFIG_SCHED_DEBUG
print_cfs_stats(struct seq_file * m,int cpu)13049 void print_cfs_stats(struct seq_file *m, int cpu)
13050 {
13051 	struct cfs_rq *cfs_rq, *pos;
13052 
13053 	rcu_read_lock();
13054 	for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
13055 		print_cfs_rq(m, cpu, cfs_rq);
13056 	rcu_read_unlock();
13057 }
13058 
13059 #ifdef CONFIG_NUMA_BALANCING
show_numa_stats(struct task_struct * p,struct seq_file * m)13060 void show_numa_stats(struct task_struct *p, struct seq_file *m)
13061 {
13062 	int node;
13063 	unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
13064 	struct numa_group *ng;
13065 
13066 	rcu_read_lock();
13067 	ng = rcu_dereference(p->numa_group);
13068 	for_each_online_node(node) {
13069 		if (p->numa_faults) {
13070 			tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
13071 			tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
13072 		}
13073 		if (ng) {
13074 			gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
13075 			gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
13076 		}
13077 		print_numa_stats(m, node, tsf, tpf, gsf, gpf);
13078 	}
13079 	rcu_read_unlock();
13080 }
13081 #endif /* CONFIG_NUMA_BALANCING */
13082 #endif /* CONFIG_SCHED_DEBUG */
13083 
init_sched_fair_class(void)13084 __init void init_sched_fair_class(void)
13085 {
13086 #ifdef CONFIG_SMP
13087 	int i;
13088 
13089 	for_each_possible_cpu(i) {
13090 		zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
13091 		zalloc_cpumask_var_node(&per_cpu(select_rq_mask,    i), GFP_KERNEL, cpu_to_node(i));
13092 		zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i),
13093 					GFP_KERNEL, cpu_to_node(i));
13094 
13095 #ifdef CONFIG_CFS_BANDWIDTH
13096 		INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
13097 		INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
13098 #endif
13099 	}
13100 
13101 	open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
13102 
13103 #ifdef CONFIG_NO_HZ_COMMON
13104 	nohz.next_balance = jiffies;
13105 	nohz.next_blocked = jiffies;
13106 	zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
13107 #endif
13108 #endif /* SMP */
13109 
13110 }
13111