xref: /linux/kernel/sched/topology.c (revision bbcd53c960713507ae764bf81970651b5577b95a)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Scheduler topology setup/handling methods
4  */
5 #include "sched.h"
6 
7 DEFINE_MUTEX(sched_domains_mutex);
8 
9 /* Protected by sched_domains_mutex: */
10 static cpumask_var_t sched_domains_tmpmask;
11 static cpumask_var_t sched_domains_tmpmask2;
12 
13 #ifdef CONFIG_SCHED_DEBUG
14 
15 static int __init sched_debug_setup(char *str)
16 {
17 	sched_debug_verbose = true;
18 
19 	return 0;
20 }
21 early_param("sched_verbose", sched_debug_setup);
22 
23 static inline bool sched_debug(void)
24 {
25 	return sched_debug_verbose;
26 }
27 
28 #define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
29 const struct sd_flag_debug sd_flag_debug[] = {
30 #include <linux/sched/sd_flags.h>
31 };
32 #undef SD_FLAG
33 
34 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
35 				  struct cpumask *groupmask)
36 {
37 	struct sched_group *group = sd->groups;
38 	unsigned long flags = sd->flags;
39 	unsigned int idx;
40 
41 	cpumask_clear(groupmask);
42 
43 	printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
44 	printk(KERN_CONT "span=%*pbl level=%s\n",
45 	       cpumask_pr_args(sched_domain_span(sd)), sd->name);
46 
47 	if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
48 		printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
49 	}
50 	if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
51 		printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
52 	}
53 
54 	for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
55 		unsigned int flag = BIT(idx);
56 		unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
57 
58 		if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
59 		    !(sd->child->flags & flag))
60 			printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
61 			       sd_flag_debug[idx].name);
62 
63 		if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
64 		    !(sd->parent->flags & flag))
65 			printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
66 			       sd_flag_debug[idx].name);
67 	}
68 
69 	printk(KERN_DEBUG "%*s groups:", level + 1, "");
70 	do {
71 		if (!group) {
72 			printk("\n");
73 			printk(KERN_ERR "ERROR: group is NULL\n");
74 			break;
75 		}
76 
77 		if (!cpumask_weight(sched_group_span(group))) {
78 			printk(KERN_CONT "\n");
79 			printk(KERN_ERR "ERROR: empty group\n");
80 			break;
81 		}
82 
83 		if (!(sd->flags & SD_OVERLAP) &&
84 		    cpumask_intersects(groupmask, sched_group_span(group))) {
85 			printk(KERN_CONT "\n");
86 			printk(KERN_ERR "ERROR: repeated CPUs\n");
87 			break;
88 		}
89 
90 		cpumask_or(groupmask, groupmask, sched_group_span(group));
91 
92 		printk(KERN_CONT " %d:{ span=%*pbl",
93 				group->sgc->id,
94 				cpumask_pr_args(sched_group_span(group)));
95 
96 		if ((sd->flags & SD_OVERLAP) &&
97 		    !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
98 			printk(KERN_CONT " mask=%*pbl",
99 				cpumask_pr_args(group_balance_mask(group)));
100 		}
101 
102 		if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
103 			printk(KERN_CONT " cap=%lu", group->sgc->capacity);
104 
105 		if (group == sd->groups && sd->child &&
106 		    !cpumask_equal(sched_domain_span(sd->child),
107 				   sched_group_span(group))) {
108 			printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
109 		}
110 
111 		printk(KERN_CONT " }");
112 
113 		group = group->next;
114 
115 		if (group != sd->groups)
116 			printk(KERN_CONT ",");
117 
118 	} while (group != sd->groups);
119 	printk(KERN_CONT "\n");
120 
121 	if (!cpumask_equal(sched_domain_span(sd), groupmask))
122 		printk(KERN_ERR "ERROR: groups don't span domain->span\n");
123 
124 	if (sd->parent &&
125 	    !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
126 		printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
127 	return 0;
128 }
129 
130 static void sched_domain_debug(struct sched_domain *sd, int cpu)
131 {
132 	int level = 0;
133 
134 	if (!sched_debug_verbose)
135 		return;
136 
137 	if (!sd) {
138 		printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
139 		return;
140 	}
141 
142 	printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
143 
144 	for (;;) {
145 		if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
146 			break;
147 		level++;
148 		sd = sd->parent;
149 		if (!sd)
150 			break;
151 	}
152 }
153 #else /* !CONFIG_SCHED_DEBUG */
154 
155 # define sched_debug_verbose 0
156 # define sched_domain_debug(sd, cpu) do { } while (0)
157 static inline bool sched_debug(void)
158 {
159 	return false;
160 }
161 #endif /* CONFIG_SCHED_DEBUG */
162 
163 /* Generate a mask of SD flags with the SDF_NEEDS_GROUPS metaflag */
164 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
165 static const unsigned int SD_DEGENERATE_GROUPS_MASK =
166 #include <linux/sched/sd_flags.h>
167 0;
168 #undef SD_FLAG
169 
170 static int sd_degenerate(struct sched_domain *sd)
171 {
172 	if (cpumask_weight(sched_domain_span(sd)) == 1)
173 		return 1;
174 
175 	/* Following flags need at least 2 groups */
176 	if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
177 	    (sd->groups != sd->groups->next))
178 		return 0;
179 
180 	/* Following flags don't use groups */
181 	if (sd->flags & (SD_WAKE_AFFINE))
182 		return 0;
183 
184 	return 1;
185 }
186 
187 static int
188 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
189 {
190 	unsigned long cflags = sd->flags, pflags = parent->flags;
191 
192 	if (sd_degenerate(parent))
193 		return 1;
194 
195 	if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
196 		return 0;
197 
198 	/* Flags needing groups don't count if only 1 group in parent */
199 	if (parent->groups == parent->groups->next)
200 		pflags &= ~SD_DEGENERATE_GROUPS_MASK;
201 
202 	if (~cflags & pflags)
203 		return 0;
204 
205 	return 1;
206 }
207 
208 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
209 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
210 unsigned int sysctl_sched_energy_aware = 1;
211 DEFINE_MUTEX(sched_energy_mutex);
212 bool sched_energy_update;
213 
214 void rebuild_sched_domains_energy(void)
215 {
216 	mutex_lock(&sched_energy_mutex);
217 	sched_energy_update = true;
218 	rebuild_sched_domains();
219 	sched_energy_update = false;
220 	mutex_unlock(&sched_energy_mutex);
221 }
222 
223 #ifdef CONFIG_PROC_SYSCTL
224 int sched_energy_aware_handler(struct ctl_table *table, int write,
225 		void *buffer, size_t *lenp, loff_t *ppos)
226 {
227 	int ret, state;
228 
229 	if (write && !capable(CAP_SYS_ADMIN))
230 		return -EPERM;
231 
232 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
233 	if (!ret && write) {
234 		state = static_branch_unlikely(&sched_energy_present);
235 		if (state != sysctl_sched_energy_aware)
236 			rebuild_sched_domains_energy();
237 	}
238 
239 	return ret;
240 }
241 #endif
242 
243 static void free_pd(struct perf_domain *pd)
244 {
245 	struct perf_domain *tmp;
246 
247 	while (pd) {
248 		tmp = pd->next;
249 		kfree(pd);
250 		pd = tmp;
251 	}
252 }
253 
254 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
255 {
256 	while (pd) {
257 		if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
258 			return pd;
259 		pd = pd->next;
260 	}
261 
262 	return NULL;
263 }
264 
265 static struct perf_domain *pd_init(int cpu)
266 {
267 	struct em_perf_domain *obj = em_cpu_get(cpu);
268 	struct perf_domain *pd;
269 
270 	if (!obj) {
271 		if (sched_debug())
272 			pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
273 		return NULL;
274 	}
275 
276 	pd = kzalloc(sizeof(*pd), GFP_KERNEL);
277 	if (!pd)
278 		return NULL;
279 	pd->em_pd = obj;
280 
281 	return pd;
282 }
283 
284 static void perf_domain_debug(const struct cpumask *cpu_map,
285 						struct perf_domain *pd)
286 {
287 	if (!sched_debug() || !pd)
288 		return;
289 
290 	printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
291 
292 	while (pd) {
293 		printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
294 				cpumask_first(perf_domain_span(pd)),
295 				cpumask_pr_args(perf_domain_span(pd)),
296 				em_pd_nr_perf_states(pd->em_pd));
297 		pd = pd->next;
298 	}
299 
300 	printk(KERN_CONT "\n");
301 }
302 
303 static void destroy_perf_domain_rcu(struct rcu_head *rp)
304 {
305 	struct perf_domain *pd;
306 
307 	pd = container_of(rp, struct perf_domain, rcu);
308 	free_pd(pd);
309 }
310 
311 static void sched_energy_set(bool has_eas)
312 {
313 	if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
314 		if (sched_debug())
315 			pr_info("%s: stopping EAS\n", __func__);
316 		static_branch_disable_cpuslocked(&sched_energy_present);
317 	} else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
318 		if (sched_debug())
319 			pr_info("%s: starting EAS\n", __func__);
320 		static_branch_enable_cpuslocked(&sched_energy_present);
321 	}
322 }
323 
324 /*
325  * EAS can be used on a root domain if it meets all the following conditions:
326  *    1. an Energy Model (EM) is available;
327  *    2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
328  *    3. no SMT is detected.
329  *    4. the EM complexity is low enough to keep scheduling overheads low;
330  *    5. schedutil is driving the frequency of all CPUs of the rd;
331  *    6. frequency invariance support is present;
332  *
333  * The complexity of the Energy Model is defined as:
334  *
335  *              C = nr_pd * (nr_cpus + nr_ps)
336  *
337  * with parameters defined as:
338  *  - nr_pd:    the number of performance domains
339  *  - nr_cpus:  the number of CPUs
340  *  - nr_ps:    the sum of the number of performance states of all performance
341  *              domains (for example, on a system with 2 performance domains,
342  *              with 10 performance states each, nr_ps = 2 * 10 = 20).
343  *
344  * It is generally not a good idea to use such a model in the wake-up path on
345  * very complex platforms because of the associated scheduling overheads. The
346  * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
347  * with per-CPU DVFS and less than 8 performance states each, for example.
348  */
349 #define EM_MAX_COMPLEXITY 2048
350 
351 extern struct cpufreq_governor schedutil_gov;
352 static bool build_perf_domains(const struct cpumask *cpu_map)
353 {
354 	int i, nr_pd = 0, nr_ps = 0, nr_cpus = cpumask_weight(cpu_map);
355 	struct perf_domain *pd = NULL, *tmp;
356 	int cpu = cpumask_first(cpu_map);
357 	struct root_domain *rd = cpu_rq(cpu)->rd;
358 	struct cpufreq_policy *policy;
359 	struct cpufreq_governor *gov;
360 
361 	if (!sysctl_sched_energy_aware)
362 		goto free;
363 
364 	/* EAS is enabled for asymmetric CPU capacity topologies. */
365 	if (!per_cpu(sd_asym_cpucapacity, cpu)) {
366 		if (sched_debug()) {
367 			pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
368 					cpumask_pr_args(cpu_map));
369 		}
370 		goto free;
371 	}
372 
373 	/* EAS definitely does *not* handle SMT */
374 	if (sched_smt_active()) {
375 		pr_warn("rd %*pbl: Disabling EAS, SMT is not supported\n",
376 			cpumask_pr_args(cpu_map));
377 		goto free;
378 	}
379 
380 	if (!arch_scale_freq_invariant()) {
381 		if (sched_debug()) {
382 			pr_warn("rd %*pbl: Disabling EAS: frequency-invariant load tracking not yet supported",
383 				cpumask_pr_args(cpu_map));
384 		}
385 		goto free;
386 	}
387 
388 	for_each_cpu(i, cpu_map) {
389 		/* Skip already covered CPUs. */
390 		if (find_pd(pd, i))
391 			continue;
392 
393 		/* Do not attempt EAS if schedutil is not being used. */
394 		policy = cpufreq_cpu_get(i);
395 		if (!policy)
396 			goto free;
397 		gov = policy->governor;
398 		cpufreq_cpu_put(policy);
399 		if (gov != &schedutil_gov) {
400 			if (rd->pd)
401 				pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
402 						cpumask_pr_args(cpu_map));
403 			goto free;
404 		}
405 
406 		/* Create the new pd and add it to the local list. */
407 		tmp = pd_init(i);
408 		if (!tmp)
409 			goto free;
410 		tmp->next = pd;
411 		pd = tmp;
412 
413 		/*
414 		 * Count performance domains and performance states for the
415 		 * complexity check.
416 		 */
417 		nr_pd++;
418 		nr_ps += em_pd_nr_perf_states(pd->em_pd);
419 	}
420 
421 	/* Bail out if the Energy Model complexity is too high. */
422 	if (nr_pd * (nr_ps + nr_cpus) > EM_MAX_COMPLEXITY) {
423 		WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
424 						cpumask_pr_args(cpu_map));
425 		goto free;
426 	}
427 
428 	perf_domain_debug(cpu_map, pd);
429 
430 	/* Attach the new list of performance domains to the root domain. */
431 	tmp = rd->pd;
432 	rcu_assign_pointer(rd->pd, pd);
433 	if (tmp)
434 		call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
435 
436 	return !!pd;
437 
438 free:
439 	free_pd(pd);
440 	tmp = rd->pd;
441 	rcu_assign_pointer(rd->pd, NULL);
442 	if (tmp)
443 		call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
444 
445 	return false;
446 }
447 #else
448 static void free_pd(struct perf_domain *pd) { }
449 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
450 
451 static void free_rootdomain(struct rcu_head *rcu)
452 {
453 	struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
454 
455 	cpupri_cleanup(&rd->cpupri);
456 	cpudl_cleanup(&rd->cpudl);
457 	free_cpumask_var(rd->dlo_mask);
458 	free_cpumask_var(rd->rto_mask);
459 	free_cpumask_var(rd->online);
460 	free_cpumask_var(rd->span);
461 	free_pd(rd->pd);
462 	kfree(rd);
463 }
464 
465 void rq_attach_root(struct rq *rq, struct root_domain *rd)
466 {
467 	struct root_domain *old_rd = NULL;
468 	unsigned long flags;
469 
470 	raw_spin_lock_irqsave(&rq->lock, flags);
471 
472 	if (rq->rd) {
473 		old_rd = rq->rd;
474 
475 		if (cpumask_test_cpu(rq->cpu, old_rd->online))
476 			set_rq_offline(rq);
477 
478 		cpumask_clear_cpu(rq->cpu, old_rd->span);
479 
480 		/*
481 		 * If we dont want to free the old_rd yet then
482 		 * set old_rd to NULL to skip the freeing later
483 		 * in this function:
484 		 */
485 		if (!atomic_dec_and_test(&old_rd->refcount))
486 			old_rd = NULL;
487 	}
488 
489 	atomic_inc(&rd->refcount);
490 	rq->rd = rd;
491 
492 	cpumask_set_cpu(rq->cpu, rd->span);
493 	if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
494 		set_rq_online(rq);
495 
496 	raw_spin_unlock_irqrestore(&rq->lock, flags);
497 
498 	if (old_rd)
499 		call_rcu(&old_rd->rcu, free_rootdomain);
500 }
501 
502 void sched_get_rd(struct root_domain *rd)
503 {
504 	atomic_inc(&rd->refcount);
505 }
506 
507 void sched_put_rd(struct root_domain *rd)
508 {
509 	if (!atomic_dec_and_test(&rd->refcount))
510 		return;
511 
512 	call_rcu(&rd->rcu, free_rootdomain);
513 }
514 
515 static int init_rootdomain(struct root_domain *rd)
516 {
517 	if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
518 		goto out;
519 	if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
520 		goto free_span;
521 	if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
522 		goto free_online;
523 	if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
524 		goto free_dlo_mask;
525 
526 #ifdef HAVE_RT_PUSH_IPI
527 	rd->rto_cpu = -1;
528 	raw_spin_lock_init(&rd->rto_lock);
529 	init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
530 #endif
531 
532 	rd->visit_gen = 0;
533 	init_dl_bw(&rd->dl_bw);
534 	if (cpudl_init(&rd->cpudl) != 0)
535 		goto free_rto_mask;
536 
537 	if (cpupri_init(&rd->cpupri) != 0)
538 		goto free_cpudl;
539 	return 0;
540 
541 free_cpudl:
542 	cpudl_cleanup(&rd->cpudl);
543 free_rto_mask:
544 	free_cpumask_var(rd->rto_mask);
545 free_dlo_mask:
546 	free_cpumask_var(rd->dlo_mask);
547 free_online:
548 	free_cpumask_var(rd->online);
549 free_span:
550 	free_cpumask_var(rd->span);
551 out:
552 	return -ENOMEM;
553 }
554 
555 /*
556  * By default the system creates a single root-domain with all CPUs as
557  * members (mimicking the global state we have today).
558  */
559 struct root_domain def_root_domain;
560 
561 void init_defrootdomain(void)
562 {
563 	init_rootdomain(&def_root_domain);
564 
565 	atomic_set(&def_root_domain.refcount, 1);
566 }
567 
568 static struct root_domain *alloc_rootdomain(void)
569 {
570 	struct root_domain *rd;
571 
572 	rd = kzalloc(sizeof(*rd), GFP_KERNEL);
573 	if (!rd)
574 		return NULL;
575 
576 	if (init_rootdomain(rd) != 0) {
577 		kfree(rd);
578 		return NULL;
579 	}
580 
581 	return rd;
582 }
583 
584 static void free_sched_groups(struct sched_group *sg, int free_sgc)
585 {
586 	struct sched_group *tmp, *first;
587 
588 	if (!sg)
589 		return;
590 
591 	first = sg;
592 	do {
593 		tmp = sg->next;
594 
595 		if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
596 			kfree(sg->sgc);
597 
598 		if (atomic_dec_and_test(&sg->ref))
599 			kfree(sg);
600 		sg = tmp;
601 	} while (sg != first);
602 }
603 
604 static void destroy_sched_domain(struct sched_domain *sd)
605 {
606 	/*
607 	 * A normal sched domain may have multiple group references, an
608 	 * overlapping domain, having private groups, only one.  Iterate,
609 	 * dropping group/capacity references, freeing where none remain.
610 	 */
611 	free_sched_groups(sd->groups, 1);
612 
613 	if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
614 		kfree(sd->shared);
615 	kfree(sd);
616 }
617 
618 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
619 {
620 	struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
621 
622 	while (sd) {
623 		struct sched_domain *parent = sd->parent;
624 		destroy_sched_domain(sd);
625 		sd = parent;
626 	}
627 }
628 
629 static void destroy_sched_domains(struct sched_domain *sd)
630 {
631 	if (sd)
632 		call_rcu(&sd->rcu, destroy_sched_domains_rcu);
633 }
634 
635 /*
636  * Keep a special pointer to the highest sched_domain that has
637  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
638  * allows us to avoid some pointer chasing select_idle_sibling().
639  *
640  * Also keep a unique ID per domain (we use the first CPU number in
641  * the cpumask of the domain), this allows us to quickly tell if
642  * two CPUs are in the same cache domain, see cpus_share_cache().
643  */
644 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
645 DEFINE_PER_CPU(int, sd_llc_size);
646 DEFINE_PER_CPU(int, sd_llc_id);
647 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
648 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
649 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
650 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
651 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
652 
653 static void update_top_cache_domain(int cpu)
654 {
655 	struct sched_domain_shared *sds = NULL;
656 	struct sched_domain *sd;
657 	int id = cpu;
658 	int size = 1;
659 
660 	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
661 	if (sd) {
662 		id = cpumask_first(sched_domain_span(sd));
663 		size = cpumask_weight(sched_domain_span(sd));
664 		sds = sd->shared;
665 	}
666 
667 	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
668 	per_cpu(sd_llc_size, cpu) = size;
669 	per_cpu(sd_llc_id, cpu) = id;
670 	rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
671 
672 	sd = lowest_flag_domain(cpu, SD_NUMA);
673 	rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
674 
675 	sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
676 	rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
677 
678 	sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY);
679 	rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
680 }
681 
682 /*
683  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
684  * hold the hotplug lock.
685  */
686 static void
687 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
688 {
689 	struct rq *rq = cpu_rq(cpu);
690 	struct sched_domain *tmp;
691 	int numa_distance = 0;
692 
693 	/* Remove the sched domains which do not contribute to scheduling. */
694 	for (tmp = sd; tmp; ) {
695 		struct sched_domain *parent = tmp->parent;
696 		if (!parent)
697 			break;
698 
699 		if (sd_parent_degenerate(tmp, parent)) {
700 			tmp->parent = parent->parent;
701 			if (parent->parent)
702 				parent->parent->child = tmp;
703 			/*
704 			 * Transfer SD_PREFER_SIBLING down in case of a
705 			 * degenerate parent; the spans match for this
706 			 * so the property transfers.
707 			 */
708 			if (parent->flags & SD_PREFER_SIBLING)
709 				tmp->flags |= SD_PREFER_SIBLING;
710 			destroy_sched_domain(parent);
711 		} else
712 			tmp = tmp->parent;
713 	}
714 
715 	if (sd && sd_degenerate(sd)) {
716 		tmp = sd;
717 		sd = sd->parent;
718 		destroy_sched_domain(tmp);
719 		if (sd)
720 			sd->child = NULL;
721 	}
722 
723 	for (tmp = sd; tmp; tmp = tmp->parent)
724 		numa_distance += !!(tmp->flags & SD_NUMA);
725 
726 	sched_domain_debug(sd, cpu);
727 
728 	rq_attach_root(rq, rd);
729 	tmp = rq->sd;
730 	rcu_assign_pointer(rq->sd, sd);
731 	dirty_sched_domain_sysctl(cpu);
732 	destroy_sched_domains(tmp);
733 
734 	update_top_cache_domain(cpu);
735 }
736 
737 struct s_data {
738 	struct sched_domain * __percpu *sd;
739 	struct root_domain	*rd;
740 };
741 
742 enum s_alloc {
743 	sa_rootdomain,
744 	sa_sd,
745 	sa_sd_storage,
746 	sa_none,
747 };
748 
749 /*
750  * Return the canonical balance CPU for this group, this is the first CPU
751  * of this group that's also in the balance mask.
752  *
753  * The balance mask are all those CPUs that could actually end up at this
754  * group. See build_balance_mask().
755  *
756  * Also see should_we_balance().
757  */
758 int group_balance_cpu(struct sched_group *sg)
759 {
760 	return cpumask_first(group_balance_mask(sg));
761 }
762 
763 
764 /*
765  * NUMA topology (first read the regular topology blurb below)
766  *
767  * Given a node-distance table, for example:
768  *
769  *   node   0   1   2   3
770  *     0:  10  20  30  20
771  *     1:  20  10  20  30
772  *     2:  30  20  10  20
773  *     3:  20  30  20  10
774  *
775  * which represents a 4 node ring topology like:
776  *
777  *   0 ----- 1
778  *   |       |
779  *   |       |
780  *   |       |
781  *   3 ----- 2
782  *
783  * We want to construct domains and groups to represent this. The way we go
784  * about doing this is to build the domains on 'hops'. For each NUMA level we
785  * construct the mask of all nodes reachable in @level hops.
786  *
787  * For the above NUMA topology that gives 3 levels:
788  *
789  * NUMA-2	0-3		0-3		0-3		0-3
790  *  groups:	{0-1,3},{1-3}	{0-2},{0,2-3}	{1-3},{0-1,3}	{0,2-3},{0-2}
791  *
792  * NUMA-1	0-1,3		0-2		1-3		0,2-3
793  *  groups:	{0},{1},{3}	{0},{1},{2}	{1},{2},{3}	{0},{2},{3}
794  *
795  * NUMA-0	0		1		2		3
796  *
797  *
798  * As can be seen; things don't nicely line up as with the regular topology.
799  * When we iterate a domain in child domain chunks some nodes can be
800  * represented multiple times -- hence the "overlap" naming for this part of
801  * the topology.
802  *
803  * In order to minimize this overlap, we only build enough groups to cover the
804  * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
805  *
806  * Because:
807  *
808  *  - the first group of each domain is its child domain; this
809  *    gets us the first 0-1,3
810  *  - the only uncovered node is 2, who's child domain is 1-3.
811  *
812  * However, because of the overlap, computing a unique CPU for each group is
813  * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
814  * groups include the CPUs of Node-0, while those CPUs would not in fact ever
815  * end up at those groups (they would end up in group: 0-1,3).
816  *
817  * To correct this we have to introduce the group balance mask. This mask
818  * will contain those CPUs in the group that can reach this group given the
819  * (child) domain tree.
820  *
821  * With this we can once again compute balance_cpu and sched_group_capacity
822  * relations.
823  *
824  * XXX include words on how balance_cpu is unique and therefore can be
825  * used for sched_group_capacity links.
826  *
827  *
828  * Another 'interesting' topology is:
829  *
830  *   node   0   1   2   3
831  *     0:  10  20  20  30
832  *     1:  20  10  20  20
833  *     2:  20  20  10  20
834  *     3:  30  20  20  10
835  *
836  * Which looks a little like:
837  *
838  *   0 ----- 1
839  *   |     / |
840  *   |   /   |
841  *   | /     |
842  *   2 ----- 3
843  *
844  * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
845  * are not.
846  *
847  * This leads to a few particularly weird cases where the sched_domain's are
848  * not of the same number for each CPU. Consider:
849  *
850  * NUMA-2	0-3						0-3
851  *  groups:	{0-2},{1-3}					{1-3},{0-2}
852  *
853  * NUMA-1	0-2		0-3		0-3		1-3
854  *
855  * NUMA-0	0		1		2		3
856  *
857  */
858 
859 
860 /*
861  * Build the balance mask; it contains only those CPUs that can arrive at this
862  * group and should be considered to continue balancing.
863  *
864  * We do this during the group creation pass, therefore the group information
865  * isn't complete yet, however since each group represents a (child) domain we
866  * can fully construct this using the sched_domain bits (which are already
867  * complete).
868  */
869 static void
870 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
871 {
872 	const struct cpumask *sg_span = sched_group_span(sg);
873 	struct sd_data *sdd = sd->private;
874 	struct sched_domain *sibling;
875 	int i;
876 
877 	cpumask_clear(mask);
878 
879 	for_each_cpu(i, sg_span) {
880 		sibling = *per_cpu_ptr(sdd->sd, i);
881 
882 		/*
883 		 * Can happen in the asymmetric case, where these siblings are
884 		 * unused. The mask will not be empty because those CPUs that
885 		 * do have the top domain _should_ span the domain.
886 		 */
887 		if (!sibling->child)
888 			continue;
889 
890 		/* If we would not end up here, we can't continue from here */
891 		if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
892 			continue;
893 
894 		cpumask_set_cpu(i, mask);
895 	}
896 
897 	/* We must not have empty masks here */
898 	WARN_ON_ONCE(cpumask_empty(mask));
899 }
900 
901 /*
902  * XXX: This creates per-node group entries; since the load-balancer will
903  * immediately access remote memory to construct this group's load-balance
904  * statistics having the groups node local is of dubious benefit.
905  */
906 static struct sched_group *
907 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
908 {
909 	struct sched_group *sg;
910 	struct cpumask *sg_span;
911 
912 	sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
913 			GFP_KERNEL, cpu_to_node(cpu));
914 
915 	if (!sg)
916 		return NULL;
917 
918 	sg_span = sched_group_span(sg);
919 	if (sd->child)
920 		cpumask_copy(sg_span, sched_domain_span(sd->child));
921 	else
922 		cpumask_copy(sg_span, sched_domain_span(sd));
923 
924 	atomic_inc(&sg->ref);
925 	return sg;
926 }
927 
928 static void init_overlap_sched_group(struct sched_domain *sd,
929 				     struct sched_group *sg)
930 {
931 	struct cpumask *mask = sched_domains_tmpmask2;
932 	struct sd_data *sdd = sd->private;
933 	struct cpumask *sg_span;
934 	int cpu;
935 
936 	build_balance_mask(sd, sg, mask);
937 	cpu = cpumask_first(mask);
938 
939 	sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
940 	if (atomic_inc_return(&sg->sgc->ref) == 1)
941 		cpumask_copy(group_balance_mask(sg), mask);
942 	else
943 		WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
944 
945 	/*
946 	 * Initialize sgc->capacity such that even if we mess up the
947 	 * domains and no possible iteration will get us here, we won't
948 	 * die on a /0 trap.
949 	 */
950 	sg_span = sched_group_span(sg);
951 	sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
952 	sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
953 	sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
954 }
955 
956 static struct sched_domain *
957 find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
958 {
959 	/*
960 	 * The proper descendant would be the one whose child won't span out
961 	 * of sd
962 	 */
963 	while (sibling->child &&
964 	       !cpumask_subset(sched_domain_span(sibling->child),
965 			       sched_domain_span(sd)))
966 		sibling = sibling->child;
967 
968 	/*
969 	 * As we are referencing sgc across different topology level, we need
970 	 * to go down to skip those sched_domains which don't contribute to
971 	 * scheduling because they will be degenerated in cpu_attach_domain
972 	 */
973 	while (sibling->child &&
974 	       cpumask_equal(sched_domain_span(sibling->child),
975 			     sched_domain_span(sibling)))
976 		sibling = sibling->child;
977 
978 	return sibling;
979 }
980 
981 static int
982 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
983 {
984 	struct sched_group *first = NULL, *last = NULL, *sg;
985 	const struct cpumask *span = sched_domain_span(sd);
986 	struct cpumask *covered = sched_domains_tmpmask;
987 	struct sd_data *sdd = sd->private;
988 	struct sched_domain *sibling;
989 	int i;
990 
991 	cpumask_clear(covered);
992 
993 	for_each_cpu_wrap(i, span, cpu) {
994 		struct cpumask *sg_span;
995 
996 		if (cpumask_test_cpu(i, covered))
997 			continue;
998 
999 		sibling = *per_cpu_ptr(sdd->sd, i);
1000 
1001 		/*
1002 		 * Asymmetric node setups can result in situations where the
1003 		 * domain tree is of unequal depth, make sure to skip domains
1004 		 * that already cover the entire range.
1005 		 *
1006 		 * In that case build_sched_domains() will have terminated the
1007 		 * iteration early and our sibling sd spans will be empty.
1008 		 * Domains should always include the CPU they're built on, so
1009 		 * check that.
1010 		 */
1011 		if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1012 			continue;
1013 
1014 		/*
1015 		 * Usually we build sched_group by sibling's child sched_domain
1016 		 * But for machines whose NUMA diameter are 3 or above, we move
1017 		 * to build sched_group by sibling's proper descendant's child
1018 		 * domain because sibling's child sched_domain will span out of
1019 		 * the sched_domain being built as below.
1020 		 *
1021 		 * Smallest diameter=3 topology is:
1022 		 *
1023 		 *   node   0   1   2   3
1024 		 *     0:  10  20  30  40
1025 		 *     1:  20  10  20  30
1026 		 *     2:  30  20  10  20
1027 		 *     3:  40  30  20  10
1028 		 *
1029 		 *   0 --- 1 --- 2 --- 3
1030 		 *
1031 		 * NUMA-3       0-3             N/A             N/A             0-3
1032 		 *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
1033 		 *
1034 		 * NUMA-2       0-2             0-3             0-3             1-3
1035 		 *  groups:     {0-1},{1-3}     {0-2},{2-3}     {1-3},{0-1}     {2-3},{0-2}
1036 		 *
1037 		 * NUMA-1       0-1             0-2             1-3             2-3
1038 		 *  groups:     {0},{1}         {1},{2},{0}     {2},{3},{1}     {3},{2}
1039 		 *
1040 		 * NUMA-0       0               1               2               3
1041 		 *
1042 		 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1043 		 * group span isn't a subset of the domain span.
1044 		 */
1045 		if (sibling->child &&
1046 		    !cpumask_subset(sched_domain_span(sibling->child), span))
1047 			sibling = find_descended_sibling(sd, sibling);
1048 
1049 		sg = build_group_from_child_sched_domain(sibling, cpu);
1050 		if (!sg)
1051 			goto fail;
1052 
1053 		sg_span = sched_group_span(sg);
1054 		cpumask_or(covered, covered, sg_span);
1055 
1056 		init_overlap_sched_group(sibling, sg);
1057 
1058 		if (!first)
1059 			first = sg;
1060 		if (last)
1061 			last->next = sg;
1062 		last = sg;
1063 		last->next = first;
1064 	}
1065 	sd->groups = first;
1066 
1067 	return 0;
1068 
1069 fail:
1070 	free_sched_groups(first, 0);
1071 
1072 	return -ENOMEM;
1073 }
1074 
1075 
1076 /*
1077  * Package topology (also see the load-balance blurb in fair.c)
1078  *
1079  * The scheduler builds a tree structure to represent a number of important
1080  * topology features. By default (default_topology[]) these include:
1081  *
1082  *  - Simultaneous multithreading (SMT)
1083  *  - Multi-Core Cache (MC)
1084  *  - Package (DIE)
1085  *
1086  * Where the last one more or less denotes everything up to a NUMA node.
1087  *
1088  * The tree consists of 3 primary data structures:
1089  *
1090  *	sched_domain -> sched_group -> sched_group_capacity
1091  *	    ^ ^             ^ ^
1092  *          `-'             `-'
1093  *
1094  * The sched_domains are per-CPU and have a two way link (parent & child) and
1095  * denote the ever growing mask of CPUs belonging to that level of topology.
1096  *
1097  * Each sched_domain has a circular (double) linked list of sched_group's, each
1098  * denoting the domains of the level below (or individual CPUs in case of the
1099  * first domain level). The sched_group linked by a sched_domain includes the
1100  * CPU of that sched_domain [*].
1101  *
1102  * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1103  *
1104  * CPU   0   1   2   3   4   5   6   7
1105  *
1106  * DIE  [                             ]
1107  * MC   [             ] [             ]
1108  * SMT  [     ] [     ] [     ] [     ]
1109  *
1110  *  - or -
1111  *
1112  * DIE  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1113  * MC	0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1114  * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1115  *
1116  * CPU   0   1   2   3   4   5   6   7
1117  *
1118  * One way to think about it is: sched_domain moves you up and down among these
1119  * topology levels, while sched_group moves you sideways through it, at child
1120  * domain granularity.
1121  *
1122  * sched_group_capacity ensures each unique sched_group has shared storage.
1123  *
1124  * There are two related construction problems, both require a CPU that
1125  * uniquely identify each group (for a given domain):
1126  *
1127  *  - The first is the balance_cpu (see should_we_balance() and the
1128  *    load-balance blub in fair.c); for each group we only want 1 CPU to
1129  *    continue balancing at a higher domain.
1130  *
1131  *  - The second is the sched_group_capacity; we want all identical groups
1132  *    to share a single sched_group_capacity.
1133  *
1134  * Since these topologies are exclusive by construction. That is, its
1135  * impossible for an SMT thread to belong to multiple cores, and cores to
1136  * be part of multiple caches. There is a very clear and unique location
1137  * for each CPU in the hierarchy.
1138  *
1139  * Therefore computing a unique CPU for each group is trivial (the iteration
1140  * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1141  * group), we can simply pick the first CPU in each group.
1142  *
1143  *
1144  * [*] in other words, the first group of each domain is its child domain.
1145  */
1146 
1147 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1148 {
1149 	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1150 	struct sched_domain *child = sd->child;
1151 	struct sched_group *sg;
1152 	bool already_visited;
1153 
1154 	if (child)
1155 		cpu = cpumask_first(sched_domain_span(child));
1156 
1157 	sg = *per_cpu_ptr(sdd->sg, cpu);
1158 	sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1159 
1160 	/* Increase refcounts for claim_allocations: */
1161 	already_visited = atomic_inc_return(&sg->ref) > 1;
1162 	/* sgc visits should follow a similar trend as sg */
1163 	WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1164 
1165 	/* If we have already visited that group, it's already initialized. */
1166 	if (already_visited)
1167 		return sg;
1168 
1169 	if (child) {
1170 		cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1171 		cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1172 	} else {
1173 		cpumask_set_cpu(cpu, sched_group_span(sg));
1174 		cpumask_set_cpu(cpu, group_balance_mask(sg));
1175 	}
1176 
1177 	sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1178 	sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1179 	sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1180 
1181 	return sg;
1182 }
1183 
1184 /*
1185  * build_sched_groups will build a circular linked list of the groups
1186  * covered by the given span, will set each group's ->cpumask correctly,
1187  * and will initialize their ->sgc.
1188  *
1189  * Assumes the sched_domain tree is fully constructed
1190  */
1191 static int
1192 build_sched_groups(struct sched_domain *sd, int cpu)
1193 {
1194 	struct sched_group *first = NULL, *last = NULL;
1195 	struct sd_data *sdd = sd->private;
1196 	const struct cpumask *span = sched_domain_span(sd);
1197 	struct cpumask *covered;
1198 	int i;
1199 
1200 	lockdep_assert_held(&sched_domains_mutex);
1201 	covered = sched_domains_tmpmask;
1202 
1203 	cpumask_clear(covered);
1204 
1205 	for_each_cpu_wrap(i, span, cpu) {
1206 		struct sched_group *sg;
1207 
1208 		if (cpumask_test_cpu(i, covered))
1209 			continue;
1210 
1211 		sg = get_group(i, sdd);
1212 
1213 		cpumask_or(covered, covered, sched_group_span(sg));
1214 
1215 		if (!first)
1216 			first = sg;
1217 		if (last)
1218 			last->next = sg;
1219 		last = sg;
1220 	}
1221 	last->next = first;
1222 	sd->groups = first;
1223 
1224 	return 0;
1225 }
1226 
1227 /*
1228  * Initialize sched groups cpu_capacity.
1229  *
1230  * cpu_capacity indicates the capacity of sched group, which is used while
1231  * distributing the load between different sched groups in a sched domain.
1232  * Typically cpu_capacity for all the groups in a sched domain will be same
1233  * unless there are asymmetries in the topology. If there are asymmetries,
1234  * group having more cpu_capacity will pickup more load compared to the
1235  * group having less cpu_capacity.
1236  */
1237 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1238 {
1239 	struct sched_group *sg = sd->groups;
1240 
1241 	WARN_ON(!sg);
1242 
1243 	do {
1244 		int cpu, max_cpu = -1;
1245 
1246 		sg->group_weight = cpumask_weight(sched_group_span(sg));
1247 
1248 		if (!(sd->flags & SD_ASYM_PACKING))
1249 			goto next;
1250 
1251 		for_each_cpu(cpu, sched_group_span(sg)) {
1252 			if (max_cpu < 0)
1253 				max_cpu = cpu;
1254 			else if (sched_asym_prefer(cpu, max_cpu))
1255 				max_cpu = cpu;
1256 		}
1257 		sg->asym_prefer_cpu = max_cpu;
1258 
1259 next:
1260 		sg = sg->next;
1261 	} while (sg != sd->groups);
1262 
1263 	if (cpu != group_balance_cpu(sg))
1264 		return;
1265 
1266 	update_group_capacity(sd, cpu);
1267 }
1268 
1269 /*
1270  * Initializers for schedule domains
1271  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1272  */
1273 
1274 static int default_relax_domain_level = -1;
1275 int sched_domain_level_max;
1276 
1277 static int __init setup_relax_domain_level(char *str)
1278 {
1279 	if (kstrtoint(str, 0, &default_relax_domain_level))
1280 		pr_warn("Unable to set relax_domain_level\n");
1281 
1282 	return 1;
1283 }
1284 __setup("relax_domain_level=", setup_relax_domain_level);
1285 
1286 static void set_domain_attribute(struct sched_domain *sd,
1287 				 struct sched_domain_attr *attr)
1288 {
1289 	int request;
1290 
1291 	if (!attr || attr->relax_domain_level < 0) {
1292 		if (default_relax_domain_level < 0)
1293 			return;
1294 		request = default_relax_domain_level;
1295 	} else
1296 		request = attr->relax_domain_level;
1297 
1298 	if (sd->level > request) {
1299 		/* Turn off idle balance on this domain: */
1300 		sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1301 	}
1302 }
1303 
1304 static void __sdt_free(const struct cpumask *cpu_map);
1305 static int __sdt_alloc(const struct cpumask *cpu_map);
1306 
1307 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1308 				 const struct cpumask *cpu_map)
1309 {
1310 	switch (what) {
1311 	case sa_rootdomain:
1312 		if (!atomic_read(&d->rd->refcount))
1313 			free_rootdomain(&d->rd->rcu);
1314 		fallthrough;
1315 	case sa_sd:
1316 		free_percpu(d->sd);
1317 		fallthrough;
1318 	case sa_sd_storage:
1319 		__sdt_free(cpu_map);
1320 		fallthrough;
1321 	case sa_none:
1322 		break;
1323 	}
1324 }
1325 
1326 static enum s_alloc
1327 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1328 {
1329 	memset(d, 0, sizeof(*d));
1330 
1331 	if (__sdt_alloc(cpu_map))
1332 		return sa_sd_storage;
1333 	d->sd = alloc_percpu(struct sched_domain *);
1334 	if (!d->sd)
1335 		return sa_sd_storage;
1336 	d->rd = alloc_rootdomain();
1337 	if (!d->rd)
1338 		return sa_sd;
1339 
1340 	return sa_rootdomain;
1341 }
1342 
1343 /*
1344  * NULL the sd_data elements we've used to build the sched_domain and
1345  * sched_group structure so that the subsequent __free_domain_allocs()
1346  * will not free the data we're using.
1347  */
1348 static void claim_allocations(int cpu, struct sched_domain *sd)
1349 {
1350 	struct sd_data *sdd = sd->private;
1351 
1352 	WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1353 	*per_cpu_ptr(sdd->sd, cpu) = NULL;
1354 
1355 	if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1356 		*per_cpu_ptr(sdd->sds, cpu) = NULL;
1357 
1358 	if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1359 		*per_cpu_ptr(sdd->sg, cpu) = NULL;
1360 
1361 	if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1362 		*per_cpu_ptr(sdd->sgc, cpu) = NULL;
1363 }
1364 
1365 #ifdef CONFIG_NUMA
1366 enum numa_topology_type sched_numa_topology_type;
1367 
1368 static int			sched_domains_numa_levels;
1369 static int			sched_domains_curr_level;
1370 
1371 int				sched_max_numa_distance;
1372 static int			*sched_domains_numa_distance;
1373 static struct cpumask		***sched_domains_numa_masks;
1374 int __read_mostly		node_reclaim_distance = RECLAIM_DISTANCE;
1375 #endif
1376 
1377 /*
1378  * SD_flags allowed in topology descriptions.
1379  *
1380  * These flags are purely descriptive of the topology and do not prescribe
1381  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1382  * function:
1383  *
1384  *   SD_SHARE_CPUCAPACITY   - describes SMT topologies
1385  *   SD_SHARE_PKG_RESOURCES - describes shared caches
1386  *   SD_NUMA                - describes NUMA topologies
1387  *
1388  * Odd one out, which beside describing the topology has a quirk also
1389  * prescribes the desired behaviour that goes along with it:
1390  *
1391  *   SD_ASYM_PACKING        - describes SMT quirks
1392  */
1393 #define TOPOLOGY_SD_FLAGS		\
1394 	(SD_SHARE_CPUCAPACITY	|	\
1395 	 SD_SHARE_PKG_RESOURCES |	\
1396 	 SD_NUMA		|	\
1397 	 SD_ASYM_PACKING)
1398 
1399 static struct sched_domain *
1400 sd_init(struct sched_domain_topology_level *tl,
1401 	const struct cpumask *cpu_map,
1402 	struct sched_domain *child, int dflags, int cpu)
1403 {
1404 	struct sd_data *sdd = &tl->data;
1405 	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1406 	int sd_id, sd_weight, sd_flags = 0;
1407 
1408 #ifdef CONFIG_NUMA
1409 	/*
1410 	 * Ugly hack to pass state to sd_numa_mask()...
1411 	 */
1412 	sched_domains_curr_level = tl->numa_level;
1413 #endif
1414 
1415 	sd_weight = cpumask_weight(tl->mask(cpu));
1416 
1417 	if (tl->sd_flags)
1418 		sd_flags = (*tl->sd_flags)();
1419 	if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1420 			"wrong sd_flags in topology description\n"))
1421 		sd_flags &= TOPOLOGY_SD_FLAGS;
1422 
1423 	/* Apply detected topology flags */
1424 	sd_flags |= dflags;
1425 
1426 	*sd = (struct sched_domain){
1427 		.min_interval		= sd_weight,
1428 		.max_interval		= 2*sd_weight,
1429 		.busy_factor		= 16,
1430 		.imbalance_pct		= 117,
1431 
1432 		.cache_nice_tries	= 0,
1433 
1434 		.flags			= 1*SD_BALANCE_NEWIDLE
1435 					| 1*SD_BALANCE_EXEC
1436 					| 1*SD_BALANCE_FORK
1437 					| 0*SD_BALANCE_WAKE
1438 					| 1*SD_WAKE_AFFINE
1439 					| 0*SD_SHARE_CPUCAPACITY
1440 					| 0*SD_SHARE_PKG_RESOURCES
1441 					| 0*SD_SERIALIZE
1442 					| 1*SD_PREFER_SIBLING
1443 					| 0*SD_NUMA
1444 					| sd_flags
1445 					,
1446 
1447 		.last_balance		= jiffies,
1448 		.balance_interval	= sd_weight,
1449 		.max_newidle_lb_cost	= 0,
1450 		.next_decay_max_lb_cost	= jiffies,
1451 		.child			= child,
1452 #ifdef CONFIG_SCHED_DEBUG
1453 		.name			= tl->name,
1454 #endif
1455 	};
1456 
1457 	cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1458 	sd_id = cpumask_first(sched_domain_span(sd));
1459 
1460 	/*
1461 	 * Convert topological properties into behaviour.
1462 	 */
1463 
1464 	/* Don't attempt to spread across CPUs of different capacities. */
1465 	if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1466 		sd->child->flags &= ~SD_PREFER_SIBLING;
1467 
1468 	if (sd->flags & SD_SHARE_CPUCAPACITY) {
1469 		sd->imbalance_pct = 110;
1470 
1471 	} else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1472 		sd->imbalance_pct = 117;
1473 		sd->cache_nice_tries = 1;
1474 
1475 #ifdef CONFIG_NUMA
1476 	} else if (sd->flags & SD_NUMA) {
1477 		sd->cache_nice_tries = 2;
1478 
1479 		sd->flags &= ~SD_PREFER_SIBLING;
1480 		sd->flags |= SD_SERIALIZE;
1481 		if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1482 			sd->flags &= ~(SD_BALANCE_EXEC |
1483 				       SD_BALANCE_FORK |
1484 				       SD_WAKE_AFFINE);
1485 		}
1486 
1487 #endif
1488 	} else {
1489 		sd->cache_nice_tries = 1;
1490 	}
1491 
1492 	/*
1493 	 * For all levels sharing cache; connect a sched_domain_shared
1494 	 * instance.
1495 	 */
1496 	if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1497 		sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1498 		atomic_inc(&sd->shared->ref);
1499 		atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1500 	}
1501 
1502 	sd->private = sdd;
1503 
1504 	return sd;
1505 }
1506 
1507 /*
1508  * Topology list, bottom-up.
1509  */
1510 static struct sched_domain_topology_level default_topology[] = {
1511 #ifdef CONFIG_SCHED_SMT
1512 	{ cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1513 #endif
1514 #ifdef CONFIG_SCHED_MC
1515 	{ cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1516 #endif
1517 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
1518 	{ NULL, },
1519 };
1520 
1521 static struct sched_domain_topology_level *sched_domain_topology =
1522 	default_topology;
1523 
1524 #define for_each_sd_topology(tl)			\
1525 	for (tl = sched_domain_topology; tl->mask; tl++)
1526 
1527 void set_sched_topology(struct sched_domain_topology_level *tl)
1528 {
1529 	if (WARN_ON_ONCE(sched_smp_initialized))
1530 		return;
1531 
1532 	sched_domain_topology = tl;
1533 }
1534 
1535 #ifdef CONFIG_NUMA
1536 
1537 static const struct cpumask *sd_numa_mask(int cpu)
1538 {
1539 	return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1540 }
1541 
1542 static void sched_numa_warn(const char *str)
1543 {
1544 	static int done = false;
1545 	int i,j;
1546 
1547 	if (done)
1548 		return;
1549 
1550 	done = true;
1551 
1552 	printk(KERN_WARNING "ERROR: %s\n\n", str);
1553 
1554 	for (i = 0; i < nr_node_ids; i++) {
1555 		printk(KERN_WARNING "  ");
1556 		for (j = 0; j < nr_node_ids; j++)
1557 			printk(KERN_CONT "%02d ", node_distance(i,j));
1558 		printk(KERN_CONT "\n");
1559 	}
1560 	printk(KERN_WARNING "\n");
1561 }
1562 
1563 bool find_numa_distance(int distance)
1564 {
1565 	int i;
1566 
1567 	if (distance == node_distance(0, 0))
1568 		return true;
1569 
1570 	for (i = 0; i < sched_domains_numa_levels; i++) {
1571 		if (sched_domains_numa_distance[i] == distance)
1572 			return true;
1573 	}
1574 
1575 	return false;
1576 }
1577 
1578 /*
1579  * A system can have three types of NUMA topology:
1580  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1581  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1582  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1583  *
1584  * The difference between a glueless mesh topology and a backplane
1585  * topology lies in whether communication between not directly
1586  * connected nodes goes through intermediary nodes (where programs
1587  * could run), or through backplane controllers. This affects
1588  * placement of programs.
1589  *
1590  * The type of topology can be discerned with the following tests:
1591  * - If the maximum distance between any nodes is 1 hop, the system
1592  *   is directly connected.
1593  * - If for two nodes A and B, located N > 1 hops away from each other,
1594  *   there is an intermediary node C, which is < N hops away from both
1595  *   nodes A and B, the system is a glueless mesh.
1596  */
1597 static void init_numa_topology_type(void)
1598 {
1599 	int a, b, c, n;
1600 
1601 	n = sched_max_numa_distance;
1602 
1603 	if (sched_domains_numa_levels <= 2) {
1604 		sched_numa_topology_type = NUMA_DIRECT;
1605 		return;
1606 	}
1607 
1608 	for_each_online_node(a) {
1609 		for_each_online_node(b) {
1610 			/* Find two nodes furthest removed from each other. */
1611 			if (node_distance(a, b) < n)
1612 				continue;
1613 
1614 			/* Is there an intermediary node between a and b? */
1615 			for_each_online_node(c) {
1616 				if (node_distance(a, c) < n &&
1617 				    node_distance(b, c) < n) {
1618 					sched_numa_topology_type =
1619 							NUMA_GLUELESS_MESH;
1620 					return;
1621 				}
1622 			}
1623 
1624 			sched_numa_topology_type = NUMA_BACKPLANE;
1625 			return;
1626 		}
1627 	}
1628 }
1629 
1630 
1631 #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1632 
1633 void sched_init_numa(void)
1634 {
1635 	struct sched_domain_topology_level *tl;
1636 	unsigned long *distance_map;
1637 	int nr_levels = 0;
1638 	int i, j;
1639 
1640 	/*
1641 	 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1642 	 * unique distances in the node_distance() table.
1643 	 */
1644 	distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1645 	if (!distance_map)
1646 		return;
1647 
1648 	bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1649 	for (i = 0; i < nr_node_ids; i++) {
1650 		for (j = 0; j < nr_node_ids; j++) {
1651 			int distance = node_distance(i, j);
1652 
1653 			if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1654 				sched_numa_warn("Invalid distance value range");
1655 				return;
1656 			}
1657 
1658 			bitmap_set(distance_map, distance, 1);
1659 		}
1660 	}
1661 	/*
1662 	 * We can now figure out how many unique distance values there are and
1663 	 * allocate memory accordingly.
1664 	 */
1665 	nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1666 
1667 	sched_domains_numa_distance = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1668 	if (!sched_domains_numa_distance) {
1669 		bitmap_free(distance_map);
1670 		return;
1671 	}
1672 
1673 	for (i = 0, j = 0; i < nr_levels; i++, j++) {
1674 		j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1675 		sched_domains_numa_distance[i] = j;
1676 	}
1677 
1678 	bitmap_free(distance_map);
1679 
1680 	/*
1681 	 * 'nr_levels' contains the number of unique distances
1682 	 *
1683 	 * The sched_domains_numa_distance[] array includes the actual distance
1684 	 * numbers.
1685 	 */
1686 
1687 	/*
1688 	 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1689 	 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1690 	 * the array will contain less then 'nr_levels' members. This could be
1691 	 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1692 	 * in other functions.
1693 	 *
1694 	 * We reset it to 'nr_levels' at the end of this function.
1695 	 */
1696 	sched_domains_numa_levels = 0;
1697 
1698 	sched_domains_numa_masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1699 	if (!sched_domains_numa_masks)
1700 		return;
1701 
1702 	/*
1703 	 * Now for each level, construct a mask per node which contains all
1704 	 * CPUs of nodes that are that many hops away from us.
1705 	 */
1706 	for (i = 0; i < nr_levels; i++) {
1707 		sched_domains_numa_masks[i] =
1708 			kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1709 		if (!sched_domains_numa_masks[i])
1710 			return;
1711 
1712 		for (j = 0; j < nr_node_ids; j++) {
1713 			struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1714 			int k;
1715 
1716 			if (!mask)
1717 				return;
1718 
1719 			sched_domains_numa_masks[i][j] = mask;
1720 
1721 			for_each_node(k) {
1722 				if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1723 					sched_numa_warn("Node-distance not symmetric");
1724 
1725 				if (node_distance(j, k) > sched_domains_numa_distance[i])
1726 					continue;
1727 
1728 				cpumask_or(mask, mask, cpumask_of_node(k));
1729 			}
1730 		}
1731 	}
1732 
1733 	/* Compute default topology size */
1734 	for (i = 0; sched_domain_topology[i].mask; i++);
1735 
1736 	tl = kzalloc((i + nr_levels + 1) *
1737 			sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1738 	if (!tl)
1739 		return;
1740 
1741 	/*
1742 	 * Copy the default topology bits..
1743 	 */
1744 	for (i = 0; sched_domain_topology[i].mask; i++)
1745 		tl[i] = sched_domain_topology[i];
1746 
1747 	/*
1748 	 * Add the NUMA identity distance, aka single NODE.
1749 	 */
1750 	tl[i++] = (struct sched_domain_topology_level){
1751 		.mask = sd_numa_mask,
1752 		.numa_level = 0,
1753 		SD_INIT_NAME(NODE)
1754 	};
1755 
1756 	/*
1757 	 * .. and append 'j' levels of NUMA goodness.
1758 	 */
1759 	for (j = 1; j < nr_levels; i++, j++) {
1760 		tl[i] = (struct sched_domain_topology_level){
1761 			.mask = sd_numa_mask,
1762 			.sd_flags = cpu_numa_flags,
1763 			.flags = SDTL_OVERLAP,
1764 			.numa_level = j,
1765 			SD_INIT_NAME(NUMA)
1766 		};
1767 	}
1768 
1769 	sched_domain_topology = tl;
1770 
1771 	sched_domains_numa_levels = nr_levels;
1772 	sched_max_numa_distance = sched_domains_numa_distance[nr_levels - 1];
1773 
1774 	init_numa_topology_type();
1775 }
1776 
1777 void sched_domains_numa_masks_set(unsigned int cpu)
1778 {
1779 	int node = cpu_to_node(cpu);
1780 	int i, j;
1781 
1782 	for (i = 0; i < sched_domains_numa_levels; i++) {
1783 		for (j = 0; j < nr_node_ids; j++) {
1784 			if (node_distance(j, node) <= sched_domains_numa_distance[i])
1785 				cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1786 		}
1787 	}
1788 }
1789 
1790 void sched_domains_numa_masks_clear(unsigned int cpu)
1791 {
1792 	int i, j;
1793 
1794 	for (i = 0; i < sched_domains_numa_levels; i++) {
1795 		for (j = 0; j < nr_node_ids; j++)
1796 			cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1797 	}
1798 }
1799 
1800 /*
1801  * sched_numa_find_closest() - given the NUMA topology, find the cpu
1802  *                             closest to @cpu from @cpumask.
1803  * cpumask: cpumask to find a cpu from
1804  * cpu: cpu to be close to
1805  *
1806  * returns: cpu, or nr_cpu_ids when nothing found.
1807  */
1808 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1809 {
1810 	int i, j = cpu_to_node(cpu);
1811 
1812 	for (i = 0; i < sched_domains_numa_levels; i++) {
1813 		cpu = cpumask_any_and(cpus, sched_domains_numa_masks[i][j]);
1814 		if (cpu < nr_cpu_ids)
1815 			return cpu;
1816 	}
1817 	return nr_cpu_ids;
1818 }
1819 
1820 #endif /* CONFIG_NUMA */
1821 
1822 static int __sdt_alloc(const struct cpumask *cpu_map)
1823 {
1824 	struct sched_domain_topology_level *tl;
1825 	int j;
1826 
1827 	for_each_sd_topology(tl) {
1828 		struct sd_data *sdd = &tl->data;
1829 
1830 		sdd->sd = alloc_percpu(struct sched_domain *);
1831 		if (!sdd->sd)
1832 			return -ENOMEM;
1833 
1834 		sdd->sds = alloc_percpu(struct sched_domain_shared *);
1835 		if (!sdd->sds)
1836 			return -ENOMEM;
1837 
1838 		sdd->sg = alloc_percpu(struct sched_group *);
1839 		if (!sdd->sg)
1840 			return -ENOMEM;
1841 
1842 		sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1843 		if (!sdd->sgc)
1844 			return -ENOMEM;
1845 
1846 		for_each_cpu(j, cpu_map) {
1847 			struct sched_domain *sd;
1848 			struct sched_domain_shared *sds;
1849 			struct sched_group *sg;
1850 			struct sched_group_capacity *sgc;
1851 
1852 			sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1853 					GFP_KERNEL, cpu_to_node(j));
1854 			if (!sd)
1855 				return -ENOMEM;
1856 
1857 			*per_cpu_ptr(sdd->sd, j) = sd;
1858 
1859 			sds = kzalloc_node(sizeof(struct sched_domain_shared),
1860 					GFP_KERNEL, cpu_to_node(j));
1861 			if (!sds)
1862 				return -ENOMEM;
1863 
1864 			*per_cpu_ptr(sdd->sds, j) = sds;
1865 
1866 			sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1867 					GFP_KERNEL, cpu_to_node(j));
1868 			if (!sg)
1869 				return -ENOMEM;
1870 
1871 			sg->next = sg;
1872 
1873 			*per_cpu_ptr(sdd->sg, j) = sg;
1874 
1875 			sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1876 					GFP_KERNEL, cpu_to_node(j));
1877 			if (!sgc)
1878 				return -ENOMEM;
1879 
1880 #ifdef CONFIG_SCHED_DEBUG
1881 			sgc->id = j;
1882 #endif
1883 
1884 			*per_cpu_ptr(sdd->sgc, j) = sgc;
1885 		}
1886 	}
1887 
1888 	return 0;
1889 }
1890 
1891 static void __sdt_free(const struct cpumask *cpu_map)
1892 {
1893 	struct sched_domain_topology_level *tl;
1894 	int j;
1895 
1896 	for_each_sd_topology(tl) {
1897 		struct sd_data *sdd = &tl->data;
1898 
1899 		for_each_cpu(j, cpu_map) {
1900 			struct sched_domain *sd;
1901 
1902 			if (sdd->sd) {
1903 				sd = *per_cpu_ptr(sdd->sd, j);
1904 				if (sd && (sd->flags & SD_OVERLAP))
1905 					free_sched_groups(sd->groups, 0);
1906 				kfree(*per_cpu_ptr(sdd->sd, j));
1907 			}
1908 
1909 			if (sdd->sds)
1910 				kfree(*per_cpu_ptr(sdd->sds, j));
1911 			if (sdd->sg)
1912 				kfree(*per_cpu_ptr(sdd->sg, j));
1913 			if (sdd->sgc)
1914 				kfree(*per_cpu_ptr(sdd->sgc, j));
1915 		}
1916 		free_percpu(sdd->sd);
1917 		sdd->sd = NULL;
1918 		free_percpu(sdd->sds);
1919 		sdd->sds = NULL;
1920 		free_percpu(sdd->sg);
1921 		sdd->sg = NULL;
1922 		free_percpu(sdd->sgc);
1923 		sdd->sgc = NULL;
1924 	}
1925 }
1926 
1927 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1928 		const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1929 		struct sched_domain *child, int dflags, int cpu)
1930 {
1931 	struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu);
1932 
1933 	if (child) {
1934 		sd->level = child->level + 1;
1935 		sched_domain_level_max = max(sched_domain_level_max, sd->level);
1936 		child->parent = sd;
1937 
1938 		if (!cpumask_subset(sched_domain_span(child),
1939 				    sched_domain_span(sd))) {
1940 			pr_err("BUG: arch topology borken\n");
1941 #ifdef CONFIG_SCHED_DEBUG
1942 			pr_err("     the %s domain not a subset of the %s domain\n",
1943 					child->name, sd->name);
1944 #endif
1945 			/* Fixup, ensure @sd has at least @child CPUs. */
1946 			cpumask_or(sched_domain_span(sd),
1947 				   sched_domain_span(sd),
1948 				   sched_domain_span(child));
1949 		}
1950 
1951 	}
1952 	set_domain_attribute(sd, attr);
1953 
1954 	return sd;
1955 }
1956 
1957 /*
1958  * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
1959  * any two given CPUs at this (non-NUMA) topology level.
1960  */
1961 static bool topology_span_sane(struct sched_domain_topology_level *tl,
1962 			      const struct cpumask *cpu_map, int cpu)
1963 {
1964 	int i;
1965 
1966 	/* NUMA levels are allowed to overlap */
1967 	if (tl->flags & SDTL_OVERLAP)
1968 		return true;
1969 
1970 	/*
1971 	 * Non-NUMA levels cannot partially overlap - they must be either
1972 	 * completely equal or completely disjoint. Otherwise we can end up
1973 	 * breaking the sched_group lists - i.e. a later get_group() pass
1974 	 * breaks the linking done for an earlier span.
1975 	 */
1976 	for_each_cpu(i, cpu_map) {
1977 		if (i == cpu)
1978 			continue;
1979 		/*
1980 		 * We should 'and' all those masks with 'cpu_map' to exactly
1981 		 * match the topology we're about to build, but that can only
1982 		 * remove CPUs, which only lessens our ability to detect
1983 		 * overlaps
1984 		 */
1985 		if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
1986 		    cpumask_intersects(tl->mask(cpu), tl->mask(i)))
1987 			return false;
1988 	}
1989 
1990 	return true;
1991 }
1992 
1993 /*
1994  * Find the sched_domain_topology_level where all CPU capacities are visible
1995  * for all CPUs.
1996  */
1997 static struct sched_domain_topology_level
1998 *asym_cpu_capacity_level(const struct cpumask *cpu_map)
1999 {
2000 	int i, j, asym_level = 0;
2001 	bool asym = false;
2002 	struct sched_domain_topology_level *tl, *asym_tl = NULL;
2003 	unsigned long cap;
2004 
2005 	/* Is there any asymmetry? */
2006 	cap = arch_scale_cpu_capacity(cpumask_first(cpu_map));
2007 
2008 	for_each_cpu(i, cpu_map) {
2009 		if (arch_scale_cpu_capacity(i) != cap) {
2010 			asym = true;
2011 			break;
2012 		}
2013 	}
2014 
2015 	if (!asym)
2016 		return NULL;
2017 
2018 	/*
2019 	 * Examine topology from all CPU's point of views to detect the lowest
2020 	 * sched_domain_topology_level where a highest capacity CPU is visible
2021 	 * to everyone.
2022 	 */
2023 	for_each_cpu(i, cpu_map) {
2024 		unsigned long max_capacity = arch_scale_cpu_capacity(i);
2025 		int tl_id = 0;
2026 
2027 		for_each_sd_topology(tl) {
2028 			if (tl_id < asym_level)
2029 				goto next_level;
2030 
2031 			for_each_cpu_and(j, tl->mask(i), cpu_map) {
2032 				unsigned long capacity;
2033 
2034 				capacity = arch_scale_cpu_capacity(j);
2035 
2036 				if (capacity <= max_capacity)
2037 					continue;
2038 
2039 				max_capacity = capacity;
2040 				asym_level = tl_id;
2041 				asym_tl = tl;
2042 			}
2043 next_level:
2044 			tl_id++;
2045 		}
2046 	}
2047 
2048 	return asym_tl;
2049 }
2050 
2051 
2052 /*
2053  * Build sched domains for a given set of CPUs and attach the sched domains
2054  * to the individual CPUs
2055  */
2056 static int
2057 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2058 {
2059 	enum s_alloc alloc_state = sa_none;
2060 	struct sched_domain *sd;
2061 	struct s_data d;
2062 	struct rq *rq = NULL;
2063 	int i, ret = -ENOMEM;
2064 	struct sched_domain_topology_level *tl_asym;
2065 	bool has_asym = false;
2066 
2067 	if (WARN_ON(cpumask_empty(cpu_map)))
2068 		goto error;
2069 
2070 	alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2071 	if (alloc_state != sa_rootdomain)
2072 		goto error;
2073 
2074 	tl_asym = asym_cpu_capacity_level(cpu_map);
2075 
2076 	/* Set up domains for CPUs specified by the cpu_map: */
2077 	for_each_cpu(i, cpu_map) {
2078 		struct sched_domain_topology_level *tl;
2079 		int dflags = 0;
2080 
2081 		sd = NULL;
2082 		for_each_sd_topology(tl) {
2083 			if (tl == tl_asym) {
2084 				dflags |= SD_ASYM_CPUCAPACITY;
2085 				has_asym = true;
2086 			}
2087 
2088 			if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2089 				goto error;
2090 
2091 			sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i);
2092 
2093 			if (tl == sched_domain_topology)
2094 				*per_cpu_ptr(d.sd, i) = sd;
2095 			if (tl->flags & SDTL_OVERLAP)
2096 				sd->flags |= SD_OVERLAP;
2097 			if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2098 				break;
2099 		}
2100 	}
2101 
2102 	/* Build the groups for the domains */
2103 	for_each_cpu(i, cpu_map) {
2104 		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2105 			sd->span_weight = cpumask_weight(sched_domain_span(sd));
2106 			if (sd->flags & SD_OVERLAP) {
2107 				if (build_overlap_sched_groups(sd, i))
2108 					goto error;
2109 			} else {
2110 				if (build_sched_groups(sd, i))
2111 					goto error;
2112 			}
2113 		}
2114 	}
2115 
2116 	/* Calculate CPU capacity for physical packages and nodes */
2117 	for (i = nr_cpumask_bits-1; i >= 0; i--) {
2118 		if (!cpumask_test_cpu(i, cpu_map))
2119 			continue;
2120 
2121 		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2122 			claim_allocations(i, sd);
2123 			init_sched_groups_capacity(i, sd);
2124 		}
2125 	}
2126 
2127 	/* Attach the domains */
2128 	rcu_read_lock();
2129 	for_each_cpu(i, cpu_map) {
2130 		rq = cpu_rq(i);
2131 		sd = *per_cpu_ptr(d.sd, i);
2132 
2133 		/* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2134 		if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2135 			WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2136 
2137 		cpu_attach_domain(sd, d.rd, i);
2138 	}
2139 	rcu_read_unlock();
2140 
2141 	if (has_asym)
2142 		static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2143 
2144 	if (rq && sched_debug_verbose) {
2145 		pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2146 			cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2147 	}
2148 
2149 	ret = 0;
2150 error:
2151 	__free_domain_allocs(&d, alloc_state, cpu_map);
2152 
2153 	return ret;
2154 }
2155 
2156 /* Current sched domains: */
2157 static cpumask_var_t			*doms_cur;
2158 
2159 /* Number of sched domains in 'doms_cur': */
2160 static int				ndoms_cur;
2161 
2162 /* Attributes of custom domains in 'doms_cur' */
2163 static struct sched_domain_attr		*dattr_cur;
2164 
2165 /*
2166  * Special case: If a kmalloc() of a doms_cur partition (array of
2167  * cpumask) fails, then fallback to a single sched domain,
2168  * as determined by the single cpumask fallback_doms.
2169  */
2170 static cpumask_var_t			fallback_doms;
2171 
2172 /*
2173  * arch_update_cpu_topology lets virtualized architectures update the
2174  * CPU core maps. It is supposed to return 1 if the topology changed
2175  * or 0 if it stayed the same.
2176  */
2177 int __weak arch_update_cpu_topology(void)
2178 {
2179 	return 0;
2180 }
2181 
2182 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2183 {
2184 	int i;
2185 	cpumask_var_t *doms;
2186 
2187 	doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2188 	if (!doms)
2189 		return NULL;
2190 	for (i = 0; i < ndoms; i++) {
2191 		if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2192 			free_sched_domains(doms, i);
2193 			return NULL;
2194 		}
2195 	}
2196 	return doms;
2197 }
2198 
2199 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2200 {
2201 	unsigned int i;
2202 	for (i = 0; i < ndoms; i++)
2203 		free_cpumask_var(doms[i]);
2204 	kfree(doms);
2205 }
2206 
2207 /*
2208  * Set up scheduler domains and groups.  For now this just excludes isolated
2209  * CPUs, but could be used to exclude other special cases in the future.
2210  */
2211 int sched_init_domains(const struct cpumask *cpu_map)
2212 {
2213 	int err;
2214 
2215 	zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2216 	zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2217 	zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2218 
2219 	arch_update_cpu_topology();
2220 	ndoms_cur = 1;
2221 	doms_cur = alloc_sched_domains(ndoms_cur);
2222 	if (!doms_cur)
2223 		doms_cur = &fallback_doms;
2224 	cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
2225 	err = build_sched_domains(doms_cur[0], NULL);
2226 
2227 	return err;
2228 }
2229 
2230 /*
2231  * Detach sched domains from a group of CPUs specified in cpu_map
2232  * These CPUs will now be attached to the NULL domain
2233  */
2234 static void detach_destroy_domains(const struct cpumask *cpu_map)
2235 {
2236 	unsigned int cpu = cpumask_any(cpu_map);
2237 	int i;
2238 
2239 	if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2240 		static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2241 
2242 	rcu_read_lock();
2243 	for_each_cpu(i, cpu_map)
2244 		cpu_attach_domain(NULL, &def_root_domain, i);
2245 	rcu_read_unlock();
2246 }
2247 
2248 /* handle null as "default" */
2249 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2250 			struct sched_domain_attr *new, int idx_new)
2251 {
2252 	struct sched_domain_attr tmp;
2253 
2254 	/* Fast path: */
2255 	if (!new && !cur)
2256 		return 1;
2257 
2258 	tmp = SD_ATTR_INIT;
2259 
2260 	return !memcmp(cur ? (cur + idx_cur) : &tmp,
2261 			new ? (new + idx_new) : &tmp,
2262 			sizeof(struct sched_domain_attr));
2263 }
2264 
2265 /*
2266  * Partition sched domains as specified by the 'ndoms_new'
2267  * cpumasks in the array doms_new[] of cpumasks. This compares
2268  * doms_new[] to the current sched domain partitioning, doms_cur[].
2269  * It destroys each deleted domain and builds each new domain.
2270  *
2271  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2272  * The masks don't intersect (don't overlap.) We should setup one
2273  * sched domain for each mask. CPUs not in any of the cpumasks will
2274  * not be load balanced. If the same cpumask appears both in the
2275  * current 'doms_cur' domains and in the new 'doms_new', we can leave
2276  * it as it is.
2277  *
2278  * The passed in 'doms_new' should be allocated using
2279  * alloc_sched_domains.  This routine takes ownership of it and will
2280  * free_sched_domains it when done with it. If the caller failed the
2281  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2282  * and partition_sched_domains() will fallback to the single partition
2283  * 'fallback_doms', it also forces the domains to be rebuilt.
2284  *
2285  * If doms_new == NULL it will be replaced with cpu_online_mask.
2286  * ndoms_new == 0 is a special case for destroying existing domains,
2287  * and it will not create the default domain.
2288  *
2289  * Call with hotplug lock and sched_domains_mutex held
2290  */
2291 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2292 				    struct sched_domain_attr *dattr_new)
2293 {
2294 	bool __maybe_unused has_eas = false;
2295 	int i, j, n;
2296 	int new_topology;
2297 
2298 	lockdep_assert_held(&sched_domains_mutex);
2299 
2300 	/* Let the architecture update CPU core mappings: */
2301 	new_topology = arch_update_cpu_topology();
2302 
2303 	if (!doms_new) {
2304 		WARN_ON_ONCE(dattr_new);
2305 		n = 0;
2306 		doms_new = alloc_sched_domains(1);
2307 		if (doms_new) {
2308 			n = 1;
2309 			cpumask_and(doms_new[0], cpu_active_mask,
2310 				    housekeeping_cpumask(HK_FLAG_DOMAIN));
2311 		}
2312 	} else {
2313 		n = ndoms_new;
2314 	}
2315 
2316 	/* Destroy deleted domains: */
2317 	for (i = 0; i < ndoms_cur; i++) {
2318 		for (j = 0; j < n && !new_topology; j++) {
2319 			if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2320 			    dattrs_equal(dattr_cur, i, dattr_new, j)) {
2321 				struct root_domain *rd;
2322 
2323 				/*
2324 				 * This domain won't be destroyed and as such
2325 				 * its dl_bw->total_bw needs to be cleared.  It
2326 				 * will be recomputed in function
2327 				 * update_tasks_root_domain().
2328 				 */
2329 				rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2330 				dl_clear_root_domain(rd);
2331 				goto match1;
2332 			}
2333 		}
2334 		/* No match - a current sched domain not in new doms_new[] */
2335 		detach_destroy_domains(doms_cur[i]);
2336 match1:
2337 		;
2338 	}
2339 
2340 	n = ndoms_cur;
2341 	if (!doms_new) {
2342 		n = 0;
2343 		doms_new = &fallback_doms;
2344 		cpumask_and(doms_new[0], cpu_active_mask,
2345 			    housekeeping_cpumask(HK_FLAG_DOMAIN));
2346 	}
2347 
2348 	/* Build new domains: */
2349 	for (i = 0; i < ndoms_new; i++) {
2350 		for (j = 0; j < n && !new_topology; j++) {
2351 			if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2352 			    dattrs_equal(dattr_new, i, dattr_cur, j))
2353 				goto match2;
2354 		}
2355 		/* No match - add a new doms_new */
2356 		build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2357 match2:
2358 		;
2359 	}
2360 
2361 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2362 	/* Build perf. domains: */
2363 	for (i = 0; i < ndoms_new; i++) {
2364 		for (j = 0; j < n && !sched_energy_update; j++) {
2365 			if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2366 			    cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2367 				has_eas = true;
2368 				goto match3;
2369 			}
2370 		}
2371 		/* No match - add perf. domains for a new rd */
2372 		has_eas |= build_perf_domains(doms_new[i]);
2373 match3:
2374 		;
2375 	}
2376 	sched_energy_set(has_eas);
2377 #endif
2378 
2379 	/* Remember the new sched domains: */
2380 	if (doms_cur != &fallback_doms)
2381 		free_sched_domains(doms_cur, ndoms_cur);
2382 
2383 	kfree(dattr_cur);
2384 	doms_cur = doms_new;
2385 	dattr_cur = dattr_new;
2386 	ndoms_cur = ndoms_new;
2387 
2388 	update_sched_domain_debugfs();
2389 }
2390 
2391 /*
2392  * Call with hotplug lock held
2393  */
2394 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2395 			     struct sched_domain_attr *dattr_new)
2396 {
2397 	mutex_lock(&sched_domains_mutex);
2398 	partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2399 	mutex_unlock(&sched_domains_mutex);
2400 }
2401