1	// SPDX-License-Identifier: GPL-2.0
2	/*
3	* Scheduler topology setup/handling methods
4	*/
5
6	#include <linux/sched/isolation.h>
7	#include <linux/bsearch.h>
8	#include "sched.h"
9
10	DEFINE_MUTEX(sched_domains_mutex);
11	void sched_domains_mutex_lock(void)
12	{
13	mutex_lock(&sched_domains_mutex);
14	}
15	void sched_domains_mutex_unlock(void)
16	{
17	mutex_unlock(lock: &sched_domains_mutex);
18	}
19
20	/* Protected by sched_domains_mutex: */
21	static cpumask_var_t sched_domains_tmpmask;
22	static cpumask_var_t sched_domains_tmpmask2;
23
24	static int __init sched_debug_setup(char *str)
25	{
26	sched_debug_verbose = true;
27
28	return 0;
29	}
30	early_param("sched_verbose", sched_debug_setup);
31
32	static inline bool sched_debug(void)
33	{
34	return sched_debug_verbose;
35	}
36
37	#define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
38	const struct sd_flag_debug sd_flag_debug[] = {
39	#include <linux/sched/sd_flags.h>
40	};
41	#undef SD_FLAG
42
43	static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
44	struct cpumask *groupmask)
45	{
46	struct sched_group *group = sd->groups;
47	unsigned long flags = sd->flags;
48	unsigned int idx;
49
50	cpumask_clear(dstp: groupmask);
51
52	printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
53	printk(KERN_CONT "span=%*pbl level=%s\n",
54	cpumask_pr_args(sched_domain_span(sd)), sd->name);
55
56	if (!cpumask_test_cpu(cpu, cpumask: sched_domain_span(sd))) {
57	printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
58	}
59	if (group && !cpumask_test_cpu(cpu, cpumask: sched_group_span(sg: group))) {
60	printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
61	}
62
63	for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
64	unsigned int flag = BIT(idx);
65	unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
66
67	if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
68	!(sd->child->flags & flag))
69	printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
70	sd_flag_debug[idx].name);
71
72	if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
73	!(sd->parent->flags & flag))
74	printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
75	sd_flag_debug[idx].name);
76	}
77
78	printk(KERN_DEBUG "%*s groups:", level + 1, "");
79	do {
80	if (!group) {
81	printk("\n");
82	printk(KERN_ERR "ERROR: group is NULL\n");
83	break;
84	}
85
86	if (cpumask_empty(srcp: sched_group_span(sg: group))) {
87	printk(KERN_CONT "\n");
88	printk(KERN_ERR "ERROR: empty group\n");
89	break;
90	}
91
92	if (!(sd->flags & SD_NUMA) &&
93	cpumask_intersects(src1p: groupmask, src2p: sched_group_span(sg: group))) {
94	printk(KERN_CONT "\n");
95	printk(KERN_ERR "ERROR: repeated CPUs\n");
96	break;
97	}
98
99	cpumask_or(dstp: groupmask, src1p: groupmask, src2p: sched_group_span(sg: group));
100
101	printk(KERN_CONT " %d:{ span=%*pbl",
102	group->sgc->id,
103	cpumask_pr_args(sched_group_span(group)));
104
105	if ((sd->flags & SD_NUMA) &&
106	!cpumask_equal(src1p: group_balance_mask(sg: group), src2p: sched_group_span(sg: group))) {
107	printk(KERN_CONT " mask=%*pbl",
108	cpumask_pr_args(group_balance_mask(group)));
109	}
110
111	if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
112	printk(KERN_CONT " cap=%lu", group->sgc->capacity);
113
114	if (group == sd->groups && sd->child &&
115	!cpumask_equal(src1p: sched_domain_span(sd: sd->child),
116	src2p: sched_group_span(sg: group))) {
117	printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
118	}
119
120	printk(KERN_CONT " }");
121
122	group = group->next;
123
124	if (group != sd->groups)
125	printk(KERN_CONT ",");
126
127	} while (group != sd->groups);
128	printk(KERN_CONT "\n");
129
130	if (!cpumask_equal(src1p: sched_domain_span(sd), src2p: groupmask))
131	printk(KERN_ERR "ERROR: groups don't span domain->span\n");
132
133	if (sd->parent &&
134	!cpumask_subset(src1p: groupmask, src2p: sched_domain_span(sd: sd->parent)))
135	printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
136	return 0;
137	}
138
139	static void sched_domain_debug(struct sched_domain *sd, int cpu)
140	{
141	int level = 0;
142
143	if (!sched_debug_verbose)
144	return;
145
146	if (!sd) {
147	printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
148	return;
149	}
150
151	printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
152
153	for (;;) {
154	if (sched_domain_debug_one(sd, cpu, level, groupmask: sched_domains_tmpmask))
155	break;
156	level++;
157	sd = sd->parent;
158	if (!sd)
159	break;
160	}
161	}
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(srcp: 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(sd: parent))
193	return 1;
194
195	if (!cpumask_equal(src1p: sched_domain_span(sd), src2p: sched_domain_span(sd: 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	static unsigned int sysctl_sched_energy_aware = 1;
211	static DEFINE_MUTEX(sched_energy_mutex);
212	static bool sched_energy_update;
213
214	static bool sched_is_eas_possible(const struct cpumask *cpu_mask)
215	{
216	bool any_asym_capacity = false;
217	int i;
218
219	/* EAS is enabled for asymmetric CPU capacity topologies. */
220	for_each_cpu(i, cpu_mask) {
221	if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, i))) {
222	any_asym_capacity = true;
223	break;
224	}
225	}
226	if (!any_asym_capacity) {
227	if (sched_debug()) {
228	pr_info("rd %*pbl: Checking EAS, CPUs do not have asymmetric capacities\n",
229	cpumask_pr_args(cpu_mask));
230	}
231	return false;
232	}
233
234	/* EAS definitely does *not* handle SMT */
235	if (sched_smt_active()) {
236	if (sched_debug()) {
237	pr_info("rd %*pbl: Checking EAS, SMT is not supported\n",
238	cpumask_pr_args(cpu_mask));
239	}
240	return false;
241	}
242
243	if (!arch_scale_freq_invariant()) {
244	if (sched_debug()) {
245	pr_info("rd %*pbl: Checking EAS: frequency-invariant load tracking not yet supported",
246	cpumask_pr_args(cpu_mask));
247	}
248	return false;
249	}
250
251	if (!cpufreq_ready_for_eas(cpu_mask)) {
252	if (sched_debug()) {
253	pr_info("rd %*pbl: Checking EAS: cpufreq is not ready\n",
254	cpumask_pr_args(cpu_mask));
255	}
256	return false;
257	}
258
259	return true;
260	}
261
262	void rebuild_sched_domains_energy(void)
263	{
264	mutex_lock(&sched_energy_mutex);
265	sched_energy_update = true;
266	rebuild_sched_domains();
267	sched_energy_update = false;
268	mutex_unlock(lock: &sched_energy_mutex);
269	}
270
271	#ifdef CONFIG_PROC_SYSCTL
272	static int sched_energy_aware_handler(const struct ctl_table *table, int write,
273	void *buffer, size_t *lenp, loff_t *ppos)
274	{
275	int ret, state;
276
277	if (write && !capable(CAP_SYS_ADMIN))
278	return -EPERM;
279
280	if (!sched_is_eas_possible(cpu_active_mask)) {
281	if (write) {
282	return -EOPNOTSUPP;
283	} else {
284	*lenp = 0;
285	return 0;
286	}
287	}
288
289	ret = proc_dointvec_minmax(table, dir: write, buffer, lenp, ppos);
290	if (!ret && write) {
291	state = static_branch_unlikely(&sched_energy_present);
292	if (state != sysctl_sched_energy_aware)
293	rebuild_sched_domains_energy();
294	}
295
296	return ret;
297	}
298
299	static const struct ctl_table sched_energy_aware_sysctls[] = {
300	{
301	.procname = "sched_energy_aware",
302	.data = &sysctl_sched_energy_aware,
303	.maxlen = sizeof(unsigned int),
304	.mode = 0644,
305	.proc_handler = sched_energy_aware_handler,
306	.extra1 = SYSCTL_ZERO,
307	.extra2 = SYSCTL_ONE,
308	},
309	};
310
311	static int __init sched_energy_aware_sysctl_init(void)
312	{
313	register_sysctl_init("kernel", sched_energy_aware_sysctls);
314	return 0;
315	}
316
317	late_initcall(sched_energy_aware_sysctl_init);
318	#endif /* CONFIG_PROC_SYSCTL */
319
320	static void free_pd(struct perf_domain *pd)
321	{
322	struct perf_domain *tmp;
323
324	while (pd) {
325	tmp = pd->next;
326	kfree(objp: pd);
327	pd = tmp;
328	}
329	}
330
331	static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
332	{
333	while (pd) {
334	if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
335	return pd;
336	pd = pd->next;
337	}
338
339	return NULL;
340	}
341
342	static struct perf_domain *pd_init(int cpu)
343	{
344	struct em_perf_domain *obj = em_cpu_get(cpu);
345	struct perf_domain *pd;
346
347	if (!obj) {
348	if (sched_debug())
349	pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
350	return NULL;
351	}
352
353	pd = kzalloc(sizeof(*pd), GFP_KERNEL);
354	if (!pd)
355	return NULL;
356	pd->em_pd = obj;
357
358	return pd;
359	}
360
361	static void perf_domain_debug(const struct cpumask *cpu_map,
362	struct perf_domain *pd)
363	{
364	if (!sched_debug() || !pd)
365	return;
366
367	printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
368
369	while (pd) {
370	printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
371	cpumask_first(perf_domain_span(pd)),
372	cpumask_pr_args(perf_domain_span(pd)),
373	em_pd_nr_perf_states(pd->em_pd));
374	pd = pd->next;
375	}
376
377	printk(KERN_CONT "\n");
378	}
379
380	static void destroy_perf_domain_rcu(struct rcu_head *rp)
381	{
382	struct perf_domain *pd;
383
384	pd = container_of(rp, struct perf_domain, rcu);
385	free_pd(pd);
386	}
387
388	static void sched_energy_set(bool has_eas)
389	{
390	if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
391	if (sched_debug())
392	pr_info("%s: stopping EAS\n", __func__);
393	static_branch_disable_cpuslocked(&sched_energy_present);
394	} else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
395	if (sched_debug())
396	pr_info("%s: starting EAS\n", __func__);
397	static_branch_enable_cpuslocked(&sched_energy_present);
398	}
399	}
400
401	/*
402	* EAS can be used on a root domain if it meets all the following conditions:
403	* 1. an Energy Model (EM) is available;
404	* 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
405	* 3. no SMT is detected.
406	* 4. schedutil is driving the frequency of all CPUs of the rd;
407	* 5. frequency invariance support is present;
408	*/
409	static bool build_perf_domains(const struct cpumask *cpu_map)
410	{
411	int i;
412	struct perf_domain *pd = NULL, *tmp;
413	int cpu = cpumask_first(srcp: cpu_map);
414	struct root_domain *rd = cpu_rq(cpu)->rd;
415
416	if (!sysctl_sched_energy_aware)
417	goto free;
418
419	if (!sched_is_eas_possible(cpu_mask: cpu_map))
420	goto free;
421
422	for_each_cpu(i, cpu_map) {
423	/* Skip already covered CPUs. */
424	if (find_pd(pd, cpu: i))
425	continue;
426
427	/* Create the new pd and add it to the local list. */
428	tmp = pd_init(cpu: i);
429	if (!tmp)
430	goto free;
431	tmp->next = pd;
432	pd = tmp;
433	}
434
435	perf_domain_debug(cpu_map, pd);
436
437	/* Attach the new list of performance domains to the root domain. */
438	tmp = rd->pd;
439	rcu_assign_pointer(rd->pd, pd);
440	if (tmp)
441	call_rcu(head: &tmp->rcu, func: destroy_perf_domain_rcu);
442
443	return !!pd;
444
445	free:
446	free_pd(pd);
447	tmp = rd->pd;
448	rcu_assign_pointer(rd->pd, NULL);
449	if (tmp)
450	call_rcu(head: &tmp->rcu, func: destroy_perf_domain_rcu);
451
452	return false;
453	}
454	#else /* !(CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL): */
455	static void free_pd(struct perf_domain *pd) { }
456	#endif /* !(CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */
457
458	static void free_rootdomain(struct rcu_head *rcu)
459	{
460	struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
461
462	cpupri_cleanup(cp: &rd->cpupri);
463	cpudl_cleanup(cp: &rd->cpudl);
464	free_cpumask_var(mask: rd->dlo_mask);
465	free_cpumask_var(mask: rd->rto_mask);
466	free_cpumask_var(mask: rd->online);
467	free_cpumask_var(mask: rd->span);
468	free_pd(pd: rd->pd);
469	kfree(objp: rd);
470	}
471
472	void rq_attach_root(struct rq *rq, struct root_domain *rd)
473	{
474	struct root_domain *old_rd = NULL;
475	struct rq_flags rf;
476
477	rq_lock_irqsave(rq, rf: &rf);
478
479	if (rq->rd) {
480	old_rd = rq->rd;
481
482	if (cpumask_test_cpu(cpu: rq->cpu, cpumask: old_rd->online))
483	set_rq_offline(rq);
484
485	cpumask_clear_cpu(cpu: rq->cpu, dstp: old_rd->span);
486
487	/*
488	* If we don't want to free the old_rd yet then
489	* set old_rd to NULL to skip the freeing later
490	* in this function:
491	*/
492	if (!atomic_dec_and_test(v: &old_rd->refcount))
493	old_rd = NULL;
494	}
495
496	atomic_inc(v: &rd->refcount);
497	rq->rd = rd;
498
499	cpumask_set_cpu(cpu: rq->cpu, dstp: rd->span);
500	if (cpumask_test_cpu(cpu: rq->cpu, cpu_active_mask))
501	set_rq_online(rq);
502
503	/*
504	* Because the rq is not a task, dl_add_task_root_domain() did not
505	* move the fair server bw to the rd if it already started.
506	* Add it now.
507	*/
508	if (rq->fair_server.dl_server)
509	__dl_server_attach_root(dl_se: &rq->fair_server, rq);
510
511	rq_unlock_irqrestore(rq, rf: &rf);
512
513	if (old_rd)
514	call_rcu(head: &old_rd->rcu, func: free_rootdomain);
515	}
516
517	void sched_get_rd(struct root_domain *rd)
518	{
519	atomic_inc(v: &rd->refcount);
520	}
521
522	void sched_put_rd(struct root_domain *rd)
523	{
524	if (!atomic_dec_and_test(v: &rd->refcount))
525	return;
526
527	call_rcu(head: &rd->rcu, func: free_rootdomain);
528	}
529
530	static int init_rootdomain(struct root_domain *rd)
531	{
532	if (!zalloc_cpumask_var(mask: &rd->span, GFP_KERNEL))
533	goto out;
534	if (!zalloc_cpumask_var(mask: &rd->online, GFP_KERNEL))
535	goto free_span;
536	if (!zalloc_cpumask_var(mask: &rd->dlo_mask, GFP_KERNEL))
537	goto free_online;
538	if (!zalloc_cpumask_var(mask: &rd->rto_mask, GFP_KERNEL))
539	goto free_dlo_mask;
540
541	#ifdef HAVE_RT_PUSH_IPI
542	rd->rto_cpu = -1;
543	raw_spin_lock_init(&rd->rto_lock);
544	rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
545	#endif
546
547	rd->visit_cookie = 0;
548	init_dl_bw(dl_b: &rd->dl_bw);
549	if (cpudl_init(cp: &rd->cpudl) != 0)
550	goto free_rto_mask;
551
552	if (cpupri_init(cp: &rd->cpupri) != 0)
553	goto free_cpudl;
554	return 0;
555
556	free_cpudl:
557	cpudl_cleanup(cp: &rd->cpudl);
558	free_rto_mask:
559	free_cpumask_var(mask: rd->rto_mask);
560	free_dlo_mask:
561	free_cpumask_var(mask: rd->dlo_mask);
562	free_online:
563	free_cpumask_var(mask: rd->online);
564	free_span:
565	free_cpumask_var(mask: rd->span);
566	out:
567	return -ENOMEM;
568	}
569
570	/*
571	* By default the system creates a single root-domain with all CPUs as
572	* members (mimicking the global state we have today).
573	*/
574	struct root_domain def_root_domain;
575
576	void __init init_defrootdomain(void)
577	{
578	init_rootdomain(rd: &def_root_domain);
579
580	atomic_set(v: &def_root_domain.refcount, i: 1);
581	}
582
583	static struct root_domain *alloc_rootdomain(void)
584	{
585	struct root_domain *rd;
586
587	rd = kzalloc(sizeof(*rd), GFP_KERNEL);
588	if (!rd)
589	return NULL;
590
591	if (init_rootdomain(rd) != 0) {
592	kfree(objp: rd);
593	return NULL;
594	}
595
596	return rd;
597	}
598
599	static void free_sched_groups(struct sched_group *sg, int free_sgc)
600	{
601	struct sched_group *tmp, *first;
602
603	if (!sg)
604	return;
605
606	first = sg;
607	do {
608	tmp = sg->next;
609
610	if (free_sgc && atomic_dec_and_test(v: &sg->sgc->ref))
611	kfree(objp: sg->sgc);
612
613	if (atomic_dec_and_test(v: &sg->ref))
614	kfree(objp: sg);
615	sg = tmp;
616	} while (sg != first);
617	}
618
619	static void destroy_sched_domain(struct sched_domain *sd)
620	{
621	/*
622	* A normal sched domain may have multiple group references, an
623	* overlapping domain, having private groups, only one. Iterate,
624	* dropping group/capacity references, freeing where none remain.
625	*/
626	free_sched_groups(sg: sd->groups, free_sgc: 1);
627
628	if (sd->shared && atomic_dec_and_test(v: &sd->shared->ref))
629	kfree(objp: sd->shared);
630	kfree(objp: sd);
631	}
632
633	static void destroy_sched_domains_rcu(struct rcu_head *rcu)
634	{
635	struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
636
637	while (sd) {
638	struct sched_domain *parent = sd->parent;
639	destroy_sched_domain(sd);
640	sd = parent;
641	}
642	}
643
644	static void destroy_sched_domains(struct sched_domain *sd)
645	{
646	if (sd)
647	call_rcu(head: &sd->rcu, func: destroy_sched_domains_rcu);
648	}
649
650	/*
651	* Keep a special pointer to the highest sched_domain that has SD_SHARE_LLC set
652	* (Last Level Cache Domain) for this allows us to avoid some pointer chasing
653	* select_idle_sibling().
654	*
655	* Also keep a unique ID per domain (we use the first CPU number in the cpumask
656	* of the domain), this allows us to quickly tell if two CPUs are in the same
657	* cache domain, see cpus_share_cache().
658	*/
659	DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
660	DEFINE_PER_CPU(int, sd_llc_size);
661	DEFINE_PER_CPU(int, sd_llc_id);
662	DEFINE_PER_CPU(int, sd_share_id);
663	DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
664	DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
665	DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
666	DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
667
668	DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
669	DEFINE_STATIC_KEY_FALSE(sched_cluster_active);
670
671	static void update_top_cache_domain(int cpu)
672	{
673	struct sched_domain_shared *sds = NULL;
674	struct sched_domain *sd;
675	int id = cpu;
676	int size = 1;
677
678	sd = highest_flag_domain(cpu, flag: SD_SHARE_LLC);
679	if (sd) {
680	id = cpumask_first(srcp: sched_domain_span(sd));
681	size = cpumask_weight(srcp: sched_domain_span(sd));
682	sds = sd->shared;
683	}
684
685	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
686	per_cpu(sd_llc_size, cpu) = size;
687	per_cpu(sd_llc_id, cpu) = id;
688	rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
689
690	sd = lowest_flag_domain(cpu, flag: SD_CLUSTER);
691	if (sd)
692	id = cpumask_first(srcp: sched_domain_span(sd));
693
694	/*
695	* This assignment should be placed after the sd_llc_id as
696	* we want this id equals to cluster id on cluster machines
697	* but equals to LLC id on non-Cluster machines.
698	*/
699	per_cpu(sd_share_id, cpu) = id;
700
701	sd = lowest_flag_domain(cpu, flag: SD_NUMA);
702	rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
703
704	sd = highest_flag_domain(cpu, flag: SD_ASYM_PACKING);
705	rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
706
707	sd = lowest_flag_domain(cpu, flag: SD_ASYM_CPUCAPACITY_FULL);
708	rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
709	}
710
711	/*
712	* Attach the domain 'sd' to 'cpu' as its base domain. Callers must
713	* hold the hotplug lock.
714	*/
715	static void
716	cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
717	{
718	struct rq *rq = cpu_rq(cpu);
719	struct sched_domain *tmp;
720
721	/* Remove the sched domains which do not contribute to scheduling. */
722	for (tmp = sd; tmp; ) {
723	struct sched_domain *parent = tmp->parent;
724	if (!parent)
725	break;
726
727	if (sd_parent_degenerate(sd: tmp, parent)) {
728	tmp->parent = parent->parent;
729
730	if (parent->parent) {
731	parent->parent->child = tmp;
732	parent->parent->groups->flags = tmp->flags;
733	}
734
735	/*
736	* Transfer SD_PREFER_SIBLING down in case of a
737	* degenerate parent; the spans match for this
738	* so the property transfers.
739	*/
740	if (parent->flags & SD_PREFER_SIBLING)
741	tmp->flags |= SD_PREFER_SIBLING;
742	destroy_sched_domain(sd: parent);
743	} else
744	tmp = tmp->parent;
745	}
746
747	if (sd && sd_degenerate(sd)) {
748	tmp = sd;
749	sd = sd->parent;
750	destroy_sched_domain(sd: tmp);
751	if (sd) {
752	struct sched_group *sg = sd->groups;
753
754	/*
755	* sched groups hold the flags of the child sched
756	* domain for convenience. Clear such flags since
757	* the child is being destroyed.
758	*/
759	do {
760	sg->flags = 0;
761	} while (sg != sd->groups);
762
763	sd->child = NULL;
764	}
765	}
766
767	sched_domain_debug(sd, cpu);
768
769	rq_attach_root(rq, rd);
770	tmp = rq->sd;
771	rcu_assign_pointer(rq->sd, sd);
772	dirty_sched_domain_sysctl(cpu);
773	destroy_sched_domains(sd: tmp);
774
775	update_top_cache_domain(cpu);
776	}
777
778	struct s_data {
779	struct sched_domain * __percpu *sd;
780	struct root_domain *rd;
781	};
782
783	enum s_alloc {
784	sa_rootdomain,
785	sa_sd,
786	sa_sd_storage,
787	sa_none,
788	};
789
790	/*
791	* Return the canonical balance CPU for this group, this is the first CPU
792	* of this group that's also in the balance mask.
793	*
794	* The balance mask are all those CPUs that could actually end up at this
795	* group. See build_balance_mask().
796	*
797	* Also see should_we_balance().
798	*/
799	int group_balance_cpu(struct sched_group *sg)
800	{
801	return cpumask_first(srcp: group_balance_mask(sg));
802	}
803
804
805	/*
806	* NUMA topology (first read the regular topology blurb below)
807	*
808	* Given a node-distance table, for example:
809	*
810	* node 0 1 2 3
811	* 0: 10 20 30 20
812	* 1: 20 10 20 30
813	* 2: 30 20 10 20
814	* 3: 20 30 20 10
815	*
816	* which represents a 4 node ring topology like:
817	*
818	* 0 ----- 1
819	* | |
820	* | |
821	* | |
822	* 3 ----- 2
823	*
824	* We want to construct domains and groups to represent this. The way we go
825	* about doing this is to build the domains on 'hops'. For each NUMA level we
826	* construct the mask of all nodes reachable in @level hops.
827	*
828	* For the above NUMA topology that gives 3 levels:
829	*
830	* NUMA-2 0-3 0-3 0-3 0-3
831	* groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
832	*
833	* NUMA-1 0-1,3 0-2 1-3 0,2-3
834	* groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
835	*
836	* NUMA-0 0 1 2 3
837	*
838	*
839	* As can be seen; things don't nicely line up as with the regular topology.
840	* When we iterate a domain in child domain chunks some nodes can be
841	* represented multiple times -- hence the "overlap" naming for this part of
842	* the topology.
843	*
844	* In order to minimize this overlap, we only build enough groups to cover the
845	* domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
846	*
847	* Because:
848	*
849	* - the first group of each domain is its child domain; this
850	* gets us the first 0-1,3
851	* - the only uncovered node is 2, who's child domain is 1-3.
852	*
853	* However, because of the overlap, computing a unique CPU for each group is
854	* more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
855	* groups include the CPUs of Node-0, while those CPUs would not in fact ever
856	* end up at those groups (they would end up in group: 0-1,3).
857	*
858	* To correct this we have to introduce the group balance mask. This mask
859	* will contain those CPUs in the group that can reach this group given the
860	* (child) domain tree.
861	*
862	* With this we can once again compute balance_cpu and sched_group_capacity
863	* relations.
864	*
865	* XXX include words on how balance_cpu is unique and therefore can be
866	* used for sched_group_capacity links.
867	*
868	*
869	* Another 'interesting' topology is:
870	*
871	* node 0 1 2 3
872	* 0: 10 20 20 30
873	* 1: 20 10 20 20
874	* 2: 20 20 10 20
875	* 3: 30 20 20 10
876	*
877	* Which looks a little like:
878	*
879	* 0 ----- 1
880	* | / |
881	* | / |
882	* | / |
883	* 2 ----- 3
884	*
885	* This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
886	* are not.
887	*
888	* This leads to a few particularly weird cases where the sched_domain's are
889	* not of the same number for each CPU. Consider:
890	*
891	* NUMA-2 0-3 0-3
892	* groups: {0-2},{1-3} {1-3},{0-2}
893	*
894	* NUMA-1 0-2 0-3 0-3 1-3
895	*
896	* NUMA-0 0 1 2 3
897	*
898	*/
899
900
901	/*
902	* Build the balance mask; it contains only those CPUs that can arrive at this
903	* group and should be considered to continue balancing.
904	*
905	* We do this during the group creation pass, therefore the group information
906	* isn't complete yet, however since each group represents a (child) domain we
907	* can fully construct this using the sched_domain bits (which are already
908	* complete).
909	*/
910	static void
911	build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
912	{
913	const struct cpumask *sg_span = sched_group_span(sg);
914	struct sd_data *sdd = sd->private;
915	struct sched_domain *sibling;
916	int i;
917
918	cpumask_clear(dstp: mask);
919
920	for_each_cpu(i, sg_span) {
921	sibling = *per_cpu_ptr(sdd->sd, i);
922
923	/*
924	* Can happen in the asymmetric case, where these siblings are
925	* unused. The mask will not be empty because those CPUs that
926	* do have the top domain _should_ span the domain.
927	*/
928	if (!sibling->child)
929	continue;
930
931	/* If we would not end up here, we can't continue from here */
932	if (!cpumask_equal(src1p: sg_span, src2p: sched_domain_span(sd: sibling->child)))
933	continue;
934
935	cpumask_set_cpu(cpu: i, dstp: mask);
936	}
937
938	/* We must not have empty masks here */
939	WARN_ON_ONCE(cpumask_empty(mask));
940	}
941
942	/*
943	* XXX: This creates per-node group entries; since the load-balancer will
944	* immediately access remote memory to construct this group's load-balance
945	* statistics having the groups node local is of dubious benefit.
946	*/
947	static struct sched_group *
948	build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
949	{
950	struct sched_group *sg;
951	struct cpumask *sg_span;
952
953	sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
954	GFP_KERNEL, cpu_to_node(cpu));
955
956	if (!sg)
957	return NULL;
958
959	sg_span = sched_group_span(sg);
960	if (sd->child) {
961	cpumask_copy(dstp: sg_span, srcp: sched_domain_span(sd: sd->child));
962	sg->flags = sd->child->flags;
963	} else {
964	cpumask_copy(dstp: sg_span, srcp: sched_domain_span(sd));
965	}
966
967	atomic_inc(v: &sg->ref);
968	return sg;
969	}
970
971	static void init_overlap_sched_group(struct sched_domain *sd,
972	struct sched_group *sg)
973	{
974	struct cpumask *mask = sched_domains_tmpmask2;
975	struct sd_data *sdd = sd->private;
976	struct cpumask *sg_span;
977	int cpu;
978
979	build_balance_mask(sd, sg, mask);
980	cpu = cpumask_first(srcp: mask);
981
982	sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
983	if (atomic_inc_return(v: &sg->sgc->ref) == 1)
984	cpumask_copy(dstp: group_balance_mask(sg), srcp: mask);
985	else
986	WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
987
988	/*
989	* Initialize sgc->capacity such that even if we mess up the
990	* domains and no possible iteration will get us here, we won't
991	* die on a /0 trap.
992	*/
993	sg_span = sched_group_span(sg);
994	sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(srcp: sg_span);
995	sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
996	sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
997	}
998
999	static struct sched_domain *
1000	find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
1001	{
1002	/*
1003	* The proper descendant would be the one whose child won't span out
1004	* of sd
1005	*/
1006	while (sibling->child &&
1007	!cpumask_subset(src1p: sched_domain_span(sd: sibling->child),
1008	src2p: sched_domain_span(sd)))
1009	sibling = sibling->child;
1010
1011	/*
1012	* As we are referencing sgc across different topology level, we need
1013	* to go down to skip those sched_domains which don't contribute to
1014	* scheduling because they will be degenerated in cpu_attach_domain
1015	*/
1016	while (sibling->child &&
1017	cpumask_equal(src1p: sched_domain_span(sd: sibling->child),
1018	src2p: sched_domain_span(sd: sibling)))
1019	sibling = sibling->child;
1020
1021	return sibling;
1022	}
1023
1024	static int
1025	build_overlap_sched_groups(struct sched_domain *sd, int cpu)
1026	{
1027	struct sched_group *first = NULL, *last = NULL, *sg;
1028	const struct cpumask *span = sched_domain_span(sd);
1029	struct cpumask *covered = sched_domains_tmpmask;
1030	struct sd_data *sdd = sd->private;
1031	struct sched_domain *sibling;
1032	int i;
1033
1034	cpumask_clear(dstp: covered);
1035
1036	for_each_cpu_wrap(i, span, cpu) {
1037	struct cpumask *sg_span;
1038
1039	if (cpumask_test_cpu(cpu: i, cpumask: covered))
1040	continue;
1041
1042	sibling = *per_cpu_ptr(sdd->sd, i);
1043
1044	/*
1045	* Asymmetric node setups can result in situations where the
1046	* domain tree is of unequal depth, make sure to skip domains
1047	* that already cover the entire range.
1048	*
1049	* In that case build_sched_domains() will have terminated the
1050	* iteration early and our sibling sd spans will be empty.
1051	* Domains should always include the CPU they're built on, so
1052	* check that.
1053	*/
1054	if (!cpumask_test_cpu(cpu: i, cpumask: sched_domain_span(sd: sibling)))
1055	continue;
1056
1057	/*
1058	* Usually we build sched_group by sibling's child sched_domain
1059	* But for machines whose NUMA diameter are 3 or above, we move
1060	* to build sched_group by sibling's proper descendant's child
1061	* domain because sibling's child sched_domain will span out of
1062	* the sched_domain being built as below.
1063	*
1064	* Smallest diameter=3 topology is:
1065	*
1066	* node 0 1 2 3
1067	* 0: 10 20 30 40
1068	* 1: 20 10 20 30
1069	* 2: 30 20 10 20
1070	* 3: 40 30 20 10
1071	*
1072	* 0 --- 1 --- 2 --- 3
1073	*
1074	* NUMA-3 0-3 N/A N/A 0-3
1075	* groups: {0-2},{1-3} {1-3},{0-2}
1076	*
1077	* NUMA-2 0-2 0-3 0-3 1-3
1078	* groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2}
1079	*
1080	* NUMA-1 0-1 0-2 1-3 2-3
1081	* groups: {0},{1} {1},{2},{0} {2},{3},{1} {3},{2}
1082	*
1083	* NUMA-0 0 1 2 3
1084	*
1085	* The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1086	* group span isn't a subset of the domain span.
1087	*/
1088	if (sibling->child &&
1089	!cpumask_subset(src1p: sched_domain_span(sd: sibling->child), src2p: span))
1090	sibling = find_descended_sibling(sd, sibling);
1091
1092	sg = build_group_from_child_sched_domain(sd: sibling, cpu);
1093	if (!sg)
1094	goto fail;
1095
1096	sg_span = sched_group_span(sg);
1097	cpumask_or(dstp: covered, src1p: covered, src2p: sg_span);
1098
1099	init_overlap_sched_group(sd: sibling, sg);
1100
1101	if (!first)
1102	first = sg;
1103	if (last)
1104	last->next = sg;
1105	last = sg;
1106	last->next = first;
1107	}
1108	sd->groups = first;
1109
1110	return 0;
1111
1112	fail:
1113	free_sched_groups(sg: first, free_sgc: 0);
1114
1115	return -ENOMEM;
1116	}
1117
1118
1119	/*
1120	* Package topology (also see the load-balance blurb in fair.c)
1121	*
1122	* The scheduler builds a tree structure to represent a number of important
1123	* topology features. By default (default_topology[]) these include:
1124	*
1125	* - Simultaneous multithreading (SMT)
1126	* - Multi-Core Cache (MC)
1127	* - Package (PKG)
1128	*
1129	* Where the last one more or less denotes everything up to a NUMA node.
1130	*
1131	* The tree consists of 3 primary data structures:
1132	*
1133	* sched_domain -> sched_group -> sched_group_capacity
1134	* ^ ^ ^ ^
1135	* `-' `-'
1136	*
1137	* The sched_domains are per-CPU and have a two way link (parent & child) and
1138	* denote the ever growing mask of CPUs belonging to that level of topology.
1139	*
1140	* Each sched_domain has a circular (double) linked list of sched_group's, each
1141	* denoting the domains of the level below (or individual CPUs in case of the
1142	* first domain level). The sched_group linked by a sched_domain includes the
1143	* CPU of that sched_domain [*].
1144	*
1145	* Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1146	*
1147	* CPU 0 1 2 3 4 5 6 7
1148	*
1149	* PKG [ ]
1150	* MC [ ] [ ]
1151	* SMT [ ] [ ] [ ] [ ]
1152	*
1153	* - or -
1154	*
1155	* PKG 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1156	* MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1157	* SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1158	*
1159	* CPU 0 1 2 3 4 5 6 7
1160	*
1161	* One way to think about it is: sched_domain moves you up and down among these
1162	* topology levels, while sched_group moves you sideways through it, at child
1163	* domain granularity.
1164	*
1165	* sched_group_capacity ensures each unique sched_group has shared storage.
1166	*
1167	* There are two related construction problems, both require a CPU that
1168	* uniquely identify each group (for a given domain):
1169	*
1170	* - The first is the balance_cpu (see should_we_balance() and the
1171	* load-balance blurb in fair.c); for each group we only want 1 CPU to
1172	* continue balancing at a higher domain.
1173	*
1174	* - The second is the sched_group_capacity; we want all identical groups
1175	* to share a single sched_group_capacity.
1176	*
1177	* Since these topologies are exclusive by construction. That is, its
1178	* impossible for an SMT thread to belong to multiple cores, and cores to
1179	* be part of multiple caches. There is a very clear and unique location
1180	* for each CPU in the hierarchy.
1181	*
1182	* Therefore computing a unique CPU for each group is trivial (the iteration
1183	* mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1184	* group), we can simply pick the first CPU in each group.
1185	*
1186	*
1187	* [*] in other words, the first group of each domain is its child domain.
1188	*/
1189
1190	static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1191	{
1192	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1193	struct sched_domain *child = sd->child;
1194	struct sched_group *sg;
1195	bool already_visited;
1196
1197	if (child)
1198	cpu = cpumask_first(srcp: sched_domain_span(sd: child));
1199
1200	sg = *per_cpu_ptr(sdd->sg, cpu);
1201	sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1202
1203	/* Increase refcounts for claim_allocations: */
1204	already_visited = atomic_inc_return(v: &sg->ref) > 1;
1205	/* sgc visits should follow a similar trend as sg */
1206	WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1207
1208	/* If we have already visited that group, it's already initialized. */
1209	if (already_visited)
1210	return sg;
1211
1212	if (child) {
1213	cpumask_copy(dstp: sched_group_span(sg), srcp: sched_domain_span(sd: child));
1214	cpumask_copy(dstp: group_balance_mask(sg), srcp: sched_group_span(sg));
1215	sg->flags = child->flags;
1216	} else {
1217	cpumask_set_cpu(cpu, dstp: sched_group_span(sg));
1218	cpumask_set_cpu(cpu, dstp: group_balance_mask(sg));
1219	}
1220
1221	sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(srcp: sched_group_span(sg));
1222	sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1223	sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1224
1225	return sg;
1226	}
1227
1228	/*
1229	* build_sched_groups will build a circular linked list of the groups
1230	* covered by the given span, will set each group's ->cpumask correctly,
1231	* and will initialize their ->sgc.
1232	*
1233	* Assumes the sched_domain tree is fully constructed
1234	*/
1235	static int
1236	build_sched_groups(struct sched_domain *sd, int cpu)
1237	{
1238	struct sched_group *first = NULL, *last = NULL;
1239	struct sd_data *sdd = sd->private;
1240	const struct cpumask *span = sched_domain_span(sd);
1241	struct cpumask *covered;
1242	int i;
1243
1244	lockdep_assert_held(&sched_domains_mutex);
1245	covered = sched_domains_tmpmask;
1246
1247	cpumask_clear(dstp: covered);
1248
1249	for_each_cpu_wrap(i, span, cpu) {
1250	struct sched_group *sg;
1251
1252	if (cpumask_test_cpu(cpu: i, cpumask: covered))
1253	continue;
1254
1255	sg = get_group(cpu: i, sdd);
1256
1257	cpumask_or(dstp: covered, src1p: covered, src2p: sched_group_span(sg));
1258
1259	if (!first)
1260	first = sg;
1261	if (last)
1262	last->next = sg;
1263	last = sg;
1264	}
1265	last->next = first;
1266	sd->groups = first;
1267
1268	return 0;
1269	}
1270
1271	/*
1272	* Initialize sched groups cpu_capacity.
1273	*
1274	* cpu_capacity indicates the capacity of sched group, which is used while
1275	* distributing the load between different sched groups in a sched domain.
1276	* Typically cpu_capacity for all the groups in a sched domain will be same
1277	* unless there are asymmetries in the topology. If there are asymmetries,
1278	* group having more cpu_capacity will pickup more load compared to the
1279	* group having less cpu_capacity.
1280	*/
1281	static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1282	{
1283	struct sched_group *sg = sd->groups;
1284	struct cpumask *mask = sched_domains_tmpmask2;
1285
1286	WARN_ON(!sg);
1287
1288	do {
1289	int cpu, cores = 0, max_cpu = -1;
1290
1291	sg->group_weight = cpumask_weight(srcp: sched_group_span(sg));
1292
1293	cpumask_copy(dstp: mask, srcp: sched_group_span(sg));
1294	for_each_cpu(cpu, mask) {
1295	cores++;
1296	#ifdef CONFIG_SCHED_SMT
1297	cpumask_andnot(dstp: mask, src1p: mask, src2p: cpu_smt_mask(cpu));
1298	#endif
1299	}
1300	sg->cores = cores;
1301
1302	if (!(sd->flags & SD_ASYM_PACKING))
1303	goto next;
1304
1305	for_each_cpu(cpu, sched_group_span(sg)) {
1306	if (max_cpu < 0)
1307	max_cpu = cpu;
1308	else if (sched_asym_prefer(a: cpu, b: max_cpu))
1309	max_cpu = cpu;
1310	}
1311	sg->asym_prefer_cpu = max_cpu;
1312
1313	next:
1314	sg = sg->next;
1315	} while (sg != sd->groups);
1316
1317	if (cpu != group_balance_cpu(sg))
1318	return;
1319
1320	update_group_capacity(sd, cpu);
1321	}
1322
1323	/* Update the "asym_prefer_cpu" when arch_asym_cpu_priority() changes. */
1324	void sched_update_asym_prefer_cpu(int cpu, int old_prio, int new_prio)
1325	{
1326	int asym_prefer_cpu = cpu;
1327	struct sched_domain *sd;
1328
1329	guard(rcu)();
1330
1331	for_each_domain(cpu, sd) {
1332	struct sched_group *sg;
1333	int group_cpu;
1334
1335	if (!(sd->flags & SD_ASYM_PACKING))
1336	continue;
1337
1338	/*
1339	* Groups of overlapping domain are replicated per NUMA
1340	* node and will require updating "asym_prefer_cpu" on
1341	* each local copy.
1342	*
1343	* If you are hitting this warning, consider moving
1344	* "sg->asym_prefer_cpu" to "sg->sgc->asym_prefer_cpu"
1345	* which is shared by all the overlapping groups.
1346	*/
1347	WARN_ON_ONCE(sd->flags & SD_NUMA);
1348
1349	sg = sd->groups;
1350	if (cpu != sg->asym_prefer_cpu) {
1351	/*
1352	* Since the parent is a superset of the current group,
1353	* if the cpu is not the "asym_prefer_cpu" at the
1354	* current level, it cannot be the preferred CPU at a
1355	* higher levels either.
1356	*/
1357	if (!sched_asym_prefer(a: cpu, b: sg->asym_prefer_cpu))
1358	return;
1359
1360	WRITE_ONCE(sg->asym_prefer_cpu, cpu);
1361	continue;
1362	}
1363
1364	/* Ranking has improved; CPU is still the preferred one. */
1365	if (new_prio >= old_prio)
1366	continue;
1367
1368	for_each_cpu(group_cpu, sched_group_span(sg)) {
1369	if (sched_asym_prefer(a: group_cpu, b: asym_prefer_cpu))
1370	asym_prefer_cpu = group_cpu;
1371	}
1372
1373	WRITE_ONCE(sg->asym_prefer_cpu, asym_prefer_cpu);
1374	}
1375	}
1376
1377	/*
1378	* Set of available CPUs grouped by their corresponding capacities
1379	* Each list entry contains a CPU mask reflecting CPUs that share the same
1380	* capacity.
1381	* The lifespan of data is unlimited.
1382	*/
1383	LIST_HEAD(asym_cap_list);
1384
1385	/*
1386	* Verify whether there is any CPU capacity asymmetry in a given sched domain.
1387	* Provides sd_flags reflecting the asymmetry scope.
1388	*/
1389	static inline int
1390	asym_cpu_capacity_classify(const struct cpumask *sd_span,
1391	const struct cpumask *cpu_map)
1392	{
1393	struct asym_cap_data *entry;
1394	int count = 0, miss = 0;
1395
1396	/*
1397	* Count how many unique CPU capacities this domain spans across
1398	* (compare sched_domain CPUs mask with ones representing available
1399	* CPUs capacities). Take into account CPUs that might be offline:
1400	* skip those.
1401	*/
1402	list_for_each_entry(entry, &asym_cap_list, link) {
1403	if (cpumask_intersects(src1p: sd_span, cpu_capacity_span(entry)))
1404	++count;
1405	else if (cpumask_intersects(src1p: cpu_map, cpu_capacity_span(entry)))
1406	++miss;
1407	}
1408
1409	WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1410
1411	/* No asymmetry detected */
1412	if (count < 2)
1413	return 0;
1414	/* Some of the available CPU capacity values have not been detected */
1415	if (miss)
1416	return SD_ASYM_CPUCAPACITY;
1417
1418	/* Full asymmetry */
1419	return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1420
1421	}
1422
1423	static void free_asym_cap_entry(struct rcu_head *head)
1424	{
1425	struct asym_cap_data *entry = container_of(head, struct asym_cap_data, rcu);
1426	kfree(objp: entry);
1427	}
1428
1429	static inline void asym_cpu_capacity_update_data(int cpu)
1430	{
1431	unsigned long capacity = arch_scale_cpu_capacity(cpu);
1432	struct asym_cap_data *insert_entry = NULL;
1433	struct asym_cap_data *entry;
1434
1435	/*
1436	* Search if capacity already exits. If not, track which the entry
1437	* where we should insert to keep the list ordered descending.
1438	*/
1439	list_for_each_entry(entry, &asym_cap_list, link) {
1440	if (capacity == entry->capacity)
1441	goto done;
1442	else if (!insert_entry && capacity > entry->capacity)
1443	insert_entry = list_prev_entry(entry, link);
1444	}
1445
1446	entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1447	if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1448	return;
1449	entry->capacity = capacity;
1450
1451	/* If NULL then the new capacity is the smallest, add last. */
1452	if (!insert_entry)
1453	list_add_tail_rcu(new: &entry->link, head: &asym_cap_list);
1454	else
1455	list_add_rcu(new: &entry->link, head: &insert_entry->link);
1456	done:
1457	__cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1458	}
1459
1460	/*
1461	* Build-up/update list of CPUs grouped by their capacities
1462	* An update requires explicit request to rebuild sched domains
1463	* with state indicating CPU topology changes.
1464	*/
1465	static void asym_cpu_capacity_scan(void)
1466	{
1467	struct asym_cap_data *entry, *next;
1468	int cpu;
1469
1470	list_for_each_entry(entry, &asym_cap_list, link)
1471	cpumask_clear(cpu_capacity_span(entry));
1472
1473	for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1474	asym_cpu_capacity_update_data(cpu);
1475
1476	list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1477	if (cpumask_empty(cpu_capacity_span(entry))) {
1478	list_del_rcu(entry: &entry->link);
1479	call_rcu(head: &entry->rcu, func: free_asym_cap_entry);
1480	}
1481	}
1482
1483	/*
1484	* Only one capacity value has been detected i.e. this system is symmetric.
1485	* No need to keep this data around.
1486	*/
1487	if (list_is_singular(head: &asym_cap_list)) {
1488	entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1489	list_del_rcu(entry: &entry->link);
1490	call_rcu(head: &entry->rcu, func: free_asym_cap_entry);
1491	}
1492	}
1493
1494	/*
1495	* Initializers for schedule domains
1496	* Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1497	*/
1498
1499	static int default_relax_domain_level = -1;
1500	int sched_domain_level_max;
1501
1502	static int __init setup_relax_domain_level(char *str)
1503	{
1504	if (kstrtoint(s: str, base: 0, res: &default_relax_domain_level))
1505	pr_warn("Unable to set relax_domain_level\n");
1506
1507	return 1;
1508	}
1509	__setup("relax_domain_level=", setup_relax_domain_level);
1510
1511	static void set_domain_attribute(struct sched_domain *sd,
1512	struct sched_domain_attr *attr)
1513	{
1514	int request;
1515
1516	if (!attr || attr->relax_domain_level < 0) {
1517	if (default_relax_domain_level < 0)
1518	return;
1519	request = default_relax_domain_level;
1520	} else
1521	request = attr->relax_domain_level;
1522
1523	if (sd->level >= request) {
1524	/* Turn off idle balance on this domain: */
1525	sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1526	}
1527	}
1528
1529	static void __sdt_free(const struct cpumask *cpu_map);
1530	static int __sdt_alloc(const struct cpumask *cpu_map);
1531
1532	static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1533	const struct cpumask *cpu_map)
1534	{
1535	switch (what) {
1536	case sa_rootdomain:
1537	if (!atomic_read(v: &d->rd->refcount))
1538	free_rootdomain(rcu: &d->rd->rcu);
1539	fallthrough;
1540	case sa_sd:
1541	free_percpu(pdata: d->sd);
1542	fallthrough;
1543	case sa_sd_storage:
1544	__sdt_free(cpu_map);
1545	fallthrough;
1546	case sa_none:
1547	break;
1548	}
1549	}
1550
1551	static enum s_alloc
1552	__visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1553	{
1554	memset(d, 0, sizeof(*d));
1555
1556	if (__sdt_alloc(cpu_map))
1557	return sa_sd_storage;
1558	d->sd = alloc_percpu(struct sched_domain *);
1559	if (!d->sd)
1560	return sa_sd_storage;
1561	d->rd = alloc_rootdomain();
1562	if (!d->rd)
1563	return sa_sd;
1564
1565	return sa_rootdomain;
1566	}
1567
1568	/*
1569	* NULL the sd_data elements we've used to build the sched_domain and
1570	* sched_group structure so that the subsequent __free_domain_allocs()
1571	* will not free the data we're using.
1572	*/
1573	static void claim_allocations(int cpu, struct sched_domain *sd)
1574	{
1575	struct sd_data *sdd = sd->private;
1576
1577	WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1578	*per_cpu_ptr(sdd->sd, cpu) = NULL;
1579
1580	if (atomic_read(v: &(*per_cpu_ptr(sdd->sds, cpu))->ref))
1581	*per_cpu_ptr(sdd->sds, cpu) = NULL;
1582
1583	if (atomic_read(v: &(*per_cpu_ptr(sdd->sg, cpu))->ref))
1584	*per_cpu_ptr(sdd->sg, cpu) = NULL;
1585
1586	if (atomic_read(v: &(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1587	*per_cpu_ptr(sdd->sgc, cpu) = NULL;
1588	}
1589
1590	#ifdef CONFIG_NUMA
1591	enum numa_topology_type sched_numa_topology_type;
1592
1593	/*
1594	* sched_domains_numa_distance is derived from sched_numa_node_distance
1595	* and provides a simplified view of NUMA distances used specifically
1596	* for building NUMA scheduling domains.
1597	*/
1598	static int sched_domains_numa_levels;
1599	static int sched_numa_node_levels;
1600
1601	int sched_max_numa_distance;
1602	static int *sched_domains_numa_distance;
1603	static int *sched_numa_node_distance;
1604	static struct cpumask ***sched_domains_numa_masks;
1605	#endif /* CONFIG_NUMA */
1606
1607	/*
1608	* SD_flags allowed in topology descriptions.
1609	*
1610	* These flags are purely descriptive of the topology and do not prescribe
1611	* behaviour. Behaviour is artificial and mapped in the below sd_init()
1612	* function. For details, see include/linux/sched/sd_flags.h.
1613	*
1614	* SD_SHARE_CPUCAPACITY
1615	* SD_SHARE_LLC
1616	* SD_CLUSTER
1617	* SD_NUMA
1618	*
1619	* Odd one out, which beside describing the topology has a quirk also
1620	* prescribes the desired behaviour that goes along with it:
1621	*
1622	* SD_ASYM_PACKING - describes SMT quirks
1623	*/
1624	#define TOPOLOGY_SD_FLAGS \
1625	(SD_SHARE_CPUCAPACITY | \
1626	SD_CLUSTER | \
1627	SD_SHARE_LLC | \
1628	SD_NUMA | \
1629	SD_ASYM_PACKING)
1630
1631	static struct sched_domain *
1632	sd_init(struct sched_domain_topology_level *tl,
1633	const struct cpumask *cpu_map,
1634	struct sched_domain *child, int cpu)
1635	{
1636	struct sd_data *sdd = &tl->data;
1637	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1638	int sd_id, sd_weight, sd_flags = 0;
1639	struct cpumask *sd_span;
1640
1641	sd_weight = cpumask_weight(srcp: tl->mask(tl, cpu));
1642
1643	if (tl->sd_flags)
1644	sd_flags = (*tl->sd_flags)();
1645	if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1646	"wrong sd_flags in topology description\n"))
1647	sd_flags &= TOPOLOGY_SD_FLAGS;
1648
1649	*sd = (struct sched_domain){
1650	.min_interval = sd_weight,
1651	.max_interval = 2*sd_weight,
1652	.busy_factor = 16,
1653	.imbalance_pct = 117,
1654
1655	.cache_nice_tries = 0,
1656
1657	.flags = 1*SD_BALANCE_NEWIDLE
1658	| 1*SD_BALANCE_EXEC
1659	| 1*SD_BALANCE_FORK
1660	| 0*SD_BALANCE_WAKE
1661	| 1*SD_WAKE_AFFINE
1662	| 0*SD_SHARE_CPUCAPACITY
1663	| 0*SD_SHARE_LLC
1664	| 0*SD_SERIALIZE
1665	| 1*SD_PREFER_SIBLING
1666	| 0*SD_NUMA
1667	| sd_flags
1668	,
1669
1670	.last_balance = jiffies,
1671	.balance_interval = sd_weight,
1672
1673	/* 50% success rate */
1674	.newidle_call = 512,
1675	.newidle_success = 256,
1676	.newidle_ratio = 512,
1677
1678	.max_newidle_lb_cost = 0,
1679	.last_decay_max_lb_cost = jiffies,
1680	.child = child,
1681	.name = tl->name,
1682	};
1683
1684	sd_span = sched_domain_span(sd);
1685	cpumask_and(dstp: sd_span, src1p: cpu_map, src2p: tl->mask(tl, cpu));
1686	sd_id = cpumask_first(srcp: sd_span);
1687
1688	sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1689
1690	WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1691	(SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1692	"CPU capacity asymmetry not supported on SMT\n");
1693
1694	/*
1695	* Convert topological properties into behaviour.
1696	*/
1697	/* Don't attempt to spread across CPUs of different capacities. */
1698	if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1699	sd->child->flags &= ~SD_PREFER_SIBLING;
1700
1701	if (sd->flags & SD_SHARE_CPUCAPACITY) {
1702	sd->imbalance_pct = 110;
1703
1704	} else if (sd->flags & SD_SHARE_LLC) {
1705	sd->imbalance_pct = 117;
1706	sd->cache_nice_tries = 1;
1707
1708	#ifdef CONFIG_NUMA
1709	} else if (sd->flags & SD_NUMA) {
1710	sd->cache_nice_tries = 2;
1711
1712	sd->flags &= ~SD_PREFER_SIBLING;
1713	sd->flags |= SD_SERIALIZE;
1714	if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1715	sd->flags &= ~(SD_BALANCE_EXEC |
1716	SD_BALANCE_FORK |
1717	SD_WAKE_AFFINE);
1718	}
1719
1720	#endif /* CONFIG_NUMA */
1721	} else {
1722	sd->cache_nice_tries = 1;
1723	}
1724
1725	/*
1726	* For all levels sharing cache; connect a sched_domain_shared
1727	* instance.
1728	*/
1729	if (sd->flags & SD_SHARE_LLC) {
1730	sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1731	atomic_inc(v: &sd->shared->ref);
1732	atomic_set(v: &sd->shared->nr_busy_cpus, i: sd_weight);
1733	}
1734
1735	sd->private = sdd;
1736
1737	return sd;
1738	}
1739
1740	#ifdef CONFIG_SCHED_SMT
1741	int cpu_smt_flags(void)
1742	{
1743	return SD_SHARE_CPUCAPACITY | SD_SHARE_LLC;
1744	}
1745
1746	const struct cpumask *tl_smt_mask(struct sched_domain_topology_level *tl, int cpu)
1747	{
1748	return cpu_smt_mask(cpu);
1749	}
1750	#endif
1751
1752	#ifdef CONFIG_SCHED_CLUSTER
1753	int cpu_cluster_flags(void)
1754	{
1755	return SD_CLUSTER | SD_SHARE_LLC;
1756	}
1757
1758	const struct cpumask *tl_cls_mask(struct sched_domain_topology_level *tl, int cpu)
1759	{
1760	return cpu_clustergroup_mask(cpu);
1761	}
1762	#endif
1763
1764	#ifdef CONFIG_SCHED_MC
1765	int cpu_core_flags(void)
1766	{
1767	return SD_SHARE_LLC;
1768	}
1769
1770	const struct cpumask *tl_mc_mask(struct sched_domain_topology_level *tl, int cpu)
1771	{
1772	return cpu_coregroup_mask(cpu);
1773	}
1774	#endif
1775
1776	const struct cpumask *tl_pkg_mask(struct sched_domain_topology_level *tl, int cpu)
1777	{
1778	return cpu_node_mask(cpu);
1779	}
1780
1781	/*
1782	* Topology list, bottom-up.
1783	*/
1784	static struct sched_domain_topology_level default_topology[] = {
1785	#ifdef CONFIG_SCHED_SMT
1786	SDTL_INIT(tl_smt_mask, cpu_smt_flags, SMT),
1787	#endif
1788
1789	#ifdef CONFIG_SCHED_CLUSTER
1790	SDTL_INIT(tl_cls_mask, cpu_cluster_flags, CLS),
1791	#endif
1792
1793	#ifdef CONFIG_SCHED_MC
1794	SDTL_INIT(tl_mc_mask, cpu_core_flags, MC),
1795	#endif
1796	SDTL_INIT(tl_pkg_mask, NULL, PKG),
1797	{ NULL, },
1798	};
1799
1800	static struct sched_domain_topology_level *sched_domain_topology =
1801	default_topology;
1802	static struct sched_domain_topology_level *sched_domain_topology_saved;
1803
1804	#define for_each_sd_topology(tl) \
1805	for (tl = sched_domain_topology; tl->mask; tl++)
1806
1807	void __init set_sched_topology(struct sched_domain_topology_level *tl)
1808	{
1809	if (WARN_ON_ONCE(sched_smp_initialized))
1810	return;
1811
1812	sched_domain_topology = tl;
1813	sched_domain_topology_saved = NULL;
1814	}
1815
1816	#ifdef CONFIG_NUMA
1817	static int cpu_numa_flags(void)
1818	{
1819	return SD_NUMA;
1820	}
1821
1822	static const struct cpumask *sd_numa_mask(struct sched_domain_topology_level *tl, int cpu)
1823	{
1824	return sched_domains_numa_masks[tl->numa_level][cpu_to_node(cpu)];
1825	}
1826
1827	static void sched_numa_warn(const char *str)
1828	{
1829	static int done = false;
1830	int i,j;
1831
1832	if (done)
1833	return;
1834
1835	done = true;
1836
1837	printk(KERN_WARNING "ERROR: %s\n\n", str);
1838
1839	for (i = 0; i < nr_node_ids; i++) {
1840	printk(KERN_WARNING " ");
1841	for (j = 0; j < nr_node_ids; j++) {
1842	if (!node_state(node: i, state: N_CPU) || !node_state(node: j, state: N_CPU))
1843	printk(KERN_CONT "(%02d) ", node_distance(i,j));
1844	else
1845	printk(KERN_CONT " %02d ", node_distance(i,j));
1846	}
1847	printk(KERN_CONT "\n");
1848	}
1849	printk(KERN_WARNING "\n");
1850	}
1851
1852	bool find_numa_distance(int distance)
1853	{
1854	bool found = false;
1855	int i, *distances;
1856
1857	if (distance == node_distance(0, 0))
1858	return true;
1859
1860	rcu_read_lock();
1861	distances = rcu_dereference(sched_numa_node_distance);
1862	if (!distances)
1863	goto unlock;
1864	for (i = 0; i < sched_numa_node_levels; i++) {
1865	if (distances[i] == distance) {
1866	found = true;
1867	break;
1868	}
1869	}
1870	unlock:
1871	rcu_read_unlock();
1872
1873	return found;
1874	}
1875
1876	#define for_each_cpu_node_but(n, nbut) \
1877	for_each_node_state(n, N_CPU) \
1878	if (n == nbut) \
1879	continue; \
1880	else
1881
1882	/*
1883	* A system can have three types of NUMA topology:
1884	* NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1885	* NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1886	* NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1887	*
1888	* The difference between a glueless mesh topology and a backplane
1889	* topology lies in whether communication between not directly
1890	* connected nodes goes through intermediary nodes (where programs
1891	* could run), or through backplane controllers. This affects
1892	* placement of programs.
1893	*
1894	* The type of topology can be discerned with the following tests:
1895	* - If the maximum distance between any nodes is 1 hop, the system
1896	* is directly connected.
1897	* - If for two nodes A and B, located N > 1 hops away from each other,
1898	* there is an intermediary node C, which is < N hops away from both
1899	* nodes A and B, the system is a glueless mesh.
1900	*/
1901	static void init_numa_topology_type(int offline_node)
1902	{
1903	int a, b, c, n;
1904
1905	n = sched_max_numa_distance;
1906
1907	if (sched_domains_numa_levels <= 2) {
1908	sched_numa_topology_type = NUMA_DIRECT;
1909	return;
1910	}
1911
1912	for_each_cpu_node_but(a, offline_node) {
1913	for_each_cpu_node_but(b, offline_node) {
1914	/* Find two nodes furthest removed from each other. */
1915	if (node_distance(a, b) < n)
1916	continue;
1917
1918	/* Is there an intermediary node between a and b? */
1919	for_each_cpu_node_but(c, offline_node) {
1920	if (node_distance(a, c) < n &&
1921	node_distance(b, c) < n) {
1922	sched_numa_topology_type =
1923	NUMA_GLUELESS_MESH;
1924	return;
1925	}
1926	}
1927
1928	sched_numa_topology_type = NUMA_BACKPLANE;
1929	return;
1930	}
1931	}
1932
1933	pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1934	sched_numa_topology_type = NUMA_DIRECT;
1935	}
1936
1937
1938	#define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1939
1940	/*
1941	* An architecture could modify its NUMA distance, to change
1942	* grouping of NUMA nodes and number of NUMA levels when creating
1943	* NUMA level sched domains.
1944	*
1945	* A NUMA level is created for each unique
1946	* arch_sched_node_distance.
1947	*/
1948	static int numa_node_dist(int i, int j)
1949	{
1950	return node_distance(i, j);
1951	}
1952
1953	int arch_sched_node_distance(int from, int to)
1954	__weak __alias(numa_node_dist);
1955
1956	static bool modified_sched_node_distance(void)
1957	{
1958	return numa_node_dist != arch_sched_node_distance;
1959	}
1960
1961	static int sched_record_numa_dist(int offline_node, int (*n_dist)(int, int),
1962	int **dist, int *levels)
1963	{
1964	unsigned long *distance_map __free(bitmap) = NULL;
1965	int nr_levels = 0;
1966	int i, j;
1967	int *distances;
1968
1969	/*
1970	* O(nr_nodes^2) de-duplicating selection sort -- in order to find the
1971	* unique distances in the node_distance() table.
1972	*/
1973	distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1974	if (!distance_map)
1975	return -ENOMEM;
1976
1977	bitmap_zero(dst: distance_map, NR_DISTANCE_VALUES);
1978	for_each_cpu_node_but(i, offline_node) {
1979	for_each_cpu_node_but(j, offline_node) {
1980	int distance = n_dist(i, j);
1981
1982	if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1983	sched_numa_warn(str: "Invalid distance value range");
1984	return -EINVAL;
1985	}
1986
1987	bitmap_set(map: distance_map, start: distance, nbits: 1);
1988	}
1989	}
1990	/*
1991	* We can now figure out how many unique distance values there are and
1992	* allocate memory accordingly.
1993	*/
1994	nr_levels = bitmap_weight(src: distance_map, NR_DISTANCE_VALUES);
1995
1996	distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1997	if (!distances)
1998	return -ENOMEM;
1999
2000	for (i = 0, j = 0; i < nr_levels; i++, j++) {
2001	j = find_next_bit(addr: distance_map, NR_DISTANCE_VALUES, offset: j);
2002	distances[i] = j;
2003	}
2004	*dist = distances;
2005	*levels = nr_levels;
2006
2007	return 0;
2008	}
2009
2010	void sched_init_numa(int offline_node)
2011	{
2012	struct sched_domain_topology_level *tl;
2013	int nr_levels, nr_node_levels;
2014	int i, j;
2015	int *distances, *domain_distances;
2016	struct cpumask ***masks;
2017
2018	/* Record the NUMA distances from SLIT table */
2019	if (sched_record_numa_dist(offline_node, n_dist: numa_node_dist, dist: &distances,
2020	levels: &nr_node_levels))
2021	return;
2022
2023	/* Record modified NUMA distances for building sched domains */
2024	if (modified_sched_node_distance()) {
2025	if (sched_record_numa_dist(offline_node, n_dist: arch_sched_node_distance,
2026	dist: &domain_distances, levels: &nr_levels)) {
2027	kfree(objp: distances);
2028	return;
2029	}
2030	} else {
2031	domain_distances = distances;
2032	nr_levels = nr_node_levels;
2033	}
2034	rcu_assign_pointer(sched_numa_node_distance, distances);
2035	WRITE_ONCE(sched_max_numa_distance, distances[nr_node_levels - 1]);
2036	WRITE_ONCE(sched_numa_node_levels, nr_node_levels);
2037
2038	/*
2039	* 'nr_levels' contains the number of unique distances
2040	*
2041	* The sched_domains_numa_distance[] array includes the actual distance
2042	* numbers.
2043	*/
2044
2045	/*
2046	* Here, we should temporarily reset sched_domains_numa_levels to 0.
2047	* If it fails to allocate memory for array sched_domains_numa_masks[][],
2048	* the array will contain less then 'nr_levels' members. This could be
2049	* dangerous when we use it to iterate array sched_domains_numa_masks[][]
2050	* in other functions.
2051	*
2052	* We reset it to 'nr_levels' at the end of this function.
2053	*/
2054	rcu_assign_pointer(sched_domains_numa_distance, domain_distances);
2055
2056	sched_domains_numa_levels = 0;
2057
2058	masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
2059	if (!masks)
2060	return;
2061
2062	/*
2063	* Now for each level, construct a mask per node which contains all
2064	* CPUs of nodes that are that many hops away from us.
2065	*/
2066	for (i = 0; i < nr_levels; i++) {
2067	masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
2068	if (!masks[i])
2069	return;
2070
2071	for_each_cpu_node_but(j, offline_node) {
2072	struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
2073	int k;
2074
2075	if (!mask)
2076	return;
2077
2078	masks[i][j] = mask;
2079
2080	for_each_cpu_node_but(k, offline_node) {
2081	if (sched_debug() &&
2082	(arch_sched_node_distance(from: j, to: k) !=
2083	arch_sched_node_distance(from: k, to: j)))
2084	sched_numa_warn(str: "Node-distance not symmetric");
2085
2086	if (arch_sched_node_distance(from: j, to: k) >
2087	sched_domains_numa_distance[i])
2088	continue;
2089
2090	cpumask_or(dstp: mask, src1p: mask, src2p: cpumask_of_node(node: k));
2091	}
2092	}
2093	}
2094	rcu_assign_pointer(sched_domains_numa_masks, masks);
2095
2096	/* Compute default topology size */
2097	for (i = 0; sched_domain_topology[i].mask; i++);
2098
2099	tl = kzalloc((i + nr_levels + 1) *
2100	sizeof(struct sched_domain_topology_level), GFP_KERNEL);
2101	if (!tl)
2102	return;
2103
2104	/*
2105	* Copy the default topology bits..
2106	*/
2107	for (i = 0; sched_domain_topology[i].mask; i++)
2108	tl[i] = sched_domain_topology[i];
2109
2110	/*
2111	* Add the NUMA identity distance, aka single NODE.
2112	*/
2113	tl[i++] = SDTL_INIT(sd_numa_mask, NULL, NODE);
2114
2115	/*
2116	* .. and append 'j' levels of NUMA goodness.
2117	*/
2118	for (j = 1; j < nr_levels; i++, j++) {
2119	tl[i] = SDTL_INIT(sd_numa_mask, cpu_numa_flags, NUMA);
2120	tl[i].numa_level = j;
2121	}
2122
2123	sched_domain_topology_saved = sched_domain_topology;
2124	sched_domain_topology = tl;
2125
2126	sched_domains_numa_levels = nr_levels;
2127
2128	init_numa_topology_type(offline_node);
2129	}
2130
2131
2132	static void sched_reset_numa(void)
2133	{
2134	int nr_levels, *distances, *dom_distances = NULL;
2135	struct cpumask ***masks;
2136
2137	nr_levels = sched_domains_numa_levels;
2138	sched_numa_node_levels = 0;
2139	sched_domains_numa_levels = 0;
2140	sched_max_numa_distance = 0;
2141	sched_numa_topology_type = NUMA_DIRECT;
2142	distances = sched_numa_node_distance;
2143	if (sched_numa_node_distance != sched_domains_numa_distance)
2144	dom_distances = sched_domains_numa_distance;
2145	rcu_assign_pointer(sched_numa_node_distance, NULL);
2146	rcu_assign_pointer(sched_domains_numa_distance, NULL);
2147	masks = sched_domains_numa_masks;
2148	rcu_assign_pointer(sched_domains_numa_masks, NULL);
2149	if (distances || masks) {
2150	int i, j;
2151
2152	synchronize_rcu();
2153	kfree(objp: distances);
2154	kfree(objp: dom_distances);
2155	for (i = 0; i < nr_levels && masks; i++) {
2156	if (!masks[i])
2157	continue;
2158	for_each_node(j)
2159	kfree(objp: masks[i][j]);
2160	kfree(objp: masks[i]);
2161	}
2162	kfree(objp: masks);
2163	}
2164	if (sched_domain_topology_saved) {
2165	kfree(objp: sched_domain_topology);
2166	sched_domain_topology = sched_domain_topology_saved;
2167	sched_domain_topology_saved = NULL;
2168	}
2169	}
2170
2171	/*
2172	* Call with hotplug lock held
2173	*/
2174	void sched_update_numa(int cpu, bool online)
2175	{
2176	int node;
2177
2178	node = cpu_to_node(cpu);
2179	/*
2180	* Scheduler NUMA topology is updated when the first CPU of a
2181	* node is onlined or the last CPU of a node is offlined.
2182	*/
2183	if (cpumask_weight(srcp: cpumask_of_node(node)) != 1)
2184	return;
2185
2186	sched_reset_numa();
2187	sched_init_numa(offline_node: online ? NUMA_NO_NODE : node);
2188	}
2189
2190	void sched_domains_numa_masks_set(unsigned int cpu)
2191	{
2192	int node = cpu_to_node(cpu);
2193	int i, j;
2194
2195	for (i = 0; i < sched_domains_numa_levels; i++) {
2196	for (j = 0; j < nr_node_ids; j++) {
2197	if (!node_state(node: j, state: N_CPU))
2198	continue;
2199
2200	/* Set ourselves in the remote node's masks */
2201	if (arch_sched_node_distance(from: j, to: node) <=
2202	sched_domains_numa_distance[i])
2203	cpumask_set_cpu(cpu, dstp: sched_domains_numa_masks[i][j]);
2204	}
2205	}
2206	}
2207
2208	void sched_domains_numa_masks_clear(unsigned int cpu)
2209	{
2210	int i, j;
2211
2212	for (i = 0; i < sched_domains_numa_levels; i++) {
2213	for (j = 0; j < nr_node_ids; j++) {
2214	if (sched_domains_numa_masks[i][j])
2215	cpumask_clear_cpu(cpu, dstp: sched_domains_numa_masks[i][j]);
2216	}
2217	}
2218	}
2219
2220	/*
2221	* sched_numa_find_closest() - given the NUMA topology, find the cpu
2222	* closest to @cpu from @cpumask.
2223	* cpumask: cpumask to find a cpu from
2224	* cpu: cpu to be close to
2225	*
2226	* returns: cpu, or nr_cpu_ids when nothing found.
2227	*/
2228	int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2229	{
2230	int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2231	struct cpumask ***masks;
2232
2233	rcu_read_lock();
2234	masks = rcu_dereference(sched_domains_numa_masks);
2235	if (!masks)
2236	goto unlock;
2237	for (i = 0; i < sched_domains_numa_levels; i++) {
2238	if (!masks[i][j])
2239	break;
2240	cpu = cpumask_any_and_distribute(src1p: cpus, src2p: masks[i][j]);
2241	if (cpu < nr_cpu_ids) {
2242	found = cpu;
2243	break;
2244	}
2245	}
2246	unlock:
2247	rcu_read_unlock();
2248
2249	return found;
2250	}
2251
2252	struct __cmp_key {
2253	const struct cpumask *cpus;
2254	struct cpumask ***masks;
2255	int node;
2256	int cpu;
2257	int w;
2258	};
2259
2260	static int hop_cmp(const void *a, const void *b)
2261	{
2262	struct cpumask **prev_hop, **cur_hop = *(struct cpumask ***)b;
2263	struct __cmp_key *k = (struct __cmp_key *)a;
2264
2265	if (cpumask_weight_and(srcp1: k->cpus, srcp2: cur_hop[k->node]) <= k->cpu)
2266	return 1;
2267
2268	if (b == k->masks) {
2269	k->w = 0;
2270	return 0;
2271	}
2272
2273	prev_hop = *((struct cpumask ***)b - 1);
2274	k->w = cpumask_weight_and(srcp1: k->cpus, srcp2: prev_hop[k->node]);
2275	if (k->w <= k->cpu)
2276	return 0;
2277
2278	return -1;
2279	}
2280
2281	/**
2282	* sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth closest CPU
2283	* from @cpus to @cpu, taking into account distance
2284	* from a given @node.
2285	* @cpus: cpumask to find a cpu from
2286	* @cpu: CPU to start searching
2287	* @node: NUMA node to order CPUs by distance
2288	*
2289	* Return: cpu, or nr_cpu_ids when nothing found.
2290	*/
2291	int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
2292	{
2293	struct __cmp_key k = { .cpus = cpus, .cpu = cpu };
2294	struct cpumask ***hop_masks;
2295	int hop, ret = nr_cpu_ids;
2296
2297	if (node == NUMA_NO_NODE)
2298	return cpumask_nth_and(cpu, srcp1: cpus, cpu_online_mask);
2299
2300	rcu_read_lock();
2301
2302	/* CPU-less node entries are uninitialized in sched_domains_numa_masks */
2303	node = numa_nearest_node(node, state: N_CPU);
2304	k.node = node;
2305
2306	k.masks = rcu_dereference(sched_domains_numa_masks);
2307	if (!k.masks)
2308	goto unlock;
2309
2310	hop_masks = bsearch(key: &k, base: k.masks, num: sched_domains_numa_levels, size: sizeof(k.masks[0]), cmp: hop_cmp);
2311	if (!hop_masks)
2312	goto unlock;
2313	hop = hop_masks - k.masks;
2314
2315	ret = hop ?
2316	cpumask_nth_and_andnot(cpu: cpu - k.w, srcp1: cpus, srcp2: k.masks[hop][node], srcp3: k.masks[hop-1][node]) :
2317	cpumask_nth_and(cpu, srcp1: cpus, srcp2: k.masks[0][node]);
2318	unlock:
2319	rcu_read_unlock();
2320	return ret;
2321	}
2322	EXPORT_SYMBOL_GPL(sched_numa_find_nth_cpu);
2323
2324	/**
2325	* sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2326	* @node
2327	* @node: The node to count hops from.
2328	* @hops: Include CPUs up to that many hops away. 0 means local node.
2329	*
2330	* Return: On success, a pointer to a cpumask of CPUs at most @hops away from
2331	* @node, an error value otherwise.
2332	*
2333	* Requires rcu_lock to be held. Returned cpumask is only valid within that
2334	* read-side section, copy it if required beyond that.
2335	*
2336	* Note that not all hops are equal in distance; see sched_init_numa() for how
2337	* distances and masks are handled.
2338	* Also note that this is a reflection of sched_domains_numa_masks, which may change
2339	* during the lifetime of the system (offline nodes are taken out of the masks).
2340	*/
2341	const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops)
2342	{
2343	struct cpumask ***masks;
2344
2345	if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
2346	return ERR_PTR(error: -EINVAL);
2347
2348	masks = rcu_dereference(sched_domains_numa_masks);
2349	if (!masks)
2350	return ERR_PTR(error: -EBUSY);
2351
2352	return masks[hops][node];
2353	}
2354	EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
2355
2356	#endif /* CONFIG_NUMA */
2357
2358	static int __sdt_alloc(const struct cpumask *cpu_map)
2359	{
2360	struct sched_domain_topology_level *tl;
2361	int j;
2362
2363	for_each_sd_topology(tl) {
2364	struct sd_data *sdd = &tl->data;
2365
2366	sdd->sd = alloc_percpu(struct sched_domain *);
2367	if (!sdd->sd)
2368	return -ENOMEM;
2369
2370	sdd->sds = alloc_percpu(struct sched_domain_shared *);
2371	if (!sdd->sds)
2372	return -ENOMEM;
2373
2374	sdd->sg = alloc_percpu(struct sched_group *);
2375	if (!sdd->sg)
2376	return -ENOMEM;
2377
2378	sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2379	if (!sdd->sgc)
2380	return -ENOMEM;
2381
2382	for_each_cpu(j, cpu_map) {
2383	struct sched_domain *sd;
2384	struct sched_domain_shared *sds;
2385	struct sched_group *sg;
2386	struct sched_group_capacity *sgc;
2387
2388	sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2389	GFP_KERNEL, cpu_to_node(j));
2390	if (!sd)
2391	return -ENOMEM;
2392
2393	*per_cpu_ptr(sdd->sd, j) = sd;
2394
2395	sds = kzalloc_node(sizeof(struct sched_domain_shared),
2396	GFP_KERNEL, cpu_to_node(j));
2397	if (!sds)
2398	return -ENOMEM;
2399
2400	*per_cpu_ptr(sdd->sds, j) = sds;
2401
2402	sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2403	GFP_KERNEL, cpu_to_node(j));
2404	if (!sg)
2405	return -ENOMEM;
2406
2407	sg->next = sg;
2408
2409	*per_cpu_ptr(sdd->sg, j) = sg;
2410
2411	sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2412	GFP_KERNEL, cpu_to_node(j));
2413	if (!sgc)
2414	return -ENOMEM;
2415
2416	sgc->id = j;
2417
2418	*per_cpu_ptr(sdd->sgc, j) = sgc;
2419	}
2420	}
2421
2422	return 0;
2423	}
2424
2425	static void __sdt_free(const struct cpumask *cpu_map)
2426	{
2427	struct sched_domain_topology_level *tl;
2428	int j;
2429
2430	for_each_sd_topology(tl) {
2431	struct sd_data *sdd = &tl->data;
2432
2433	for_each_cpu(j, cpu_map) {
2434	struct sched_domain *sd;
2435
2436	if (sdd->sd) {
2437	sd = *per_cpu_ptr(sdd->sd, j);
2438	if (sd && (sd->flags & SD_NUMA))
2439	free_sched_groups(sg: sd->groups, free_sgc: 0);
2440	kfree(objp: *per_cpu_ptr(sdd->sd, j));
2441	}
2442
2443	if (sdd->sds)
2444	kfree(objp: *per_cpu_ptr(sdd->sds, j));
2445	if (sdd->sg)
2446	kfree(objp: *per_cpu_ptr(sdd->sg, j));
2447	if (sdd->sgc)
2448	kfree(objp: *per_cpu_ptr(sdd->sgc, j));
2449	}
2450	free_percpu(pdata: sdd->sd);
2451	sdd->sd = NULL;
2452	free_percpu(pdata: sdd->sds);
2453	sdd->sds = NULL;
2454	free_percpu(pdata: sdd->sg);
2455	sdd->sg = NULL;
2456	free_percpu(pdata: sdd->sgc);
2457	sdd->sgc = NULL;
2458	}
2459	}
2460
2461	static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2462	const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2463	struct sched_domain *child, int cpu)
2464	{
2465	struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2466
2467	if (child) {
2468	sd->level = child->level + 1;
2469	sched_domain_level_max = max(sched_domain_level_max, sd->level);
2470	child->parent = sd;
2471
2472	if (!cpumask_subset(src1p: sched_domain_span(sd: child),
2473	src2p: sched_domain_span(sd))) {
2474	pr_err("BUG: arch topology borken\n");
2475	pr_err(" the %s domain not a subset of the %s domain\n",
2476	child->name, sd->name);
2477	/* Fixup, ensure @sd has at least @child CPUs. */
2478	cpumask_or(dstp: sched_domain_span(sd),
2479	src1p: sched_domain_span(sd),
2480	src2p: sched_domain_span(sd: child));
2481	}
2482
2483	}
2484	set_domain_attribute(sd, attr);
2485
2486	return sd;
2487	}
2488
2489	/*
2490	* Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2491	* any two given CPUs on non-NUMA topology levels.
2492	*/
2493	static bool topology_span_sane(const struct cpumask *cpu_map)
2494	{
2495	struct sched_domain_topology_level *tl;
2496	struct cpumask *covered, *id_seen;
2497	int cpu;
2498
2499	lockdep_assert_held(&sched_domains_mutex);
2500	covered = sched_domains_tmpmask;
2501	id_seen = sched_domains_tmpmask2;
2502
2503	for_each_sd_topology(tl) {
2504	int tl_common_flags = 0;
2505
2506	if (tl->sd_flags)
2507	tl_common_flags = (*tl->sd_flags)();
2508
2509	/* NUMA levels are allowed to overlap */
2510	if (tl_common_flags & SD_NUMA)
2511	continue;
2512
2513	cpumask_clear(dstp: covered);
2514	cpumask_clear(dstp: id_seen);
2515
2516	/*
2517	* Non-NUMA levels cannot partially overlap - they must be either
2518	* completely equal or completely disjoint. Otherwise we can end up
2519	* breaking the sched_group lists - i.e. a later get_group() pass
2520	* breaks the linking done for an earlier span.
2521	*/
2522	for_each_cpu(cpu, cpu_map) {
2523	const struct cpumask *tl_cpu_mask = tl->mask(tl, cpu);
2524	int id;
2525
2526	/* lowest bit set in this mask is used as a unique id */
2527	id = cpumask_first(srcp: tl_cpu_mask);
2528
2529	if (cpumask_test_cpu(cpu: id, cpumask: id_seen)) {
2530	/* First CPU has already been seen, ensure identical spans */
2531	if (!cpumask_equal(src1p: tl->mask(tl, id), src2p: tl_cpu_mask))
2532	return false;
2533	} else {
2534	/* First CPU hasn't been seen before, ensure it's a completely new span */
2535	if (cpumask_intersects(src1p: tl_cpu_mask, src2p: covered))
2536	return false;
2537
2538	cpumask_or(dstp: covered, src1p: covered, src2p: tl_cpu_mask);
2539	cpumask_set_cpu(cpu: id, dstp: id_seen);
2540	}
2541	}
2542	}
2543	return true;
2544	}
2545
2546	/*
2547	* Build sched domains for a given set of CPUs and attach the sched domains
2548	* to the individual CPUs
2549	*/
2550	static int
2551	build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2552	{
2553	enum s_alloc alloc_state = sa_none;
2554	struct sched_domain *sd;
2555	struct s_data d;
2556	struct rq *rq = NULL;
2557	int i, ret = -ENOMEM;
2558	bool has_asym = false;
2559	bool has_cluster = false;
2560
2561	if (WARN_ON(cpumask_empty(cpu_map)))
2562	goto error;
2563
2564	alloc_state = __visit_domain_allocation_hell(d: &d, cpu_map);
2565	if (alloc_state != sa_rootdomain)
2566	goto error;
2567
2568	/* Set up domains for CPUs specified by the cpu_map: */
2569	for_each_cpu(i, cpu_map) {
2570	struct sched_domain_topology_level *tl;
2571
2572	sd = NULL;
2573	for_each_sd_topology(tl) {
2574
2575	sd = build_sched_domain(tl, cpu_map, attr, child: sd, cpu: i);
2576
2577	has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2578
2579	if (tl == sched_domain_topology)
2580	*per_cpu_ptr(d.sd, i) = sd;
2581	if (cpumask_equal(src1p: cpu_map, src2p: sched_domain_span(sd)))
2582	break;
2583	}
2584	}
2585
2586	if (WARN_ON(!topology_span_sane(cpu_map)))
2587	goto error;
2588
2589	/* Build the groups for the domains */
2590	for_each_cpu(i, cpu_map) {
2591	for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2592	sd->span_weight = cpumask_weight(srcp: sched_domain_span(sd));
2593	if (sd->flags & SD_NUMA) {
2594	if (build_overlap_sched_groups(sd, cpu: i))
2595	goto error;
2596	} else {
2597	if (build_sched_groups(sd, cpu: i))
2598	goto error;
2599	}
2600	}
2601	}
2602
2603	/*
2604	* Calculate an allowed NUMA imbalance such that LLCs do not get
2605	* imbalanced.
2606	*/
2607	for_each_cpu(i, cpu_map) {
2608	unsigned int imb = 0;
2609	unsigned int imb_span = 1;
2610
2611	for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2612	struct sched_domain *child = sd->child;
2613
2614	if (!(sd->flags & SD_SHARE_LLC) && child &&
2615	(child->flags & SD_SHARE_LLC)) {
2616	struct sched_domain __rcu *top_p;
2617	unsigned int nr_llcs;
2618
2619	/*
2620	* For a single LLC per node, allow an
2621	* imbalance up to 12.5% of the node. This is
2622	* arbitrary cutoff based two factors -- SMT and
2623	* memory channels. For SMT-2, the intent is to
2624	* avoid premature sharing of HT resources but
2625	* SMT-4 or SMT-8 *may* benefit from a different
2626	* cutoff. For memory channels, this is a very
2627	* rough estimate of how many channels may be
2628	* active and is based on recent CPUs with
2629	* many cores.
2630	*
2631	* For multiple LLCs, allow an imbalance
2632	* until multiple tasks would share an LLC
2633	* on one node while LLCs on another node
2634	* remain idle. This assumes that there are
2635	* enough logical CPUs per LLC to avoid SMT
2636	* factors and that there is a correlation
2637	* between LLCs and memory channels.
2638	*/
2639	nr_llcs = sd->span_weight / child->span_weight;
2640	if (nr_llcs == 1)
2641	imb = sd->span_weight >> 3;
2642	else
2643	imb = nr_llcs;
2644	imb = max(1U, imb);
2645	sd->imb_numa_nr = imb;
2646
2647	/* Set span based on the first NUMA domain. */
2648	top_p = sd->parent;
2649	while (top_p && !(top_p->flags & SD_NUMA)) {
2650	top_p = top_p->parent;
2651	}
2652	imb_span = top_p ? top_p->span_weight : sd->span_weight;
2653	} else {
2654	int factor = max(1U, (sd->span_weight / imb_span));
2655
2656	sd->imb_numa_nr = imb * factor;
2657	}
2658	}
2659	}
2660
2661	/* Calculate CPU capacity for physical packages and nodes */
2662	for (i = nr_cpumask_bits-1; i >= 0; i--) {
2663	if (!cpumask_test_cpu(cpu: i, cpumask: cpu_map))
2664	continue;
2665
2666	for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2667	claim_allocations(cpu: i, sd);
2668	init_sched_groups_capacity(cpu: i, sd);
2669	}
2670	}
2671
2672	/* Attach the domains */
2673	rcu_read_lock();
2674	for_each_cpu(i, cpu_map) {
2675	rq = cpu_rq(i);
2676	sd = *per_cpu_ptr(d.sd, i);
2677
2678	cpu_attach_domain(sd, rd: d.rd, cpu: i);
2679
2680	if (lowest_flag_domain(cpu: i, flag: SD_CLUSTER))
2681	has_cluster = true;
2682	}
2683	rcu_read_unlock();
2684
2685	if (has_asym)
2686	static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2687
2688	if (has_cluster)
2689	static_branch_inc_cpuslocked(&sched_cluster_active);
2690
2691	if (rq && sched_debug_verbose)
2692	pr_info("root domain span: %*pbl\n", cpumask_pr_args(cpu_map));
2693
2694	ret = 0;
2695	error:
2696	__free_domain_allocs(d: &d, what: alloc_state, cpu_map);
2697
2698	return ret;
2699	}
2700
2701	/* Current sched domains: */
2702	static cpumask_var_t *doms_cur;
2703
2704	/* Number of sched domains in 'doms_cur': */
2705	static int ndoms_cur;
2706
2707	/* Attributes of custom domains in 'doms_cur' */
2708	static struct sched_domain_attr *dattr_cur;
2709
2710	/*
2711	* Special case: If a kmalloc() of a doms_cur partition (array of
2712	* cpumask) fails, then fallback to a single sched domain,
2713	* as determined by the single cpumask fallback_doms.
2714	*/
2715	static cpumask_var_t fallback_doms;
2716
2717	/*
2718	* arch_update_cpu_topology lets virtualized architectures update the
2719	* CPU core maps. It is supposed to return 1 if the topology changed
2720	* or 0 if it stayed the same.
2721	*/
2722	int __weak arch_update_cpu_topology(void)
2723	{
2724	return 0;
2725	}
2726
2727	cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2728	{
2729	int i;
2730	cpumask_var_t *doms;
2731
2732	doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2733	if (!doms)
2734	return NULL;
2735	for (i = 0; i < ndoms; i++) {
2736	if (!alloc_cpumask_var(mask: &doms[i], GFP_KERNEL)) {
2737	free_sched_domains(doms, ndoms: i);
2738	return NULL;
2739	}
2740	}
2741	return doms;
2742	}
2743
2744	void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2745	{
2746	unsigned int i;
2747	for (i = 0; i < ndoms; i++)
2748	free_cpumask_var(mask: doms[i]);
2749	kfree(objp: doms);
2750	}
2751
2752	/*
2753	* Set up scheduler domains and groups. For now this just excludes isolated
2754	* CPUs, but could be used to exclude other special cases in the future.
2755	*/
2756	int __init sched_init_domains(const struct cpumask *cpu_map)
2757	{
2758	int err;
2759
2760	zalloc_cpumask_var(mask: &sched_domains_tmpmask, GFP_KERNEL);
2761	zalloc_cpumask_var(mask: &sched_domains_tmpmask2, GFP_KERNEL);
2762	zalloc_cpumask_var(mask: &fallback_doms, GFP_KERNEL);
2763
2764	arch_update_cpu_topology();
2765	asym_cpu_capacity_scan();
2766	ndoms_cur = 1;
2767	doms_cur = alloc_sched_domains(ndoms: ndoms_cur);
2768	if (!doms_cur)
2769	doms_cur = &fallback_doms;
2770	cpumask_and(dstp: doms_cur[0], src1p: cpu_map, src2p: housekeeping_cpumask(type: HK_TYPE_DOMAIN));
2771	err = build_sched_domains(cpu_map: doms_cur[0], NULL);
2772
2773	return err;
2774	}
2775
2776	/*
2777	* Detach sched domains from a group of CPUs specified in cpu_map
2778	* These CPUs will now be attached to the NULL domain
2779	*/
2780	static void detach_destroy_domains(const struct cpumask *cpu_map)
2781	{
2782	unsigned int cpu = cpumask_any(cpu_map);
2783	int i;
2784
2785	if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2786	static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2787
2788	if (static_branch_unlikely(&sched_cluster_active))
2789	static_branch_dec_cpuslocked(&sched_cluster_active);
2790
2791	rcu_read_lock();
2792	for_each_cpu(i, cpu_map)
2793	cpu_attach_domain(NULL, rd: &def_root_domain, cpu: i);
2794	rcu_read_unlock();
2795	}
2796
2797	/* handle null as "default" */
2798	static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2799	struct sched_domain_attr *new, int idx_new)
2800	{
2801	struct sched_domain_attr tmp;
2802
2803	/* Fast path: */
2804	if (!new && !cur)
2805	return 1;
2806
2807	tmp = SD_ATTR_INIT;
2808
2809	return !memcmp(p: cur ? (cur + idx_cur) : &tmp,
2810	q: new ? (new + idx_new) : &tmp,
2811	size: sizeof(struct sched_domain_attr));
2812	}
2813
2814	/*
2815	* Partition sched domains as specified by the 'ndoms_new'
2816	* cpumasks in the array doms_new[] of cpumasks. This compares
2817	* doms_new[] to the current sched domain partitioning, doms_cur[].
2818	* It destroys each deleted domain and builds each new domain.
2819	*
2820	* 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2821	* The masks don't intersect (don't overlap.) We should setup one
2822	* sched domain for each mask. CPUs not in any of the cpumasks will
2823	* not be load balanced. If the same cpumask appears both in the
2824	* current 'doms_cur' domains and in the new 'doms_new', we can leave
2825	* it as it is.
2826	*
2827	* The passed in 'doms_new' should be allocated using
2828	* alloc_sched_domains. This routine takes ownership of it and will
2829	* free_sched_domains it when done with it. If the caller failed the
2830	* alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2831	* and partition_sched_domains() will fallback to the single partition
2832	* 'fallback_doms', it also forces the domains to be rebuilt.
2833	*
2834	* If doms_new == NULL it will be replaced with cpu_online_mask.
2835	* ndoms_new == 0 is a special case for destroying existing domains,
2836	* and it will not create the default domain.
2837	*
2838	* Call with hotplug lock and sched_domains_mutex held
2839	*/
2840	static void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2841	struct sched_domain_attr *dattr_new)
2842	{
2843	bool __maybe_unused has_eas = false;
2844	int i, j, n;
2845	int new_topology;
2846
2847	lockdep_assert_held(&sched_domains_mutex);
2848
2849	/* Let the architecture update CPU core mappings: */
2850	new_topology = arch_update_cpu_topology();
2851	/* Trigger rebuilding CPU capacity asymmetry data */
2852	if (new_topology)
2853	asym_cpu_capacity_scan();
2854
2855	if (!doms_new) {
2856	WARN_ON_ONCE(dattr_new);
2857	n = 0;
2858	doms_new = alloc_sched_domains(ndoms: 1);
2859	if (doms_new) {
2860	n = 1;
2861	cpumask_and(dstp: doms_new[0], cpu_active_mask,
2862	src2p: housekeeping_cpumask(type: HK_TYPE_DOMAIN));
2863	}
2864	} else {
2865	n = ndoms_new;
2866	}
2867
2868	/* Destroy deleted domains: */
2869	for (i = 0; i < ndoms_cur; i++) {
2870	for (j = 0; j < n && !new_topology; j++) {
2871	if (cpumask_equal(src1p: doms_cur[i], src2p: doms_new[j]) &&
2872	dattrs_equal(cur: dattr_cur, idx_cur: i, new: dattr_new, idx_new: j))
2873	goto match1;
2874	}
2875	/* No match - a current sched domain not in new doms_new[] */
2876	detach_destroy_domains(cpu_map: doms_cur[i]);
2877	match1:
2878	;
2879	}
2880
2881	n = ndoms_cur;
2882	if (!doms_new) {
2883	n = 0;
2884	doms_new = &fallback_doms;
2885	cpumask_and(dstp: doms_new[0], cpu_active_mask,
2886	src2p: housekeeping_cpumask(type: HK_TYPE_DOMAIN));
2887	}
2888
2889	/* Build new domains: */
2890	for (i = 0; i < ndoms_new; i++) {
2891	for (j = 0; j < n && !new_topology; j++) {
2892	if (cpumask_equal(src1p: doms_new[i], src2p: doms_cur[j]) &&
2893	dattrs_equal(cur: dattr_new, idx_cur: i, new: dattr_cur, idx_new: j))
2894	goto match2;
2895	}
2896	/* No match - add a new doms_new */
2897	build_sched_domains(cpu_map: doms_new[i], attr: dattr_new ? dattr_new + i : NULL);
2898	match2:
2899	;
2900	}
2901
2902	#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2903	/* Build perf domains: */
2904	for (i = 0; i < ndoms_new; i++) {
2905	for (j = 0; j < n && !sched_energy_update; j++) {
2906	if (cpumask_equal(src1p: doms_new[i], src2p: doms_cur[j]) &&
2907	cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2908	has_eas = true;
2909	goto match3;
2910	}
2911	}
2912	/* No match - add perf domains for a new rd */
2913	has_eas |= build_perf_domains(cpu_map: doms_new[i]);
2914	match3:
2915	;
2916	}
2917	sched_energy_set(has_eas);
2918	#endif
2919
2920	/* Remember the new sched domains: */
2921	if (doms_cur != &fallback_doms)
2922	free_sched_domains(doms: doms_cur, ndoms: ndoms_cur);
2923
2924	kfree(objp: dattr_cur);
2925	doms_cur = doms_new;
2926	dattr_cur = dattr_new;
2927	ndoms_cur = ndoms_new;
2928
2929	update_sched_domain_debugfs();
2930	dl_rebuild_rd_accounting();
2931	}
2932
2933	/*
2934	* Call with hotplug lock held
2935	*/
2936	void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2937	struct sched_domain_attr *dattr_new)
2938	{
2939	sched_domains_mutex_lock();
2940	partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2941	sched_domains_mutex_unlock();
2942	}
2943