xref: /openbmc/linux/kernel/sched/topology.c (revision 97f38170)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Scheduler topology setup/handling methods
4  */
5 
6 #include <linux/bsearch.h>
7 
8 DEFINE_MUTEX(sched_domains_mutex);
9 
10 /* Protected by sched_domains_mutex: */
11 static cpumask_var_t sched_domains_tmpmask;
12 static cpumask_var_t sched_domains_tmpmask2;
13 
14 #ifdef CONFIG_SCHED_DEBUG
15 
sched_debug_setup(char * str)16 static int __init sched_debug_setup(char *str)
17 {
18 	sched_debug_verbose = true;
19 
20 	return 0;
21 }
22 early_param("sched_verbose", sched_debug_setup);
23 
sched_debug(void)24 static inline bool sched_debug(void)
25 {
26 	return sched_debug_verbose;
27 }
28 
29 #define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
30 const struct sd_flag_debug sd_flag_debug[] = {
31 #include <linux/sched/sd_flags.h>
32 };
33 #undef SD_FLAG
34 
sched_domain_debug_one(struct sched_domain * sd,int cpu,int level,struct cpumask * groupmask)35 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
36 				  struct cpumask *groupmask)
37 {
38 	struct sched_group *group = sd->groups;
39 	unsigned long flags = sd->flags;
40 	unsigned int idx;
41 
42 	cpumask_clear(groupmask);
43 
44 	printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
45 	printk(KERN_CONT "span=%*pbl level=%s\n",
46 	       cpumask_pr_args(sched_domain_span(sd)), sd->name);
47 
48 	if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
49 		printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
50 	}
51 	if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
52 		printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
53 	}
54 
55 	for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
56 		unsigned int flag = BIT(idx);
57 		unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
58 
59 		if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
60 		    !(sd->child->flags & flag))
61 			printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
62 			       sd_flag_debug[idx].name);
63 
64 		if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
65 		    !(sd->parent->flags & flag))
66 			printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
67 			       sd_flag_debug[idx].name);
68 	}
69 
70 	printk(KERN_DEBUG "%*s groups:", level + 1, "");
71 	do {
72 		if (!group) {
73 			printk("\n");
74 			printk(KERN_ERR "ERROR: group is NULL\n");
75 			break;
76 		}
77 
78 		if (cpumask_empty(sched_group_span(group))) {
79 			printk(KERN_CONT "\n");
80 			printk(KERN_ERR "ERROR: empty group\n");
81 			break;
82 		}
83 
84 		if (!(sd->flags & SD_OVERLAP) &&
85 		    cpumask_intersects(groupmask, sched_group_span(group))) {
86 			printk(KERN_CONT "\n");
87 			printk(KERN_ERR "ERROR: repeated CPUs\n");
88 			break;
89 		}
90 
91 		cpumask_or(groupmask, groupmask, sched_group_span(group));
92 
93 		printk(KERN_CONT " %d:{ span=%*pbl",
94 				group->sgc->id,
95 				cpumask_pr_args(sched_group_span(group)));
96 
97 		if ((sd->flags & SD_OVERLAP) &&
98 		    !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
99 			printk(KERN_CONT " mask=%*pbl",
100 				cpumask_pr_args(group_balance_mask(group)));
101 		}
102 
103 		if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
104 			printk(KERN_CONT " cap=%lu", group->sgc->capacity);
105 
106 		if (group == sd->groups && sd->child &&
107 		    !cpumask_equal(sched_domain_span(sd->child),
108 				   sched_group_span(group))) {
109 			printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
110 		}
111 
112 		printk(KERN_CONT " }");
113 
114 		group = group->next;
115 
116 		if (group != sd->groups)
117 			printk(KERN_CONT ",");
118 
119 	} while (group != sd->groups);
120 	printk(KERN_CONT "\n");
121 
122 	if (!cpumask_equal(sched_domain_span(sd), groupmask))
123 		printk(KERN_ERR "ERROR: groups don't span domain->span\n");
124 
125 	if (sd->parent &&
126 	    !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
127 		printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
128 	return 0;
129 }
130 
sched_domain_debug(struct sched_domain * sd,int cpu)131 static void sched_domain_debug(struct sched_domain *sd, int cpu)
132 {
133 	int level = 0;
134 
135 	if (!sched_debug_verbose)
136 		return;
137 
138 	if (!sd) {
139 		printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
140 		return;
141 	}
142 
143 	printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
144 
145 	for (;;) {
146 		if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
147 			break;
148 		level++;
149 		sd = sd->parent;
150 		if (!sd)
151 			break;
152 	}
153 }
154 #else /* !CONFIG_SCHED_DEBUG */
155 
156 # define sched_debug_verbose 0
157 # define sched_domain_debug(sd, cpu) do { } while (0)
sched_debug(void)158 static inline bool sched_debug(void)
159 {
160 	return false;
161 }
162 #endif /* CONFIG_SCHED_DEBUG */
163 
164 /* Generate a mask of SD flags with the SDF_NEEDS_GROUPS metaflag */
165 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
166 static const unsigned int SD_DEGENERATE_GROUPS_MASK =
167 #include <linux/sched/sd_flags.h>
168 0;
169 #undef SD_FLAG
170 
sd_degenerate(struct sched_domain * sd)171 static int sd_degenerate(struct sched_domain *sd)
172 {
173 	if (cpumask_weight(sched_domain_span(sd)) == 1)
174 		return 1;
175 
176 	/* Following flags need at least 2 groups */
177 	if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
178 	    (sd->groups != sd->groups->next))
179 		return 0;
180 
181 	/* Following flags don't use groups */
182 	if (sd->flags & (SD_WAKE_AFFINE))
183 		return 0;
184 
185 	return 1;
186 }
187 
188 static int
sd_parent_degenerate(struct sched_domain * sd,struct sched_domain * parent)189 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
190 {
191 	unsigned long cflags = sd->flags, pflags = parent->flags;
192 
193 	if (sd_degenerate(parent))
194 		return 1;
195 
196 	if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
197 		return 0;
198 
199 	/* Flags needing groups don't count if only 1 group in parent */
200 	if (parent->groups == parent->groups->next)
201 		pflags &= ~SD_DEGENERATE_GROUPS_MASK;
202 
203 	if (~cflags & pflags)
204 		return 0;
205 
206 	return 1;
207 }
208 
209 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
210 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
211 static unsigned int sysctl_sched_energy_aware = 1;
212 static DEFINE_MUTEX(sched_energy_mutex);
213 static bool sched_energy_update;
214 
rebuild_sched_domains_energy(void)215 void rebuild_sched_domains_energy(void)
216 {
217 	mutex_lock(&sched_energy_mutex);
218 	sched_energy_update = true;
219 	rebuild_sched_domains();
220 	sched_energy_update = false;
221 	mutex_unlock(&sched_energy_mutex);
222 }
223 
224 #ifdef CONFIG_PROC_SYSCTL
sched_energy_aware_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)225 static int sched_energy_aware_handler(struct ctl_table *table, int write,
226 		void *buffer, size_t *lenp, loff_t *ppos)
227 {
228 	int ret, state;
229 
230 	if (write && !capable(CAP_SYS_ADMIN))
231 		return -EPERM;
232 
233 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
234 	if (!ret && write) {
235 		state = static_branch_unlikely(&sched_energy_present);
236 		if (state != sysctl_sched_energy_aware)
237 			rebuild_sched_domains_energy();
238 	}
239 
240 	return ret;
241 }
242 
243 static struct ctl_table sched_energy_aware_sysctls[] = {
244 	{
245 		.procname       = "sched_energy_aware",
246 		.data           = &sysctl_sched_energy_aware,
247 		.maxlen         = sizeof(unsigned int),
248 		.mode           = 0644,
249 		.proc_handler   = sched_energy_aware_handler,
250 		.extra1         = SYSCTL_ZERO,
251 		.extra2         = SYSCTL_ONE,
252 	},
253 	{}
254 };
255 
sched_energy_aware_sysctl_init(void)256 static int __init sched_energy_aware_sysctl_init(void)
257 {
258 	register_sysctl_init("kernel", sched_energy_aware_sysctls);
259 	return 0;
260 }
261 
262 late_initcall(sched_energy_aware_sysctl_init);
263 #endif
264 
free_pd(struct perf_domain * pd)265 static void free_pd(struct perf_domain *pd)
266 {
267 	struct perf_domain *tmp;
268 
269 	while (pd) {
270 		tmp = pd->next;
271 		kfree(pd);
272 		pd = tmp;
273 	}
274 }
275 
find_pd(struct perf_domain * pd,int cpu)276 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
277 {
278 	while (pd) {
279 		if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
280 			return pd;
281 		pd = pd->next;
282 	}
283 
284 	return NULL;
285 }
286 
pd_init(int cpu)287 static struct perf_domain *pd_init(int cpu)
288 {
289 	struct em_perf_domain *obj = em_cpu_get(cpu);
290 	struct perf_domain *pd;
291 
292 	if (!obj) {
293 		if (sched_debug())
294 			pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
295 		return NULL;
296 	}
297 
298 	pd = kzalloc(sizeof(*pd), GFP_KERNEL);
299 	if (!pd)
300 		return NULL;
301 	pd->em_pd = obj;
302 
303 	return pd;
304 }
305 
perf_domain_debug(const struct cpumask * cpu_map,struct perf_domain * pd)306 static void perf_domain_debug(const struct cpumask *cpu_map,
307 						struct perf_domain *pd)
308 {
309 	if (!sched_debug() || !pd)
310 		return;
311 
312 	printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
313 
314 	while (pd) {
315 		printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
316 				cpumask_first(perf_domain_span(pd)),
317 				cpumask_pr_args(perf_domain_span(pd)),
318 				em_pd_nr_perf_states(pd->em_pd));
319 		pd = pd->next;
320 	}
321 
322 	printk(KERN_CONT "\n");
323 }
324 
destroy_perf_domain_rcu(struct rcu_head * rp)325 static void destroy_perf_domain_rcu(struct rcu_head *rp)
326 {
327 	struct perf_domain *pd;
328 
329 	pd = container_of(rp, struct perf_domain, rcu);
330 	free_pd(pd);
331 }
332 
sched_energy_set(bool has_eas)333 static void sched_energy_set(bool has_eas)
334 {
335 	if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
336 		if (sched_debug())
337 			pr_info("%s: stopping EAS\n", __func__);
338 		static_branch_disable_cpuslocked(&sched_energy_present);
339 	} else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
340 		if (sched_debug())
341 			pr_info("%s: starting EAS\n", __func__);
342 		static_branch_enable_cpuslocked(&sched_energy_present);
343 	}
344 }
345 
346 /*
347  * EAS can be used on a root domain if it meets all the following conditions:
348  *    1. an Energy Model (EM) is available;
349  *    2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
350  *    3. no SMT is detected.
351  *    4. the EM complexity is low enough to keep scheduling overheads low;
352  *    5. schedutil is driving the frequency of all CPUs of the rd;
353  *    6. frequency invariance support is present;
354  *
355  * The complexity of the Energy Model is defined as:
356  *
357  *              C = nr_pd * (nr_cpus + nr_ps)
358  *
359  * with parameters defined as:
360  *  - nr_pd:    the number of performance domains
361  *  - nr_cpus:  the number of CPUs
362  *  - nr_ps:    the sum of the number of performance states of all performance
363  *              domains (for example, on a system with 2 performance domains,
364  *              with 10 performance states each, nr_ps = 2 * 10 = 20).
365  *
366  * It is generally not a good idea to use such a model in the wake-up path on
367  * very complex platforms because of the associated scheduling overheads. The
368  * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
369  * with per-CPU DVFS and less than 8 performance states each, for example.
370  */
371 #define EM_MAX_COMPLEXITY 2048
372 
373 extern struct cpufreq_governor schedutil_gov;
build_perf_domains(const struct cpumask * cpu_map)374 static bool build_perf_domains(const struct cpumask *cpu_map)
375 {
376 	int i, nr_pd = 0, nr_ps = 0, nr_cpus = cpumask_weight(cpu_map);
377 	struct perf_domain *pd = NULL, *tmp;
378 	int cpu = cpumask_first(cpu_map);
379 	struct root_domain *rd = cpu_rq(cpu)->rd;
380 	struct cpufreq_policy *policy;
381 	struct cpufreq_governor *gov;
382 
383 	if (!sysctl_sched_energy_aware)
384 		goto free;
385 
386 	/* EAS is enabled for asymmetric CPU capacity topologies. */
387 	if (!per_cpu(sd_asym_cpucapacity, cpu)) {
388 		if (sched_debug()) {
389 			pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
390 					cpumask_pr_args(cpu_map));
391 		}
392 		goto free;
393 	}
394 
395 	/* EAS definitely does *not* handle SMT */
396 	if (sched_smt_active()) {
397 		pr_warn("rd %*pbl: Disabling EAS, SMT is not supported\n",
398 			cpumask_pr_args(cpu_map));
399 		goto free;
400 	}
401 
402 	if (!arch_scale_freq_invariant()) {
403 		if (sched_debug()) {
404 			pr_warn("rd %*pbl: Disabling EAS: frequency-invariant load tracking not yet supported",
405 				cpumask_pr_args(cpu_map));
406 		}
407 		goto free;
408 	}
409 
410 	for_each_cpu(i, cpu_map) {
411 		/* Skip already covered CPUs. */
412 		if (find_pd(pd, i))
413 			continue;
414 
415 		/* Do not attempt EAS if schedutil is not being used. */
416 		policy = cpufreq_cpu_get(i);
417 		if (!policy)
418 			goto free;
419 		gov = policy->governor;
420 		cpufreq_cpu_put(policy);
421 		if (gov != &schedutil_gov) {
422 			if (rd->pd)
423 				pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
424 						cpumask_pr_args(cpu_map));
425 			goto free;
426 		}
427 
428 		/* Create the new pd and add it to the local list. */
429 		tmp = pd_init(i);
430 		if (!tmp)
431 			goto free;
432 		tmp->next = pd;
433 		pd = tmp;
434 
435 		/*
436 		 * Count performance domains and performance states for the
437 		 * complexity check.
438 		 */
439 		nr_pd++;
440 		nr_ps += em_pd_nr_perf_states(pd->em_pd);
441 	}
442 
443 	/* Bail out if the Energy Model complexity is too high. */
444 	if (nr_pd * (nr_ps + nr_cpus) > EM_MAX_COMPLEXITY) {
445 		WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
446 						cpumask_pr_args(cpu_map));
447 		goto free;
448 	}
449 
450 	perf_domain_debug(cpu_map, pd);
451 
452 	/* Attach the new list of performance domains to the root domain. */
453 	tmp = rd->pd;
454 	rcu_assign_pointer(rd->pd, pd);
455 	if (tmp)
456 		call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
457 
458 	return !!pd;
459 
460 free:
461 	free_pd(pd);
462 	tmp = rd->pd;
463 	rcu_assign_pointer(rd->pd, NULL);
464 	if (tmp)
465 		call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
466 
467 	return false;
468 }
469 #else
free_pd(struct perf_domain * pd)470 static void free_pd(struct perf_domain *pd) { }
471 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
472 
free_rootdomain(struct rcu_head * rcu)473 static void free_rootdomain(struct rcu_head *rcu)
474 {
475 	struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
476 
477 	cpupri_cleanup(&rd->cpupri);
478 	cpudl_cleanup(&rd->cpudl);
479 	free_cpumask_var(rd->dlo_mask);
480 	free_cpumask_var(rd->rto_mask);
481 	free_cpumask_var(rd->online);
482 	free_cpumask_var(rd->span);
483 	free_pd(rd->pd);
484 	kfree(rd);
485 }
486 
rq_attach_root(struct rq * rq,struct root_domain * rd)487 void rq_attach_root(struct rq *rq, struct root_domain *rd)
488 {
489 	struct root_domain *old_rd = NULL;
490 	struct rq_flags rf;
491 
492 	rq_lock_irqsave(rq, &rf);
493 
494 	if (rq->rd) {
495 		old_rd = rq->rd;
496 
497 		if (cpumask_test_cpu(rq->cpu, old_rd->online))
498 			set_rq_offline(rq);
499 
500 		cpumask_clear_cpu(rq->cpu, old_rd->span);
501 
502 		/*
503 		 * If we dont want to free the old_rd yet then
504 		 * set old_rd to NULL to skip the freeing later
505 		 * in this function:
506 		 */
507 		if (!atomic_dec_and_test(&old_rd->refcount))
508 			old_rd = NULL;
509 	}
510 
511 	atomic_inc(&rd->refcount);
512 	rq->rd = rd;
513 
514 	cpumask_set_cpu(rq->cpu, rd->span);
515 	if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
516 		set_rq_online(rq);
517 
518 	rq_unlock_irqrestore(rq, &rf);
519 
520 	if (old_rd)
521 		call_rcu(&old_rd->rcu, free_rootdomain);
522 }
523 
sched_get_rd(struct root_domain * rd)524 void sched_get_rd(struct root_domain *rd)
525 {
526 	atomic_inc(&rd->refcount);
527 }
528 
sched_put_rd(struct root_domain * rd)529 void sched_put_rd(struct root_domain *rd)
530 {
531 	if (!atomic_dec_and_test(&rd->refcount))
532 		return;
533 
534 	call_rcu(&rd->rcu, free_rootdomain);
535 }
536 
init_rootdomain(struct root_domain * rd)537 static int init_rootdomain(struct root_domain *rd)
538 {
539 	if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
540 		goto out;
541 	if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
542 		goto free_span;
543 	if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
544 		goto free_online;
545 	if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
546 		goto free_dlo_mask;
547 
548 #ifdef HAVE_RT_PUSH_IPI
549 	rd->rto_cpu = -1;
550 	raw_spin_lock_init(&rd->rto_lock);
551 	rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
552 #endif
553 
554 	rd->visit_gen = 0;
555 	init_dl_bw(&rd->dl_bw);
556 	if (cpudl_init(&rd->cpudl) != 0)
557 		goto free_rto_mask;
558 
559 	if (cpupri_init(&rd->cpupri) != 0)
560 		goto free_cpudl;
561 	return 0;
562 
563 free_cpudl:
564 	cpudl_cleanup(&rd->cpudl);
565 free_rto_mask:
566 	free_cpumask_var(rd->rto_mask);
567 free_dlo_mask:
568 	free_cpumask_var(rd->dlo_mask);
569 free_online:
570 	free_cpumask_var(rd->online);
571 free_span:
572 	free_cpumask_var(rd->span);
573 out:
574 	return -ENOMEM;
575 }
576 
577 /*
578  * By default the system creates a single root-domain with all CPUs as
579  * members (mimicking the global state we have today).
580  */
581 struct root_domain def_root_domain;
582 
init_defrootdomain(void)583 void __init init_defrootdomain(void)
584 {
585 	init_rootdomain(&def_root_domain);
586 
587 	atomic_set(&def_root_domain.refcount, 1);
588 }
589 
alloc_rootdomain(void)590 static struct root_domain *alloc_rootdomain(void)
591 {
592 	struct root_domain *rd;
593 
594 	rd = kzalloc(sizeof(*rd), GFP_KERNEL);
595 	if (!rd)
596 		return NULL;
597 
598 	if (init_rootdomain(rd) != 0) {
599 		kfree(rd);
600 		return NULL;
601 	}
602 
603 	return rd;
604 }
605 
free_sched_groups(struct sched_group * sg,int free_sgc)606 static void free_sched_groups(struct sched_group *sg, int free_sgc)
607 {
608 	struct sched_group *tmp, *first;
609 
610 	if (!sg)
611 		return;
612 
613 	first = sg;
614 	do {
615 		tmp = sg->next;
616 
617 		if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
618 			kfree(sg->sgc);
619 
620 		if (atomic_dec_and_test(&sg->ref))
621 			kfree(sg);
622 		sg = tmp;
623 	} while (sg != first);
624 }
625 
destroy_sched_domain(struct sched_domain * sd)626 static void destroy_sched_domain(struct sched_domain *sd)
627 {
628 	/*
629 	 * A normal sched domain may have multiple group references, an
630 	 * overlapping domain, having private groups, only one.  Iterate,
631 	 * dropping group/capacity references, freeing where none remain.
632 	 */
633 	free_sched_groups(sd->groups, 1);
634 
635 	if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
636 		kfree(sd->shared);
637 	kfree(sd);
638 }
639 
destroy_sched_domains_rcu(struct rcu_head * rcu)640 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
641 {
642 	struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
643 
644 	while (sd) {
645 		struct sched_domain *parent = sd->parent;
646 		destroy_sched_domain(sd);
647 		sd = parent;
648 	}
649 }
650 
destroy_sched_domains(struct sched_domain * sd)651 static void destroy_sched_domains(struct sched_domain *sd)
652 {
653 	if (sd)
654 		call_rcu(&sd->rcu, destroy_sched_domains_rcu);
655 }
656 
657 /*
658  * Keep a special pointer to the highest sched_domain that has
659  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
660  * allows us to avoid some pointer chasing select_idle_sibling().
661  *
662  * Also keep a unique ID per domain (we use the first CPU number in
663  * the cpumask of the domain), this allows us to quickly tell if
664  * two CPUs are in the same cache domain, see cpus_share_cache().
665  */
666 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
667 DEFINE_PER_CPU(int, sd_llc_size);
668 DEFINE_PER_CPU(int, sd_llc_id);
669 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
670 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
671 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
672 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
673 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
674 
update_top_cache_domain(int cpu)675 static void update_top_cache_domain(int cpu)
676 {
677 	struct sched_domain_shared *sds = NULL;
678 	struct sched_domain *sd;
679 	int id = cpu;
680 	int size = 1;
681 
682 	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
683 	if (sd) {
684 		id = cpumask_first(sched_domain_span(sd));
685 		size = cpumask_weight(sched_domain_span(sd));
686 		sds = sd->shared;
687 	}
688 
689 	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
690 	per_cpu(sd_llc_size, cpu) = size;
691 	per_cpu(sd_llc_id, cpu) = id;
692 	rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
693 
694 	sd = lowest_flag_domain(cpu, SD_NUMA);
695 	rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
696 
697 	sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
698 	rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
699 
700 	sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
701 	rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
702 }
703 
704 /*
705  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
706  * hold the hotplug lock.
707  */
708 static void
cpu_attach_domain(struct sched_domain * sd,struct root_domain * rd,int cpu)709 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
710 {
711 	struct rq *rq = cpu_rq(cpu);
712 	struct sched_domain *tmp;
713 
714 	/* Remove the sched domains which do not contribute to scheduling. */
715 	for (tmp = sd; tmp; ) {
716 		struct sched_domain *parent = tmp->parent;
717 		if (!parent)
718 			break;
719 
720 		if (sd_parent_degenerate(tmp, parent)) {
721 			tmp->parent = parent->parent;
722 
723 			if (parent->parent) {
724 				parent->parent->child = tmp;
725 				parent->parent->groups->flags = tmp->flags;
726 			}
727 
728 			/*
729 			 * Transfer SD_PREFER_SIBLING down in case of a
730 			 * degenerate parent; the spans match for this
731 			 * so the property transfers.
732 			 */
733 			if (parent->flags & SD_PREFER_SIBLING)
734 				tmp->flags |= SD_PREFER_SIBLING;
735 			destroy_sched_domain(parent);
736 		} else
737 			tmp = tmp->parent;
738 	}
739 
740 	if (sd && sd_degenerate(sd)) {
741 		tmp = sd;
742 		sd = sd->parent;
743 		destroy_sched_domain(tmp);
744 		if (sd) {
745 			struct sched_group *sg = sd->groups;
746 
747 			/*
748 			 * sched groups hold the flags of the child sched
749 			 * domain for convenience. Clear such flags since
750 			 * the child is being destroyed.
751 			 */
752 			do {
753 				sg->flags = 0;
754 			} while (sg != sd->groups);
755 
756 			sd->child = NULL;
757 		}
758 	}
759 
760 	sched_domain_debug(sd, cpu);
761 
762 	rq_attach_root(rq, rd);
763 	tmp = rq->sd;
764 	rcu_assign_pointer(rq->sd, sd);
765 	dirty_sched_domain_sysctl(cpu);
766 	destroy_sched_domains(tmp);
767 
768 	update_top_cache_domain(cpu);
769 }
770 
771 struct s_data {
772 	struct sched_domain * __percpu *sd;
773 	struct root_domain	*rd;
774 };
775 
776 enum s_alloc {
777 	sa_rootdomain,
778 	sa_sd,
779 	sa_sd_storage,
780 	sa_none,
781 };
782 
783 /*
784  * Return the canonical balance CPU for this group, this is the first CPU
785  * of this group that's also in the balance mask.
786  *
787  * The balance mask are all those CPUs that could actually end up at this
788  * group. See build_balance_mask().
789  *
790  * Also see should_we_balance().
791  */
group_balance_cpu(struct sched_group * sg)792 int group_balance_cpu(struct sched_group *sg)
793 {
794 	return cpumask_first(group_balance_mask(sg));
795 }
796 
797 
798 /*
799  * NUMA topology (first read the regular topology blurb below)
800  *
801  * Given a node-distance table, for example:
802  *
803  *   node   0   1   2   3
804  *     0:  10  20  30  20
805  *     1:  20  10  20  30
806  *     2:  30  20  10  20
807  *     3:  20  30  20  10
808  *
809  * which represents a 4 node ring topology like:
810  *
811  *   0 ----- 1
812  *   |       |
813  *   |       |
814  *   |       |
815  *   3 ----- 2
816  *
817  * We want to construct domains and groups to represent this. The way we go
818  * about doing this is to build the domains on 'hops'. For each NUMA level we
819  * construct the mask of all nodes reachable in @level hops.
820  *
821  * For the above NUMA topology that gives 3 levels:
822  *
823  * NUMA-2	0-3		0-3		0-3		0-3
824  *  groups:	{0-1,3},{1-3}	{0-2},{0,2-3}	{1-3},{0-1,3}	{0,2-3},{0-2}
825  *
826  * NUMA-1	0-1,3		0-2		1-3		0,2-3
827  *  groups:	{0},{1},{3}	{0},{1},{2}	{1},{2},{3}	{0},{2},{3}
828  *
829  * NUMA-0	0		1		2		3
830  *
831  *
832  * As can be seen; things don't nicely line up as with the regular topology.
833  * When we iterate a domain in child domain chunks some nodes can be
834  * represented multiple times -- hence the "overlap" naming for this part of
835  * the topology.
836  *
837  * In order to minimize this overlap, we only build enough groups to cover the
838  * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
839  *
840  * Because:
841  *
842  *  - the first group of each domain is its child domain; this
843  *    gets us the first 0-1,3
844  *  - the only uncovered node is 2, who's child domain is 1-3.
845  *
846  * However, because of the overlap, computing a unique CPU for each group is
847  * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
848  * groups include the CPUs of Node-0, while those CPUs would not in fact ever
849  * end up at those groups (they would end up in group: 0-1,3).
850  *
851  * To correct this we have to introduce the group balance mask. This mask
852  * will contain those CPUs in the group that can reach this group given the
853  * (child) domain tree.
854  *
855  * With this we can once again compute balance_cpu and sched_group_capacity
856  * relations.
857  *
858  * XXX include words on how balance_cpu is unique and therefore can be
859  * used for sched_group_capacity links.
860  *
861  *
862  * Another 'interesting' topology is:
863  *
864  *   node   0   1   2   3
865  *     0:  10  20  20  30
866  *     1:  20  10  20  20
867  *     2:  20  20  10  20
868  *     3:  30  20  20  10
869  *
870  * Which looks a little like:
871  *
872  *   0 ----- 1
873  *   |     / |
874  *   |   /   |
875  *   | /     |
876  *   2 ----- 3
877  *
878  * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
879  * are not.
880  *
881  * This leads to a few particularly weird cases where the sched_domain's are
882  * not of the same number for each CPU. Consider:
883  *
884  * NUMA-2	0-3						0-3
885  *  groups:	{0-2},{1-3}					{1-3},{0-2}
886  *
887  * NUMA-1	0-2		0-3		0-3		1-3
888  *
889  * NUMA-0	0		1		2		3
890  *
891  */
892 
893 
894 /*
895  * Build the balance mask; it contains only those CPUs that can arrive at this
896  * group and should be considered to continue balancing.
897  *
898  * We do this during the group creation pass, therefore the group information
899  * isn't complete yet, however since each group represents a (child) domain we
900  * can fully construct this using the sched_domain bits (which are already
901  * complete).
902  */
903 static void
build_balance_mask(struct sched_domain * sd,struct sched_group * sg,struct cpumask * mask)904 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
905 {
906 	const struct cpumask *sg_span = sched_group_span(sg);
907 	struct sd_data *sdd = sd->private;
908 	struct sched_domain *sibling;
909 	int i;
910 
911 	cpumask_clear(mask);
912 
913 	for_each_cpu(i, sg_span) {
914 		sibling = *per_cpu_ptr(sdd->sd, i);
915 
916 		/*
917 		 * Can happen in the asymmetric case, where these siblings are
918 		 * unused. The mask will not be empty because those CPUs that
919 		 * do have the top domain _should_ span the domain.
920 		 */
921 		if (!sibling->child)
922 			continue;
923 
924 		/* If we would not end up here, we can't continue from here */
925 		if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
926 			continue;
927 
928 		cpumask_set_cpu(i, mask);
929 	}
930 
931 	/* We must not have empty masks here */
932 	WARN_ON_ONCE(cpumask_empty(mask));
933 }
934 
935 /*
936  * XXX: This creates per-node group entries; since the load-balancer will
937  * immediately access remote memory to construct this group's load-balance
938  * statistics having the groups node local is of dubious benefit.
939  */
940 static struct sched_group *
build_group_from_child_sched_domain(struct sched_domain * sd,int cpu)941 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
942 {
943 	struct sched_group *sg;
944 	struct cpumask *sg_span;
945 
946 	sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
947 			GFP_KERNEL, cpu_to_node(cpu));
948 
949 	if (!sg)
950 		return NULL;
951 
952 	sg_span = sched_group_span(sg);
953 	if (sd->child) {
954 		cpumask_copy(sg_span, sched_domain_span(sd->child));
955 		sg->flags = sd->child->flags;
956 	} else {
957 		cpumask_copy(sg_span, sched_domain_span(sd));
958 	}
959 
960 	atomic_inc(&sg->ref);
961 	return sg;
962 }
963 
init_overlap_sched_group(struct sched_domain * sd,struct sched_group * sg)964 static void init_overlap_sched_group(struct sched_domain *sd,
965 				     struct sched_group *sg)
966 {
967 	struct cpumask *mask = sched_domains_tmpmask2;
968 	struct sd_data *sdd = sd->private;
969 	struct cpumask *sg_span;
970 	int cpu;
971 
972 	build_balance_mask(sd, sg, mask);
973 	cpu = cpumask_first(mask);
974 
975 	sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
976 	if (atomic_inc_return(&sg->sgc->ref) == 1)
977 		cpumask_copy(group_balance_mask(sg), mask);
978 	else
979 		WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
980 
981 	/*
982 	 * Initialize sgc->capacity such that even if we mess up the
983 	 * domains and no possible iteration will get us here, we won't
984 	 * die on a /0 trap.
985 	 */
986 	sg_span = sched_group_span(sg);
987 	sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
988 	sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
989 	sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
990 }
991 
992 static struct sched_domain *
find_descended_sibling(struct sched_domain * sd,struct sched_domain * sibling)993 find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
994 {
995 	/*
996 	 * The proper descendant would be the one whose child won't span out
997 	 * of sd
998 	 */
999 	while (sibling->child &&
1000 	       !cpumask_subset(sched_domain_span(sibling->child),
1001 			       sched_domain_span(sd)))
1002 		sibling = sibling->child;
1003 
1004 	/*
1005 	 * As we are referencing sgc across different topology level, we need
1006 	 * to go down to skip those sched_domains which don't contribute to
1007 	 * scheduling because they will be degenerated in cpu_attach_domain
1008 	 */
1009 	while (sibling->child &&
1010 	       cpumask_equal(sched_domain_span(sibling->child),
1011 			     sched_domain_span(sibling)))
1012 		sibling = sibling->child;
1013 
1014 	return sibling;
1015 }
1016 
1017 static int
build_overlap_sched_groups(struct sched_domain * sd,int cpu)1018 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
1019 {
1020 	struct sched_group *first = NULL, *last = NULL, *sg;
1021 	const struct cpumask *span = sched_domain_span(sd);
1022 	struct cpumask *covered = sched_domains_tmpmask;
1023 	struct sd_data *sdd = sd->private;
1024 	struct sched_domain *sibling;
1025 	int i;
1026 
1027 	cpumask_clear(covered);
1028 
1029 	for_each_cpu_wrap(i, span, cpu) {
1030 		struct cpumask *sg_span;
1031 
1032 		if (cpumask_test_cpu(i, covered))
1033 			continue;
1034 
1035 		sibling = *per_cpu_ptr(sdd->sd, i);
1036 
1037 		/*
1038 		 * Asymmetric node setups can result in situations where the
1039 		 * domain tree is of unequal depth, make sure to skip domains
1040 		 * that already cover the entire range.
1041 		 *
1042 		 * In that case build_sched_domains() will have terminated the
1043 		 * iteration early and our sibling sd spans will be empty.
1044 		 * Domains should always include the CPU they're built on, so
1045 		 * check that.
1046 		 */
1047 		if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1048 			continue;
1049 
1050 		/*
1051 		 * Usually we build sched_group by sibling's child sched_domain
1052 		 * But for machines whose NUMA diameter are 3 or above, we move
1053 		 * to build sched_group by sibling's proper descendant's child
1054 		 * domain because sibling's child sched_domain will span out of
1055 		 * the sched_domain being built as below.
1056 		 *
1057 		 * Smallest diameter=3 topology is:
1058 		 *
1059 		 *   node   0   1   2   3
1060 		 *     0:  10  20  30  40
1061 		 *     1:  20  10  20  30
1062 		 *     2:  30  20  10  20
1063 		 *     3:  40  30  20  10
1064 		 *
1065 		 *   0 --- 1 --- 2 --- 3
1066 		 *
1067 		 * NUMA-3       0-3             N/A             N/A             0-3
1068 		 *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
1069 		 *
1070 		 * NUMA-2       0-2             0-3             0-3             1-3
1071 		 *  groups:     {0-1},{1-3}     {0-2},{2-3}     {1-3},{0-1}     {2-3},{0-2}
1072 		 *
1073 		 * NUMA-1       0-1             0-2             1-3             2-3
1074 		 *  groups:     {0},{1}         {1},{2},{0}     {2},{3},{1}     {3},{2}
1075 		 *
1076 		 * NUMA-0       0               1               2               3
1077 		 *
1078 		 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1079 		 * group span isn't a subset of the domain span.
1080 		 */
1081 		if (sibling->child &&
1082 		    !cpumask_subset(sched_domain_span(sibling->child), span))
1083 			sibling = find_descended_sibling(sd, sibling);
1084 
1085 		sg = build_group_from_child_sched_domain(sibling, cpu);
1086 		if (!sg)
1087 			goto fail;
1088 
1089 		sg_span = sched_group_span(sg);
1090 		cpumask_or(covered, covered, sg_span);
1091 
1092 		init_overlap_sched_group(sibling, sg);
1093 
1094 		if (!first)
1095 			first = sg;
1096 		if (last)
1097 			last->next = sg;
1098 		last = sg;
1099 		last->next = first;
1100 	}
1101 	sd->groups = first;
1102 
1103 	return 0;
1104 
1105 fail:
1106 	free_sched_groups(first, 0);
1107 
1108 	return -ENOMEM;
1109 }
1110 
1111 
1112 /*
1113  * Package topology (also see the load-balance blurb in fair.c)
1114  *
1115  * The scheduler builds a tree structure to represent a number of important
1116  * topology features. By default (default_topology[]) these include:
1117  *
1118  *  - Simultaneous multithreading (SMT)
1119  *  - Multi-Core Cache (MC)
1120  *  - Package (DIE)
1121  *
1122  * Where the last one more or less denotes everything up to a NUMA node.
1123  *
1124  * The tree consists of 3 primary data structures:
1125  *
1126  *	sched_domain -> sched_group -> sched_group_capacity
1127  *	    ^ ^             ^ ^
1128  *          `-'             `-'
1129  *
1130  * The sched_domains are per-CPU and have a two way link (parent & child) and
1131  * denote the ever growing mask of CPUs belonging to that level of topology.
1132  *
1133  * Each sched_domain has a circular (double) linked list of sched_group's, each
1134  * denoting the domains of the level below (or individual CPUs in case of the
1135  * first domain level). The sched_group linked by a sched_domain includes the
1136  * CPU of that sched_domain [*].
1137  *
1138  * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1139  *
1140  * CPU   0   1   2   3   4   5   6   7
1141  *
1142  * DIE  [                             ]
1143  * MC   [             ] [             ]
1144  * SMT  [     ] [     ] [     ] [     ]
1145  *
1146  *  - or -
1147  *
1148  * DIE  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1149  * MC	0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1150  * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1151  *
1152  * CPU   0   1   2   3   4   5   6   7
1153  *
1154  * One way to think about it is: sched_domain moves you up and down among these
1155  * topology levels, while sched_group moves you sideways through it, at child
1156  * domain granularity.
1157  *
1158  * sched_group_capacity ensures each unique sched_group has shared storage.
1159  *
1160  * There are two related construction problems, both require a CPU that
1161  * uniquely identify each group (for a given domain):
1162  *
1163  *  - The first is the balance_cpu (see should_we_balance() and the
1164  *    load-balance blub in fair.c); for each group we only want 1 CPU to
1165  *    continue balancing at a higher domain.
1166  *
1167  *  - The second is the sched_group_capacity; we want all identical groups
1168  *    to share a single sched_group_capacity.
1169  *
1170  * Since these topologies are exclusive by construction. That is, its
1171  * impossible for an SMT thread to belong to multiple cores, and cores to
1172  * be part of multiple caches. There is a very clear and unique location
1173  * for each CPU in the hierarchy.
1174  *
1175  * Therefore computing a unique CPU for each group is trivial (the iteration
1176  * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1177  * group), we can simply pick the first CPU in each group.
1178  *
1179  *
1180  * [*] in other words, the first group of each domain is its child domain.
1181  */
1182 
get_group(int cpu,struct sd_data * sdd)1183 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1184 {
1185 	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1186 	struct sched_domain *child = sd->child;
1187 	struct sched_group *sg;
1188 	bool already_visited;
1189 
1190 	if (child)
1191 		cpu = cpumask_first(sched_domain_span(child));
1192 
1193 	sg = *per_cpu_ptr(sdd->sg, cpu);
1194 	sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1195 
1196 	/* Increase refcounts for claim_allocations: */
1197 	already_visited = atomic_inc_return(&sg->ref) > 1;
1198 	/* sgc visits should follow a similar trend as sg */
1199 	WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1200 
1201 	/* If we have already visited that group, it's already initialized. */
1202 	if (already_visited)
1203 		return sg;
1204 
1205 	if (child) {
1206 		cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1207 		cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1208 		sg->flags = child->flags;
1209 	} else {
1210 		cpumask_set_cpu(cpu, sched_group_span(sg));
1211 		cpumask_set_cpu(cpu, group_balance_mask(sg));
1212 	}
1213 
1214 	sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1215 	sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1216 	sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1217 
1218 	return sg;
1219 }
1220 
1221 /*
1222  * build_sched_groups will build a circular linked list of the groups
1223  * covered by the given span, will set each group's ->cpumask correctly,
1224  * and will initialize their ->sgc.
1225  *
1226  * Assumes the sched_domain tree is fully constructed
1227  */
1228 static int
build_sched_groups(struct sched_domain * sd,int cpu)1229 build_sched_groups(struct sched_domain *sd, int cpu)
1230 {
1231 	struct sched_group *first = NULL, *last = NULL;
1232 	struct sd_data *sdd = sd->private;
1233 	const struct cpumask *span = sched_domain_span(sd);
1234 	struct cpumask *covered;
1235 	int i;
1236 
1237 	lockdep_assert_held(&sched_domains_mutex);
1238 	covered = sched_domains_tmpmask;
1239 
1240 	cpumask_clear(covered);
1241 
1242 	for_each_cpu_wrap(i, span, cpu) {
1243 		struct sched_group *sg;
1244 
1245 		if (cpumask_test_cpu(i, covered))
1246 			continue;
1247 
1248 		sg = get_group(i, sdd);
1249 
1250 		cpumask_or(covered, covered, sched_group_span(sg));
1251 
1252 		if (!first)
1253 			first = sg;
1254 		if (last)
1255 			last->next = sg;
1256 		last = sg;
1257 	}
1258 	last->next = first;
1259 	sd->groups = first;
1260 
1261 	return 0;
1262 }
1263 
1264 /*
1265  * Initialize sched groups cpu_capacity.
1266  *
1267  * cpu_capacity indicates the capacity of sched group, which is used while
1268  * distributing the load between different sched groups in a sched domain.
1269  * Typically cpu_capacity for all the groups in a sched domain will be same
1270  * unless there are asymmetries in the topology. If there are asymmetries,
1271  * group having more cpu_capacity will pickup more load compared to the
1272  * group having less cpu_capacity.
1273  */
init_sched_groups_capacity(int cpu,struct sched_domain * sd)1274 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1275 {
1276 	struct sched_group *sg = sd->groups;
1277 	struct cpumask *mask = sched_domains_tmpmask2;
1278 
1279 	WARN_ON(!sg);
1280 
1281 	do {
1282 		int cpu, cores = 0, max_cpu = -1;
1283 
1284 		sg->group_weight = cpumask_weight(sched_group_span(sg));
1285 
1286 		cpumask_copy(mask, sched_group_span(sg));
1287 		for_each_cpu(cpu, mask) {
1288 			cores++;
1289 #ifdef CONFIG_SCHED_SMT
1290 			cpumask_andnot(mask, mask, cpu_smt_mask(cpu));
1291 #endif
1292 		}
1293 		sg->cores = cores;
1294 
1295 		if (!(sd->flags & SD_ASYM_PACKING))
1296 			goto next;
1297 
1298 		for_each_cpu(cpu, sched_group_span(sg)) {
1299 			if (max_cpu < 0)
1300 				max_cpu = cpu;
1301 			else if (sched_asym_prefer(cpu, max_cpu))
1302 				max_cpu = cpu;
1303 		}
1304 		sg->asym_prefer_cpu = max_cpu;
1305 
1306 next:
1307 		sg = sg->next;
1308 	} while (sg != sd->groups);
1309 
1310 	if (cpu != group_balance_cpu(sg))
1311 		return;
1312 
1313 	update_group_capacity(sd, cpu);
1314 }
1315 
1316 /*
1317  * Asymmetric CPU capacity bits
1318  */
1319 struct asym_cap_data {
1320 	struct list_head link;
1321 	unsigned long capacity;
1322 	unsigned long cpus[];
1323 };
1324 
1325 /*
1326  * Set of available CPUs grouped by their corresponding capacities
1327  * Each list entry contains a CPU mask reflecting CPUs that share the same
1328  * capacity.
1329  * The lifespan of data is unlimited.
1330  */
1331 static LIST_HEAD(asym_cap_list);
1332 
1333 #define cpu_capacity_span(asym_data) to_cpumask((asym_data)->cpus)
1334 
1335 /*
1336  * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1337  * Provides sd_flags reflecting the asymmetry scope.
1338  */
1339 static inline int
asym_cpu_capacity_classify(const struct cpumask * sd_span,const struct cpumask * cpu_map)1340 asym_cpu_capacity_classify(const struct cpumask *sd_span,
1341 			   const struct cpumask *cpu_map)
1342 {
1343 	struct asym_cap_data *entry;
1344 	int count = 0, miss = 0;
1345 
1346 	/*
1347 	 * Count how many unique CPU capacities this domain spans across
1348 	 * (compare sched_domain CPUs mask with ones representing  available
1349 	 * CPUs capacities). Take into account CPUs that might be offline:
1350 	 * skip those.
1351 	 */
1352 	list_for_each_entry(entry, &asym_cap_list, link) {
1353 		if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1354 			++count;
1355 		else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1356 			++miss;
1357 	}
1358 
1359 	WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1360 
1361 	/* No asymmetry detected */
1362 	if (count < 2)
1363 		return 0;
1364 	/* Some of the available CPU capacity values have not been detected */
1365 	if (miss)
1366 		return SD_ASYM_CPUCAPACITY;
1367 
1368 	/* Full asymmetry */
1369 	return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1370 
1371 }
1372 
asym_cpu_capacity_update_data(int cpu)1373 static inline void asym_cpu_capacity_update_data(int cpu)
1374 {
1375 	unsigned long capacity = arch_scale_cpu_capacity(cpu);
1376 	struct asym_cap_data *entry = NULL;
1377 
1378 	list_for_each_entry(entry, &asym_cap_list, link) {
1379 		if (capacity == entry->capacity)
1380 			goto done;
1381 	}
1382 
1383 	entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1384 	if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1385 		return;
1386 	entry->capacity = capacity;
1387 	list_add(&entry->link, &asym_cap_list);
1388 done:
1389 	__cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1390 }
1391 
1392 /*
1393  * Build-up/update list of CPUs grouped by their capacities
1394  * An update requires explicit request to rebuild sched domains
1395  * with state indicating CPU topology changes.
1396  */
asym_cpu_capacity_scan(void)1397 static void asym_cpu_capacity_scan(void)
1398 {
1399 	struct asym_cap_data *entry, *next;
1400 	int cpu;
1401 
1402 	list_for_each_entry(entry, &asym_cap_list, link)
1403 		cpumask_clear(cpu_capacity_span(entry));
1404 
1405 	for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1406 		asym_cpu_capacity_update_data(cpu);
1407 
1408 	list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1409 		if (cpumask_empty(cpu_capacity_span(entry))) {
1410 			list_del(&entry->link);
1411 			kfree(entry);
1412 		}
1413 	}
1414 
1415 	/*
1416 	 * Only one capacity value has been detected i.e. this system is symmetric.
1417 	 * No need to keep this data around.
1418 	 */
1419 	if (list_is_singular(&asym_cap_list)) {
1420 		entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1421 		list_del(&entry->link);
1422 		kfree(entry);
1423 	}
1424 }
1425 
1426 /*
1427  * Initializers for schedule domains
1428  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1429  */
1430 
1431 static int default_relax_domain_level = -1;
1432 int sched_domain_level_max;
1433 
setup_relax_domain_level(char * str)1434 static int __init setup_relax_domain_level(char *str)
1435 {
1436 	if (kstrtoint(str, 0, &default_relax_domain_level))
1437 		pr_warn("Unable to set relax_domain_level\n");
1438 
1439 	return 1;
1440 }
1441 __setup("relax_domain_level=", setup_relax_domain_level);
1442 
set_domain_attribute(struct sched_domain * sd,struct sched_domain_attr * attr)1443 static void set_domain_attribute(struct sched_domain *sd,
1444 				 struct sched_domain_attr *attr)
1445 {
1446 	int request;
1447 
1448 	if (!attr || attr->relax_domain_level < 0) {
1449 		if (default_relax_domain_level < 0)
1450 			return;
1451 		request = default_relax_domain_level;
1452 	} else
1453 		request = attr->relax_domain_level;
1454 
1455 	if (sd->level >= request) {
1456 		/* Turn off idle balance on this domain: */
1457 		sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1458 	}
1459 }
1460 
1461 static void __sdt_free(const struct cpumask *cpu_map);
1462 static int __sdt_alloc(const struct cpumask *cpu_map);
1463 
__free_domain_allocs(struct s_data * d,enum s_alloc what,const struct cpumask * cpu_map)1464 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1465 				 const struct cpumask *cpu_map)
1466 {
1467 	switch (what) {
1468 	case sa_rootdomain:
1469 		if (!atomic_read(&d->rd->refcount))
1470 			free_rootdomain(&d->rd->rcu);
1471 		fallthrough;
1472 	case sa_sd:
1473 		free_percpu(d->sd);
1474 		fallthrough;
1475 	case sa_sd_storage:
1476 		__sdt_free(cpu_map);
1477 		fallthrough;
1478 	case sa_none:
1479 		break;
1480 	}
1481 }
1482 
1483 static enum s_alloc
__visit_domain_allocation_hell(struct s_data * d,const struct cpumask * cpu_map)1484 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1485 {
1486 	memset(d, 0, sizeof(*d));
1487 
1488 	if (__sdt_alloc(cpu_map))
1489 		return sa_sd_storage;
1490 	d->sd = alloc_percpu(struct sched_domain *);
1491 	if (!d->sd)
1492 		return sa_sd_storage;
1493 	d->rd = alloc_rootdomain();
1494 	if (!d->rd)
1495 		return sa_sd;
1496 
1497 	return sa_rootdomain;
1498 }
1499 
1500 /*
1501  * NULL the sd_data elements we've used to build the sched_domain and
1502  * sched_group structure so that the subsequent __free_domain_allocs()
1503  * will not free the data we're using.
1504  */
claim_allocations(int cpu,struct sched_domain * sd)1505 static void claim_allocations(int cpu, struct sched_domain *sd)
1506 {
1507 	struct sd_data *sdd = sd->private;
1508 
1509 	WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1510 	*per_cpu_ptr(sdd->sd, cpu) = NULL;
1511 
1512 	if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1513 		*per_cpu_ptr(sdd->sds, cpu) = NULL;
1514 
1515 	if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1516 		*per_cpu_ptr(sdd->sg, cpu) = NULL;
1517 
1518 	if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1519 		*per_cpu_ptr(sdd->sgc, cpu) = NULL;
1520 }
1521 
1522 #ifdef CONFIG_NUMA
1523 enum numa_topology_type sched_numa_topology_type;
1524 
1525 static int			sched_domains_numa_levels;
1526 static int			sched_domains_curr_level;
1527 
1528 int				sched_max_numa_distance;
1529 static int			*sched_domains_numa_distance;
1530 static struct cpumask		***sched_domains_numa_masks;
1531 #endif
1532 
1533 /*
1534  * SD_flags allowed in topology descriptions.
1535  *
1536  * These flags are purely descriptive of the topology and do not prescribe
1537  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1538  * function:
1539  *
1540  *   SD_SHARE_CPUCAPACITY   - describes SMT topologies
1541  *   SD_SHARE_PKG_RESOURCES - describes shared caches
1542  *   SD_NUMA                - describes NUMA topologies
1543  *
1544  * Odd one out, which beside describing the topology has a quirk also
1545  * prescribes the desired behaviour that goes along with it:
1546  *
1547  *   SD_ASYM_PACKING        - describes SMT quirks
1548  */
1549 #define TOPOLOGY_SD_FLAGS		\
1550 	(SD_SHARE_CPUCAPACITY	|	\
1551 	 SD_SHARE_PKG_RESOURCES |	\
1552 	 SD_NUMA		|	\
1553 	 SD_ASYM_PACKING)
1554 
1555 static struct sched_domain *
sd_init(struct sched_domain_topology_level * tl,const struct cpumask * cpu_map,struct sched_domain * child,int cpu)1556 sd_init(struct sched_domain_topology_level *tl,
1557 	const struct cpumask *cpu_map,
1558 	struct sched_domain *child, int cpu)
1559 {
1560 	struct sd_data *sdd = &tl->data;
1561 	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1562 	int sd_id, sd_weight, sd_flags = 0;
1563 	struct cpumask *sd_span;
1564 
1565 #ifdef CONFIG_NUMA
1566 	/*
1567 	 * Ugly hack to pass state to sd_numa_mask()...
1568 	 */
1569 	sched_domains_curr_level = tl->numa_level;
1570 #endif
1571 
1572 	sd_weight = cpumask_weight(tl->mask(cpu));
1573 
1574 	if (tl->sd_flags)
1575 		sd_flags = (*tl->sd_flags)();
1576 	if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1577 			"wrong sd_flags in topology description\n"))
1578 		sd_flags &= TOPOLOGY_SD_FLAGS;
1579 
1580 	*sd = (struct sched_domain){
1581 		.min_interval		= sd_weight,
1582 		.max_interval		= 2*sd_weight,
1583 		.busy_factor		= 16,
1584 		.imbalance_pct		= 117,
1585 
1586 		.cache_nice_tries	= 0,
1587 
1588 		.flags			= 1*SD_BALANCE_NEWIDLE
1589 					| 1*SD_BALANCE_EXEC
1590 					| 1*SD_BALANCE_FORK
1591 					| 0*SD_BALANCE_WAKE
1592 					| 1*SD_WAKE_AFFINE
1593 					| 0*SD_SHARE_CPUCAPACITY
1594 					| 0*SD_SHARE_PKG_RESOURCES
1595 					| 0*SD_SERIALIZE
1596 					| 1*SD_PREFER_SIBLING
1597 					| 0*SD_NUMA
1598 					| sd_flags
1599 					,
1600 
1601 		.last_balance		= jiffies,
1602 		.balance_interval	= sd_weight,
1603 		.max_newidle_lb_cost	= 0,
1604 		.last_decay_max_lb_cost	= jiffies,
1605 		.child			= child,
1606 #ifdef CONFIG_SCHED_DEBUG
1607 		.name			= tl->name,
1608 #endif
1609 	};
1610 
1611 	sd_span = sched_domain_span(sd);
1612 	cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1613 	sd_id = cpumask_first(sd_span);
1614 
1615 	sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1616 
1617 	WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1618 		  (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1619 		  "CPU capacity asymmetry not supported on SMT\n");
1620 
1621 	/*
1622 	 * Convert topological properties into behaviour.
1623 	 */
1624 	/* Don't attempt to spread across CPUs of different capacities. */
1625 	if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1626 		sd->child->flags &= ~SD_PREFER_SIBLING;
1627 
1628 	if (sd->flags & SD_SHARE_CPUCAPACITY) {
1629 		sd->imbalance_pct = 110;
1630 
1631 	} else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1632 		sd->imbalance_pct = 117;
1633 		sd->cache_nice_tries = 1;
1634 
1635 #ifdef CONFIG_NUMA
1636 	} else if (sd->flags & SD_NUMA) {
1637 		sd->cache_nice_tries = 2;
1638 
1639 		sd->flags &= ~SD_PREFER_SIBLING;
1640 		sd->flags |= SD_SERIALIZE;
1641 		if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1642 			sd->flags &= ~(SD_BALANCE_EXEC |
1643 				       SD_BALANCE_FORK |
1644 				       SD_WAKE_AFFINE);
1645 		}
1646 
1647 #endif
1648 	} else {
1649 		sd->cache_nice_tries = 1;
1650 	}
1651 
1652 	/*
1653 	 * For all levels sharing cache; connect a sched_domain_shared
1654 	 * instance.
1655 	 */
1656 	if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1657 		sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1658 		atomic_inc(&sd->shared->ref);
1659 		atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1660 	}
1661 
1662 	sd->private = sdd;
1663 
1664 	return sd;
1665 }
1666 
1667 /*
1668  * Topology list, bottom-up.
1669  */
1670 static struct sched_domain_topology_level default_topology[] = {
1671 #ifdef CONFIG_SCHED_SMT
1672 	{ cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1673 #endif
1674 
1675 #ifdef CONFIG_SCHED_CLUSTER
1676 	{ cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1677 #endif
1678 
1679 #ifdef CONFIG_SCHED_MC
1680 	{ cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1681 #endif
1682 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
1683 	{ NULL, },
1684 };
1685 
1686 static struct sched_domain_topology_level *sched_domain_topology =
1687 	default_topology;
1688 static struct sched_domain_topology_level *sched_domain_topology_saved;
1689 
1690 #define for_each_sd_topology(tl)			\
1691 	for (tl = sched_domain_topology; tl->mask; tl++)
1692 
set_sched_topology(struct sched_domain_topology_level * tl)1693 void __init set_sched_topology(struct sched_domain_topology_level *tl)
1694 {
1695 	if (WARN_ON_ONCE(sched_smp_initialized))
1696 		return;
1697 
1698 	sched_domain_topology = tl;
1699 	sched_domain_topology_saved = NULL;
1700 }
1701 
1702 #ifdef CONFIG_NUMA
1703 
sd_numa_mask(int cpu)1704 static const struct cpumask *sd_numa_mask(int cpu)
1705 {
1706 	return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1707 }
1708 
sched_numa_warn(const char * str)1709 static void sched_numa_warn(const char *str)
1710 {
1711 	static int done = false;
1712 	int i,j;
1713 
1714 	if (done)
1715 		return;
1716 
1717 	done = true;
1718 
1719 	printk(KERN_WARNING "ERROR: %s\n\n", str);
1720 
1721 	for (i = 0; i < nr_node_ids; i++) {
1722 		printk(KERN_WARNING "  ");
1723 		for (j = 0; j < nr_node_ids; j++) {
1724 			if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
1725 				printk(KERN_CONT "(%02d) ", node_distance(i,j));
1726 			else
1727 				printk(KERN_CONT " %02d  ", node_distance(i,j));
1728 		}
1729 		printk(KERN_CONT "\n");
1730 	}
1731 	printk(KERN_WARNING "\n");
1732 }
1733 
find_numa_distance(int distance)1734 bool find_numa_distance(int distance)
1735 {
1736 	bool found = false;
1737 	int i, *distances;
1738 
1739 	if (distance == node_distance(0, 0))
1740 		return true;
1741 
1742 	rcu_read_lock();
1743 	distances = rcu_dereference(sched_domains_numa_distance);
1744 	if (!distances)
1745 		goto unlock;
1746 	for (i = 0; i < sched_domains_numa_levels; i++) {
1747 		if (distances[i] == distance) {
1748 			found = true;
1749 			break;
1750 		}
1751 	}
1752 unlock:
1753 	rcu_read_unlock();
1754 
1755 	return found;
1756 }
1757 
1758 #define for_each_cpu_node_but(n, nbut)		\
1759 	for_each_node_state(n, N_CPU)		\
1760 		if (n == nbut)			\
1761 			continue;		\
1762 		else
1763 
1764 /*
1765  * A system can have three types of NUMA topology:
1766  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1767  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1768  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1769  *
1770  * The difference between a glueless mesh topology and a backplane
1771  * topology lies in whether communication between not directly
1772  * connected nodes goes through intermediary nodes (where programs
1773  * could run), or through backplane controllers. This affects
1774  * placement of programs.
1775  *
1776  * The type of topology can be discerned with the following tests:
1777  * - If the maximum distance between any nodes is 1 hop, the system
1778  *   is directly connected.
1779  * - If for two nodes A and B, located N > 1 hops away from each other,
1780  *   there is an intermediary node C, which is < N hops away from both
1781  *   nodes A and B, the system is a glueless mesh.
1782  */
init_numa_topology_type(int offline_node)1783 static void init_numa_topology_type(int offline_node)
1784 {
1785 	int a, b, c, n;
1786 
1787 	n = sched_max_numa_distance;
1788 
1789 	if (sched_domains_numa_levels <= 2) {
1790 		sched_numa_topology_type = NUMA_DIRECT;
1791 		return;
1792 	}
1793 
1794 	for_each_cpu_node_but(a, offline_node) {
1795 		for_each_cpu_node_but(b, offline_node) {
1796 			/* Find two nodes furthest removed from each other. */
1797 			if (node_distance(a, b) < n)
1798 				continue;
1799 
1800 			/* Is there an intermediary node between a and b? */
1801 			for_each_cpu_node_but(c, offline_node) {
1802 				if (node_distance(a, c) < n &&
1803 				    node_distance(b, c) < n) {
1804 					sched_numa_topology_type =
1805 							NUMA_GLUELESS_MESH;
1806 					return;
1807 				}
1808 			}
1809 
1810 			sched_numa_topology_type = NUMA_BACKPLANE;
1811 			return;
1812 		}
1813 	}
1814 
1815 	pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1816 	sched_numa_topology_type = NUMA_DIRECT;
1817 }
1818 
1819 
1820 #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1821 
sched_init_numa(int offline_node)1822 void sched_init_numa(int offline_node)
1823 {
1824 	struct sched_domain_topology_level *tl;
1825 	unsigned long *distance_map;
1826 	int nr_levels = 0;
1827 	int i, j;
1828 	int *distances;
1829 	struct cpumask ***masks;
1830 
1831 	/*
1832 	 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1833 	 * unique distances in the node_distance() table.
1834 	 */
1835 	distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1836 	if (!distance_map)
1837 		return;
1838 
1839 	bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1840 	for_each_cpu_node_but(i, offline_node) {
1841 		for_each_cpu_node_but(j, offline_node) {
1842 			int distance = node_distance(i, j);
1843 
1844 			if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1845 				sched_numa_warn("Invalid distance value range");
1846 				bitmap_free(distance_map);
1847 				return;
1848 			}
1849 
1850 			bitmap_set(distance_map, distance, 1);
1851 		}
1852 	}
1853 	/*
1854 	 * We can now figure out how many unique distance values there are and
1855 	 * allocate memory accordingly.
1856 	 */
1857 	nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1858 
1859 	distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1860 	if (!distances) {
1861 		bitmap_free(distance_map);
1862 		return;
1863 	}
1864 
1865 	for (i = 0, j = 0; i < nr_levels; i++, j++) {
1866 		j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1867 		distances[i] = j;
1868 	}
1869 	rcu_assign_pointer(sched_domains_numa_distance, distances);
1870 
1871 	bitmap_free(distance_map);
1872 
1873 	/*
1874 	 * 'nr_levels' contains the number of unique distances
1875 	 *
1876 	 * The sched_domains_numa_distance[] array includes the actual distance
1877 	 * numbers.
1878 	 */
1879 
1880 	/*
1881 	 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1882 	 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1883 	 * the array will contain less then 'nr_levels' members. This could be
1884 	 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1885 	 * in other functions.
1886 	 *
1887 	 * We reset it to 'nr_levels' at the end of this function.
1888 	 */
1889 	sched_domains_numa_levels = 0;
1890 
1891 	masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1892 	if (!masks)
1893 		return;
1894 
1895 	/*
1896 	 * Now for each level, construct a mask per node which contains all
1897 	 * CPUs of nodes that are that many hops away from us.
1898 	 */
1899 	for (i = 0; i < nr_levels; i++) {
1900 		masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1901 		if (!masks[i])
1902 			return;
1903 
1904 		for_each_cpu_node_but(j, offline_node) {
1905 			struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1906 			int k;
1907 
1908 			if (!mask)
1909 				return;
1910 
1911 			masks[i][j] = mask;
1912 
1913 			for_each_cpu_node_but(k, offline_node) {
1914 				if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1915 					sched_numa_warn("Node-distance not symmetric");
1916 
1917 				if (node_distance(j, k) > sched_domains_numa_distance[i])
1918 					continue;
1919 
1920 				cpumask_or(mask, mask, cpumask_of_node(k));
1921 			}
1922 		}
1923 	}
1924 	rcu_assign_pointer(sched_domains_numa_masks, masks);
1925 
1926 	/* Compute default topology size */
1927 	for (i = 0; sched_domain_topology[i].mask; i++);
1928 
1929 	tl = kzalloc((i + nr_levels + 1) *
1930 			sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1931 	if (!tl)
1932 		return;
1933 
1934 	/*
1935 	 * Copy the default topology bits..
1936 	 */
1937 	for (i = 0; sched_domain_topology[i].mask; i++)
1938 		tl[i] = sched_domain_topology[i];
1939 
1940 	/*
1941 	 * Add the NUMA identity distance, aka single NODE.
1942 	 */
1943 	tl[i++] = (struct sched_domain_topology_level){
1944 		.mask = sd_numa_mask,
1945 		.numa_level = 0,
1946 		SD_INIT_NAME(NODE)
1947 	};
1948 
1949 	/*
1950 	 * .. and append 'j' levels of NUMA goodness.
1951 	 */
1952 	for (j = 1; j < nr_levels; i++, j++) {
1953 		tl[i] = (struct sched_domain_topology_level){
1954 			.mask = sd_numa_mask,
1955 			.sd_flags = cpu_numa_flags,
1956 			.flags = SDTL_OVERLAP,
1957 			.numa_level = j,
1958 			SD_INIT_NAME(NUMA)
1959 		};
1960 	}
1961 
1962 	sched_domain_topology_saved = sched_domain_topology;
1963 	sched_domain_topology = tl;
1964 
1965 	sched_domains_numa_levels = nr_levels;
1966 	WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]);
1967 
1968 	init_numa_topology_type(offline_node);
1969 }
1970 
1971 
sched_reset_numa(void)1972 static void sched_reset_numa(void)
1973 {
1974 	int nr_levels, *distances;
1975 	struct cpumask ***masks;
1976 
1977 	nr_levels = sched_domains_numa_levels;
1978 	sched_domains_numa_levels = 0;
1979 	sched_max_numa_distance = 0;
1980 	sched_numa_topology_type = NUMA_DIRECT;
1981 	distances = sched_domains_numa_distance;
1982 	rcu_assign_pointer(sched_domains_numa_distance, NULL);
1983 	masks = sched_domains_numa_masks;
1984 	rcu_assign_pointer(sched_domains_numa_masks, NULL);
1985 	if (distances || masks) {
1986 		int i, j;
1987 
1988 		synchronize_rcu();
1989 		kfree(distances);
1990 		for (i = 0; i < nr_levels && masks; i++) {
1991 			if (!masks[i])
1992 				continue;
1993 			for_each_node(j)
1994 				kfree(masks[i][j]);
1995 			kfree(masks[i]);
1996 		}
1997 		kfree(masks);
1998 	}
1999 	if (sched_domain_topology_saved) {
2000 		kfree(sched_domain_topology);
2001 		sched_domain_topology = sched_domain_topology_saved;
2002 		sched_domain_topology_saved = NULL;
2003 	}
2004 }
2005 
2006 /*
2007  * Call with hotplug lock held
2008  */
sched_update_numa(int cpu,bool online)2009 void sched_update_numa(int cpu, bool online)
2010 {
2011 	int node;
2012 
2013 	node = cpu_to_node(cpu);
2014 	/*
2015 	 * Scheduler NUMA topology is updated when the first CPU of a
2016 	 * node is onlined or the last CPU of a node is offlined.
2017 	 */
2018 	if (cpumask_weight(cpumask_of_node(node)) != 1)
2019 		return;
2020 
2021 	sched_reset_numa();
2022 	sched_init_numa(online ? NUMA_NO_NODE : node);
2023 }
2024 
sched_domains_numa_masks_set(unsigned int cpu)2025 void sched_domains_numa_masks_set(unsigned int cpu)
2026 {
2027 	int node = cpu_to_node(cpu);
2028 	int i, j;
2029 
2030 	for (i = 0; i < sched_domains_numa_levels; i++) {
2031 		for (j = 0; j < nr_node_ids; j++) {
2032 			if (!node_state(j, N_CPU))
2033 				continue;
2034 
2035 			/* Set ourselves in the remote node's masks */
2036 			if (node_distance(j, node) <= sched_domains_numa_distance[i])
2037 				cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
2038 		}
2039 	}
2040 }
2041 
sched_domains_numa_masks_clear(unsigned int cpu)2042 void sched_domains_numa_masks_clear(unsigned int cpu)
2043 {
2044 	int i, j;
2045 
2046 	for (i = 0; i < sched_domains_numa_levels; i++) {
2047 		for (j = 0; j < nr_node_ids; j++) {
2048 			if (sched_domains_numa_masks[i][j])
2049 				cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
2050 		}
2051 	}
2052 }
2053 
2054 /*
2055  * sched_numa_find_closest() - given the NUMA topology, find the cpu
2056  *                             closest to @cpu from @cpumask.
2057  * cpumask: cpumask to find a cpu from
2058  * cpu: cpu to be close to
2059  *
2060  * returns: cpu, or nr_cpu_ids when nothing found.
2061  */
sched_numa_find_closest(const struct cpumask * cpus,int cpu)2062 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2063 {
2064 	int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2065 	struct cpumask ***masks;
2066 
2067 	rcu_read_lock();
2068 	masks = rcu_dereference(sched_domains_numa_masks);
2069 	if (!masks)
2070 		goto unlock;
2071 	for (i = 0; i < sched_domains_numa_levels; i++) {
2072 		if (!masks[i][j])
2073 			break;
2074 		cpu = cpumask_any_and(cpus, masks[i][j]);
2075 		if (cpu < nr_cpu_ids) {
2076 			found = cpu;
2077 			break;
2078 		}
2079 	}
2080 unlock:
2081 	rcu_read_unlock();
2082 
2083 	return found;
2084 }
2085 
2086 struct __cmp_key {
2087 	const struct cpumask *cpus;
2088 	struct cpumask ***masks;
2089 	int node;
2090 	int cpu;
2091 	int w;
2092 };
2093 
hop_cmp(const void * a,const void * b)2094 static int hop_cmp(const void *a, const void *b)
2095 {
2096 	struct cpumask **prev_hop, **cur_hop = *(struct cpumask ***)b;
2097 	struct __cmp_key *k = (struct __cmp_key *)a;
2098 
2099 	if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu)
2100 		return 1;
2101 
2102 	if (b == k->masks) {
2103 		k->w = 0;
2104 		return 0;
2105 	}
2106 
2107 	prev_hop = *((struct cpumask ***)b - 1);
2108 	k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]);
2109 	if (k->w <= k->cpu)
2110 		return 0;
2111 
2112 	return -1;
2113 }
2114 
2115 /*
2116  * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth next cpu
2117  *                             closest to @cpu from @cpumask.
2118  * cpumask: cpumask to find a cpu from
2119  * cpu: Nth cpu to find
2120  *
2121  * returns: cpu, or nr_cpu_ids when nothing found.
2122  */
sched_numa_find_nth_cpu(const struct cpumask * cpus,int cpu,int node)2123 int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
2124 {
2125 	struct __cmp_key k = { .cpus = cpus, .cpu = cpu };
2126 	struct cpumask ***hop_masks;
2127 	int hop, ret = nr_cpu_ids;
2128 
2129 	if (node == NUMA_NO_NODE)
2130 		return cpumask_nth_and(cpu, cpus, cpu_online_mask);
2131 
2132 	rcu_read_lock();
2133 
2134 	/* CPU-less node entries are uninitialized in sched_domains_numa_masks */
2135 	node = numa_nearest_node(node, N_CPU);
2136 	k.node = node;
2137 
2138 	k.masks = rcu_dereference(sched_domains_numa_masks);
2139 	if (!k.masks)
2140 		goto unlock;
2141 
2142 	hop_masks = bsearch(&k, k.masks, sched_domains_numa_levels, sizeof(k.masks[0]), hop_cmp);
2143 	hop = hop_masks	- k.masks;
2144 
2145 	ret = hop ?
2146 		cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) :
2147 		cpumask_nth_and(cpu, cpus, k.masks[0][node]);
2148 unlock:
2149 	rcu_read_unlock();
2150 	return ret;
2151 }
2152 EXPORT_SYMBOL_GPL(sched_numa_find_nth_cpu);
2153 
2154 /**
2155  * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2156  *                         @node
2157  * @node: The node to count hops from.
2158  * @hops: Include CPUs up to that many hops away. 0 means local node.
2159  *
2160  * Return: On success, a pointer to a cpumask of CPUs at most @hops away from
2161  * @node, an error value otherwise.
2162  *
2163  * Requires rcu_lock to be held. Returned cpumask is only valid within that
2164  * read-side section, copy it if required beyond that.
2165  *
2166  * Note that not all hops are equal in distance; see sched_init_numa() for how
2167  * distances and masks are handled.
2168  * Also note that this is a reflection of sched_domains_numa_masks, which may change
2169  * during the lifetime of the system (offline nodes are taken out of the masks).
2170  */
sched_numa_hop_mask(unsigned int node,unsigned int hops)2171 const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops)
2172 {
2173 	struct cpumask ***masks;
2174 
2175 	if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
2176 		return ERR_PTR(-EINVAL);
2177 
2178 	masks = rcu_dereference(sched_domains_numa_masks);
2179 	if (!masks)
2180 		return ERR_PTR(-EBUSY);
2181 
2182 	return masks[hops][node];
2183 }
2184 EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
2185 
2186 #endif /* CONFIG_NUMA */
2187 
__sdt_alloc(const struct cpumask * cpu_map)2188 static int __sdt_alloc(const struct cpumask *cpu_map)
2189 {
2190 	struct sched_domain_topology_level *tl;
2191 	int j;
2192 
2193 	for_each_sd_topology(tl) {
2194 		struct sd_data *sdd = &tl->data;
2195 
2196 		sdd->sd = alloc_percpu(struct sched_domain *);
2197 		if (!sdd->sd)
2198 			return -ENOMEM;
2199 
2200 		sdd->sds = alloc_percpu(struct sched_domain_shared *);
2201 		if (!sdd->sds)
2202 			return -ENOMEM;
2203 
2204 		sdd->sg = alloc_percpu(struct sched_group *);
2205 		if (!sdd->sg)
2206 			return -ENOMEM;
2207 
2208 		sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2209 		if (!sdd->sgc)
2210 			return -ENOMEM;
2211 
2212 		for_each_cpu(j, cpu_map) {
2213 			struct sched_domain *sd;
2214 			struct sched_domain_shared *sds;
2215 			struct sched_group *sg;
2216 			struct sched_group_capacity *sgc;
2217 
2218 			sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2219 					GFP_KERNEL, cpu_to_node(j));
2220 			if (!sd)
2221 				return -ENOMEM;
2222 
2223 			*per_cpu_ptr(sdd->sd, j) = sd;
2224 
2225 			sds = kzalloc_node(sizeof(struct sched_domain_shared),
2226 					GFP_KERNEL, cpu_to_node(j));
2227 			if (!sds)
2228 				return -ENOMEM;
2229 
2230 			*per_cpu_ptr(sdd->sds, j) = sds;
2231 
2232 			sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2233 					GFP_KERNEL, cpu_to_node(j));
2234 			if (!sg)
2235 				return -ENOMEM;
2236 
2237 			sg->next = sg;
2238 
2239 			*per_cpu_ptr(sdd->sg, j) = sg;
2240 
2241 			sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2242 					GFP_KERNEL, cpu_to_node(j));
2243 			if (!sgc)
2244 				return -ENOMEM;
2245 
2246 #ifdef CONFIG_SCHED_DEBUG
2247 			sgc->id = j;
2248 #endif
2249 
2250 			*per_cpu_ptr(sdd->sgc, j) = sgc;
2251 		}
2252 	}
2253 
2254 	return 0;
2255 }
2256 
__sdt_free(const struct cpumask * cpu_map)2257 static void __sdt_free(const struct cpumask *cpu_map)
2258 {
2259 	struct sched_domain_topology_level *tl;
2260 	int j;
2261 
2262 	for_each_sd_topology(tl) {
2263 		struct sd_data *sdd = &tl->data;
2264 
2265 		for_each_cpu(j, cpu_map) {
2266 			struct sched_domain *sd;
2267 
2268 			if (sdd->sd) {
2269 				sd = *per_cpu_ptr(sdd->sd, j);
2270 				if (sd && (sd->flags & SD_OVERLAP))
2271 					free_sched_groups(sd->groups, 0);
2272 				kfree(*per_cpu_ptr(sdd->sd, j));
2273 			}
2274 
2275 			if (sdd->sds)
2276 				kfree(*per_cpu_ptr(sdd->sds, j));
2277 			if (sdd->sg)
2278 				kfree(*per_cpu_ptr(sdd->sg, j));
2279 			if (sdd->sgc)
2280 				kfree(*per_cpu_ptr(sdd->sgc, j));
2281 		}
2282 		free_percpu(sdd->sd);
2283 		sdd->sd = NULL;
2284 		free_percpu(sdd->sds);
2285 		sdd->sds = NULL;
2286 		free_percpu(sdd->sg);
2287 		sdd->sg = NULL;
2288 		free_percpu(sdd->sgc);
2289 		sdd->sgc = NULL;
2290 	}
2291 }
2292 
build_sched_domain(struct sched_domain_topology_level * tl,const struct cpumask * cpu_map,struct sched_domain_attr * attr,struct sched_domain * child,int cpu)2293 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2294 		const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2295 		struct sched_domain *child, int cpu)
2296 {
2297 	struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2298 
2299 	if (child) {
2300 		sd->level = child->level + 1;
2301 		sched_domain_level_max = max(sched_domain_level_max, sd->level);
2302 		child->parent = sd;
2303 
2304 		if (!cpumask_subset(sched_domain_span(child),
2305 				    sched_domain_span(sd))) {
2306 			pr_err("BUG: arch topology borken\n");
2307 #ifdef CONFIG_SCHED_DEBUG
2308 			pr_err("     the %s domain not a subset of the %s domain\n",
2309 					child->name, sd->name);
2310 #endif
2311 			/* Fixup, ensure @sd has at least @child CPUs. */
2312 			cpumask_or(sched_domain_span(sd),
2313 				   sched_domain_span(sd),
2314 				   sched_domain_span(child));
2315 		}
2316 
2317 	}
2318 	set_domain_attribute(sd, attr);
2319 
2320 	return sd;
2321 }
2322 
2323 /*
2324  * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2325  * any two given CPUs at this (non-NUMA) topology level.
2326  */
topology_span_sane(struct sched_domain_topology_level * tl,const struct cpumask * cpu_map,int cpu)2327 static bool topology_span_sane(struct sched_domain_topology_level *tl,
2328 			      const struct cpumask *cpu_map, int cpu)
2329 {
2330 	int i;
2331 
2332 	/* NUMA levels are allowed to overlap */
2333 	if (tl->flags & SDTL_OVERLAP)
2334 		return true;
2335 
2336 	/*
2337 	 * Non-NUMA levels cannot partially overlap - they must be either
2338 	 * completely equal or completely disjoint. Otherwise we can end up
2339 	 * breaking the sched_group lists - i.e. a later get_group() pass
2340 	 * breaks the linking done for an earlier span.
2341 	 */
2342 	for_each_cpu(i, cpu_map) {
2343 		if (i == cpu)
2344 			continue;
2345 		/*
2346 		 * We should 'and' all those masks with 'cpu_map' to exactly
2347 		 * match the topology we're about to build, but that can only
2348 		 * remove CPUs, which only lessens our ability to detect
2349 		 * overlaps
2350 		 */
2351 		if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2352 		    cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2353 			return false;
2354 	}
2355 
2356 	return true;
2357 }
2358 
2359 /*
2360  * Build sched domains for a given set of CPUs and attach the sched domains
2361  * to the individual CPUs
2362  */
2363 static int
build_sched_domains(const struct cpumask * cpu_map,struct sched_domain_attr * attr)2364 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2365 {
2366 	enum s_alloc alloc_state = sa_none;
2367 	struct sched_domain *sd;
2368 	struct s_data d;
2369 	struct rq *rq = NULL;
2370 	int i, ret = -ENOMEM;
2371 	bool has_asym = false;
2372 
2373 	if (WARN_ON(cpumask_empty(cpu_map)))
2374 		goto error;
2375 
2376 	alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2377 	if (alloc_state != sa_rootdomain)
2378 		goto error;
2379 
2380 	/* Set up domains for CPUs specified by the cpu_map: */
2381 	for_each_cpu(i, cpu_map) {
2382 		struct sched_domain_topology_level *tl;
2383 
2384 		sd = NULL;
2385 		for_each_sd_topology(tl) {
2386 
2387 			if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2388 				goto error;
2389 
2390 			sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2391 
2392 			has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2393 
2394 			if (tl == sched_domain_topology)
2395 				*per_cpu_ptr(d.sd, i) = sd;
2396 			if (tl->flags & SDTL_OVERLAP)
2397 				sd->flags |= SD_OVERLAP;
2398 			if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2399 				break;
2400 		}
2401 	}
2402 
2403 	/* Build the groups for the domains */
2404 	for_each_cpu(i, cpu_map) {
2405 		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2406 			sd->span_weight = cpumask_weight(sched_domain_span(sd));
2407 			if (sd->flags & SD_OVERLAP) {
2408 				if (build_overlap_sched_groups(sd, i))
2409 					goto error;
2410 			} else {
2411 				if (build_sched_groups(sd, i))
2412 					goto error;
2413 			}
2414 		}
2415 	}
2416 
2417 	/*
2418 	 * Calculate an allowed NUMA imbalance such that LLCs do not get
2419 	 * imbalanced.
2420 	 */
2421 	for_each_cpu(i, cpu_map) {
2422 		unsigned int imb = 0;
2423 		unsigned int imb_span = 1;
2424 
2425 		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2426 			struct sched_domain *child = sd->child;
2427 
2428 			if (!(sd->flags & SD_SHARE_PKG_RESOURCES) && child &&
2429 			    (child->flags & SD_SHARE_PKG_RESOURCES)) {
2430 				struct sched_domain __rcu *top_p;
2431 				unsigned int nr_llcs;
2432 
2433 				/*
2434 				 * For a single LLC per node, allow an
2435 				 * imbalance up to 12.5% of the node. This is
2436 				 * arbitrary cutoff based two factors -- SMT and
2437 				 * memory channels. For SMT-2, the intent is to
2438 				 * avoid premature sharing of HT resources but
2439 				 * SMT-4 or SMT-8 *may* benefit from a different
2440 				 * cutoff. For memory channels, this is a very
2441 				 * rough estimate of how many channels may be
2442 				 * active and is based on recent CPUs with
2443 				 * many cores.
2444 				 *
2445 				 * For multiple LLCs, allow an imbalance
2446 				 * until multiple tasks would share an LLC
2447 				 * on one node while LLCs on another node
2448 				 * remain idle. This assumes that there are
2449 				 * enough logical CPUs per LLC to avoid SMT
2450 				 * factors and that there is a correlation
2451 				 * between LLCs and memory channels.
2452 				 */
2453 				nr_llcs = sd->span_weight / child->span_weight;
2454 				if (nr_llcs == 1)
2455 					imb = sd->span_weight >> 3;
2456 				else
2457 					imb = nr_llcs;
2458 				imb = max(1U, imb);
2459 				sd->imb_numa_nr = imb;
2460 
2461 				/* Set span based on the first NUMA domain. */
2462 				top_p = sd->parent;
2463 				while (top_p && !(top_p->flags & SD_NUMA)) {
2464 					top_p = top_p->parent;
2465 				}
2466 				imb_span = top_p ? top_p->span_weight : sd->span_weight;
2467 			} else {
2468 				int factor = max(1U, (sd->span_weight / imb_span));
2469 
2470 				sd->imb_numa_nr = imb * factor;
2471 			}
2472 		}
2473 	}
2474 
2475 	/* Calculate CPU capacity for physical packages and nodes */
2476 	for (i = nr_cpumask_bits-1; i >= 0; i--) {
2477 		if (!cpumask_test_cpu(i, cpu_map))
2478 			continue;
2479 
2480 		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2481 			claim_allocations(i, sd);
2482 			init_sched_groups_capacity(i, sd);
2483 		}
2484 	}
2485 
2486 	/* Attach the domains */
2487 	rcu_read_lock();
2488 	for_each_cpu(i, cpu_map) {
2489 		rq = cpu_rq(i);
2490 		sd = *per_cpu_ptr(d.sd, i);
2491 
2492 		/* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2493 		if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2494 			WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2495 
2496 		cpu_attach_domain(sd, d.rd, i);
2497 	}
2498 	rcu_read_unlock();
2499 
2500 	if (has_asym)
2501 		static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2502 
2503 	if (rq && sched_debug_verbose) {
2504 		pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2505 			cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2506 	}
2507 
2508 	ret = 0;
2509 error:
2510 	__free_domain_allocs(&d, alloc_state, cpu_map);
2511 
2512 	return ret;
2513 }
2514 
2515 /* Current sched domains: */
2516 static cpumask_var_t			*doms_cur;
2517 
2518 /* Number of sched domains in 'doms_cur': */
2519 static int				ndoms_cur;
2520 
2521 /* Attributes of custom domains in 'doms_cur' */
2522 static struct sched_domain_attr		*dattr_cur;
2523 
2524 /*
2525  * Special case: If a kmalloc() of a doms_cur partition (array of
2526  * cpumask) fails, then fallback to a single sched domain,
2527  * as determined by the single cpumask fallback_doms.
2528  */
2529 static cpumask_var_t			fallback_doms;
2530 
2531 /*
2532  * arch_update_cpu_topology lets virtualized architectures update the
2533  * CPU core maps. It is supposed to return 1 if the topology changed
2534  * or 0 if it stayed the same.
2535  */
arch_update_cpu_topology(void)2536 int __weak arch_update_cpu_topology(void)
2537 {
2538 	return 0;
2539 }
2540 
alloc_sched_domains(unsigned int ndoms)2541 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2542 {
2543 	int i;
2544 	cpumask_var_t *doms;
2545 
2546 	doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2547 	if (!doms)
2548 		return NULL;
2549 	for (i = 0; i < ndoms; i++) {
2550 		if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2551 			free_sched_domains(doms, i);
2552 			return NULL;
2553 		}
2554 	}
2555 	return doms;
2556 }
2557 
free_sched_domains(cpumask_var_t doms[],unsigned int ndoms)2558 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2559 {
2560 	unsigned int i;
2561 	for (i = 0; i < ndoms; i++)
2562 		free_cpumask_var(doms[i]);
2563 	kfree(doms);
2564 }
2565 
2566 /*
2567  * Set up scheduler domains and groups.  For now this just excludes isolated
2568  * CPUs, but could be used to exclude other special cases in the future.
2569  */
sched_init_domains(const struct cpumask * cpu_map)2570 int __init sched_init_domains(const struct cpumask *cpu_map)
2571 {
2572 	int err;
2573 
2574 	zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2575 	zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2576 	zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2577 
2578 	arch_update_cpu_topology();
2579 	asym_cpu_capacity_scan();
2580 	ndoms_cur = 1;
2581 	doms_cur = alloc_sched_domains(ndoms_cur);
2582 	if (!doms_cur)
2583 		doms_cur = &fallback_doms;
2584 	cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
2585 	err = build_sched_domains(doms_cur[0], NULL);
2586 
2587 	return err;
2588 }
2589 
2590 /*
2591  * Detach sched domains from a group of CPUs specified in cpu_map
2592  * These CPUs will now be attached to the NULL domain
2593  */
detach_destroy_domains(const struct cpumask * cpu_map)2594 static void detach_destroy_domains(const struct cpumask *cpu_map)
2595 {
2596 	unsigned int cpu = cpumask_any(cpu_map);
2597 	int i;
2598 
2599 	if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2600 		static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2601 
2602 	rcu_read_lock();
2603 	for_each_cpu(i, cpu_map)
2604 		cpu_attach_domain(NULL, &def_root_domain, i);
2605 	rcu_read_unlock();
2606 }
2607 
2608 /* handle null as "default" */
dattrs_equal(struct sched_domain_attr * cur,int idx_cur,struct sched_domain_attr * new,int idx_new)2609 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2610 			struct sched_domain_attr *new, int idx_new)
2611 {
2612 	struct sched_domain_attr tmp;
2613 
2614 	/* Fast path: */
2615 	if (!new && !cur)
2616 		return 1;
2617 
2618 	tmp = SD_ATTR_INIT;
2619 
2620 	return !memcmp(cur ? (cur + idx_cur) : &tmp,
2621 			new ? (new + idx_new) : &tmp,
2622 			sizeof(struct sched_domain_attr));
2623 }
2624 
2625 /*
2626  * Partition sched domains as specified by the 'ndoms_new'
2627  * cpumasks in the array doms_new[] of cpumasks. This compares
2628  * doms_new[] to the current sched domain partitioning, doms_cur[].
2629  * It destroys each deleted domain and builds each new domain.
2630  *
2631  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2632  * The masks don't intersect (don't overlap.) We should setup one
2633  * sched domain for each mask. CPUs not in any of the cpumasks will
2634  * not be load balanced. If the same cpumask appears both in the
2635  * current 'doms_cur' domains and in the new 'doms_new', we can leave
2636  * it as it is.
2637  *
2638  * The passed in 'doms_new' should be allocated using
2639  * alloc_sched_domains.  This routine takes ownership of it and will
2640  * free_sched_domains it when done with it. If the caller failed the
2641  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2642  * and partition_sched_domains() will fallback to the single partition
2643  * 'fallback_doms', it also forces the domains to be rebuilt.
2644  *
2645  * If doms_new == NULL it will be replaced with cpu_online_mask.
2646  * ndoms_new == 0 is a special case for destroying existing domains,
2647  * and it will not create the default domain.
2648  *
2649  * Call with hotplug lock and sched_domains_mutex held
2650  */
partition_sched_domains_locked(int ndoms_new,cpumask_var_t doms_new[],struct sched_domain_attr * dattr_new)2651 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2652 				    struct sched_domain_attr *dattr_new)
2653 {
2654 	bool __maybe_unused has_eas = false;
2655 	int i, j, n;
2656 	int new_topology;
2657 
2658 	lockdep_assert_held(&sched_domains_mutex);
2659 
2660 	/* Let the architecture update CPU core mappings: */
2661 	new_topology = arch_update_cpu_topology();
2662 	/* Trigger rebuilding CPU capacity asymmetry data */
2663 	if (new_topology)
2664 		asym_cpu_capacity_scan();
2665 
2666 	if (!doms_new) {
2667 		WARN_ON_ONCE(dattr_new);
2668 		n = 0;
2669 		doms_new = alloc_sched_domains(1);
2670 		if (doms_new) {
2671 			n = 1;
2672 			cpumask_and(doms_new[0], cpu_active_mask,
2673 				    housekeeping_cpumask(HK_TYPE_DOMAIN));
2674 		}
2675 	} else {
2676 		n = ndoms_new;
2677 	}
2678 
2679 	/* Destroy deleted domains: */
2680 	for (i = 0; i < ndoms_cur; i++) {
2681 		for (j = 0; j < n && !new_topology; j++) {
2682 			if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2683 			    dattrs_equal(dattr_cur, i, dattr_new, j)) {
2684 				struct root_domain *rd;
2685 
2686 				/*
2687 				 * This domain won't be destroyed and as such
2688 				 * its dl_bw->total_bw needs to be cleared.  It
2689 				 * will be recomputed in function
2690 				 * update_tasks_root_domain().
2691 				 */
2692 				rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2693 				dl_clear_root_domain(rd);
2694 				goto match1;
2695 			}
2696 		}
2697 		/* No match - a current sched domain not in new doms_new[] */
2698 		detach_destroy_domains(doms_cur[i]);
2699 match1:
2700 		;
2701 	}
2702 
2703 	n = ndoms_cur;
2704 	if (!doms_new) {
2705 		n = 0;
2706 		doms_new = &fallback_doms;
2707 		cpumask_and(doms_new[0], cpu_active_mask,
2708 			    housekeeping_cpumask(HK_TYPE_DOMAIN));
2709 	}
2710 
2711 	/* Build new domains: */
2712 	for (i = 0; i < ndoms_new; i++) {
2713 		for (j = 0; j < n && !new_topology; j++) {
2714 			if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2715 			    dattrs_equal(dattr_new, i, dattr_cur, j))
2716 				goto match2;
2717 		}
2718 		/* No match - add a new doms_new */
2719 		build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2720 match2:
2721 		;
2722 	}
2723 
2724 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2725 	/* Build perf. domains: */
2726 	for (i = 0; i < ndoms_new; i++) {
2727 		for (j = 0; j < n && !sched_energy_update; j++) {
2728 			if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2729 			    cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2730 				has_eas = true;
2731 				goto match3;
2732 			}
2733 		}
2734 		/* No match - add perf. domains for a new rd */
2735 		has_eas |= build_perf_domains(doms_new[i]);
2736 match3:
2737 		;
2738 	}
2739 	sched_energy_set(has_eas);
2740 #endif
2741 
2742 	/* Remember the new sched domains: */
2743 	if (doms_cur != &fallback_doms)
2744 		free_sched_domains(doms_cur, ndoms_cur);
2745 
2746 	kfree(dattr_cur);
2747 	doms_cur = doms_new;
2748 	dattr_cur = dattr_new;
2749 	ndoms_cur = ndoms_new;
2750 
2751 	update_sched_domain_debugfs();
2752 }
2753 
2754 /*
2755  * Call with hotplug lock held
2756  */
partition_sched_domains(int ndoms_new,cpumask_var_t doms_new[],struct sched_domain_attr * dattr_new)2757 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2758 			     struct sched_domain_attr *dattr_new)
2759 {
2760 	mutex_lock(&sched_domains_mutex);
2761 	partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2762 	mutex_unlock(&sched_domains_mutex);
2763 }
2764