Loading...
Note: File does not exist in v3.5.6.
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Scheduler topology setup/handling methods
4 */
5
6DEFINE_MUTEX(sched_domains_mutex);
7
8/* Protected by sched_domains_mutex: */
9static cpumask_var_t sched_domains_tmpmask;
10static cpumask_var_t sched_domains_tmpmask2;
11
12#ifdef CONFIG_SCHED_DEBUG
13
14static int __init sched_debug_setup(char *str)
15{
16 sched_debug_verbose = true;
17
18 return 0;
19}
20early_param("sched_verbose", sched_debug_setup);
21
22static inline bool sched_debug(void)
23{
24 return sched_debug_verbose;
25}
26
27#define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
28const struct sd_flag_debug sd_flag_debug[] = {
29#include <linux/sched/sd_flags.h>
30};
31#undef SD_FLAG
32
33static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
34 struct cpumask *groupmask)
35{
36 struct sched_group *group = sd->groups;
37 unsigned long flags = sd->flags;
38 unsigned int idx;
39
40 cpumask_clear(groupmask);
41
42 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
43 printk(KERN_CONT "span=%*pbl level=%s\n",
44 cpumask_pr_args(sched_domain_span(sd)), sd->name);
45
46 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
47 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
48 }
49 if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
50 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
51 }
52
53 for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
54 unsigned int flag = BIT(idx);
55 unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
56
57 if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
58 !(sd->child->flags & flag))
59 printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
60 sd_flag_debug[idx].name);
61
62 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
63 !(sd->parent->flags & flag))
64 printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
65 sd_flag_debug[idx].name);
66 }
67
68 printk(KERN_DEBUG "%*s groups:", level + 1, "");
69 do {
70 if (!group) {
71 printk("\n");
72 printk(KERN_ERR "ERROR: group is NULL\n");
73 break;
74 }
75
76 if (cpumask_empty(sched_group_span(group))) {
77 printk(KERN_CONT "\n");
78 printk(KERN_ERR "ERROR: empty group\n");
79 break;
80 }
81
82 if (!(sd->flags & SD_OVERLAP) &&
83 cpumask_intersects(groupmask, sched_group_span(group))) {
84 printk(KERN_CONT "\n");
85 printk(KERN_ERR "ERROR: repeated CPUs\n");
86 break;
87 }
88
89 cpumask_or(groupmask, groupmask, sched_group_span(group));
90
91 printk(KERN_CONT " %d:{ span=%*pbl",
92 group->sgc->id,
93 cpumask_pr_args(sched_group_span(group)));
94
95 if ((sd->flags & SD_OVERLAP) &&
96 !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
97 printk(KERN_CONT " mask=%*pbl",
98 cpumask_pr_args(group_balance_mask(group)));
99 }
100
101 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
102 printk(KERN_CONT " cap=%lu", group->sgc->capacity);
103
104 if (group == sd->groups && sd->child &&
105 !cpumask_equal(sched_domain_span(sd->child),
106 sched_group_span(group))) {
107 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
108 }
109
110 printk(KERN_CONT " }");
111
112 group = group->next;
113
114 if (group != sd->groups)
115 printk(KERN_CONT ",");
116
117 } while (group != sd->groups);
118 printk(KERN_CONT "\n");
119
120 if (!cpumask_equal(sched_domain_span(sd), groupmask))
121 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
122
123 if (sd->parent &&
124 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
125 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
126 return 0;
127}
128
129static void sched_domain_debug(struct sched_domain *sd, int cpu)
130{
131 int level = 0;
132
133 if (!sched_debug_verbose)
134 return;
135
136 if (!sd) {
137 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
138 return;
139 }
140
141 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
142
143 for (;;) {
144 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
145 break;
146 level++;
147 sd = sd->parent;
148 if (!sd)
149 break;
150 }
151}
152#else /* !CONFIG_SCHED_DEBUG */
153
154# define sched_debug_verbose 0
155# define sched_domain_debug(sd, cpu) do { } while (0)
156static inline bool sched_debug(void)
157{
158 return false;
159}
160#endif /* CONFIG_SCHED_DEBUG */
161
162/* Generate a mask of SD flags with the SDF_NEEDS_GROUPS metaflag */
163#define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
164static const unsigned int SD_DEGENERATE_GROUPS_MASK =
165#include <linux/sched/sd_flags.h>
1660;
167#undef SD_FLAG
168
169static int sd_degenerate(struct sched_domain *sd)
170{
171 if (cpumask_weight(sched_domain_span(sd)) == 1)
172 return 1;
173
174 /* Following flags need at least 2 groups */
175 if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
176 (sd->groups != sd->groups->next))
177 return 0;
178
179 /* Following flags don't use groups */
180 if (sd->flags & (SD_WAKE_AFFINE))
181 return 0;
182
183 return 1;
184}
185
186static int
187sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
188{
189 unsigned long cflags = sd->flags, pflags = parent->flags;
190
191 if (sd_degenerate(parent))
192 return 1;
193
194 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
195 return 0;
196
197 /* Flags needing groups don't count if only 1 group in parent */
198 if (parent->groups == parent->groups->next)
199 pflags &= ~SD_DEGENERATE_GROUPS_MASK;
200
201 if (~cflags & pflags)
202 return 0;
203
204 return 1;
205}
206
207#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
208DEFINE_STATIC_KEY_FALSE(sched_energy_present);
209static unsigned int sysctl_sched_energy_aware = 1;
210DEFINE_MUTEX(sched_energy_mutex);
211bool sched_energy_update;
212
213void rebuild_sched_domains_energy(void)
214{
215 mutex_lock(&sched_energy_mutex);
216 sched_energy_update = true;
217 rebuild_sched_domains();
218 sched_energy_update = false;
219 mutex_unlock(&sched_energy_mutex);
220}
221
222#ifdef CONFIG_PROC_SYSCTL
223static int sched_energy_aware_handler(struct ctl_table *table, int write,
224 void *buffer, size_t *lenp, loff_t *ppos)
225{
226 int ret, state;
227
228 if (write && !capable(CAP_SYS_ADMIN))
229 return -EPERM;
230
231 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
232 if (!ret && write) {
233 state = static_branch_unlikely(&sched_energy_present);
234 if (state != sysctl_sched_energy_aware)
235 rebuild_sched_domains_energy();
236 }
237
238 return ret;
239}
240
241static struct ctl_table sched_energy_aware_sysctls[] = {
242 {
243 .procname = "sched_energy_aware",
244 .data = &sysctl_sched_energy_aware,
245 .maxlen = sizeof(unsigned int),
246 .mode = 0644,
247 .proc_handler = sched_energy_aware_handler,
248 .extra1 = SYSCTL_ZERO,
249 .extra2 = SYSCTL_ONE,
250 },
251 {}
252};
253
254static int __init sched_energy_aware_sysctl_init(void)
255{
256 register_sysctl_init("kernel", sched_energy_aware_sysctls);
257 return 0;
258}
259
260late_initcall(sched_energy_aware_sysctl_init);
261#endif
262
263static void free_pd(struct perf_domain *pd)
264{
265 struct perf_domain *tmp;
266
267 while (pd) {
268 tmp = pd->next;
269 kfree(pd);
270 pd = tmp;
271 }
272}
273
274static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
275{
276 while (pd) {
277 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
278 return pd;
279 pd = pd->next;
280 }
281
282 return NULL;
283}
284
285static struct perf_domain *pd_init(int cpu)
286{
287 struct em_perf_domain *obj = em_cpu_get(cpu);
288 struct perf_domain *pd;
289
290 if (!obj) {
291 if (sched_debug())
292 pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
293 return NULL;
294 }
295
296 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
297 if (!pd)
298 return NULL;
299 pd->em_pd = obj;
300
301 return pd;
302}
303
304static void perf_domain_debug(const struct cpumask *cpu_map,
305 struct perf_domain *pd)
306{
307 if (!sched_debug() || !pd)
308 return;
309
310 printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
311
312 while (pd) {
313 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
314 cpumask_first(perf_domain_span(pd)),
315 cpumask_pr_args(perf_domain_span(pd)),
316 em_pd_nr_perf_states(pd->em_pd));
317 pd = pd->next;
318 }
319
320 printk(KERN_CONT "\n");
321}
322
323static void destroy_perf_domain_rcu(struct rcu_head *rp)
324{
325 struct perf_domain *pd;
326
327 pd = container_of(rp, struct perf_domain, rcu);
328 free_pd(pd);
329}
330
331static void sched_energy_set(bool has_eas)
332{
333 if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
334 if (sched_debug())
335 pr_info("%s: stopping EAS\n", __func__);
336 static_branch_disable_cpuslocked(&sched_energy_present);
337 } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
338 if (sched_debug())
339 pr_info("%s: starting EAS\n", __func__);
340 static_branch_enable_cpuslocked(&sched_energy_present);
341 }
342}
343
344/*
345 * EAS can be used on a root domain if it meets all the following conditions:
346 * 1. an Energy Model (EM) is available;
347 * 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
348 * 3. no SMT is detected.
349 * 4. the EM complexity is low enough to keep scheduling overheads low;
350 * 5. schedutil is driving the frequency of all CPUs of the rd;
351 * 6. frequency invariance support is present;
352 *
353 * The complexity of the Energy Model is defined as:
354 *
355 * C = nr_pd * (nr_cpus + nr_ps)
356 *
357 * with parameters defined as:
358 * - nr_pd: the number of performance domains
359 * - nr_cpus: the number of CPUs
360 * - nr_ps: the sum of the number of performance states of all performance
361 * domains (for example, on a system with 2 performance domains,
362 * with 10 performance states each, nr_ps = 2 * 10 = 20).
363 *
364 * It is generally not a good idea to use such a model in the wake-up path on
365 * very complex platforms because of the associated scheduling overheads. The
366 * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
367 * with per-CPU DVFS and less than 8 performance states each, for example.
368 */
369#define EM_MAX_COMPLEXITY 2048
370
371extern struct cpufreq_governor schedutil_gov;
372static bool build_perf_domains(const struct cpumask *cpu_map)
373{
374 int i, nr_pd = 0, nr_ps = 0, nr_cpus = cpumask_weight(cpu_map);
375 struct perf_domain *pd = NULL, *tmp;
376 int cpu = cpumask_first(cpu_map);
377 struct root_domain *rd = cpu_rq(cpu)->rd;
378 struct cpufreq_policy *policy;
379 struct cpufreq_governor *gov;
380
381 if (!sysctl_sched_energy_aware)
382 goto free;
383
384 /* EAS is enabled for asymmetric CPU capacity topologies. */
385 if (!per_cpu(sd_asym_cpucapacity, cpu)) {
386 if (sched_debug()) {
387 pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
388 cpumask_pr_args(cpu_map));
389 }
390 goto free;
391 }
392
393 /* EAS definitely does *not* handle SMT */
394 if (sched_smt_active()) {
395 pr_warn("rd %*pbl: Disabling EAS, SMT is not supported\n",
396 cpumask_pr_args(cpu_map));
397 goto free;
398 }
399
400 if (!arch_scale_freq_invariant()) {
401 if (sched_debug()) {
402 pr_warn("rd %*pbl: Disabling EAS: frequency-invariant load tracking not yet supported",
403 cpumask_pr_args(cpu_map));
404 }
405 goto free;
406 }
407
408 for_each_cpu(i, cpu_map) {
409 /* Skip already covered CPUs. */
410 if (find_pd(pd, i))
411 continue;
412
413 /* Do not attempt EAS if schedutil is not being used. */
414 policy = cpufreq_cpu_get(i);
415 if (!policy)
416 goto free;
417 gov = policy->governor;
418 cpufreq_cpu_put(policy);
419 if (gov != &schedutil_gov) {
420 if (rd->pd)
421 pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
422 cpumask_pr_args(cpu_map));
423 goto free;
424 }
425
426 /* Create the new pd and add it to the local list. */
427 tmp = pd_init(i);
428 if (!tmp)
429 goto free;
430 tmp->next = pd;
431 pd = tmp;
432
433 /*
434 * Count performance domains and performance states for the
435 * complexity check.
436 */
437 nr_pd++;
438 nr_ps += em_pd_nr_perf_states(pd->em_pd);
439 }
440
441 /* Bail out if the Energy Model complexity is too high. */
442 if (nr_pd * (nr_ps + nr_cpus) > EM_MAX_COMPLEXITY) {
443 WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
444 cpumask_pr_args(cpu_map));
445 goto free;
446 }
447
448 perf_domain_debug(cpu_map, pd);
449
450 /* Attach the new list of performance domains to the root domain. */
451 tmp = rd->pd;
452 rcu_assign_pointer(rd->pd, pd);
453 if (tmp)
454 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
455
456 return !!pd;
457
458free:
459 free_pd(pd);
460 tmp = rd->pd;
461 rcu_assign_pointer(rd->pd, NULL);
462 if (tmp)
463 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
464
465 return false;
466}
467#else
468static void free_pd(struct perf_domain *pd) { }
469#endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
470
471static void free_rootdomain(struct rcu_head *rcu)
472{
473 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
474
475 cpupri_cleanup(&rd->cpupri);
476 cpudl_cleanup(&rd->cpudl);
477 free_cpumask_var(rd->dlo_mask);
478 free_cpumask_var(rd->rto_mask);
479 free_cpumask_var(rd->online);
480 free_cpumask_var(rd->span);
481 free_pd(rd->pd);
482 kfree(rd);
483}
484
485void rq_attach_root(struct rq *rq, struct root_domain *rd)
486{
487 struct root_domain *old_rd = NULL;
488 unsigned long flags;
489
490 raw_spin_rq_lock_irqsave(rq, flags);
491
492 if (rq->rd) {
493 old_rd = rq->rd;
494
495 if (cpumask_test_cpu(rq->cpu, old_rd->online))
496 set_rq_offline(rq);
497
498 cpumask_clear_cpu(rq->cpu, old_rd->span);
499
500 /*
501 * If we dont want to free the old_rd yet then
502 * set old_rd to NULL to skip the freeing later
503 * in this function:
504 */
505 if (!atomic_dec_and_test(&old_rd->refcount))
506 old_rd = NULL;
507 }
508
509 atomic_inc(&rd->refcount);
510 rq->rd = rd;
511
512 cpumask_set_cpu(rq->cpu, rd->span);
513 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
514 set_rq_online(rq);
515
516 raw_spin_rq_unlock_irqrestore(rq, flags);
517
518 if (old_rd)
519 call_rcu(&old_rd->rcu, free_rootdomain);
520}
521
522void sched_get_rd(struct root_domain *rd)
523{
524 atomic_inc(&rd->refcount);
525}
526
527void sched_put_rd(struct root_domain *rd)
528{
529 if (!atomic_dec_and_test(&rd->refcount))
530 return;
531
532 call_rcu(&rd->rcu, free_rootdomain);
533}
534
535static int init_rootdomain(struct root_domain *rd)
536{
537 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
538 goto out;
539 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
540 goto free_span;
541 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
542 goto free_online;
543 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
544 goto free_dlo_mask;
545
546#ifdef HAVE_RT_PUSH_IPI
547 rd->rto_cpu = -1;
548 raw_spin_lock_init(&rd->rto_lock);
549 rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
550#endif
551
552 rd->visit_gen = 0;
553 init_dl_bw(&rd->dl_bw);
554 if (cpudl_init(&rd->cpudl) != 0)
555 goto free_rto_mask;
556
557 if (cpupri_init(&rd->cpupri) != 0)
558 goto free_cpudl;
559 return 0;
560
561free_cpudl:
562 cpudl_cleanup(&rd->cpudl);
563free_rto_mask:
564 free_cpumask_var(rd->rto_mask);
565free_dlo_mask:
566 free_cpumask_var(rd->dlo_mask);
567free_online:
568 free_cpumask_var(rd->online);
569free_span:
570 free_cpumask_var(rd->span);
571out:
572 return -ENOMEM;
573}
574
575/*
576 * By default the system creates a single root-domain with all CPUs as
577 * members (mimicking the global state we have today).
578 */
579struct root_domain def_root_domain;
580
581void init_defrootdomain(void)
582{
583 init_rootdomain(&def_root_domain);
584
585 atomic_set(&def_root_domain.refcount, 1);
586}
587
588static struct root_domain *alloc_rootdomain(void)
589{
590 struct root_domain *rd;
591
592 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
593 if (!rd)
594 return NULL;
595
596 if (init_rootdomain(rd) != 0) {
597 kfree(rd);
598 return NULL;
599 }
600
601 return rd;
602}
603
604static void free_sched_groups(struct sched_group *sg, int free_sgc)
605{
606 struct sched_group *tmp, *first;
607
608 if (!sg)
609 return;
610
611 first = sg;
612 do {
613 tmp = sg->next;
614
615 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
616 kfree(sg->sgc);
617
618 if (atomic_dec_and_test(&sg->ref))
619 kfree(sg);
620 sg = tmp;
621 } while (sg != first);
622}
623
624static void destroy_sched_domain(struct sched_domain *sd)
625{
626 /*
627 * A normal sched domain may have multiple group references, an
628 * overlapping domain, having private groups, only one. Iterate,
629 * dropping group/capacity references, freeing where none remain.
630 */
631 free_sched_groups(sd->groups, 1);
632
633 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
634 kfree(sd->shared);
635 kfree(sd);
636}
637
638static void destroy_sched_domains_rcu(struct rcu_head *rcu)
639{
640 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
641
642 while (sd) {
643 struct sched_domain *parent = sd->parent;
644 destroy_sched_domain(sd);
645 sd = parent;
646 }
647}
648
649static void destroy_sched_domains(struct sched_domain *sd)
650{
651 if (sd)
652 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
653}
654
655/*
656 * Keep a special pointer to the highest sched_domain that has
657 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
658 * allows us to avoid some pointer chasing select_idle_sibling().
659 *
660 * Also keep a unique ID per domain (we use the first CPU number in
661 * the cpumask of the domain), this allows us to quickly tell if
662 * two CPUs are in the same cache domain, see cpus_share_cache().
663 */
664DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
665DEFINE_PER_CPU(int, sd_llc_size);
666DEFINE_PER_CPU(int, sd_llc_id);
667DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
668DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
669DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
670DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
671DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
672
673static void update_top_cache_domain(int cpu)
674{
675 struct sched_domain_shared *sds = NULL;
676 struct sched_domain *sd;
677 int id = cpu;
678 int size = 1;
679
680 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
681 if (sd) {
682 id = cpumask_first(sched_domain_span(sd));
683 size = cpumask_weight(sched_domain_span(sd));
684 sds = sd->shared;
685 }
686
687 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
688 per_cpu(sd_llc_size, cpu) = size;
689 per_cpu(sd_llc_id, cpu) = id;
690 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
691
692 sd = lowest_flag_domain(cpu, SD_NUMA);
693 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
694
695 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
696 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
697
698 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
699 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
700}
701
702/*
703 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
704 * hold the hotplug lock.
705 */
706static void
707cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
708{
709 struct rq *rq = cpu_rq(cpu);
710 struct sched_domain *tmp;
711
712 /* Remove the sched domains which do not contribute to scheduling. */
713 for (tmp = sd; tmp; ) {
714 struct sched_domain *parent = tmp->parent;
715 if (!parent)
716 break;
717
718 if (sd_parent_degenerate(tmp, parent)) {
719 tmp->parent = parent->parent;
720 if (parent->parent)
721 parent->parent->child = tmp;
722 /*
723 * Transfer SD_PREFER_SIBLING down in case of a
724 * degenerate parent; the spans match for this
725 * so the property transfers.
726 */
727 if (parent->flags & SD_PREFER_SIBLING)
728 tmp->flags |= SD_PREFER_SIBLING;
729 destroy_sched_domain(parent);
730 } else
731 tmp = tmp->parent;
732 }
733
734 if (sd && sd_degenerate(sd)) {
735 tmp = sd;
736 sd = sd->parent;
737 destroy_sched_domain(tmp);
738 if (sd) {
739 struct sched_group *sg = sd->groups;
740
741 /*
742 * sched groups hold the flags of the child sched
743 * domain for convenience. Clear such flags since
744 * the child is being destroyed.
745 */
746 do {
747 sg->flags = 0;
748 } while (sg != sd->groups);
749
750 sd->child = NULL;
751 }
752 }
753
754 sched_domain_debug(sd, cpu);
755
756 rq_attach_root(rq, rd);
757 tmp = rq->sd;
758 rcu_assign_pointer(rq->sd, sd);
759 dirty_sched_domain_sysctl(cpu);
760 destroy_sched_domains(tmp);
761
762 update_top_cache_domain(cpu);
763}
764
765struct s_data {
766 struct sched_domain * __percpu *sd;
767 struct root_domain *rd;
768};
769
770enum s_alloc {
771 sa_rootdomain,
772 sa_sd,
773 sa_sd_storage,
774 sa_none,
775};
776
777/*
778 * Return the canonical balance CPU for this group, this is the first CPU
779 * of this group that's also in the balance mask.
780 *
781 * The balance mask are all those CPUs that could actually end up at this
782 * group. See build_balance_mask().
783 *
784 * Also see should_we_balance().
785 */
786int group_balance_cpu(struct sched_group *sg)
787{
788 return cpumask_first(group_balance_mask(sg));
789}
790
791
792/*
793 * NUMA topology (first read the regular topology blurb below)
794 *
795 * Given a node-distance table, for example:
796 *
797 * node 0 1 2 3
798 * 0: 10 20 30 20
799 * 1: 20 10 20 30
800 * 2: 30 20 10 20
801 * 3: 20 30 20 10
802 *
803 * which represents a 4 node ring topology like:
804 *
805 * 0 ----- 1
806 * | |
807 * | |
808 * | |
809 * 3 ----- 2
810 *
811 * We want to construct domains and groups to represent this. The way we go
812 * about doing this is to build the domains on 'hops'. For each NUMA level we
813 * construct the mask of all nodes reachable in @level hops.
814 *
815 * For the above NUMA topology that gives 3 levels:
816 *
817 * NUMA-2 0-3 0-3 0-3 0-3
818 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
819 *
820 * NUMA-1 0-1,3 0-2 1-3 0,2-3
821 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
822 *
823 * NUMA-0 0 1 2 3
824 *
825 *
826 * As can be seen; things don't nicely line up as with the regular topology.
827 * When we iterate a domain in child domain chunks some nodes can be
828 * represented multiple times -- hence the "overlap" naming for this part of
829 * the topology.
830 *
831 * In order to minimize this overlap, we only build enough groups to cover the
832 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
833 *
834 * Because:
835 *
836 * - the first group of each domain is its child domain; this
837 * gets us the first 0-1,3
838 * - the only uncovered node is 2, who's child domain is 1-3.
839 *
840 * However, because of the overlap, computing a unique CPU for each group is
841 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
842 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
843 * end up at those groups (they would end up in group: 0-1,3).
844 *
845 * To correct this we have to introduce the group balance mask. This mask
846 * will contain those CPUs in the group that can reach this group given the
847 * (child) domain tree.
848 *
849 * With this we can once again compute balance_cpu and sched_group_capacity
850 * relations.
851 *
852 * XXX include words on how balance_cpu is unique and therefore can be
853 * used for sched_group_capacity links.
854 *
855 *
856 * Another 'interesting' topology is:
857 *
858 * node 0 1 2 3
859 * 0: 10 20 20 30
860 * 1: 20 10 20 20
861 * 2: 20 20 10 20
862 * 3: 30 20 20 10
863 *
864 * Which looks a little like:
865 *
866 * 0 ----- 1
867 * | / |
868 * | / |
869 * | / |
870 * 2 ----- 3
871 *
872 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
873 * are not.
874 *
875 * This leads to a few particularly weird cases where the sched_domain's are
876 * not of the same number for each CPU. Consider:
877 *
878 * NUMA-2 0-3 0-3
879 * groups: {0-2},{1-3} {1-3},{0-2}
880 *
881 * NUMA-1 0-2 0-3 0-3 1-3
882 *
883 * NUMA-0 0 1 2 3
884 *
885 */
886
887
888/*
889 * Build the balance mask; it contains only those CPUs that can arrive at this
890 * group and should be considered to continue balancing.
891 *
892 * We do this during the group creation pass, therefore the group information
893 * isn't complete yet, however since each group represents a (child) domain we
894 * can fully construct this using the sched_domain bits (which are already
895 * complete).
896 */
897static void
898build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
899{
900 const struct cpumask *sg_span = sched_group_span(sg);
901 struct sd_data *sdd = sd->private;
902 struct sched_domain *sibling;
903 int i;
904
905 cpumask_clear(mask);
906
907 for_each_cpu(i, sg_span) {
908 sibling = *per_cpu_ptr(sdd->sd, i);
909
910 /*
911 * Can happen in the asymmetric case, where these siblings are
912 * unused. The mask will not be empty because those CPUs that
913 * do have the top domain _should_ span the domain.
914 */
915 if (!sibling->child)
916 continue;
917
918 /* If we would not end up here, we can't continue from here */
919 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
920 continue;
921
922 cpumask_set_cpu(i, mask);
923 }
924
925 /* We must not have empty masks here */
926 WARN_ON_ONCE(cpumask_empty(mask));
927}
928
929/*
930 * XXX: This creates per-node group entries; since the load-balancer will
931 * immediately access remote memory to construct this group's load-balance
932 * statistics having the groups node local is of dubious benefit.
933 */
934static struct sched_group *
935build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
936{
937 struct sched_group *sg;
938 struct cpumask *sg_span;
939
940 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
941 GFP_KERNEL, cpu_to_node(cpu));
942
943 if (!sg)
944 return NULL;
945
946 sg_span = sched_group_span(sg);
947 if (sd->child) {
948 cpumask_copy(sg_span, sched_domain_span(sd->child));
949 sg->flags = sd->child->flags;
950 } else {
951 cpumask_copy(sg_span, sched_domain_span(sd));
952 }
953
954 atomic_inc(&sg->ref);
955 return sg;
956}
957
958static void init_overlap_sched_group(struct sched_domain *sd,
959 struct sched_group *sg)
960{
961 struct cpumask *mask = sched_domains_tmpmask2;
962 struct sd_data *sdd = sd->private;
963 struct cpumask *sg_span;
964 int cpu;
965
966 build_balance_mask(sd, sg, mask);
967 cpu = cpumask_first(mask);
968
969 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
970 if (atomic_inc_return(&sg->sgc->ref) == 1)
971 cpumask_copy(group_balance_mask(sg), mask);
972 else
973 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
974
975 /*
976 * Initialize sgc->capacity such that even if we mess up the
977 * domains and no possible iteration will get us here, we won't
978 * die on a /0 trap.
979 */
980 sg_span = sched_group_span(sg);
981 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
982 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
983 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
984}
985
986static struct sched_domain *
987find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
988{
989 /*
990 * The proper descendant would be the one whose child won't span out
991 * of sd
992 */
993 while (sibling->child &&
994 !cpumask_subset(sched_domain_span(sibling->child),
995 sched_domain_span(sd)))
996 sibling = sibling->child;
997
998 /*
999 * As we are referencing sgc across different topology level, we need
1000 * to go down to skip those sched_domains which don't contribute to
1001 * scheduling because they will be degenerated in cpu_attach_domain
1002 */
1003 while (sibling->child &&
1004 cpumask_equal(sched_domain_span(sibling->child),
1005 sched_domain_span(sibling)))
1006 sibling = sibling->child;
1007
1008 return sibling;
1009}
1010
1011static int
1012build_overlap_sched_groups(struct sched_domain *sd, int cpu)
1013{
1014 struct sched_group *first = NULL, *last = NULL, *sg;
1015 const struct cpumask *span = sched_domain_span(sd);
1016 struct cpumask *covered = sched_domains_tmpmask;
1017 struct sd_data *sdd = sd->private;
1018 struct sched_domain *sibling;
1019 int i;
1020
1021 cpumask_clear(covered);
1022
1023 for_each_cpu_wrap(i, span, cpu) {
1024 struct cpumask *sg_span;
1025
1026 if (cpumask_test_cpu(i, covered))
1027 continue;
1028
1029 sibling = *per_cpu_ptr(sdd->sd, i);
1030
1031 /*
1032 * Asymmetric node setups can result in situations where the
1033 * domain tree is of unequal depth, make sure to skip domains
1034 * that already cover the entire range.
1035 *
1036 * In that case build_sched_domains() will have terminated the
1037 * iteration early and our sibling sd spans will be empty.
1038 * Domains should always include the CPU they're built on, so
1039 * check that.
1040 */
1041 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1042 continue;
1043
1044 /*
1045 * Usually we build sched_group by sibling's child sched_domain
1046 * But for machines whose NUMA diameter are 3 or above, we move
1047 * to build sched_group by sibling's proper descendant's child
1048 * domain because sibling's child sched_domain will span out of
1049 * the sched_domain being built as below.
1050 *
1051 * Smallest diameter=3 topology is:
1052 *
1053 * node 0 1 2 3
1054 * 0: 10 20 30 40
1055 * 1: 20 10 20 30
1056 * 2: 30 20 10 20
1057 * 3: 40 30 20 10
1058 *
1059 * 0 --- 1 --- 2 --- 3
1060 *
1061 * NUMA-3 0-3 N/A N/A 0-3
1062 * groups: {0-2},{1-3} {1-3},{0-2}
1063 *
1064 * NUMA-2 0-2 0-3 0-3 1-3
1065 * groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2}
1066 *
1067 * NUMA-1 0-1 0-2 1-3 2-3
1068 * groups: {0},{1} {1},{2},{0} {2},{3},{1} {3},{2}
1069 *
1070 * NUMA-0 0 1 2 3
1071 *
1072 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1073 * group span isn't a subset of the domain span.
1074 */
1075 if (sibling->child &&
1076 !cpumask_subset(sched_domain_span(sibling->child), span))
1077 sibling = find_descended_sibling(sd, sibling);
1078
1079 sg = build_group_from_child_sched_domain(sibling, cpu);
1080 if (!sg)
1081 goto fail;
1082
1083 sg_span = sched_group_span(sg);
1084 cpumask_or(covered, covered, sg_span);
1085
1086 init_overlap_sched_group(sibling, sg);
1087
1088 if (!first)
1089 first = sg;
1090 if (last)
1091 last->next = sg;
1092 last = sg;
1093 last->next = first;
1094 }
1095 sd->groups = first;
1096
1097 return 0;
1098
1099fail:
1100 free_sched_groups(first, 0);
1101
1102 return -ENOMEM;
1103}
1104
1105
1106/*
1107 * Package topology (also see the load-balance blurb in fair.c)
1108 *
1109 * The scheduler builds a tree structure to represent a number of important
1110 * topology features. By default (default_topology[]) these include:
1111 *
1112 * - Simultaneous multithreading (SMT)
1113 * - Multi-Core Cache (MC)
1114 * - Package (DIE)
1115 *
1116 * Where the last one more or less denotes everything up to a NUMA node.
1117 *
1118 * The tree consists of 3 primary data structures:
1119 *
1120 * sched_domain -> sched_group -> sched_group_capacity
1121 * ^ ^ ^ ^
1122 * `-' `-'
1123 *
1124 * The sched_domains are per-CPU and have a two way link (parent & child) and
1125 * denote the ever growing mask of CPUs belonging to that level of topology.
1126 *
1127 * Each sched_domain has a circular (double) linked list of sched_group's, each
1128 * denoting the domains of the level below (or individual CPUs in case of the
1129 * first domain level). The sched_group linked by a sched_domain includes the
1130 * CPU of that sched_domain [*].
1131 *
1132 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1133 *
1134 * CPU 0 1 2 3 4 5 6 7
1135 *
1136 * DIE [ ]
1137 * MC [ ] [ ]
1138 * SMT [ ] [ ] [ ] [ ]
1139 *
1140 * - or -
1141 *
1142 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1143 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1144 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1145 *
1146 * CPU 0 1 2 3 4 5 6 7
1147 *
1148 * One way to think about it is: sched_domain moves you up and down among these
1149 * topology levels, while sched_group moves you sideways through it, at child
1150 * domain granularity.
1151 *
1152 * sched_group_capacity ensures each unique sched_group has shared storage.
1153 *
1154 * There are two related construction problems, both require a CPU that
1155 * uniquely identify each group (for a given domain):
1156 *
1157 * - The first is the balance_cpu (see should_we_balance() and the
1158 * load-balance blub in fair.c); for each group we only want 1 CPU to
1159 * continue balancing at a higher domain.
1160 *
1161 * - The second is the sched_group_capacity; we want all identical groups
1162 * to share a single sched_group_capacity.
1163 *
1164 * Since these topologies are exclusive by construction. That is, its
1165 * impossible for an SMT thread to belong to multiple cores, and cores to
1166 * be part of multiple caches. There is a very clear and unique location
1167 * for each CPU in the hierarchy.
1168 *
1169 * Therefore computing a unique CPU for each group is trivial (the iteration
1170 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1171 * group), we can simply pick the first CPU in each group.
1172 *
1173 *
1174 * [*] in other words, the first group of each domain is its child domain.
1175 */
1176
1177static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1178{
1179 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1180 struct sched_domain *child = sd->child;
1181 struct sched_group *sg;
1182 bool already_visited;
1183
1184 if (child)
1185 cpu = cpumask_first(sched_domain_span(child));
1186
1187 sg = *per_cpu_ptr(sdd->sg, cpu);
1188 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1189
1190 /* Increase refcounts for claim_allocations: */
1191 already_visited = atomic_inc_return(&sg->ref) > 1;
1192 /* sgc visits should follow a similar trend as sg */
1193 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1194
1195 /* If we have already visited that group, it's already initialized. */
1196 if (already_visited)
1197 return sg;
1198
1199 if (child) {
1200 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1201 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1202 sg->flags = child->flags;
1203 } else {
1204 cpumask_set_cpu(cpu, sched_group_span(sg));
1205 cpumask_set_cpu(cpu, group_balance_mask(sg));
1206 }
1207
1208 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1209 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1210 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1211
1212 return sg;
1213}
1214
1215/*
1216 * build_sched_groups will build a circular linked list of the groups
1217 * covered by the given span, will set each group's ->cpumask correctly,
1218 * and will initialize their ->sgc.
1219 *
1220 * Assumes the sched_domain tree is fully constructed
1221 */
1222static int
1223build_sched_groups(struct sched_domain *sd, int cpu)
1224{
1225 struct sched_group *first = NULL, *last = NULL;
1226 struct sd_data *sdd = sd->private;
1227 const struct cpumask *span = sched_domain_span(sd);
1228 struct cpumask *covered;
1229 int i;
1230
1231 lockdep_assert_held(&sched_domains_mutex);
1232 covered = sched_domains_tmpmask;
1233
1234 cpumask_clear(covered);
1235
1236 for_each_cpu_wrap(i, span, cpu) {
1237 struct sched_group *sg;
1238
1239 if (cpumask_test_cpu(i, covered))
1240 continue;
1241
1242 sg = get_group(i, sdd);
1243
1244 cpumask_or(covered, covered, sched_group_span(sg));
1245
1246 if (!first)
1247 first = sg;
1248 if (last)
1249 last->next = sg;
1250 last = sg;
1251 }
1252 last->next = first;
1253 sd->groups = first;
1254
1255 return 0;
1256}
1257
1258/*
1259 * Initialize sched groups cpu_capacity.
1260 *
1261 * cpu_capacity indicates the capacity of sched group, which is used while
1262 * distributing the load between different sched groups in a sched domain.
1263 * Typically cpu_capacity for all the groups in a sched domain will be same
1264 * unless there are asymmetries in the topology. If there are asymmetries,
1265 * group having more cpu_capacity will pickup more load compared to the
1266 * group having less cpu_capacity.
1267 */
1268static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1269{
1270 struct sched_group *sg = sd->groups;
1271
1272 WARN_ON(!sg);
1273
1274 do {
1275 int cpu, max_cpu = -1;
1276
1277 sg->group_weight = cpumask_weight(sched_group_span(sg));
1278
1279 if (!(sd->flags & SD_ASYM_PACKING))
1280 goto next;
1281
1282 for_each_cpu(cpu, sched_group_span(sg)) {
1283 if (max_cpu < 0)
1284 max_cpu = cpu;
1285 else if (sched_asym_prefer(cpu, max_cpu))
1286 max_cpu = cpu;
1287 }
1288 sg->asym_prefer_cpu = max_cpu;
1289
1290next:
1291 sg = sg->next;
1292 } while (sg != sd->groups);
1293
1294 if (cpu != group_balance_cpu(sg))
1295 return;
1296
1297 update_group_capacity(sd, cpu);
1298}
1299
1300/*
1301 * Asymmetric CPU capacity bits
1302 */
1303struct asym_cap_data {
1304 struct list_head link;
1305 unsigned long capacity;
1306 unsigned long cpus[];
1307};
1308
1309/*
1310 * Set of available CPUs grouped by their corresponding capacities
1311 * Each list entry contains a CPU mask reflecting CPUs that share the same
1312 * capacity.
1313 * The lifespan of data is unlimited.
1314 */
1315static LIST_HEAD(asym_cap_list);
1316
1317#define cpu_capacity_span(asym_data) to_cpumask((asym_data)->cpus)
1318
1319/*
1320 * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1321 * Provides sd_flags reflecting the asymmetry scope.
1322 */
1323static inline int
1324asym_cpu_capacity_classify(const struct cpumask *sd_span,
1325 const struct cpumask *cpu_map)
1326{
1327 struct asym_cap_data *entry;
1328 int count = 0, miss = 0;
1329
1330 /*
1331 * Count how many unique CPU capacities this domain spans across
1332 * (compare sched_domain CPUs mask with ones representing available
1333 * CPUs capacities). Take into account CPUs that might be offline:
1334 * skip those.
1335 */
1336 list_for_each_entry(entry, &asym_cap_list, link) {
1337 if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1338 ++count;
1339 else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1340 ++miss;
1341 }
1342
1343 WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1344
1345 /* No asymmetry detected */
1346 if (count < 2)
1347 return 0;
1348 /* Some of the available CPU capacity values have not been detected */
1349 if (miss)
1350 return SD_ASYM_CPUCAPACITY;
1351
1352 /* Full asymmetry */
1353 return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1354
1355}
1356
1357static inline void asym_cpu_capacity_update_data(int cpu)
1358{
1359 unsigned long capacity = arch_scale_cpu_capacity(cpu);
1360 struct asym_cap_data *entry = NULL;
1361
1362 list_for_each_entry(entry, &asym_cap_list, link) {
1363 if (capacity == entry->capacity)
1364 goto done;
1365 }
1366
1367 entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1368 if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1369 return;
1370 entry->capacity = capacity;
1371 list_add(&entry->link, &asym_cap_list);
1372done:
1373 __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1374}
1375
1376/*
1377 * Build-up/update list of CPUs grouped by their capacities
1378 * An update requires explicit request to rebuild sched domains
1379 * with state indicating CPU topology changes.
1380 */
1381static void asym_cpu_capacity_scan(void)
1382{
1383 struct asym_cap_data *entry, *next;
1384 int cpu;
1385
1386 list_for_each_entry(entry, &asym_cap_list, link)
1387 cpumask_clear(cpu_capacity_span(entry));
1388
1389 for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1390 asym_cpu_capacity_update_data(cpu);
1391
1392 list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1393 if (cpumask_empty(cpu_capacity_span(entry))) {
1394 list_del(&entry->link);
1395 kfree(entry);
1396 }
1397 }
1398
1399 /*
1400 * Only one capacity value has been detected i.e. this system is symmetric.
1401 * No need to keep this data around.
1402 */
1403 if (list_is_singular(&asym_cap_list)) {
1404 entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1405 list_del(&entry->link);
1406 kfree(entry);
1407 }
1408}
1409
1410/*
1411 * Initializers for schedule domains
1412 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1413 */
1414
1415static int default_relax_domain_level = -1;
1416int sched_domain_level_max;
1417
1418static int __init setup_relax_domain_level(char *str)
1419{
1420 if (kstrtoint(str, 0, &default_relax_domain_level))
1421 pr_warn("Unable to set relax_domain_level\n");
1422
1423 return 1;
1424}
1425__setup("relax_domain_level=", setup_relax_domain_level);
1426
1427static void set_domain_attribute(struct sched_domain *sd,
1428 struct sched_domain_attr *attr)
1429{
1430 int request;
1431
1432 if (!attr || attr->relax_domain_level < 0) {
1433 if (default_relax_domain_level < 0)
1434 return;
1435 request = default_relax_domain_level;
1436 } else
1437 request = attr->relax_domain_level;
1438
1439 if (sd->level > request) {
1440 /* Turn off idle balance on this domain: */
1441 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1442 }
1443}
1444
1445static void __sdt_free(const struct cpumask *cpu_map);
1446static int __sdt_alloc(const struct cpumask *cpu_map);
1447
1448static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1449 const struct cpumask *cpu_map)
1450{
1451 switch (what) {
1452 case sa_rootdomain:
1453 if (!atomic_read(&d->rd->refcount))
1454 free_rootdomain(&d->rd->rcu);
1455 fallthrough;
1456 case sa_sd:
1457 free_percpu(d->sd);
1458 fallthrough;
1459 case sa_sd_storage:
1460 __sdt_free(cpu_map);
1461 fallthrough;
1462 case sa_none:
1463 break;
1464 }
1465}
1466
1467static enum s_alloc
1468__visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1469{
1470 memset(d, 0, sizeof(*d));
1471
1472 if (__sdt_alloc(cpu_map))
1473 return sa_sd_storage;
1474 d->sd = alloc_percpu(struct sched_domain *);
1475 if (!d->sd)
1476 return sa_sd_storage;
1477 d->rd = alloc_rootdomain();
1478 if (!d->rd)
1479 return sa_sd;
1480
1481 return sa_rootdomain;
1482}
1483
1484/*
1485 * NULL the sd_data elements we've used to build the sched_domain and
1486 * sched_group structure so that the subsequent __free_domain_allocs()
1487 * will not free the data we're using.
1488 */
1489static void claim_allocations(int cpu, struct sched_domain *sd)
1490{
1491 struct sd_data *sdd = sd->private;
1492
1493 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1494 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1495
1496 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1497 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1498
1499 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1500 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1501
1502 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1503 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1504}
1505
1506#ifdef CONFIG_NUMA
1507enum numa_topology_type sched_numa_topology_type;
1508
1509static int sched_domains_numa_levels;
1510static int sched_domains_curr_level;
1511
1512int sched_max_numa_distance;
1513static int *sched_domains_numa_distance;
1514static struct cpumask ***sched_domains_numa_masks;
1515#endif
1516
1517/*
1518 * SD_flags allowed in topology descriptions.
1519 *
1520 * These flags are purely descriptive of the topology and do not prescribe
1521 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1522 * function:
1523 *
1524 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1525 * SD_SHARE_PKG_RESOURCES - describes shared caches
1526 * SD_NUMA - describes NUMA topologies
1527 *
1528 * Odd one out, which beside describing the topology has a quirk also
1529 * prescribes the desired behaviour that goes along with it:
1530 *
1531 * SD_ASYM_PACKING - describes SMT quirks
1532 */
1533#define TOPOLOGY_SD_FLAGS \
1534 (SD_SHARE_CPUCAPACITY | \
1535 SD_SHARE_PKG_RESOURCES | \
1536 SD_NUMA | \
1537 SD_ASYM_PACKING)
1538
1539static struct sched_domain *
1540sd_init(struct sched_domain_topology_level *tl,
1541 const struct cpumask *cpu_map,
1542 struct sched_domain *child, int cpu)
1543{
1544 struct sd_data *sdd = &tl->data;
1545 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1546 int sd_id, sd_weight, sd_flags = 0;
1547 struct cpumask *sd_span;
1548
1549#ifdef CONFIG_NUMA
1550 /*
1551 * Ugly hack to pass state to sd_numa_mask()...
1552 */
1553 sched_domains_curr_level = tl->numa_level;
1554#endif
1555
1556 sd_weight = cpumask_weight(tl->mask(cpu));
1557
1558 if (tl->sd_flags)
1559 sd_flags = (*tl->sd_flags)();
1560 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1561 "wrong sd_flags in topology description\n"))
1562 sd_flags &= TOPOLOGY_SD_FLAGS;
1563
1564 *sd = (struct sched_domain){
1565 .min_interval = sd_weight,
1566 .max_interval = 2*sd_weight,
1567 .busy_factor = 16,
1568 .imbalance_pct = 117,
1569
1570 .cache_nice_tries = 0,
1571
1572 .flags = 1*SD_BALANCE_NEWIDLE
1573 | 1*SD_BALANCE_EXEC
1574 | 1*SD_BALANCE_FORK
1575 | 0*SD_BALANCE_WAKE
1576 | 1*SD_WAKE_AFFINE
1577 | 0*SD_SHARE_CPUCAPACITY
1578 | 0*SD_SHARE_PKG_RESOURCES
1579 | 0*SD_SERIALIZE
1580 | 1*SD_PREFER_SIBLING
1581 | 0*SD_NUMA
1582 | sd_flags
1583 ,
1584
1585 .last_balance = jiffies,
1586 .balance_interval = sd_weight,
1587 .max_newidle_lb_cost = 0,
1588 .last_decay_max_lb_cost = jiffies,
1589 .child = child,
1590#ifdef CONFIG_SCHED_DEBUG
1591 .name = tl->name,
1592#endif
1593 };
1594
1595 sd_span = sched_domain_span(sd);
1596 cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1597 sd_id = cpumask_first(sd_span);
1598
1599 sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1600
1601 WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1602 (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1603 "CPU capacity asymmetry not supported on SMT\n");
1604
1605 /*
1606 * Convert topological properties into behaviour.
1607 */
1608 /* Don't attempt to spread across CPUs of different capacities. */
1609 if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1610 sd->child->flags &= ~SD_PREFER_SIBLING;
1611
1612 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1613 sd->imbalance_pct = 110;
1614
1615 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1616 sd->imbalance_pct = 117;
1617 sd->cache_nice_tries = 1;
1618
1619#ifdef CONFIG_NUMA
1620 } else if (sd->flags & SD_NUMA) {
1621 sd->cache_nice_tries = 2;
1622
1623 sd->flags &= ~SD_PREFER_SIBLING;
1624 sd->flags |= SD_SERIALIZE;
1625 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1626 sd->flags &= ~(SD_BALANCE_EXEC |
1627 SD_BALANCE_FORK |
1628 SD_WAKE_AFFINE);
1629 }
1630
1631#endif
1632 } else {
1633 sd->cache_nice_tries = 1;
1634 }
1635
1636 /*
1637 * For all levels sharing cache; connect a sched_domain_shared
1638 * instance.
1639 */
1640 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1641 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1642 atomic_inc(&sd->shared->ref);
1643 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1644 }
1645
1646 sd->private = sdd;
1647
1648 return sd;
1649}
1650
1651/*
1652 * Topology list, bottom-up.
1653 */
1654static struct sched_domain_topology_level default_topology[] = {
1655#ifdef CONFIG_SCHED_SMT
1656 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1657#endif
1658
1659#ifdef CONFIG_SCHED_CLUSTER
1660 { cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1661#endif
1662
1663#ifdef CONFIG_SCHED_MC
1664 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1665#endif
1666 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1667 { NULL, },
1668};
1669
1670static struct sched_domain_topology_level *sched_domain_topology =
1671 default_topology;
1672static struct sched_domain_topology_level *sched_domain_topology_saved;
1673
1674#define for_each_sd_topology(tl) \
1675 for (tl = sched_domain_topology; tl->mask; tl++)
1676
1677void set_sched_topology(struct sched_domain_topology_level *tl)
1678{
1679 if (WARN_ON_ONCE(sched_smp_initialized))
1680 return;
1681
1682 sched_domain_topology = tl;
1683 sched_domain_topology_saved = NULL;
1684}
1685
1686#ifdef CONFIG_NUMA
1687
1688static const struct cpumask *sd_numa_mask(int cpu)
1689{
1690 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1691}
1692
1693static void sched_numa_warn(const char *str)
1694{
1695 static int done = false;
1696 int i,j;
1697
1698 if (done)
1699 return;
1700
1701 done = true;
1702
1703 printk(KERN_WARNING "ERROR: %s\n\n", str);
1704
1705 for (i = 0; i < nr_node_ids; i++) {
1706 printk(KERN_WARNING " ");
1707 for (j = 0; j < nr_node_ids; j++) {
1708 if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
1709 printk(KERN_CONT "(%02d) ", node_distance(i,j));
1710 else
1711 printk(KERN_CONT " %02d ", node_distance(i,j));
1712 }
1713 printk(KERN_CONT "\n");
1714 }
1715 printk(KERN_WARNING "\n");
1716}
1717
1718bool find_numa_distance(int distance)
1719{
1720 bool found = false;
1721 int i, *distances;
1722
1723 if (distance == node_distance(0, 0))
1724 return true;
1725
1726 rcu_read_lock();
1727 distances = rcu_dereference(sched_domains_numa_distance);
1728 if (!distances)
1729 goto unlock;
1730 for (i = 0; i < sched_domains_numa_levels; i++) {
1731 if (distances[i] == distance) {
1732 found = true;
1733 break;
1734 }
1735 }
1736unlock:
1737 rcu_read_unlock();
1738
1739 return found;
1740}
1741
1742#define for_each_cpu_node_but(n, nbut) \
1743 for_each_node_state(n, N_CPU) \
1744 if (n == nbut) \
1745 continue; \
1746 else
1747
1748/*
1749 * A system can have three types of NUMA topology:
1750 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1751 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1752 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1753 *
1754 * The difference between a glueless mesh topology and a backplane
1755 * topology lies in whether communication between not directly
1756 * connected nodes goes through intermediary nodes (where programs
1757 * could run), or through backplane controllers. This affects
1758 * placement of programs.
1759 *
1760 * The type of topology can be discerned with the following tests:
1761 * - If the maximum distance between any nodes is 1 hop, the system
1762 * is directly connected.
1763 * - If for two nodes A and B, located N > 1 hops away from each other,
1764 * there is an intermediary node C, which is < N hops away from both
1765 * nodes A and B, the system is a glueless mesh.
1766 */
1767static void init_numa_topology_type(int offline_node)
1768{
1769 int a, b, c, n;
1770
1771 n = sched_max_numa_distance;
1772
1773 if (sched_domains_numa_levels <= 2) {
1774 sched_numa_topology_type = NUMA_DIRECT;
1775 return;
1776 }
1777
1778 for_each_cpu_node_but(a, offline_node) {
1779 for_each_cpu_node_but(b, offline_node) {
1780 /* Find two nodes furthest removed from each other. */
1781 if (node_distance(a, b) < n)
1782 continue;
1783
1784 /* Is there an intermediary node between a and b? */
1785 for_each_cpu_node_but(c, offline_node) {
1786 if (node_distance(a, c) < n &&
1787 node_distance(b, c) < n) {
1788 sched_numa_topology_type =
1789 NUMA_GLUELESS_MESH;
1790 return;
1791 }
1792 }
1793
1794 sched_numa_topology_type = NUMA_BACKPLANE;
1795 return;
1796 }
1797 }
1798
1799 pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1800 sched_numa_topology_type = NUMA_DIRECT;
1801}
1802
1803
1804#define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1805
1806void sched_init_numa(int offline_node)
1807{
1808 struct sched_domain_topology_level *tl;
1809 unsigned long *distance_map;
1810 int nr_levels = 0;
1811 int i, j;
1812 int *distances;
1813 struct cpumask ***masks;
1814
1815 /*
1816 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1817 * unique distances in the node_distance() table.
1818 */
1819 distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1820 if (!distance_map)
1821 return;
1822
1823 bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1824 for_each_cpu_node_but(i, offline_node) {
1825 for_each_cpu_node_but(j, offline_node) {
1826 int distance = node_distance(i, j);
1827
1828 if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1829 sched_numa_warn("Invalid distance value range");
1830 bitmap_free(distance_map);
1831 return;
1832 }
1833
1834 bitmap_set(distance_map, distance, 1);
1835 }
1836 }
1837 /*
1838 * We can now figure out how many unique distance values there are and
1839 * allocate memory accordingly.
1840 */
1841 nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1842
1843 distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1844 if (!distances) {
1845 bitmap_free(distance_map);
1846 return;
1847 }
1848
1849 for (i = 0, j = 0; i < nr_levels; i++, j++) {
1850 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1851 distances[i] = j;
1852 }
1853 rcu_assign_pointer(sched_domains_numa_distance, distances);
1854
1855 bitmap_free(distance_map);
1856
1857 /*
1858 * 'nr_levels' contains the number of unique distances
1859 *
1860 * The sched_domains_numa_distance[] array includes the actual distance
1861 * numbers.
1862 */
1863
1864 /*
1865 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1866 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1867 * the array will contain less then 'nr_levels' members. This could be
1868 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1869 * in other functions.
1870 *
1871 * We reset it to 'nr_levels' at the end of this function.
1872 */
1873 sched_domains_numa_levels = 0;
1874
1875 masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1876 if (!masks)
1877 return;
1878
1879 /*
1880 * Now for each level, construct a mask per node which contains all
1881 * CPUs of nodes that are that many hops away from us.
1882 */
1883 for (i = 0; i < nr_levels; i++) {
1884 masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1885 if (!masks[i])
1886 return;
1887
1888 for_each_cpu_node_but(j, offline_node) {
1889 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1890 int k;
1891
1892 if (!mask)
1893 return;
1894
1895 masks[i][j] = mask;
1896
1897 for_each_cpu_node_but(k, offline_node) {
1898 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1899 sched_numa_warn("Node-distance not symmetric");
1900
1901 if (node_distance(j, k) > sched_domains_numa_distance[i])
1902 continue;
1903
1904 cpumask_or(mask, mask, cpumask_of_node(k));
1905 }
1906 }
1907 }
1908 rcu_assign_pointer(sched_domains_numa_masks, masks);
1909
1910 /* Compute default topology size */
1911 for (i = 0; sched_domain_topology[i].mask; i++);
1912
1913 tl = kzalloc((i + nr_levels + 1) *
1914 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1915 if (!tl)
1916 return;
1917
1918 /*
1919 * Copy the default topology bits..
1920 */
1921 for (i = 0; sched_domain_topology[i].mask; i++)
1922 tl[i] = sched_domain_topology[i];
1923
1924 /*
1925 * Add the NUMA identity distance, aka single NODE.
1926 */
1927 tl[i++] = (struct sched_domain_topology_level){
1928 .mask = sd_numa_mask,
1929 .numa_level = 0,
1930 SD_INIT_NAME(NODE)
1931 };
1932
1933 /*
1934 * .. and append 'j' levels of NUMA goodness.
1935 */
1936 for (j = 1; j < nr_levels; i++, j++) {
1937 tl[i] = (struct sched_domain_topology_level){
1938 .mask = sd_numa_mask,
1939 .sd_flags = cpu_numa_flags,
1940 .flags = SDTL_OVERLAP,
1941 .numa_level = j,
1942 SD_INIT_NAME(NUMA)
1943 };
1944 }
1945
1946 sched_domain_topology_saved = sched_domain_topology;
1947 sched_domain_topology = tl;
1948
1949 sched_domains_numa_levels = nr_levels;
1950 WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]);
1951
1952 init_numa_topology_type(offline_node);
1953}
1954
1955
1956static void sched_reset_numa(void)
1957{
1958 int nr_levels, *distances;
1959 struct cpumask ***masks;
1960
1961 nr_levels = sched_domains_numa_levels;
1962 sched_domains_numa_levels = 0;
1963 sched_max_numa_distance = 0;
1964 sched_numa_topology_type = NUMA_DIRECT;
1965 distances = sched_domains_numa_distance;
1966 rcu_assign_pointer(sched_domains_numa_distance, NULL);
1967 masks = sched_domains_numa_masks;
1968 rcu_assign_pointer(sched_domains_numa_masks, NULL);
1969 if (distances || masks) {
1970 int i, j;
1971
1972 synchronize_rcu();
1973 kfree(distances);
1974 for (i = 0; i < nr_levels && masks; i++) {
1975 if (!masks[i])
1976 continue;
1977 for_each_node(j)
1978 kfree(masks[i][j]);
1979 kfree(masks[i]);
1980 }
1981 kfree(masks);
1982 }
1983 if (sched_domain_topology_saved) {
1984 kfree(sched_domain_topology);
1985 sched_domain_topology = sched_domain_topology_saved;
1986 sched_domain_topology_saved = NULL;
1987 }
1988}
1989
1990/*
1991 * Call with hotplug lock held
1992 */
1993void sched_update_numa(int cpu, bool online)
1994{
1995 int node;
1996
1997 node = cpu_to_node(cpu);
1998 /*
1999 * Scheduler NUMA topology is updated when the first CPU of a
2000 * node is onlined or the last CPU of a node is offlined.
2001 */
2002 if (cpumask_weight(cpumask_of_node(node)) != 1)
2003 return;
2004
2005 sched_reset_numa();
2006 sched_init_numa(online ? NUMA_NO_NODE : node);
2007}
2008
2009void sched_domains_numa_masks_set(unsigned int cpu)
2010{
2011 int node = cpu_to_node(cpu);
2012 int i, j;
2013
2014 for (i = 0; i < sched_domains_numa_levels; i++) {
2015 for (j = 0; j < nr_node_ids; j++) {
2016 if (!node_state(j, N_CPU))
2017 continue;
2018
2019 /* Set ourselves in the remote node's masks */
2020 if (node_distance(j, node) <= sched_domains_numa_distance[i])
2021 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
2022 }
2023 }
2024}
2025
2026void sched_domains_numa_masks_clear(unsigned int cpu)
2027{
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 (sched_domains_numa_masks[i][j])
2033 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
2034 }
2035 }
2036}
2037
2038/*
2039 * sched_numa_find_closest() - given the NUMA topology, find the cpu
2040 * closest to @cpu from @cpumask.
2041 * cpumask: cpumask to find a cpu from
2042 * cpu: cpu to be close to
2043 *
2044 * returns: cpu, or nr_cpu_ids when nothing found.
2045 */
2046int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2047{
2048 int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2049 struct cpumask ***masks;
2050
2051 rcu_read_lock();
2052 masks = rcu_dereference(sched_domains_numa_masks);
2053 if (!masks)
2054 goto unlock;
2055 for (i = 0; i < sched_domains_numa_levels; i++) {
2056 if (!masks[i][j])
2057 break;
2058 cpu = cpumask_any_and(cpus, masks[i][j]);
2059 if (cpu < nr_cpu_ids) {
2060 found = cpu;
2061 break;
2062 }
2063 }
2064unlock:
2065 rcu_read_unlock();
2066
2067 return found;
2068}
2069
2070#endif /* CONFIG_NUMA */
2071
2072static int __sdt_alloc(const struct cpumask *cpu_map)
2073{
2074 struct sched_domain_topology_level *tl;
2075 int j;
2076
2077 for_each_sd_topology(tl) {
2078 struct sd_data *sdd = &tl->data;
2079
2080 sdd->sd = alloc_percpu(struct sched_domain *);
2081 if (!sdd->sd)
2082 return -ENOMEM;
2083
2084 sdd->sds = alloc_percpu(struct sched_domain_shared *);
2085 if (!sdd->sds)
2086 return -ENOMEM;
2087
2088 sdd->sg = alloc_percpu(struct sched_group *);
2089 if (!sdd->sg)
2090 return -ENOMEM;
2091
2092 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2093 if (!sdd->sgc)
2094 return -ENOMEM;
2095
2096 for_each_cpu(j, cpu_map) {
2097 struct sched_domain *sd;
2098 struct sched_domain_shared *sds;
2099 struct sched_group *sg;
2100 struct sched_group_capacity *sgc;
2101
2102 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2103 GFP_KERNEL, cpu_to_node(j));
2104 if (!sd)
2105 return -ENOMEM;
2106
2107 *per_cpu_ptr(sdd->sd, j) = sd;
2108
2109 sds = kzalloc_node(sizeof(struct sched_domain_shared),
2110 GFP_KERNEL, cpu_to_node(j));
2111 if (!sds)
2112 return -ENOMEM;
2113
2114 *per_cpu_ptr(sdd->sds, j) = sds;
2115
2116 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2117 GFP_KERNEL, cpu_to_node(j));
2118 if (!sg)
2119 return -ENOMEM;
2120
2121 sg->next = sg;
2122
2123 *per_cpu_ptr(sdd->sg, j) = sg;
2124
2125 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2126 GFP_KERNEL, cpu_to_node(j));
2127 if (!sgc)
2128 return -ENOMEM;
2129
2130#ifdef CONFIG_SCHED_DEBUG
2131 sgc->id = j;
2132#endif
2133
2134 *per_cpu_ptr(sdd->sgc, j) = sgc;
2135 }
2136 }
2137
2138 return 0;
2139}
2140
2141static void __sdt_free(const struct cpumask *cpu_map)
2142{
2143 struct sched_domain_topology_level *tl;
2144 int j;
2145
2146 for_each_sd_topology(tl) {
2147 struct sd_data *sdd = &tl->data;
2148
2149 for_each_cpu(j, cpu_map) {
2150 struct sched_domain *sd;
2151
2152 if (sdd->sd) {
2153 sd = *per_cpu_ptr(sdd->sd, j);
2154 if (sd && (sd->flags & SD_OVERLAP))
2155 free_sched_groups(sd->groups, 0);
2156 kfree(*per_cpu_ptr(sdd->sd, j));
2157 }
2158
2159 if (sdd->sds)
2160 kfree(*per_cpu_ptr(sdd->sds, j));
2161 if (sdd->sg)
2162 kfree(*per_cpu_ptr(sdd->sg, j));
2163 if (sdd->sgc)
2164 kfree(*per_cpu_ptr(sdd->sgc, j));
2165 }
2166 free_percpu(sdd->sd);
2167 sdd->sd = NULL;
2168 free_percpu(sdd->sds);
2169 sdd->sds = NULL;
2170 free_percpu(sdd->sg);
2171 sdd->sg = NULL;
2172 free_percpu(sdd->sgc);
2173 sdd->sgc = NULL;
2174 }
2175}
2176
2177static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2178 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2179 struct sched_domain *child, int cpu)
2180{
2181 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2182
2183 if (child) {
2184 sd->level = child->level + 1;
2185 sched_domain_level_max = max(sched_domain_level_max, sd->level);
2186 child->parent = sd;
2187
2188 if (!cpumask_subset(sched_domain_span(child),
2189 sched_domain_span(sd))) {
2190 pr_err("BUG: arch topology borken\n");
2191#ifdef CONFIG_SCHED_DEBUG
2192 pr_err(" the %s domain not a subset of the %s domain\n",
2193 child->name, sd->name);
2194#endif
2195 /* Fixup, ensure @sd has at least @child CPUs. */
2196 cpumask_or(sched_domain_span(sd),
2197 sched_domain_span(sd),
2198 sched_domain_span(child));
2199 }
2200
2201 }
2202 set_domain_attribute(sd, attr);
2203
2204 return sd;
2205}
2206
2207/*
2208 * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2209 * any two given CPUs at this (non-NUMA) topology level.
2210 */
2211static bool topology_span_sane(struct sched_domain_topology_level *tl,
2212 const struct cpumask *cpu_map, int cpu)
2213{
2214 int i;
2215
2216 /* NUMA levels are allowed to overlap */
2217 if (tl->flags & SDTL_OVERLAP)
2218 return true;
2219
2220 /*
2221 * Non-NUMA levels cannot partially overlap - they must be either
2222 * completely equal or completely disjoint. Otherwise we can end up
2223 * breaking the sched_group lists - i.e. a later get_group() pass
2224 * breaks the linking done for an earlier span.
2225 */
2226 for_each_cpu(i, cpu_map) {
2227 if (i == cpu)
2228 continue;
2229 /*
2230 * We should 'and' all those masks with 'cpu_map' to exactly
2231 * match the topology we're about to build, but that can only
2232 * remove CPUs, which only lessens our ability to detect
2233 * overlaps
2234 */
2235 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2236 cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2237 return false;
2238 }
2239
2240 return true;
2241}
2242
2243/*
2244 * Build sched domains for a given set of CPUs and attach the sched domains
2245 * to the individual CPUs
2246 */
2247static int
2248build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2249{
2250 enum s_alloc alloc_state = sa_none;
2251 struct sched_domain *sd;
2252 struct s_data d;
2253 struct rq *rq = NULL;
2254 int i, ret = -ENOMEM;
2255 bool has_asym = false;
2256
2257 if (WARN_ON(cpumask_empty(cpu_map)))
2258 goto error;
2259
2260 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2261 if (alloc_state != sa_rootdomain)
2262 goto error;
2263
2264 /* Set up domains for CPUs specified by the cpu_map: */
2265 for_each_cpu(i, cpu_map) {
2266 struct sched_domain_topology_level *tl;
2267
2268 sd = NULL;
2269 for_each_sd_topology(tl) {
2270
2271 if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2272 goto error;
2273
2274 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2275
2276 has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2277
2278 if (tl == sched_domain_topology)
2279 *per_cpu_ptr(d.sd, i) = sd;
2280 if (tl->flags & SDTL_OVERLAP)
2281 sd->flags |= SD_OVERLAP;
2282 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2283 break;
2284 }
2285 }
2286
2287 /* Build the groups for the domains */
2288 for_each_cpu(i, cpu_map) {
2289 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2290 sd->span_weight = cpumask_weight(sched_domain_span(sd));
2291 if (sd->flags & SD_OVERLAP) {
2292 if (build_overlap_sched_groups(sd, i))
2293 goto error;
2294 } else {
2295 if (build_sched_groups(sd, i))
2296 goto error;
2297 }
2298 }
2299 }
2300
2301 /*
2302 * Calculate an allowed NUMA imbalance such that LLCs do not get
2303 * imbalanced.
2304 */
2305 for_each_cpu(i, cpu_map) {
2306 unsigned int imb = 0;
2307 unsigned int imb_span = 1;
2308
2309 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2310 struct sched_domain *child = sd->child;
2311
2312 if (!(sd->flags & SD_SHARE_PKG_RESOURCES) && child &&
2313 (child->flags & SD_SHARE_PKG_RESOURCES)) {
2314 struct sched_domain __rcu *top_p;
2315 unsigned int nr_llcs;
2316
2317 /*
2318 * For a single LLC per node, allow an
2319 * imbalance up to 12.5% of the node. This is
2320 * arbitrary cutoff based two factors -- SMT and
2321 * memory channels. For SMT-2, the intent is to
2322 * avoid premature sharing of HT resources but
2323 * SMT-4 or SMT-8 *may* benefit from a different
2324 * cutoff. For memory channels, this is a very
2325 * rough estimate of how many channels may be
2326 * active and is based on recent CPUs with
2327 * many cores.
2328 *
2329 * For multiple LLCs, allow an imbalance
2330 * until multiple tasks would share an LLC
2331 * on one node while LLCs on another node
2332 * remain idle. This assumes that there are
2333 * enough logical CPUs per LLC to avoid SMT
2334 * factors and that there is a correlation
2335 * between LLCs and memory channels.
2336 */
2337 nr_llcs = sd->span_weight / child->span_weight;
2338 if (nr_llcs == 1)
2339 imb = sd->span_weight >> 3;
2340 else
2341 imb = nr_llcs;
2342 imb = max(1U, imb);
2343 sd->imb_numa_nr = imb;
2344
2345 /* Set span based on the first NUMA domain. */
2346 top_p = sd->parent;
2347 while (top_p && !(top_p->flags & SD_NUMA)) {
2348 top_p = top_p->parent;
2349 }
2350 imb_span = top_p ? top_p->span_weight : sd->span_weight;
2351 } else {
2352 int factor = max(1U, (sd->span_weight / imb_span));
2353
2354 sd->imb_numa_nr = imb * factor;
2355 }
2356 }
2357 }
2358
2359 /* Calculate CPU capacity for physical packages and nodes */
2360 for (i = nr_cpumask_bits-1; i >= 0; i--) {
2361 if (!cpumask_test_cpu(i, cpu_map))
2362 continue;
2363
2364 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2365 claim_allocations(i, sd);
2366 init_sched_groups_capacity(i, sd);
2367 }
2368 }
2369
2370 /* Attach the domains */
2371 rcu_read_lock();
2372 for_each_cpu(i, cpu_map) {
2373 rq = cpu_rq(i);
2374 sd = *per_cpu_ptr(d.sd, i);
2375
2376 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2377 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2378 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2379
2380 cpu_attach_domain(sd, d.rd, i);
2381 }
2382 rcu_read_unlock();
2383
2384 if (has_asym)
2385 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2386
2387 if (rq && sched_debug_verbose) {
2388 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2389 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2390 }
2391
2392 ret = 0;
2393error:
2394 __free_domain_allocs(&d, alloc_state, cpu_map);
2395
2396 return ret;
2397}
2398
2399/* Current sched domains: */
2400static cpumask_var_t *doms_cur;
2401
2402/* Number of sched domains in 'doms_cur': */
2403static int ndoms_cur;
2404
2405/* Attributes of custom domains in 'doms_cur' */
2406static struct sched_domain_attr *dattr_cur;
2407
2408/*
2409 * Special case: If a kmalloc() of a doms_cur partition (array of
2410 * cpumask) fails, then fallback to a single sched domain,
2411 * as determined by the single cpumask fallback_doms.
2412 */
2413static cpumask_var_t fallback_doms;
2414
2415/*
2416 * arch_update_cpu_topology lets virtualized architectures update the
2417 * CPU core maps. It is supposed to return 1 if the topology changed
2418 * or 0 if it stayed the same.
2419 */
2420int __weak arch_update_cpu_topology(void)
2421{
2422 return 0;
2423}
2424
2425cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2426{
2427 int i;
2428 cpumask_var_t *doms;
2429
2430 doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2431 if (!doms)
2432 return NULL;
2433 for (i = 0; i < ndoms; i++) {
2434 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2435 free_sched_domains(doms, i);
2436 return NULL;
2437 }
2438 }
2439 return doms;
2440}
2441
2442void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2443{
2444 unsigned int i;
2445 for (i = 0; i < ndoms; i++)
2446 free_cpumask_var(doms[i]);
2447 kfree(doms);
2448}
2449
2450/*
2451 * Set up scheduler domains and groups. For now this just excludes isolated
2452 * CPUs, but could be used to exclude other special cases in the future.
2453 */
2454int sched_init_domains(const struct cpumask *cpu_map)
2455{
2456 int err;
2457
2458 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2459 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2460 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2461
2462 arch_update_cpu_topology();
2463 asym_cpu_capacity_scan();
2464 ndoms_cur = 1;
2465 doms_cur = alloc_sched_domains(ndoms_cur);
2466 if (!doms_cur)
2467 doms_cur = &fallback_doms;
2468 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
2469 err = build_sched_domains(doms_cur[0], NULL);
2470
2471 return err;
2472}
2473
2474/*
2475 * Detach sched domains from a group of CPUs specified in cpu_map
2476 * These CPUs will now be attached to the NULL domain
2477 */
2478static void detach_destroy_domains(const struct cpumask *cpu_map)
2479{
2480 unsigned int cpu = cpumask_any(cpu_map);
2481 int i;
2482
2483 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2484 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2485
2486 rcu_read_lock();
2487 for_each_cpu(i, cpu_map)
2488 cpu_attach_domain(NULL, &def_root_domain, i);
2489 rcu_read_unlock();
2490}
2491
2492/* handle null as "default" */
2493static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2494 struct sched_domain_attr *new, int idx_new)
2495{
2496 struct sched_domain_attr tmp;
2497
2498 /* Fast path: */
2499 if (!new && !cur)
2500 return 1;
2501
2502 tmp = SD_ATTR_INIT;
2503
2504 return !memcmp(cur ? (cur + idx_cur) : &tmp,
2505 new ? (new + idx_new) : &tmp,
2506 sizeof(struct sched_domain_attr));
2507}
2508
2509/*
2510 * Partition sched domains as specified by the 'ndoms_new'
2511 * cpumasks in the array doms_new[] of cpumasks. This compares
2512 * doms_new[] to the current sched domain partitioning, doms_cur[].
2513 * It destroys each deleted domain and builds each new domain.
2514 *
2515 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2516 * The masks don't intersect (don't overlap.) We should setup one
2517 * sched domain for each mask. CPUs not in any of the cpumasks will
2518 * not be load balanced. If the same cpumask appears both in the
2519 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2520 * it as it is.
2521 *
2522 * The passed in 'doms_new' should be allocated using
2523 * alloc_sched_domains. This routine takes ownership of it and will
2524 * free_sched_domains it when done with it. If the caller failed the
2525 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2526 * and partition_sched_domains() will fallback to the single partition
2527 * 'fallback_doms', it also forces the domains to be rebuilt.
2528 *
2529 * If doms_new == NULL it will be replaced with cpu_online_mask.
2530 * ndoms_new == 0 is a special case for destroying existing domains,
2531 * and it will not create the default domain.
2532 *
2533 * Call with hotplug lock and sched_domains_mutex held
2534 */
2535void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2536 struct sched_domain_attr *dattr_new)
2537{
2538 bool __maybe_unused has_eas = false;
2539 int i, j, n;
2540 int new_topology;
2541
2542 lockdep_assert_held(&sched_domains_mutex);
2543
2544 /* Let the architecture update CPU core mappings: */
2545 new_topology = arch_update_cpu_topology();
2546 /* Trigger rebuilding CPU capacity asymmetry data */
2547 if (new_topology)
2548 asym_cpu_capacity_scan();
2549
2550 if (!doms_new) {
2551 WARN_ON_ONCE(dattr_new);
2552 n = 0;
2553 doms_new = alloc_sched_domains(1);
2554 if (doms_new) {
2555 n = 1;
2556 cpumask_and(doms_new[0], cpu_active_mask,
2557 housekeeping_cpumask(HK_TYPE_DOMAIN));
2558 }
2559 } else {
2560 n = ndoms_new;
2561 }
2562
2563 /* Destroy deleted domains: */
2564 for (i = 0; i < ndoms_cur; i++) {
2565 for (j = 0; j < n && !new_topology; j++) {
2566 if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2567 dattrs_equal(dattr_cur, i, dattr_new, j)) {
2568 struct root_domain *rd;
2569
2570 /*
2571 * This domain won't be destroyed and as such
2572 * its dl_bw->total_bw needs to be cleared. It
2573 * will be recomputed in function
2574 * update_tasks_root_domain().
2575 */
2576 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2577 dl_clear_root_domain(rd);
2578 goto match1;
2579 }
2580 }
2581 /* No match - a current sched domain not in new doms_new[] */
2582 detach_destroy_domains(doms_cur[i]);
2583match1:
2584 ;
2585 }
2586
2587 n = ndoms_cur;
2588 if (!doms_new) {
2589 n = 0;
2590 doms_new = &fallback_doms;
2591 cpumask_and(doms_new[0], cpu_active_mask,
2592 housekeeping_cpumask(HK_TYPE_DOMAIN));
2593 }
2594
2595 /* Build new domains: */
2596 for (i = 0; i < ndoms_new; i++) {
2597 for (j = 0; j < n && !new_topology; j++) {
2598 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2599 dattrs_equal(dattr_new, i, dattr_cur, j))
2600 goto match2;
2601 }
2602 /* No match - add a new doms_new */
2603 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2604match2:
2605 ;
2606 }
2607
2608#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2609 /* Build perf. domains: */
2610 for (i = 0; i < ndoms_new; i++) {
2611 for (j = 0; j < n && !sched_energy_update; j++) {
2612 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2613 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2614 has_eas = true;
2615 goto match3;
2616 }
2617 }
2618 /* No match - add perf. domains for a new rd */
2619 has_eas |= build_perf_domains(doms_new[i]);
2620match3:
2621 ;
2622 }
2623 sched_energy_set(has_eas);
2624#endif
2625
2626 /* Remember the new sched domains: */
2627 if (doms_cur != &fallback_doms)
2628 free_sched_domains(doms_cur, ndoms_cur);
2629
2630 kfree(dattr_cur);
2631 doms_cur = doms_new;
2632 dattr_cur = dattr_new;
2633 ndoms_cur = ndoms_new;
2634
2635 update_sched_domain_debugfs();
2636}
2637
2638/*
2639 * Call with hotplug lock held
2640 */
2641void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2642 struct sched_domain_attr *dattr_new)
2643{
2644 mutex_lock(&sched_domains_mutex);
2645 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2646 mutex_unlock(&sched_domains_mutex);
2647}