Loading...
Note: File does not exist in v3.15.
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Scheduler topology setup/handling methods
4 */
5
6#include <linux/bsearch.h>
7
8DEFINE_MUTEX(sched_domains_mutex);
9
10/* Protected by sched_domains_mutex: */
11static cpumask_var_t sched_domains_tmpmask;
12static cpumask_var_t sched_domains_tmpmask2;
13
14#ifdef CONFIG_SCHED_DEBUG
15
16static int __init sched_debug_setup(char *str)
17{
18 sched_debug_verbose = true;
19
20 return 0;
21}
22early_param("sched_verbose", sched_debug_setup);
23
24static 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 },
30const struct sd_flag_debug sd_flag_debug[] = {
31#include <linux/sched/sd_flags.h>
32};
33#undef SD_FLAG
34
35static 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
131static 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)
158static 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)) |
166static const unsigned int SD_DEGENERATE_GROUPS_MASK =
167#include <linux/sched/sd_flags.h>
1680;
169#undef SD_FLAG
170
171static 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
188static int
189sd_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)
210DEFINE_STATIC_KEY_FALSE(sched_energy_present);
211static unsigned int sysctl_sched_energy_aware = 1;
212static DEFINE_MUTEX(sched_energy_mutex);
213static bool sched_energy_update;
214
215static bool sched_is_eas_possible(const struct cpumask *cpu_mask)
216{
217 bool any_asym_capacity = false;
218 struct cpufreq_policy *policy;
219 struct cpufreq_governor *gov;
220 int i;
221
222 /* EAS is enabled for asymmetric CPU capacity topologies. */
223 for_each_cpu(i, cpu_mask) {
224 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, i))) {
225 any_asym_capacity = true;
226 break;
227 }
228 }
229 if (!any_asym_capacity) {
230 if (sched_debug()) {
231 pr_info("rd %*pbl: Checking EAS, CPUs do not have asymmetric capacities\n",
232 cpumask_pr_args(cpu_mask));
233 }
234 return false;
235 }
236
237 /* EAS definitely does *not* handle SMT */
238 if (sched_smt_active()) {
239 if (sched_debug()) {
240 pr_info("rd %*pbl: Checking EAS, SMT is not supported\n",
241 cpumask_pr_args(cpu_mask));
242 }
243 return false;
244 }
245
246 if (!arch_scale_freq_invariant()) {
247 if (sched_debug()) {
248 pr_info("rd %*pbl: Checking EAS: frequency-invariant load tracking not yet supported",
249 cpumask_pr_args(cpu_mask));
250 }
251 return false;
252 }
253
254 /* Do not attempt EAS if schedutil is not being used. */
255 for_each_cpu(i, cpu_mask) {
256 policy = cpufreq_cpu_get(i);
257 if (!policy) {
258 if (sched_debug()) {
259 pr_info("rd %*pbl: Checking EAS, cpufreq policy not set for CPU: %d",
260 cpumask_pr_args(cpu_mask), i);
261 }
262 return false;
263 }
264 gov = policy->governor;
265 cpufreq_cpu_put(policy);
266 if (gov != &schedutil_gov) {
267 if (sched_debug()) {
268 pr_info("rd %*pbl: Checking EAS, schedutil is mandatory\n",
269 cpumask_pr_args(cpu_mask));
270 }
271 return false;
272 }
273 }
274
275 return true;
276}
277
278void rebuild_sched_domains_energy(void)
279{
280 mutex_lock(&sched_energy_mutex);
281 sched_energy_update = true;
282 rebuild_sched_domains();
283 sched_energy_update = false;
284 mutex_unlock(&sched_energy_mutex);
285}
286
287#ifdef CONFIG_PROC_SYSCTL
288static int sched_energy_aware_handler(struct ctl_table *table, int write,
289 void *buffer, size_t *lenp, loff_t *ppos)
290{
291 int ret, state;
292
293 if (write && !capable(CAP_SYS_ADMIN))
294 return -EPERM;
295
296 if (!sched_is_eas_possible(cpu_active_mask)) {
297 if (write) {
298 return -EOPNOTSUPP;
299 } else {
300 *lenp = 0;
301 return 0;
302 }
303 }
304
305 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
306 if (!ret && write) {
307 state = static_branch_unlikely(&sched_energy_present);
308 if (state != sysctl_sched_energy_aware)
309 rebuild_sched_domains_energy();
310 }
311
312 return ret;
313}
314
315static struct ctl_table sched_energy_aware_sysctls[] = {
316 {
317 .procname = "sched_energy_aware",
318 .data = &sysctl_sched_energy_aware,
319 .maxlen = sizeof(unsigned int),
320 .mode = 0644,
321 .proc_handler = sched_energy_aware_handler,
322 .extra1 = SYSCTL_ZERO,
323 .extra2 = SYSCTL_ONE,
324 },
325 {}
326};
327
328static int __init sched_energy_aware_sysctl_init(void)
329{
330 register_sysctl_init("kernel", sched_energy_aware_sysctls);
331 return 0;
332}
333
334late_initcall(sched_energy_aware_sysctl_init);
335#endif
336
337static void free_pd(struct perf_domain *pd)
338{
339 struct perf_domain *tmp;
340
341 while (pd) {
342 tmp = pd->next;
343 kfree(pd);
344 pd = tmp;
345 }
346}
347
348static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
349{
350 while (pd) {
351 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
352 return pd;
353 pd = pd->next;
354 }
355
356 return NULL;
357}
358
359static struct perf_domain *pd_init(int cpu)
360{
361 struct em_perf_domain *obj = em_cpu_get(cpu);
362 struct perf_domain *pd;
363
364 if (!obj) {
365 if (sched_debug())
366 pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
367 return NULL;
368 }
369
370 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
371 if (!pd)
372 return NULL;
373 pd->em_pd = obj;
374
375 return pd;
376}
377
378static void perf_domain_debug(const struct cpumask *cpu_map,
379 struct perf_domain *pd)
380{
381 if (!sched_debug() || !pd)
382 return;
383
384 printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
385
386 while (pd) {
387 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
388 cpumask_first(perf_domain_span(pd)),
389 cpumask_pr_args(perf_domain_span(pd)),
390 em_pd_nr_perf_states(pd->em_pd));
391 pd = pd->next;
392 }
393
394 printk(KERN_CONT "\n");
395}
396
397static void destroy_perf_domain_rcu(struct rcu_head *rp)
398{
399 struct perf_domain *pd;
400
401 pd = container_of(rp, struct perf_domain, rcu);
402 free_pd(pd);
403}
404
405static void sched_energy_set(bool has_eas)
406{
407 if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
408 if (sched_debug())
409 pr_info("%s: stopping EAS\n", __func__);
410 static_branch_disable_cpuslocked(&sched_energy_present);
411 } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
412 if (sched_debug())
413 pr_info("%s: starting EAS\n", __func__);
414 static_branch_enable_cpuslocked(&sched_energy_present);
415 }
416}
417
418/*
419 * EAS can be used on a root domain if it meets all the following conditions:
420 * 1. an Energy Model (EM) is available;
421 * 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
422 * 3. no SMT is detected.
423 * 4. schedutil is driving the frequency of all CPUs of the rd;
424 * 5. frequency invariance support is present;
425 */
426static bool build_perf_domains(const struct cpumask *cpu_map)
427{
428 int i;
429 struct perf_domain *pd = NULL, *tmp;
430 int cpu = cpumask_first(cpu_map);
431 struct root_domain *rd = cpu_rq(cpu)->rd;
432
433 if (!sysctl_sched_energy_aware)
434 goto free;
435
436 if (!sched_is_eas_possible(cpu_map))
437 goto free;
438
439 for_each_cpu(i, cpu_map) {
440 /* Skip already covered CPUs. */
441 if (find_pd(pd, i))
442 continue;
443
444 /* Create the new pd and add it to the local list. */
445 tmp = pd_init(i);
446 if (!tmp)
447 goto free;
448 tmp->next = pd;
449 pd = tmp;
450 }
451
452 perf_domain_debug(cpu_map, pd);
453
454 /* Attach the new list of performance domains to the root domain. */
455 tmp = rd->pd;
456 rcu_assign_pointer(rd->pd, pd);
457 if (tmp)
458 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
459
460 return !!pd;
461
462free:
463 free_pd(pd);
464 tmp = rd->pd;
465 rcu_assign_pointer(rd->pd, NULL);
466 if (tmp)
467 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
468
469 return false;
470}
471#else
472static void free_pd(struct perf_domain *pd) { }
473#endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
474
475static void free_rootdomain(struct rcu_head *rcu)
476{
477 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
478
479 cpupri_cleanup(&rd->cpupri);
480 cpudl_cleanup(&rd->cpudl);
481 free_cpumask_var(rd->dlo_mask);
482 free_cpumask_var(rd->rto_mask);
483 free_cpumask_var(rd->online);
484 free_cpumask_var(rd->span);
485 free_pd(rd->pd);
486 kfree(rd);
487}
488
489void rq_attach_root(struct rq *rq, struct root_domain *rd)
490{
491 struct root_domain *old_rd = NULL;
492 struct rq_flags rf;
493
494 rq_lock_irqsave(rq, &rf);
495
496 if (rq->rd) {
497 old_rd = rq->rd;
498
499 if (cpumask_test_cpu(rq->cpu, old_rd->online))
500 set_rq_offline(rq);
501
502 cpumask_clear_cpu(rq->cpu, old_rd->span);
503
504 /*
505 * If we dont want to free the old_rd yet then
506 * set old_rd to NULL to skip the freeing later
507 * in this function:
508 */
509 if (!atomic_dec_and_test(&old_rd->refcount))
510 old_rd = NULL;
511 }
512
513 atomic_inc(&rd->refcount);
514 rq->rd = rd;
515
516 cpumask_set_cpu(rq->cpu, rd->span);
517 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
518 set_rq_online(rq);
519
520 rq_unlock_irqrestore(rq, &rf);
521
522 if (old_rd)
523 call_rcu(&old_rd->rcu, free_rootdomain);
524}
525
526void sched_get_rd(struct root_domain *rd)
527{
528 atomic_inc(&rd->refcount);
529}
530
531void sched_put_rd(struct root_domain *rd)
532{
533 if (!atomic_dec_and_test(&rd->refcount))
534 return;
535
536 call_rcu(&rd->rcu, free_rootdomain);
537}
538
539static int init_rootdomain(struct root_domain *rd)
540{
541 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
542 goto out;
543 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
544 goto free_span;
545 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
546 goto free_online;
547 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
548 goto free_dlo_mask;
549
550#ifdef HAVE_RT_PUSH_IPI
551 rd->rto_cpu = -1;
552 raw_spin_lock_init(&rd->rto_lock);
553 rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
554#endif
555
556 rd->visit_gen = 0;
557 init_dl_bw(&rd->dl_bw);
558 if (cpudl_init(&rd->cpudl) != 0)
559 goto free_rto_mask;
560
561 if (cpupri_init(&rd->cpupri) != 0)
562 goto free_cpudl;
563 return 0;
564
565free_cpudl:
566 cpudl_cleanup(&rd->cpudl);
567free_rto_mask:
568 free_cpumask_var(rd->rto_mask);
569free_dlo_mask:
570 free_cpumask_var(rd->dlo_mask);
571free_online:
572 free_cpumask_var(rd->online);
573free_span:
574 free_cpumask_var(rd->span);
575out:
576 return -ENOMEM;
577}
578
579/*
580 * By default the system creates a single root-domain with all CPUs as
581 * members (mimicking the global state we have today).
582 */
583struct root_domain def_root_domain;
584
585void __init init_defrootdomain(void)
586{
587 init_rootdomain(&def_root_domain);
588
589 atomic_set(&def_root_domain.refcount, 1);
590}
591
592static struct root_domain *alloc_rootdomain(void)
593{
594 struct root_domain *rd;
595
596 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
597 if (!rd)
598 return NULL;
599
600 if (init_rootdomain(rd) != 0) {
601 kfree(rd);
602 return NULL;
603 }
604
605 return rd;
606}
607
608static void free_sched_groups(struct sched_group *sg, int free_sgc)
609{
610 struct sched_group *tmp, *first;
611
612 if (!sg)
613 return;
614
615 first = sg;
616 do {
617 tmp = sg->next;
618
619 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
620 kfree(sg->sgc);
621
622 if (atomic_dec_and_test(&sg->ref))
623 kfree(sg);
624 sg = tmp;
625 } while (sg != first);
626}
627
628static void destroy_sched_domain(struct sched_domain *sd)
629{
630 /*
631 * A normal sched domain may have multiple group references, an
632 * overlapping domain, having private groups, only one. Iterate,
633 * dropping group/capacity references, freeing where none remain.
634 */
635 free_sched_groups(sd->groups, 1);
636
637 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
638 kfree(sd->shared);
639 kfree(sd);
640}
641
642static void destroy_sched_domains_rcu(struct rcu_head *rcu)
643{
644 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
645
646 while (sd) {
647 struct sched_domain *parent = sd->parent;
648 destroy_sched_domain(sd);
649 sd = parent;
650 }
651}
652
653static void destroy_sched_domains(struct sched_domain *sd)
654{
655 if (sd)
656 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
657}
658
659/*
660 * Keep a special pointer to the highest sched_domain that has
661 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
662 * allows us to avoid some pointer chasing select_idle_sibling().
663 *
664 * Also keep a unique ID per domain (we use the first CPU number in
665 * the cpumask of the domain), this allows us to quickly tell if
666 * two CPUs are in the same cache domain, see cpus_share_cache().
667 */
668DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
669DEFINE_PER_CPU(int, sd_llc_size);
670DEFINE_PER_CPU(int, sd_llc_id);
671DEFINE_PER_CPU(int, sd_share_id);
672DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
673DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
674DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
675DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
676
677DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
678DEFINE_STATIC_KEY_FALSE(sched_cluster_active);
679
680static void update_top_cache_domain(int cpu)
681{
682 struct sched_domain_shared *sds = NULL;
683 struct sched_domain *sd;
684 int id = cpu;
685 int size = 1;
686
687 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
688 if (sd) {
689 id = cpumask_first(sched_domain_span(sd));
690 size = cpumask_weight(sched_domain_span(sd));
691 sds = sd->shared;
692 }
693
694 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
695 per_cpu(sd_llc_size, cpu) = size;
696 per_cpu(sd_llc_id, cpu) = id;
697 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
698
699 sd = lowest_flag_domain(cpu, SD_CLUSTER);
700 if (sd)
701 id = cpumask_first(sched_domain_span(sd));
702
703 /*
704 * This assignment should be placed after the sd_llc_id as
705 * we want this id equals to cluster id on cluster machines
706 * but equals to LLC id on non-Cluster machines.
707 */
708 per_cpu(sd_share_id, cpu) = id;
709
710 sd = lowest_flag_domain(cpu, SD_NUMA);
711 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
712
713 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
714 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
715
716 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
717 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
718}
719
720/*
721 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
722 * hold the hotplug lock.
723 */
724static void
725cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
726{
727 struct rq *rq = cpu_rq(cpu);
728 struct sched_domain *tmp;
729
730 /* Remove the sched domains which do not contribute to scheduling. */
731 for (tmp = sd; tmp; ) {
732 struct sched_domain *parent = tmp->parent;
733 if (!parent)
734 break;
735
736 if (sd_parent_degenerate(tmp, parent)) {
737 tmp->parent = parent->parent;
738
739 if (parent->parent) {
740 parent->parent->child = tmp;
741 parent->parent->groups->flags = tmp->flags;
742 }
743
744 /*
745 * Transfer SD_PREFER_SIBLING down in case of a
746 * degenerate parent; the spans match for this
747 * so the property transfers.
748 */
749 if (parent->flags & SD_PREFER_SIBLING)
750 tmp->flags |= SD_PREFER_SIBLING;
751 destroy_sched_domain(parent);
752 } else
753 tmp = tmp->parent;
754 }
755
756 if (sd && sd_degenerate(sd)) {
757 tmp = sd;
758 sd = sd->parent;
759 destroy_sched_domain(tmp);
760 if (sd) {
761 struct sched_group *sg = sd->groups;
762
763 /*
764 * sched groups hold the flags of the child sched
765 * domain for convenience. Clear such flags since
766 * the child is being destroyed.
767 */
768 do {
769 sg->flags = 0;
770 } while (sg != sd->groups);
771
772 sd->child = NULL;
773 }
774 }
775
776 sched_domain_debug(sd, cpu);
777
778 rq_attach_root(rq, rd);
779 tmp = rq->sd;
780 rcu_assign_pointer(rq->sd, sd);
781 dirty_sched_domain_sysctl(cpu);
782 destroy_sched_domains(tmp);
783
784 update_top_cache_domain(cpu);
785}
786
787struct s_data {
788 struct sched_domain * __percpu *sd;
789 struct root_domain *rd;
790};
791
792enum s_alloc {
793 sa_rootdomain,
794 sa_sd,
795 sa_sd_storage,
796 sa_none,
797};
798
799/*
800 * Return the canonical balance CPU for this group, this is the first CPU
801 * of this group that's also in the balance mask.
802 *
803 * The balance mask are all those CPUs that could actually end up at this
804 * group. See build_balance_mask().
805 *
806 * Also see should_we_balance().
807 */
808int group_balance_cpu(struct sched_group *sg)
809{
810 return cpumask_first(group_balance_mask(sg));
811}
812
813
814/*
815 * NUMA topology (first read the regular topology blurb below)
816 *
817 * Given a node-distance table, for example:
818 *
819 * node 0 1 2 3
820 * 0: 10 20 30 20
821 * 1: 20 10 20 30
822 * 2: 30 20 10 20
823 * 3: 20 30 20 10
824 *
825 * which represents a 4 node ring topology like:
826 *
827 * 0 ----- 1
828 * | |
829 * | |
830 * | |
831 * 3 ----- 2
832 *
833 * We want to construct domains and groups to represent this. The way we go
834 * about doing this is to build the domains on 'hops'. For each NUMA level we
835 * construct the mask of all nodes reachable in @level hops.
836 *
837 * For the above NUMA topology that gives 3 levels:
838 *
839 * NUMA-2 0-3 0-3 0-3 0-3
840 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
841 *
842 * NUMA-1 0-1,3 0-2 1-3 0,2-3
843 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
844 *
845 * NUMA-0 0 1 2 3
846 *
847 *
848 * As can be seen; things don't nicely line up as with the regular topology.
849 * When we iterate a domain in child domain chunks some nodes can be
850 * represented multiple times -- hence the "overlap" naming for this part of
851 * the topology.
852 *
853 * In order to minimize this overlap, we only build enough groups to cover the
854 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
855 *
856 * Because:
857 *
858 * - the first group of each domain is its child domain; this
859 * gets us the first 0-1,3
860 * - the only uncovered node is 2, who's child domain is 1-3.
861 *
862 * However, because of the overlap, computing a unique CPU for each group is
863 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
864 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
865 * end up at those groups (they would end up in group: 0-1,3).
866 *
867 * To correct this we have to introduce the group balance mask. This mask
868 * will contain those CPUs in the group that can reach this group given the
869 * (child) domain tree.
870 *
871 * With this we can once again compute balance_cpu and sched_group_capacity
872 * relations.
873 *
874 * XXX include words on how balance_cpu is unique and therefore can be
875 * used for sched_group_capacity links.
876 *
877 *
878 * Another 'interesting' topology is:
879 *
880 * node 0 1 2 3
881 * 0: 10 20 20 30
882 * 1: 20 10 20 20
883 * 2: 20 20 10 20
884 * 3: 30 20 20 10
885 *
886 * Which looks a little like:
887 *
888 * 0 ----- 1
889 * | / |
890 * | / |
891 * | / |
892 * 2 ----- 3
893 *
894 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
895 * are not.
896 *
897 * This leads to a few particularly weird cases where the sched_domain's are
898 * not of the same number for each CPU. Consider:
899 *
900 * NUMA-2 0-3 0-3
901 * groups: {0-2},{1-3} {1-3},{0-2}
902 *
903 * NUMA-1 0-2 0-3 0-3 1-3
904 *
905 * NUMA-0 0 1 2 3
906 *
907 */
908
909
910/*
911 * Build the balance mask; it contains only those CPUs that can arrive at this
912 * group and should be considered to continue balancing.
913 *
914 * We do this during the group creation pass, therefore the group information
915 * isn't complete yet, however since each group represents a (child) domain we
916 * can fully construct this using the sched_domain bits (which are already
917 * complete).
918 */
919static void
920build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
921{
922 const struct cpumask *sg_span = sched_group_span(sg);
923 struct sd_data *sdd = sd->private;
924 struct sched_domain *sibling;
925 int i;
926
927 cpumask_clear(mask);
928
929 for_each_cpu(i, sg_span) {
930 sibling = *per_cpu_ptr(sdd->sd, i);
931
932 /*
933 * Can happen in the asymmetric case, where these siblings are
934 * unused. The mask will not be empty because those CPUs that
935 * do have the top domain _should_ span the domain.
936 */
937 if (!sibling->child)
938 continue;
939
940 /* If we would not end up here, we can't continue from here */
941 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
942 continue;
943
944 cpumask_set_cpu(i, mask);
945 }
946
947 /* We must not have empty masks here */
948 WARN_ON_ONCE(cpumask_empty(mask));
949}
950
951/*
952 * XXX: This creates per-node group entries; since the load-balancer will
953 * immediately access remote memory to construct this group's load-balance
954 * statistics having the groups node local is of dubious benefit.
955 */
956static struct sched_group *
957build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
958{
959 struct sched_group *sg;
960 struct cpumask *sg_span;
961
962 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
963 GFP_KERNEL, cpu_to_node(cpu));
964
965 if (!sg)
966 return NULL;
967
968 sg_span = sched_group_span(sg);
969 if (sd->child) {
970 cpumask_copy(sg_span, sched_domain_span(sd->child));
971 sg->flags = sd->child->flags;
972 } else {
973 cpumask_copy(sg_span, sched_domain_span(sd));
974 }
975
976 atomic_inc(&sg->ref);
977 return sg;
978}
979
980static void init_overlap_sched_group(struct sched_domain *sd,
981 struct sched_group *sg)
982{
983 struct cpumask *mask = sched_domains_tmpmask2;
984 struct sd_data *sdd = sd->private;
985 struct cpumask *sg_span;
986 int cpu;
987
988 build_balance_mask(sd, sg, mask);
989 cpu = cpumask_first(mask);
990
991 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
992 if (atomic_inc_return(&sg->sgc->ref) == 1)
993 cpumask_copy(group_balance_mask(sg), mask);
994 else
995 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
996
997 /*
998 * Initialize sgc->capacity such that even if we mess up the
999 * domains and no possible iteration will get us here, we won't
1000 * die on a /0 trap.
1001 */
1002 sg_span = sched_group_span(sg);
1003 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
1004 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1005 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1006}
1007
1008static struct sched_domain *
1009find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
1010{
1011 /*
1012 * The proper descendant would be the one whose child won't span out
1013 * of sd
1014 */
1015 while (sibling->child &&
1016 !cpumask_subset(sched_domain_span(sibling->child),
1017 sched_domain_span(sd)))
1018 sibling = sibling->child;
1019
1020 /*
1021 * As we are referencing sgc across different topology level, we need
1022 * to go down to skip those sched_domains which don't contribute to
1023 * scheduling because they will be degenerated in cpu_attach_domain
1024 */
1025 while (sibling->child &&
1026 cpumask_equal(sched_domain_span(sibling->child),
1027 sched_domain_span(sibling)))
1028 sibling = sibling->child;
1029
1030 return sibling;
1031}
1032
1033static int
1034build_overlap_sched_groups(struct sched_domain *sd, int cpu)
1035{
1036 struct sched_group *first = NULL, *last = NULL, *sg;
1037 const struct cpumask *span = sched_domain_span(sd);
1038 struct cpumask *covered = sched_domains_tmpmask;
1039 struct sd_data *sdd = sd->private;
1040 struct sched_domain *sibling;
1041 int i;
1042
1043 cpumask_clear(covered);
1044
1045 for_each_cpu_wrap(i, span, cpu) {
1046 struct cpumask *sg_span;
1047
1048 if (cpumask_test_cpu(i, covered))
1049 continue;
1050
1051 sibling = *per_cpu_ptr(sdd->sd, i);
1052
1053 /*
1054 * Asymmetric node setups can result in situations where the
1055 * domain tree is of unequal depth, make sure to skip domains
1056 * that already cover the entire range.
1057 *
1058 * In that case build_sched_domains() will have terminated the
1059 * iteration early and our sibling sd spans will be empty.
1060 * Domains should always include the CPU they're built on, so
1061 * check that.
1062 */
1063 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1064 continue;
1065
1066 /*
1067 * Usually we build sched_group by sibling's child sched_domain
1068 * But for machines whose NUMA diameter are 3 or above, we move
1069 * to build sched_group by sibling's proper descendant's child
1070 * domain because sibling's child sched_domain will span out of
1071 * the sched_domain being built as below.
1072 *
1073 * Smallest diameter=3 topology is:
1074 *
1075 * node 0 1 2 3
1076 * 0: 10 20 30 40
1077 * 1: 20 10 20 30
1078 * 2: 30 20 10 20
1079 * 3: 40 30 20 10
1080 *
1081 * 0 --- 1 --- 2 --- 3
1082 *
1083 * NUMA-3 0-3 N/A N/A 0-3
1084 * groups: {0-2},{1-3} {1-3},{0-2}
1085 *
1086 * NUMA-2 0-2 0-3 0-3 1-3
1087 * groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2}
1088 *
1089 * NUMA-1 0-1 0-2 1-3 2-3
1090 * groups: {0},{1} {1},{2},{0} {2},{3},{1} {3},{2}
1091 *
1092 * NUMA-0 0 1 2 3
1093 *
1094 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1095 * group span isn't a subset of the domain span.
1096 */
1097 if (sibling->child &&
1098 !cpumask_subset(sched_domain_span(sibling->child), span))
1099 sibling = find_descended_sibling(sd, sibling);
1100
1101 sg = build_group_from_child_sched_domain(sibling, cpu);
1102 if (!sg)
1103 goto fail;
1104
1105 sg_span = sched_group_span(sg);
1106 cpumask_or(covered, covered, sg_span);
1107
1108 init_overlap_sched_group(sibling, sg);
1109
1110 if (!first)
1111 first = sg;
1112 if (last)
1113 last->next = sg;
1114 last = sg;
1115 last->next = first;
1116 }
1117 sd->groups = first;
1118
1119 return 0;
1120
1121fail:
1122 free_sched_groups(first, 0);
1123
1124 return -ENOMEM;
1125}
1126
1127
1128/*
1129 * Package topology (also see the load-balance blurb in fair.c)
1130 *
1131 * The scheduler builds a tree structure to represent a number of important
1132 * topology features. By default (default_topology[]) these include:
1133 *
1134 * - Simultaneous multithreading (SMT)
1135 * - Multi-Core Cache (MC)
1136 * - Package (PKG)
1137 *
1138 * Where the last one more or less denotes everything up to a NUMA node.
1139 *
1140 * The tree consists of 3 primary data structures:
1141 *
1142 * sched_domain -> sched_group -> sched_group_capacity
1143 * ^ ^ ^ ^
1144 * `-' `-'
1145 *
1146 * The sched_domains are per-CPU and have a two way link (parent & child) and
1147 * denote the ever growing mask of CPUs belonging to that level of topology.
1148 *
1149 * Each sched_domain has a circular (double) linked list of sched_group's, each
1150 * denoting the domains of the level below (or individual CPUs in case of the
1151 * first domain level). The sched_group linked by a sched_domain includes the
1152 * CPU of that sched_domain [*].
1153 *
1154 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1155 *
1156 * CPU 0 1 2 3 4 5 6 7
1157 *
1158 * PKG [ ]
1159 * MC [ ] [ ]
1160 * SMT [ ] [ ] [ ] [ ]
1161 *
1162 * - or -
1163 *
1164 * PKG 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1165 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1166 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1167 *
1168 * CPU 0 1 2 3 4 5 6 7
1169 *
1170 * One way to think about it is: sched_domain moves you up and down among these
1171 * topology levels, while sched_group moves you sideways through it, at child
1172 * domain granularity.
1173 *
1174 * sched_group_capacity ensures each unique sched_group has shared storage.
1175 *
1176 * There are two related construction problems, both require a CPU that
1177 * uniquely identify each group (for a given domain):
1178 *
1179 * - The first is the balance_cpu (see should_we_balance() and the
1180 * load-balance blub in fair.c); for each group we only want 1 CPU to
1181 * continue balancing at a higher domain.
1182 *
1183 * - The second is the sched_group_capacity; we want all identical groups
1184 * to share a single sched_group_capacity.
1185 *
1186 * Since these topologies are exclusive by construction. That is, its
1187 * impossible for an SMT thread to belong to multiple cores, and cores to
1188 * be part of multiple caches. There is a very clear and unique location
1189 * for each CPU in the hierarchy.
1190 *
1191 * Therefore computing a unique CPU for each group is trivial (the iteration
1192 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1193 * group), we can simply pick the first CPU in each group.
1194 *
1195 *
1196 * [*] in other words, the first group of each domain is its child domain.
1197 */
1198
1199static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1200{
1201 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1202 struct sched_domain *child = sd->child;
1203 struct sched_group *sg;
1204 bool already_visited;
1205
1206 if (child)
1207 cpu = cpumask_first(sched_domain_span(child));
1208
1209 sg = *per_cpu_ptr(sdd->sg, cpu);
1210 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1211
1212 /* Increase refcounts for claim_allocations: */
1213 already_visited = atomic_inc_return(&sg->ref) > 1;
1214 /* sgc visits should follow a similar trend as sg */
1215 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1216
1217 /* If we have already visited that group, it's already initialized. */
1218 if (already_visited)
1219 return sg;
1220
1221 if (child) {
1222 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1223 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1224 sg->flags = child->flags;
1225 } else {
1226 cpumask_set_cpu(cpu, sched_group_span(sg));
1227 cpumask_set_cpu(cpu, group_balance_mask(sg));
1228 }
1229
1230 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1231 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1232 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1233
1234 return sg;
1235}
1236
1237/*
1238 * build_sched_groups will build a circular linked list of the groups
1239 * covered by the given span, will set each group's ->cpumask correctly,
1240 * and will initialize their ->sgc.
1241 *
1242 * Assumes the sched_domain tree is fully constructed
1243 */
1244static int
1245build_sched_groups(struct sched_domain *sd, int cpu)
1246{
1247 struct sched_group *first = NULL, *last = NULL;
1248 struct sd_data *sdd = sd->private;
1249 const struct cpumask *span = sched_domain_span(sd);
1250 struct cpumask *covered;
1251 int i;
1252
1253 lockdep_assert_held(&sched_domains_mutex);
1254 covered = sched_domains_tmpmask;
1255
1256 cpumask_clear(covered);
1257
1258 for_each_cpu_wrap(i, span, cpu) {
1259 struct sched_group *sg;
1260
1261 if (cpumask_test_cpu(i, covered))
1262 continue;
1263
1264 sg = get_group(i, sdd);
1265
1266 cpumask_or(covered, covered, sched_group_span(sg));
1267
1268 if (!first)
1269 first = sg;
1270 if (last)
1271 last->next = sg;
1272 last = sg;
1273 }
1274 last->next = first;
1275 sd->groups = first;
1276
1277 return 0;
1278}
1279
1280/*
1281 * Initialize sched groups cpu_capacity.
1282 *
1283 * cpu_capacity indicates the capacity of sched group, which is used while
1284 * distributing the load between different sched groups in a sched domain.
1285 * Typically cpu_capacity for all the groups in a sched domain will be same
1286 * unless there are asymmetries in the topology. If there are asymmetries,
1287 * group having more cpu_capacity will pickup more load compared to the
1288 * group having less cpu_capacity.
1289 */
1290static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1291{
1292 struct sched_group *sg = sd->groups;
1293 struct cpumask *mask = sched_domains_tmpmask2;
1294
1295 WARN_ON(!sg);
1296
1297 do {
1298 int cpu, cores = 0, max_cpu = -1;
1299
1300 sg->group_weight = cpumask_weight(sched_group_span(sg));
1301
1302 cpumask_copy(mask, sched_group_span(sg));
1303 for_each_cpu(cpu, mask) {
1304 cores++;
1305#ifdef CONFIG_SCHED_SMT
1306 cpumask_andnot(mask, mask, cpu_smt_mask(cpu));
1307#endif
1308 }
1309 sg->cores = cores;
1310
1311 if (!(sd->flags & SD_ASYM_PACKING))
1312 goto next;
1313
1314 for_each_cpu(cpu, sched_group_span(sg)) {
1315 if (max_cpu < 0)
1316 max_cpu = cpu;
1317 else if (sched_asym_prefer(cpu, max_cpu))
1318 max_cpu = cpu;
1319 }
1320 sg->asym_prefer_cpu = max_cpu;
1321
1322next:
1323 sg = sg->next;
1324 } while (sg != sd->groups);
1325
1326 if (cpu != group_balance_cpu(sg))
1327 return;
1328
1329 update_group_capacity(sd, cpu);
1330}
1331
1332/*
1333 * Asymmetric CPU capacity bits
1334 */
1335struct asym_cap_data {
1336 struct list_head link;
1337 unsigned long capacity;
1338 unsigned long cpus[];
1339};
1340
1341/*
1342 * Set of available CPUs grouped by their corresponding capacities
1343 * Each list entry contains a CPU mask reflecting CPUs that share the same
1344 * capacity.
1345 * The lifespan of data is unlimited.
1346 */
1347static LIST_HEAD(asym_cap_list);
1348
1349#define cpu_capacity_span(asym_data) to_cpumask((asym_data)->cpus)
1350
1351/*
1352 * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1353 * Provides sd_flags reflecting the asymmetry scope.
1354 */
1355static inline int
1356asym_cpu_capacity_classify(const struct cpumask *sd_span,
1357 const struct cpumask *cpu_map)
1358{
1359 struct asym_cap_data *entry;
1360 int count = 0, miss = 0;
1361
1362 /*
1363 * Count how many unique CPU capacities this domain spans across
1364 * (compare sched_domain CPUs mask with ones representing available
1365 * CPUs capacities). Take into account CPUs that might be offline:
1366 * skip those.
1367 */
1368 list_for_each_entry(entry, &asym_cap_list, link) {
1369 if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1370 ++count;
1371 else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1372 ++miss;
1373 }
1374
1375 WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1376
1377 /* No asymmetry detected */
1378 if (count < 2)
1379 return 0;
1380 /* Some of the available CPU capacity values have not been detected */
1381 if (miss)
1382 return SD_ASYM_CPUCAPACITY;
1383
1384 /* Full asymmetry */
1385 return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1386
1387}
1388
1389static inline void asym_cpu_capacity_update_data(int cpu)
1390{
1391 unsigned long capacity = arch_scale_cpu_capacity(cpu);
1392 struct asym_cap_data *entry = NULL;
1393
1394 list_for_each_entry(entry, &asym_cap_list, link) {
1395 if (capacity == entry->capacity)
1396 goto done;
1397 }
1398
1399 entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1400 if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1401 return;
1402 entry->capacity = capacity;
1403 list_add(&entry->link, &asym_cap_list);
1404done:
1405 __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1406}
1407
1408/*
1409 * Build-up/update list of CPUs grouped by their capacities
1410 * An update requires explicit request to rebuild sched domains
1411 * with state indicating CPU topology changes.
1412 */
1413static void asym_cpu_capacity_scan(void)
1414{
1415 struct asym_cap_data *entry, *next;
1416 int cpu;
1417
1418 list_for_each_entry(entry, &asym_cap_list, link)
1419 cpumask_clear(cpu_capacity_span(entry));
1420
1421 for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1422 asym_cpu_capacity_update_data(cpu);
1423
1424 list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1425 if (cpumask_empty(cpu_capacity_span(entry))) {
1426 list_del(&entry->link);
1427 kfree(entry);
1428 }
1429 }
1430
1431 /*
1432 * Only one capacity value has been detected i.e. this system is symmetric.
1433 * No need to keep this data around.
1434 */
1435 if (list_is_singular(&asym_cap_list)) {
1436 entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1437 list_del(&entry->link);
1438 kfree(entry);
1439 }
1440}
1441
1442/*
1443 * Initializers for schedule domains
1444 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1445 */
1446
1447static int default_relax_domain_level = -1;
1448int sched_domain_level_max;
1449
1450static int __init setup_relax_domain_level(char *str)
1451{
1452 if (kstrtoint(str, 0, &default_relax_domain_level))
1453 pr_warn("Unable to set relax_domain_level\n");
1454
1455 return 1;
1456}
1457__setup("relax_domain_level=", setup_relax_domain_level);
1458
1459static void set_domain_attribute(struct sched_domain *sd,
1460 struct sched_domain_attr *attr)
1461{
1462 int request;
1463
1464 if (!attr || attr->relax_domain_level < 0) {
1465 if (default_relax_domain_level < 0)
1466 return;
1467 request = default_relax_domain_level;
1468 } else
1469 request = attr->relax_domain_level;
1470
1471 if (sd->level > request) {
1472 /* Turn off idle balance on this domain: */
1473 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1474 }
1475}
1476
1477static void __sdt_free(const struct cpumask *cpu_map);
1478static int __sdt_alloc(const struct cpumask *cpu_map);
1479
1480static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1481 const struct cpumask *cpu_map)
1482{
1483 switch (what) {
1484 case sa_rootdomain:
1485 if (!atomic_read(&d->rd->refcount))
1486 free_rootdomain(&d->rd->rcu);
1487 fallthrough;
1488 case sa_sd:
1489 free_percpu(d->sd);
1490 fallthrough;
1491 case sa_sd_storage:
1492 __sdt_free(cpu_map);
1493 fallthrough;
1494 case sa_none:
1495 break;
1496 }
1497}
1498
1499static enum s_alloc
1500__visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1501{
1502 memset(d, 0, sizeof(*d));
1503
1504 if (__sdt_alloc(cpu_map))
1505 return sa_sd_storage;
1506 d->sd = alloc_percpu(struct sched_domain *);
1507 if (!d->sd)
1508 return sa_sd_storage;
1509 d->rd = alloc_rootdomain();
1510 if (!d->rd)
1511 return sa_sd;
1512
1513 return sa_rootdomain;
1514}
1515
1516/*
1517 * NULL the sd_data elements we've used to build the sched_domain and
1518 * sched_group structure so that the subsequent __free_domain_allocs()
1519 * will not free the data we're using.
1520 */
1521static void claim_allocations(int cpu, struct sched_domain *sd)
1522{
1523 struct sd_data *sdd = sd->private;
1524
1525 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1526 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1527
1528 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1529 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1530
1531 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1532 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1533
1534 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1535 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1536}
1537
1538#ifdef CONFIG_NUMA
1539enum numa_topology_type sched_numa_topology_type;
1540
1541static int sched_domains_numa_levels;
1542static int sched_domains_curr_level;
1543
1544int sched_max_numa_distance;
1545static int *sched_domains_numa_distance;
1546static struct cpumask ***sched_domains_numa_masks;
1547#endif
1548
1549/*
1550 * SD_flags allowed in topology descriptions.
1551 *
1552 * These flags are purely descriptive of the topology and do not prescribe
1553 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1554 * function:
1555 *
1556 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1557 * SD_SHARE_PKG_RESOURCES - describes shared caches
1558 * SD_NUMA - describes NUMA topologies
1559 *
1560 * Odd one out, which beside describing the topology has a quirk also
1561 * prescribes the desired behaviour that goes along with it:
1562 *
1563 * SD_ASYM_PACKING - describes SMT quirks
1564 */
1565#define TOPOLOGY_SD_FLAGS \
1566 (SD_SHARE_CPUCAPACITY | \
1567 SD_CLUSTER | \
1568 SD_SHARE_PKG_RESOURCES | \
1569 SD_NUMA | \
1570 SD_ASYM_PACKING)
1571
1572static struct sched_domain *
1573sd_init(struct sched_domain_topology_level *tl,
1574 const struct cpumask *cpu_map,
1575 struct sched_domain *child, int cpu)
1576{
1577 struct sd_data *sdd = &tl->data;
1578 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1579 int sd_id, sd_weight, sd_flags = 0;
1580 struct cpumask *sd_span;
1581
1582#ifdef CONFIG_NUMA
1583 /*
1584 * Ugly hack to pass state to sd_numa_mask()...
1585 */
1586 sched_domains_curr_level = tl->numa_level;
1587#endif
1588
1589 sd_weight = cpumask_weight(tl->mask(cpu));
1590
1591 if (tl->sd_flags)
1592 sd_flags = (*tl->sd_flags)();
1593 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1594 "wrong sd_flags in topology description\n"))
1595 sd_flags &= TOPOLOGY_SD_FLAGS;
1596
1597 *sd = (struct sched_domain){
1598 .min_interval = sd_weight,
1599 .max_interval = 2*sd_weight,
1600 .busy_factor = 16,
1601 .imbalance_pct = 117,
1602
1603 .cache_nice_tries = 0,
1604
1605 .flags = 1*SD_BALANCE_NEWIDLE
1606 | 1*SD_BALANCE_EXEC
1607 | 1*SD_BALANCE_FORK
1608 | 0*SD_BALANCE_WAKE
1609 | 1*SD_WAKE_AFFINE
1610 | 0*SD_SHARE_CPUCAPACITY
1611 | 0*SD_SHARE_PKG_RESOURCES
1612 | 0*SD_SERIALIZE
1613 | 1*SD_PREFER_SIBLING
1614 | 0*SD_NUMA
1615 | sd_flags
1616 ,
1617
1618 .last_balance = jiffies,
1619 .balance_interval = sd_weight,
1620 .max_newidle_lb_cost = 0,
1621 .last_decay_max_lb_cost = jiffies,
1622 .child = child,
1623#ifdef CONFIG_SCHED_DEBUG
1624 .name = tl->name,
1625#endif
1626 };
1627
1628 sd_span = sched_domain_span(sd);
1629 cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1630 sd_id = cpumask_first(sd_span);
1631
1632 sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1633
1634 WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1635 (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1636 "CPU capacity asymmetry not supported on SMT\n");
1637
1638 /*
1639 * Convert topological properties into behaviour.
1640 */
1641 /* Don't attempt to spread across CPUs of different capacities. */
1642 if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1643 sd->child->flags &= ~SD_PREFER_SIBLING;
1644
1645 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1646 sd->imbalance_pct = 110;
1647
1648 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1649 sd->imbalance_pct = 117;
1650 sd->cache_nice_tries = 1;
1651
1652#ifdef CONFIG_NUMA
1653 } else if (sd->flags & SD_NUMA) {
1654 sd->cache_nice_tries = 2;
1655
1656 sd->flags &= ~SD_PREFER_SIBLING;
1657 sd->flags |= SD_SERIALIZE;
1658 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1659 sd->flags &= ~(SD_BALANCE_EXEC |
1660 SD_BALANCE_FORK |
1661 SD_WAKE_AFFINE);
1662 }
1663
1664#endif
1665 } else {
1666 sd->cache_nice_tries = 1;
1667 }
1668
1669 /*
1670 * For all levels sharing cache; connect a sched_domain_shared
1671 * instance.
1672 */
1673 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1674 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1675 atomic_inc(&sd->shared->ref);
1676 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1677 }
1678
1679 sd->private = sdd;
1680
1681 return sd;
1682}
1683
1684/*
1685 * Topology list, bottom-up.
1686 */
1687static struct sched_domain_topology_level default_topology[] = {
1688#ifdef CONFIG_SCHED_SMT
1689 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1690#endif
1691
1692#ifdef CONFIG_SCHED_CLUSTER
1693 { cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1694#endif
1695
1696#ifdef CONFIG_SCHED_MC
1697 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1698#endif
1699 { cpu_cpu_mask, SD_INIT_NAME(PKG) },
1700 { NULL, },
1701};
1702
1703static struct sched_domain_topology_level *sched_domain_topology =
1704 default_topology;
1705static struct sched_domain_topology_level *sched_domain_topology_saved;
1706
1707#define for_each_sd_topology(tl) \
1708 for (tl = sched_domain_topology; tl->mask; tl++)
1709
1710void __init set_sched_topology(struct sched_domain_topology_level *tl)
1711{
1712 if (WARN_ON_ONCE(sched_smp_initialized))
1713 return;
1714
1715 sched_domain_topology = tl;
1716 sched_domain_topology_saved = NULL;
1717}
1718
1719#ifdef CONFIG_NUMA
1720
1721static const struct cpumask *sd_numa_mask(int cpu)
1722{
1723 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1724}
1725
1726static void sched_numa_warn(const char *str)
1727{
1728 static int done = false;
1729 int i,j;
1730
1731 if (done)
1732 return;
1733
1734 done = true;
1735
1736 printk(KERN_WARNING "ERROR: %s\n\n", str);
1737
1738 for (i = 0; i < nr_node_ids; i++) {
1739 printk(KERN_WARNING " ");
1740 for (j = 0; j < nr_node_ids; j++) {
1741 if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
1742 printk(KERN_CONT "(%02d) ", node_distance(i,j));
1743 else
1744 printk(KERN_CONT " %02d ", node_distance(i,j));
1745 }
1746 printk(KERN_CONT "\n");
1747 }
1748 printk(KERN_WARNING "\n");
1749}
1750
1751bool find_numa_distance(int distance)
1752{
1753 bool found = false;
1754 int i, *distances;
1755
1756 if (distance == node_distance(0, 0))
1757 return true;
1758
1759 rcu_read_lock();
1760 distances = rcu_dereference(sched_domains_numa_distance);
1761 if (!distances)
1762 goto unlock;
1763 for (i = 0; i < sched_domains_numa_levels; i++) {
1764 if (distances[i] == distance) {
1765 found = true;
1766 break;
1767 }
1768 }
1769unlock:
1770 rcu_read_unlock();
1771
1772 return found;
1773}
1774
1775#define for_each_cpu_node_but(n, nbut) \
1776 for_each_node_state(n, N_CPU) \
1777 if (n == nbut) \
1778 continue; \
1779 else
1780
1781/*
1782 * A system can have three types of NUMA topology:
1783 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1784 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1785 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1786 *
1787 * The difference between a glueless mesh topology and a backplane
1788 * topology lies in whether communication between not directly
1789 * connected nodes goes through intermediary nodes (where programs
1790 * could run), or through backplane controllers. This affects
1791 * placement of programs.
1792 *
1793 * The type of topology can be discerned with the following tests:
1794 * - If the maximum distance between any nodes is 1 hop, the system
1795 * is directly connected.
1796 * - If for two nodes A and B, located N > 1 hops away from each other,
1797 * there is an intermediary node C, which is < N hops away from both
1798 * nodes A and B, the system is a glueless mesh.
1799 */
1800static void init_numa_topology_type(int offline_node)
1801{
1802 int a, b, c, n;
1803
1804 n = sched_max_numa_distance;
1805
1806 if (sched_domains_numa_levels <= 2) {
1807 sched_numa_topology_type = NUMA_DIRECT;
1808 return;
1809 }
1810
1811 for_each_cpu_node_but(a, offline_node) {
1812 for_each_cpu_node_but(b, offline_node) {
1813 /* Find two nodes furthest removed from each other. */
1814 if (node_distance(a, b) < n)
1815 continue;
1816
1817 /* Is there an intermediary node between a and b? */
1818 for_each_cpu_node_but(c, offline_node) {
1819 if (node_distance(a, c) < n &&
1820 node_distance(b, c) < n) {
1821 sched_numa_topology_type =
1822 NUMA_GLUELESS_MESH;
1823 return;
1824 }
1825 }
1826
1827 sched_numa_topology_type = NUMA_BACKPLANE;
1828 return;
1829 }
1830 }
1831
1832 pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1833 sched_numa_topology_type = NUMA_DIRECT;
1834}
1835
1836
1837#define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1838
1839void sched_init_numa(int offline_node)
1840{
1841 struct sched_domain_topology_level *tl;
1842 unsigned long *distance_map;
1843 int nr_levels = 0;
1844 int i, j;
1845 int *distances;
1846 struct cpumask ***masks;
1847
1848 /*
1849 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1850 * unique distances in the node_distance() table.
1851 */
1852 distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1853 if (!distance_map)
1854 return;
1855
1856 bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1857 for_each_cpu_node_but(i, offline_node) {
1858 for_each_cpu_node_but(j, offline_node) {
1859 int distance = node_distance(i, j);
1860
1861 if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1862 sched_numa_warn("Invalid distance value range");
1863 bitmap_free(distance_map);
1864 return;
1865 }
1866
1867 bitmap_set(distance_map, distance, 1);
1868 }
1869 }
1870 /*
1871 * We can now figure out how many unique distance values there are and
1872 * allocate memory accordingly.
1873 */
1874 nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1875
1876 distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1877 if (!distances) {
1878 bitmap_free(distance_map);
1879 return;
1880 }
1881
1882 for (i = 0, j = 0; i < nr_levels; i++, j++) {
1883 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1884 distances[i] = j;
1885 }
1886 rcu_assign_pointer(sched_domains_numa_distance, distances);
1887
1888 bitmap_free(distance_map);
1889
1890 /*
1891 * 'nr_levels' contains the number of unique distances
1892 *
1893 * The sched_domains_numa_distance[] array includes the actual distance
1894 * numbers.
1895 */
1896
1897 /*
1898 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1899 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1900 * the array will contain less then 'nr_levels' members. This could be
1901 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1902 * in other functions.
1903 *
1904 * We reset it to 'nr_levels' at the end of this function.
1905 */
1906 sched_domains_numa_levels = 0;
1907
1908 masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1909 if (!masks)
1910 return;
1911
1912 /*
1913 * Now for each level, construct a mask per node which contains all
1914 * CPUs of nodes that are that many hops away from us.
1915 */
1916 for (i = 0; i < nr_levels; i++) {
1917 masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1918 if (!masks[i])
1919 return;
1920
1921 for_each_cpu_node_but(j, offline_node) {
1922 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1923 int k;
1924
1925 if (!mask)
1926 return;
1927
1928 masks[i][j] = mask;
1929
1930 for_each_cpu_node_but(k, offline_node) {
1931 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1932 sched_numa_warn("Node-distance not symmetric");
1933
1934 if (node_distance(j, k) > sched_domains_numa_distance[i])
1935 continue;
1936
1937 cpumask_or(mask, mask, cpumask_of_node(k));
1938 }
1939 }
1940 }
1941 rcu_assign_pointer(sched_domains_numa_masks, masks);
1942
1943 /* Compute default topology size */
1944 for (i = 0; sched_domain_topology[i].mask; i++);
1945
1946 tl = kzalloc((i + nr_levels + 1) *
1947 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1948 if (!tl)
1949 return;
1950
1951 /*
1952 * Copy the default topology bits..
1953 */
1954 for (i = 0; sched_domain_topology[i].mask; i++)
1955 tl[i] = sched_domain_topology[i];
1956
1957 /*
1958 * Add the NUMA identity distance, aka single NODE.
1959 */
1960 tl[i++] = (struct sched_domain_topology_level){
1961 .mask = sd_numa_mask,
1962 .numa_level = 0,
1963 SD_INIT_NAME(NODE)
1964 };
1965
1966 /*
1967 * .. and append 'j' levels of NUMA goodness.
1968 */
1969 for (j = 1; j < nr_levels; i++, j++) {
1970 tl[i] = (struct sched_domain_topology_level){
1971 .mask = sd_numa_mask,
1972 .sd_flags = cpu_numa_flags,
1973 .flags = SDTL_OVERLAP,
1974 .numa_level = j,
1975 SD_INIT_NAME(NUMA)
1976 };
1977 }
1978
1979 sched_domain_topology_saved = sched_domain_topology;
1980 sched_domain_topology = tl;
1981
1982 sched_domains_numa_levels = nr_levels;
1983 WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]);
1984
1985 init_numa_topology_type(offline_node);
1986}
1987
1988
1989static void sched_reset_numa(void)
1990{
1991 int nr_levels, *distances;
1992 struct cpumask ***masks;
1993
1994 nr_levels = sched_domains_numa_levels;
1995 sched_domains_numa_levels = 0;
1996 sched_max_numa_distance = 0;
1997 sched_numa_topology_type = NUMA_DIRECT;
1998 distances = sched_domains_numa_distance;
1999 rcu_assign_pointer(sched_domains_numa_distance, NULL);
2000 masks = sched_domains_numa_masks;
2001 rcu_assign_pointer(sched_domains_numa_masks, NULL);
2002 if (distances || masks) {
2003 int i, j;
2004
2005 synchronize_rcu();
2006 kfree(distances);
2007 for (i = 0; i < nr_levels && masks; i++) {
2008 if (!masks[i])
2009 continue;
2010 for_each_node(j)
2011 kfree(masks[i][j]);
2012 kfree(masks[i]);
2013 }
2014 kfree(masks);
2015 }
2016 if (sched_domain_topology_saved) {
2017 kfree(sched_domain_topology);
2018 sched_domain_topology = sched_domain_topology_saved;
2019 sched_domain_topology_saved = NULL;
2020 }
2021}
2022
2023/*
2024 * Call with hotplug lock held
2025 */
2026void sched_update_numa(int cpu, bool online)
2027{
2028 int node;
2029
2030 node = cpu_to_node(cpu);
2031 /*
2032 * Scheduler NUMA topology is updated when the first CPU of a
2033 * node is onlined or the last CPU of a node is offlined.
2034 */
2035 if (cpumask_weight(cpumask_of_node(node)) != 1)
2036 return;
2037
2038 sched_reset_numa();
2039 sched_init_numa(online ? NUMA_NO_NODE : node);
2040}
2041
2042void sched_domains_numa_masks_set(unsigned int cpu)
2043{
2044 int node = cpu_to_node(cpu);
2045 int i, j;
2046
2047 for (i = 0; i < sched_domains_numa_levels; i++) {
2048 for (j = 0; j < nr_node_ids; j++) {
2049 if (!node_state(j, N_CPU))
2050 continue;
2051
2052 /* Set ourselves in the remote node's masks */
2053 if (node_distance(j, node) <= sched_domains_numa_distance[i])
2054 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
2055 }
2056 }
2057}
2058
2059void sched_domains_numa_masks_clear(unsigned int cpu)
2060{
2061 int i, j;
2062
2063 for (i = 0; i < sched_domains_numa_levels; i++) {
2064 for (j = 0; j < nr_node_ids; j++) {
2065 if (sched_domains_numa_masks[i][j])
2066 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
2067 }
2068 }
2069}
2070
2071/*
2072 * sched_numa_find_closest() - given the NUMA topology, find the cpu
2073 * closest to @cpu from @cpumask.
2074 * cpumask: cpumask to find a cpu from
2075 * cpu: cpu to be close to
2076 *
2077 * returns: cpu, or nr_cpu_ids when nothing found.
2078 */
2079int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2080{
2081 int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2082 struct cpumask ***masks;
2083
2084 rcu_read_lock();
2085 masks = rcu_dereference(sched_domains_numa_masks);
2086 if (!masks)
2087 goto unlock;
2088 for (i = 0; i < sched_domains_numa_levels; i++) {
2089 if (!masks[i][j])
2090 break;
2091 cpu = cpumask_any_and(cpus, masks[i][j]);
2092 if (cpu < nr_cpu_ids) {
2093 found = cpu;
2094 break;
2095 }
2096 }
2097unlock:
2098 rcu_read_unlock();
2099
2100 return found;
2101}
2102
2103struct __cmp_key {
2104 const struct cpumask *cpus;
2105 struct cpumask ***masks;
2106 int node;
2107 int cpu;
2108 int w;
2109};
2110
2111static int hop_cmp(const void *a, const void *b)
2112{
2113 struct cpumask **prev_hop, **cur_hop = *(struct cpumask ***)b;
2114 struct __cmp_key *k = (struct __cmp_key *)a;
2115
2116 if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu)
2117 return 1;
2118
2119 if (b == k->masks) {
2120 k->w = 0;
2121 return 0;
2122 }
2123
2124 prev_hop = *((struct cpumask ***)b - 1);
2125 k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]);
2126 if (k->w <= k->cpu)
2127 return 0;
2128
2129 return -1;
2130}
2131
2132/**
2133 * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth closest CPU
2134 * from @cpus to @cpu, taking into account distance
2135 * from a given @node.
2136 * @cpus: cpumask to find a cpu from
2137 * @cpu: CPU to start searching
2138 * @node: NUMA node to order CPUs by distance
2139 *
2140 * Return: cpu, or nr_cpu_ids when nothing found.
2141 */
2142int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
2143{
2144 struct __cmp_key k = { .cpus = cpus, .cpu = cpu };
2145 struct cpumask ***hop_masks;
2146 int hop, ret = nr_cpu_ids;
2147
2148 if (node == NUMA_NO_NODE)
2149 return cpumask_nth_and(cpu, cpus, cpu_online_mask);
2150
2151 rcu_read_lock();
2152
2153 /* CPU-less node entries are uninitialized in sched_domains_numa_masks */
2154 node = numa_nearest_node(node, N_CPU);
2155 k.node = node;
2156
2157 k.masks = rcu_dereference(sched_domains_numa_masks);
2158 if (!k.masks)
2159 goto unlock;
2160
2161 hop_masks = bsearch(&k, k.masks, sched_domains_numa_levels, sizeof(k.masks[0]), hop_cmp);
2162 hop = hop_masks - k.masks;
2163
2164 ret = hop ?
2165 cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) :
2166 cpumask_nth_and(cpu, cpus, k.masks[0][node]);
2167unlock:
2168 rcu_read_unlock();
2169 return ret;
2170}
2171EXPORT_SYMBOL_GPL(sched_numa_find_nth_cpu);
2172
2173/**
2174 * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2175 * @node
2176 * @node: The node to count hops from.
2177 * @hops: Include CPUs up to that many hops away. 0 means local node.
2178 *
2179 * Return: On success, a pointer to a cpumask of CPUs at most @hops away from
2180 * @node, an error value otherwise.
2181 *
2182 * Requires rcu_lock to be held. Returned cpumask is only valid within that
2183 * read-side section, copy it if required beyond that.
2184 *
2185 * Note that not all hops are equal in distance; see sched_init_numa() for how
2186 * distances and masks are handled.
2187 * Also note that this is a reflection of sched_domains_numa_masks, which may change
2188 * during the lifetime of the system (offline nodes are taken out of the masks).
2189 */
2190const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops)
2191{
2192 struct cpumask ***masks;
2193
2194 if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
2195 return ERR_PTR(-EINVAL);
2196
2197 masks = rcu_dereference(sched_domains_numa_masks);
2198 if (!masks)
2199 return ERR_PTR(-EBUSY);
2200
2201 return masks[hops][node];
2202}
2203EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
2204
2205#endif /* CONFIG_NUMA */
2206
2207static int __sdt_alloc(const struct cpumask *cpu_map)
2208{
2209 struct sched_domain_topology_level *tl;
2210 int j;
2211
2212 for_each_sd_topology(tl) {
2213 struct sd_data *sdd = &tl->data;
2214
2215 sdd->sd = alloc_percpu(struct sched_domain *);
2216 if (!sdd->sd)
2217 return -ENOMEM;
2218
2219 sdd->sds = alloc_percpu(struct sched_domain_shared *);
2220 if (!sdd->sds)
2221 return -ENOMEM;
2222
2223 sdd->sg = alloc_percpu(struct sched_group *);
2224 if (!sdd->sg)
2225 return -ENOMEM;
2226
2227 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2228 if (!sdd->sgc)
2229 return -ENOMEM;
2230
2231 for_each_cpu(j, cpu_map) {
2232 struct sched_domain *sd;
2233 struct sched_domain_shared *sds;
2234 struct sched_group *sg;
2235 struct sched_group_capacity *sgc;
2236
2237 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2238 GFP_KERNEL, cpu_to_node(j));
2239 if (!sd)
2240 return -ENOMEM;
2241
2242 *per_cpu_ptr(sdd->sd, j) = sd;
2243
2244 sds = kzalloc_node(sizeof(struct sched_domain_shared),
2245 GFP_KERNEL, cpu_to_node(j));
2246 if (!sds)
2247 return -ENOMEM;
2248
2249 *per_cpu_ptr(sdd->sds, j) = sds;
2250
2251 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2252 GFP_KERNEL, cpu_to_node(j));
2253 if (!sg)
2254 return -ENOMEM;
2255
2256 sg->next = sg;
2257
2258 *per_cpu_ptr(sdd->sg, j) = sg;
2259
2260 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2261 GFP_KERNEL, cpu_to_node(j));
2262 if (!sgc)
2263 return -ENOMEM;
2264
2265#ifdef CONFIG_SCHED_DEBUG
2266 sgc->id = j;
2267#endif
2268
2269 *per_cpu_ptr(sdd->sgc, j) = sgc;
2270 }
2271 }
2272
2273 return 0;
2274}
2275
2276static void __sdt_free(const struct cpumask *cpu_map)
2277{
2278 struct sched_domain_topology_level *tl;
2279 int j;
2280
2281 for_each_sd_topology(tl) {
2282 struct sd_data *sdd = &tl->data;
2283
2284 for_each_cpu(j, cpu_map) {
2285 struct sched_domain *sd;
2286
2287 if (sdd->sd) {
2288 sd = *per_cpu_ptr(sdd->sd, j);
2289 if (sd && (sd->flags & SD_OVERLAP))
2290 free_sched_groups(sd->groups, 0);
2291 kfree(*per_cpu_ptr(sdd->sd, j));
2292 }
2293
2294 if (sdd->sds)
2295 kfree(*per_cpu_ptr(sdd->sds, j));
2296 if (sdd->sg)
2297 kfree(*per_cpu_ptr(sdd->sg, j));
2298 if (sdd->sgc)
2299 kfree(*per_cpu_ptr(sdd->sgc, j));
2300 }
2301 free_percpu(sdd->sd);
2302 sdd->sd = NULL;
2303 free_percpu(sdd->sds);
2304 sdd->sds = NULL;
2305 free_percpu(sdd->sg);
2306 sdd->sg = NULL;
2307 free_percpu(sdd->sgc);
2308 sdd->sgc = NULL;
2309 }
2310}
2311
2312static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2313 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2314 struct sched_domain *child, int cpu)
2315{
2316 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2317
2318 if (child) {
2319 sd->level = child->level + 1;
2320 sched_domain_level_max = max(sched_domain_level_max, sd->level);
2321 child->parent = sd;
2322
2323 if (!cpumask_subset(sched_domain_span(child),
2324 sched_domain_span(sd))) {
2325 pr_err("BUG: arch topology borken\n");
2326#ifdef CONFIG_SCHED_DEBUG
2327 pr_err(" the %s domain not a subset of the %s domain\n",
2328 child->name, sd->name);
2329#endif
2330 /* Fixup, ensure @sd has at least @child CPUs. */
2331 cpumask_or(sched_domain_span(sd),
2332 sched_domain_span(sd),
2333 sched_domain_span(child));
2334 }
2335
2336 }
2337 set_domain_attribute(sd, attr);
2338
2339 return sd;
2340}
2341
2342/*
2343 * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2344 * any two given CPUs at this (non-NUMA) topology level.
2345 */
2346static bool topology_span_sane(struct sched_domain_topology_level *tl,
2347 const struct cpumask *cpu_map, int cpu)
2348{
2349 int i;
2350
2351 /* NUMA levels are allowed to overlap */
2352 if (tl->flags & SDTL_OVERLAP)
2353 return true;
2354
2355 /*
2356 * Non-NUMA levels cannot partially overlap - they must be either
2357 * completely equal or completely disjoint. Otherwise we can end up
2358 * breaking the sched_group lists - i.e. a later get_group() pass
2359 * breaks the linking done for an earlier span.
2360 */
2361 for_each_cpu(i, cpu_map) {
2362 if (i == cpu)
2363 continue;
2364 /*
2365 * We should 'and' all those masks with 'cpu_map' to exactly
2366 * match the topology we're about to build, but that can only
2367 * remove CPUs, which only lessens our ability to detect
2368 * overlaps
2369 */
2370 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2371 cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2372 return false;
2373 }
2374
2375 return true;
2376}
2377
2378/*
2379 * Build sched domains for a given set of CPUs and attach the sched domains
2380 * to the individual CPUs
2381 */
2382static int
2383build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2384{
2385 enum s_alloc alloc_state = sa_none;
2386 struct sched_domain *sd;
2387 struct s_data d;
2388 struct rq *rq = NULL;
2389 int i, ret = -ENOMEM;
2390 bool has_asym = false;
2391 bool has_cluster = false;
2392
2393 if (WARN_ON(cpumask_empty(cpu_map)))
2394 goto error;
2395
2396 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2397 if (alloc_state != sa_rootdomain)
2398 goto error;
2399
2400 /* Set up domains for CPUs specified by the cpu_map: */
2401 for_each_cpu(i, cpu_map) {
2402 struct sched_domain_topology_level *tl;
2403
2404 sd = NULL;
2405 for_each_sd_topology(tl) {
2406
2407 if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2408 goto error;
2409
2410 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2411
2412 has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2413
2414 if (tl == sched_domain_topology)
2415 *per_cpu_ptr(d.sd, i) = sd;
2416 if (tl->flags & SDTL_OVERLAP)
2417 sd->flags |= SD_OVERLAP;
2418 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2419 break;
2420 }
2421 }
2422
2423 /* Build the groups for the domains */
2424 for_each_cpu(i, cpu_map) {
2425 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2426 sd->span_weight = cpumask_weight(sched_domain_span(sd));
2427 if (sd->flags & SD_OVERLAP) {
2428 if (build_overlap_sched_groups(sd, i))
2429 goto error;
2430 } else {
2431 if (build_sched_groups(sd, i))
2432 goto error;
2433 }
2434 }
2435 }
2436
2437 /*
2438 * Calculate an allowed NUMA imbalance such that LLCs do not get
2439 * imbalanced.
2440 */
2441 for_each_cpu(i, cpu_map) {
2442 unsigned int imb = 0;
2443 unsigned int imb_span = 1;
2444
2445 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2446 struct sched_domain *child = sd->child;
2447
2448 if (!(sd->flags & SD_SHARE_PKG_RESOURCES) && child &&
2449 (child->flags & SD_SHARE_PKG_RESOURCES)) {
2450 struct sched_domain __rcu *top_p;
2451 unsigned int nr_llcs;
2452
2453 /*
2454 * For a single LLC per node, allow an
2455 * imbalance up to 12.5% of the node. This is
2456 * arbitrary cutoff based two factors -- SMT and
2457 * memory channels. For SMT-2, the intent is to
2458 * avoid premature sharing of HT resources but
2459 * SMT-4 or SMT-8 *may* benefit from a different
2460 * cutoff. For memory channels, this is a very
2461 * rough estimate of how many channels may be
2462 * active and is based on recent CPUs with
2463 * many cores.
2464 *
2465 * For multiple LLCs, allow an imbalance
2466 * until multiple tasks would share an LLC
2467 * on one node while LLCs on another node
2468 * remain idle. This assumes that there are
2469 * enough logical CPUs per LLC to avoid SMT
2470 * factors and that there is a correlation
2471 * between LLCs and memory channels.
2472 */
2473 nr_llcs = sd->span_weight / child->span_weight;
2474 if (nr_llcs == 1)
2475 imb = sd->span_weight >> 3;
2476 else
2477 imb = nr_llcs;
2478 imb = max(1U, imb);
2479 sd->imb_numa_nr = imb;
2480
2481 /* Set span based on the first NUMA domain. */
2482 top_p = sd->parent;
2483 while (top_p && !(top_p->flags & SD_NUMA)) {
2484 top_p = top_p->parent;
2485 }
2486 imb_span = top_p ? top_p->span_weight : sd->span_weight;
2487 } else {
2488 int factor = max(1U, (sd->span_weight / imb_span));
2489
2490 sd->imb_numa_nr = imb * factor;
2491 }
2492 }
2493 }
2494
2495 /* Calculate CPU capacity for physical packages and nodes */
2496 for (i = nr_cpumask_bits-1; i >= 0; i--) {
2497 if (!cpumask_test_cpu(i, cpu_map))
2498 continue;
2499
2500 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2501 claim_allocations(i, sd);
2502 init_sched_groups_capacity(i, sd);
2503 }
2504 }
2505
2506 /* Attach the domains */
2507 rcu_read_lock();
2508 for_each_cpu(i, cpu_map) {
2509 unsigned long capacity;
2510
2511 rq = cpu_rq(i);
2512 sd = *per_cpu_ptr(d.sd, i);
2513
2514 capacity = arch_scale_cpu_capacity(i);
2515 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2516 if (capacity > READ_ONCE(d.rd->max_cpu_capacity))
2517 WRITE_ONCE(d.rd->max_cpu_capacity, capacity);
2518
2519 cpu_attach_domain(sd, d.rd, i);
2520
2521 if (lowest_flag_domain(i, SD_CLUSTER))
2522 has_cluster = true;
2523 }
2524 rcu_read_unlock();
2525
2526 if (has_asym)
2527 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2528
2529 if (has_cluster)
2530 static_branch_inc_cpuslocked(&sched_cluster_active);
2531
2532 if (rq && sched_debug_verbose) {
2533 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2534 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2535 }
2536
2537 ret = 0;
2538error:
2539 __free_domain_allocs(&d, alloc_state, cpu_map);
2540
2541 return ret;
2542}
2543
2544/* Current sched domains: */
2545static cpumask_var_t *doms_cur;
2546
2547/* Number of sched domains in 'doms_cur': */
2548static int ndoms_cur;
2549
2550/* Attributes of custom domains in 'doms_cur' */
2551static struct sched_domain_attr *dattr_cur;
2552
2553/*
2554 * Special case: If a kmalloc() of a doms_cur partition (array of
2555 * cpumask) fails, then fallback to a single sched domain,
2556 * as determined by the single cpumask fallback_doms.
2557 */
2558static cpumask_var_t fallback_doms;
2559
2560/*
2561 * arch_update_cpu_topology lets virtualized architectures update the
2562 * CPU core maps. It is supposed to return 1 if the topology changed
2563 * or 0 if it stayed the same.
2564 */
2565int __weak arch_update_cpu_topology(void)
2566{
2567 return 0;
2568}
2569
2570cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2571{
2572 int i;
2573 cpumask_var_t *doms;
2574
2575 doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2576 if (!doms)
2577 return NULL;
2578 for (i = 0; i < ndoms; i++) {
2579 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2580 free_sched_domains(doms, i);
2581 return NULL;
2582 }
2583 }
2584 return doms;
2585}
2586
2587void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2588{
2589 unsigned int i;
2590 for (i = 0; i < ndoms; i++)
2591 free_cpumask_var(doms[i]);
2592 kfree(doms);
2593}
2594
2595/*
2596 * Set up scheduler domains and groups. For now this just excludes isolated
2597 * CPUs, but could be used to exclude other special cases in the future.
2598 */
2599int __init sched_init_domains(const struct cpumask *cpu_map)
2600{
2601 int err;
2602
2603 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2604 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2605 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2606
2607 arch_update_cpu_topology();
2608 asym_cpu_capacity_scan();
2609 ndoms_cur = 1;
2610 doms_cur = alloc_sched_domains(ndoms_cur);
2611 if (!doms_cur)
2612 doms_cur = &fallback_doms;
2613 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
2614 err = build_sched_domains(doms_cur[0], NULL);
2615
2616 return err;
2617}
2618
2619/*
2620 * Detach sched domains from a group of CPUs specified in cpu_map
2621 * These CPUs will now be attached to the NULL domain
2622 */
2623static void detach_destroy_domains(const struct cpumask *cpu_map)
2624{
2625 unsigned int cpu = cpumask_any(cpu_map);
2626 int i;
2627
2628 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2629 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2630
2631 if (static_branch_unlikely(&sched_cluster_active))
2632 static_branch_dec_cpuslocked(&sched_cluster_active);
2633
2634 rcu_read_lock();
2635 for_each_cpu(i, cpu_map)
2636 cpu_attach_domain(NULL, &def_root_domain, i);
2637 rcu_read_unlock();
2638}
2639
2640/* handle null as "default" */
2641static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2642 struct sched_domain_attr *new, int idx_new)
2643{
2644 struct sched_domain_attr tmp;
2645
2646 /* Fast path: */
2647 if (!new && !cur)
2648 return 1;
2649
2650 tmp = SD_ATTR_INIT;
2651
2652 return !memcmp(cur ? (cur + idx_cur) : &tmp,
2653 new ? (new + idx_new) : &tmp,
2654 sizeof(struct sched_domain_attr));
2655}
2656
2657/*
2658 * Partition sched domains as specified by the 'ndoms_new'
2659 * cpumasks in the array doms_new[] of cpumasks. This compares
2660 * doms_new[] to the current sched domain partitioning, doms_cur[].
2661 * It destroys each deleted domain and builds each new domain.
2662 *
2663 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2664 * The masks don't intersect (don't overlap.) We should setup one
2665 * sched domain for each mask. CPUs not in any of the cpumasks will
2666 * not be load balanced. If the same cpumask appears both in the
2667 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2668 * it as it is.
2669 *
2670 * The passed in 'doms_new' should be allocated using
2671 * alloc_sched_domains. This routine takes ownership of it and will
2672 * free_sched_domains it when done with it. If the caller failed the
2673 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2674 * and partition_sched_domains() will fallback to the single partition
2675 * 'fallback_doms', it also forces the domains to be rebuilt.
2676 *
2677 * If doms_new == NULL it will be replaced with cpu_online_mask.
2678 * ndoms_new == 0 is a special case for destroying existing domains,
2679 * and it will not create the default domain.
2680 *
2681 * Call with hotplug lock and sched_domains_mutex held
2682 */
2683void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2684 struct sched_domain_attr *dattr_new)
2685{
2686 bool __maybe_unused has_eas = false;
2687 int i, j, n;
2688 int new_topology;
2689
2690 lockdep_assert_held(&sched_domains_mutex);
2691
2692 /* Let the architecture update CPU core mappings: */
2693 new_topology = arch_update_cpu_topology();
2694 /* Trigger rebuilding CPU capacity asymmetry data */
2695 if (new_topology)
2696 asym_cpu_capacity_scan();
2697
2698 if (!doms_new) {
2699 WARN_ON_ONCE(dattr_new);
2700 n = 0;
2701 doms_new = alloc_sched_domains(1);
2702 if (doms_new) {
2703 n = 1;
2704 cpumask_and(doms_new[0], cpu_active_mask,
2705 housekeeping_cpumask(HK_TYPE_DOMAIN));
2706 }
2707 } else {
2708 n = ndoms_new;
2709 }
2710
2711 /* Destroy deleted domains: */
2712 for (i = 0; i < ndoms_cur; i++) {
2713 for (j = 0; j < n && !new_topology; j++) {
2714 if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2715 dattrs_equal(dattr_cur, i, dattr_new, j)) {
2716 struct root_domain *rd;
2717
2718 /*
2719 * This domain won't be destroyed and as such
2720 * its dl_bw->total_bw needs to be cleared. It
2721 * will be recomputed in function
2722 * update_tasks_root_domain().
2723 */
2724 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2725 dl_clear_root_domain(rd);
2726 goto match1;
2727 }
2728 }
2729 /* No match - a current sched domain not in new doms_new[] */
2730 detach_destroy_domains(doms_cur[i]);
2731match1:
2732 ;
2733 }
2734
2735 n = ndoms_cur;
2736 if (!doms_new) {
2737 n = 0;
2738 doms_new = &fallback_doms;
2739 cpumask_and(doms_new[0], cpu_active_mask,
2740 housekeeping_cpumask(HK_TYPE_DOMAIN));
2741 }
2742
2743 /* Build new domains: */
2744 for (i = 0; i < ndoms_new; i++) {
2745 for (j = 0; j < n && !new_topology; j++) {
2746 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2747 dattrs_equal(dattr_new, i, dattr_cur, j))
2748 goto match2;
2749 }
2750 /* No match - add a new doms_new */
2751 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2752match2:
2753 ;
2754 }
2755
2756#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2757 /* Build perf. domains: */
2758 for (i = 0; i < ndoms_new; i++) {
2759 for (j = 0; j < n && !sched_energy_update; j++) {
2760 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2761 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2762 has_eas = true;
2763 goto match3;
2764 }
2765 }
2766 /* No match - add perf. domains for a new rd */
2767 has_eas |= build_perf_domains(doms_new[i]);
2768match3:
2769 ;
2770 }
2771 sched_energy_set(has_eas);
2772#endif
2773
2774 /* Remember the new sched domains: */
2775 if (doms_cur != &fallback_doms)
2776 free_sched_domains(doms_cur, ndoms_cur);
2777
2778 kfree(dattr_cur);
2779 doms_cur = doms_new;
2780 dattr_cur = dattr_new;
2781 ndoms_cur = ndoms_new;
2782
2783 update_sched_domain_debugfs();
2784}
2785
2786/*
2787 * Call with hotplug lock held
2788 */
2789void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2790 struct sched_domain_attr *dattr_new)
2791{
2792 mutex_lock(&sched_domains_mutex);
2793 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2794 mutex_unlock(&sched_domains_mutex);
2795}