2 * Scheduler topology setup/handling methods
4 #include <linux/sched.h>
5 #include <linux/mutex.h>
6 #include <linux/sched/isolation.h>
10 DEFINE_MUTEX(sched_domains_mutex);
12 /* Protected by sched_domains_mutex: */
13 cpumask_var_t sched_domains_tmpmask;
14 cpumask_var_t sched_domains_tmpmask2;
16 #ifdef CONFIG_SCHED_DEBUG
18 static int __init sched_debug_setup(char *str)
20 sched_debug_enabled = true;
24 early_param("sched_debug", sched_debug_setup);
26 static inline bool sched_debug(void)
28 return sched_debug_enabled;
31 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
32 struct cpumask *groupmask)
34 struct sched_group *group = sd->groups;
36 cpumask_clear(groupmask);
38 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
40 if (!(sd->flags & SD_LOAD_BALANCE)) {
41 printk("does not load-balance\n");
43 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
48 printk(KERN_CONT "span=%*pbl level=%s\n",
49 cpumask_pr_args(sched_domain_span(sd)), sd->name);
51 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
52 printk(KERN_ERR "ERROR: domain->span does not contain "
55 if (!cpumask_test_cpu(cpu, sched_group_span(group))) {
56 printk(KERN_ERR "ERROR: domain->groups does not contain"
60 printk(KERN_DEBUG "%*s groups:", level + 1, "");
64 printk(KERN_ERR "ERROR: group is NULL\n");
68 if (!cpumask_weight(sched_group_span(group))) {
69 printk(KERN_CONT "\n");
70 printk(KERN_ERR "ERROR: empty group\n");
74 if (!(sd->flags & SD_OVERLAP) &&
75 cpumask_intersects(groupmask, sched_group_span(group))) {
76 printk(KERN_CONT "\n");
77 printk(KERN_ERR "ERROR: repeated CPUs\n");
81 cpumask_or(groupmask, groupmask, sched_group_span(group));
83 printk(KERN_CONT " %d:{ span=%*pbl",
85 cpumask_pr_args(sched_group_span(group)));
87 if ((sd->flags & SD_OVERLAP) &&
88 !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
89 printk(KERN_CONT " mask=%*pbl",
90 cpumask_pr_args(group_balance_mask(group)));
93 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
94 printk(KERN_CONT " cap=%lu", group->sgc->capacity);
96 if (group == sd->groups && sd->child &&
97 !cpumask_equal(sched_domain_span(sd->child),
98 sched_group_span(group))) {
99 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
102 printk(KERN_CONT " }");
106 if (group != sd->groups)
107 printk(KERN_CONT ",");
109 } while (group != sd->groups);
110 printk(KERN_CONT "\n");
112 if (!cpumask_equal(sched_domain_span(sd), groupmask))
113 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
116 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
117 printk(KERN_ERR "ERROR: parent span is not a superset "
118 "of domain->span\n");
122 static void sched_domain_debug(struct sched_domain *sd, int cpu)
126 if (!sched_debug_enabled)
130 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
134 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
137 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
145 #else /* !CONFIG_SCHED_DEBUG */
147 # define sched_debug_enabled 0
148 # define sched_domain_debug(sd, cpu) do { } while (0)
149 static inline bool sched_debug(void)
153 #endif /* CONFIG_SCHED_DEBUG */
155 static int sd_degenerate(struct sched_domain *sd)
157 if (cpumask_weight(sched_domain_span(sd)) == 1)
160 /* Following flags need at least 2 groups */
161 if (sd->flags & (SD_LOAD_BALANCE |
165 SD_SHARE_CPUCAPACITY |
166 SD_ASYM_CPUCAPACITY |
167 SD_SHARE_PKG_RESOURCES |
168 SD_SHARE_POWERDOMAIN)) {
169 if (sd->groups != sd->groups->next)
173 /* Following flags don't use groups */
174 if (sd->flags & (SD_WAKE_AFFINE))
181 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
183 unsigned long cflags = sd->flags, pflags = parent->flags;
185 if (sd_degenerate(parent))
188 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
191 /* Flags needing groups don't count if only 1 group in parent */
192 if (parent->groups == parent->groups->next) {
193 pflags &= ~(SD_LOAD_BALANCE |
197 SD_ASYM_CPUCAPACITY |
198 SD_SHARE_CPUCAPACITY |
199 SD_SHARE_PKG_RESOURCES |
201 SD_SHARE_POWERDOMAIN);
202 if (nr_node_ids == 1)
203 pflags &= ~SD_SERIALIZE;
205 if (~cflags & pflags)
211 static void free_rootdomain(struct rcu_head *rcu)
213 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
215 cpupri_cleanup(&rd->cpupri);
216 cpudl_cleanup(&rd->cpudl);
217 free_cpumask_var(rd->dlo_mask);
218 free_cpumask_var(rd->rto_mask);
219 free_cpumask_var(rd->online);
220 free_cpumask_var(rd->span);
224 void rq_attach_root(struct rq *rq, struct root_domain *rd)
226 struct root_domain *old_rd = NULL;
229 raw_spin_lock_irqsave(&rq->lock, flags);
234 if (cpumask_test_cpu(rq->cpu, old_rd->online))
237 cpumask_clear_cpu(rq->cpu, old_rd->span);
240 * If we dont want to free the old_rd yet then
241 * set old_rd to NULL to skip the freeing later
244 if (!atomic_dec_and_test(&old_rd->refcount))
248 atomic_inc(&rd->refcount);
251 cpumask_set_cpu(rq->cpu, rd->span);
252 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
255 raw_spin_unlock_irqrestore(&rq->lock, flags);
258 call_rcu_sched(&old_rd->rcu, free_rootdomain);
261 static int init_rootdomain(struct root_domain *rd)
263 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
265 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
267 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
269 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
272 #ifdef HAVE_RT_PUSH_IPI
274 raw_spin_lock_init(&rd->rto_lock);
275 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
278 init_dl_bw(&rd->dl_bw);
279 if (cpudl_init(&rd->cpudl) != 0)
282 if (cpupri_init(&rd->cpupri) != 0)
287 cpudl_cleanup(&rd->cpudl);
289 free_cpumask_var(rd->rto_mask);
291 free_cpumask_var(rd->dlo_mask);
293 free_cpumask_var(rd->online);
295 free_cpumask_var(rd->span);
301 * By default the system creates a single root-domain with all CPUs as
302 * members (mimicking the global state we have today).
304 struct root_domain def_root_domain;
306 void init_defrootdomain(void)
308 init_rootdomain(&def_root_domain);
310 atomic_set(&def_root_domain.refcount, 1);
313 static struct root_domain *alloc_rootdomain(void)
315 struct root_domain *rd;
317 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
321 if (init_rootdomain(rd) != 0) {
329 static void free_sched_groups(struct sched_group *sg, int free_sgc)
331 struct sched_group *tmp, *first;
340 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
343 if (atomic_dec_and_test(&sg->ref))
346 } while (sg != first);
349 static void destroy_sched_domain(struct sched_domain *sd)
352 * A normal sched domain may have multiple group references, an
353 * overlapping domain, having private groups, only one. Iterate,
354 * dropping group/capacity references, freeing where none remain.
356 free_sched_groups(sd->groups, 1);
358 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
363 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
365 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
368 struct sched_domain *parent = sd->parent;
369 destroy_sched_domain(sd);
374 static void destroy_sched_domains(struct sched_domain *sd)
377 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
381 * Keep a special pointer to the highest sched_domain that has
382 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
383 * allows us to avoid some pointer chasing select_idle_sibling().
385 * Also keep a unique ID per domain (we use the first CPU number in
386 * the cpumask of the domain), this allows us to quickly tell if
387 * two CPUs are in the same cache domain, see cpus_share_cache().
389 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
390 DEFINE_PER_CPU(int, sd_llc_size);
391 DEFINE_PER_CPU(int, sd_llc_id);
392 DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
393 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
394 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
396 static void update_top_cache_domain(int cpu)
398 struct sched_domain_shared *sds = NULL;
399 struct sched_domain *sd;
403 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
405 id = cpumask_first(sched_domain_span(sd));
406 size = cpumask_weight(sched_domain_span(sd));
410 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
411 per_cpu(sd_llc_size, cpu) = size;
412 per_cpu(sd_llc_id, cpu) = id;
413 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
415 sd = lowest_flag_domain(cpu, SD_NUMA);
416 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
418 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
419 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
423 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
424 * hold the hotplug lock.
427 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
429 struct rq *rq = cpu_rq(cpu);
430 struct sched_domain *tmp;
432 /* Remove the sched domains which do not contribute to scheduling. */
433 for (tmp = sd; tmp; ) {
434 struct sched_domain *parent = tmp->parent;
438 if (sd_parent_degenerate(tmp, parent)) {
439 tmp->parent = parent->parent;
441 parent->parent->child = tmp;
443 * Transfer SD_PREFER_SIBLING down in case of a
444 * degenerate parent; the spans match for this
445 * so the property transfers.
447 if (parent->flags & SD_PREFER_SIBLING)
448 tmp->flags |= SD_PREFER_SIBLING;
449 destroy_sched_domain(parent);
454 if (sd && sd_degenerate(sd)) {
457 destroy_sched_domain(tmp);
462 sched_domain_debug(sd, cpu);
464 rq_attach_root(rq, rd);
466 rcu_assign_pointer(rq->sd, sd);
467 dirty_sched_domain_sysctl(cpu);
468 destroy_sched_domains(tmp);
470 update_top_cache_domain(cpu);
474 struct sched_domain ** __percpu sd;
475 struct root_domain *rd;
486 * Return the canonical balance CPU for this group, this is the first CPU
487 * of this group that's also in the balance mask.
489 * The balance mask are all those CPUs that could actually end up at this
490 * group. See build_balance_mask().
492 * Also see should_we_balance().
494 int group_balance_cpu(struct sched_group *sg)
496 return cpumask_first(group_balance_mask(sg));
501 * NUMA topology (first read the regular topology blurb below)
503 * Given a node-distance table, for example:
511 * which represents a 4 node ring topology like:
519 * We want to construct domains and groups to represent this. The way we go
520 * about doing this is to build the domains on 'hops'. For each NUMA level we
521 * construct the mask of all nodes reachable in @level hops.
523 * For the above NUMA topology that gives 3 levels:
525 * NUMA-2 0-3 0-3 0-3 0-3
526 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
528 * NUMA-1 0-1,3 0-2 1-3 0,2-3
529 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
534 * As can be seen; things don't nicely line up as with the regular topology.
535 * When we iterate a domain in child domain chunks some nodes can be
536 * represented multiple times -- hence the "overlap" naming for this part of
539 * In order to minimize this overlap, we only build enough groups to cover the
540 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
544 * - the first group of each domain is its child domain; this
545 * gets us the first 0-1,3
546 * - the only uncovered node is 2, who's child domain is 1-3.
548 * However, because of the overlap, computing a unique CPU for each group is
549 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
550 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
551 * end up at those groups (they would end up in group: 0-1,3).
553 * To correct this we have to introduce the group balance mask. This mask
554 * will contain those CPUs in the group that can reach this group given the
555 * (child) domain tree.
557 * With this we can once again compute balance_cpu and sched_group_capacity
560 * XXX include words on how balance_cpu is unique and therefore can be
561 * used for sched_group_capacity links.
564 * Another 'interesting' topology is:
572 * Which looks a little like:
580 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
583 * This leads to a few particularly weird cases where the sched_domain's are
584 * not of the same number for each cpu. Consider:
587 * groups: {0-2},{1-3} {1-3},{0-2}
589 * NUMA-1 0-2 0-3 0-3 1-3
597 * Build the balance mask; it contains only those CPUs that can arrive at this
598 * group and should be considered to continue balancing.
600 * We do this during the group creation pass, therefore the group information
601 * isn't complete yet, however since each group represents a (child) domain we
602 * can fully construct this using the sched_domain bits (which are already
606 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
608 const struct cpumask *sg_span = sched_group_span(sg);
609 struct sd_data *sdd = sd->private;
610 struct sched_domain *sibling;
615 for_each_cpu(i, sg_span) {
616 sibling = *per_cpu_ptr(sdd->sd, i);
619 * Can happen in the asymmetric case, where these siblings are
620 * unused. The mask will not be empty because those CPUs that
621 * do have the top domain _should_ span the domain.
626 /* If we would not end up here, we can't continue from here */
627 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
630 cpumask_set_cpu(i, mask);
633 /* We must not have empty masks here */
634 WARN_ON_ONCE(cpumask_empty(mask));
638 * XXX: This creates per-node group entries; since the load-balancer will
639 * immediately access remote memory to construct this group's load-balance
640 * statistics having the groups node local is of dubious benefit.
642 static struct sched_group *
643 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
645 struct sched_group *sg;
646 struct cpumask *sg_span;
648 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
649 GFP_KERNEL, cpu_to_node(cpu));
654 sg_span = sched_group_span(sg);
656 cpumask_copy(sg_span, sched_domain_span(sd->child));
658 cpumask_copy(sg_span, sched_domain_span(sd));
660 atomic_inc(&sg->ref);
664 static void init_overlap_sched_group(struct sched_domain *sd,
665 struct sched_group *sg)
667 struct cpumask *mask = sched_domains_tmpmask2;
668 struct sd_data *sdd = sd->private;
669 struct cpumask *sg_span;
672 build_balance_mask(sd, sg, mask);
673 cpu = cpumask_first_and(sched_group_span(sg), mask);
675 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
676 if (atomic_inc_return(&sg->sgc->ref) == 1)
677 cpumask_copy(group_balance_mask(sg), mask);
679 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
682 * Initialize sgc->capacity such that even if we mess up the
683 * domains and no possible iteration will get us here, we won't
686 sg_span = sched_group_span(sg);
687 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
688 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
692 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
694 struct sched_group *first = NULL, *last = NULL, *sg;
695 const struct cpumask *span = sched_domain_span(sd);
696 struct cpumask *covered = sched_domains_tmpmask;
697 struct sd_data *sdd = sd->private;
698 struct sched_domain *sibling;
701 cpumask_clear(covered);
703 for_each_cpu_wrap(i, span, cpu) {
704 struct cpumask *sg_span;
706 if (cpumask_test_cpu(i, covered))
709 sibling = *per_cpu_ptr(sdd->sd, i);
712 * Asymmetric node setups can result in situations where the
713 * domain tree is of unequal depth, make sure to skip domains
714 * that already cover the entire range.
716 * In that case build_sched_domains() will have terminated the
717 * iteration early and our sibling sd spans will be empty.
718 * Domains should always include the CPU they're built on, so
721 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
724 sg = build_group_from_child_sched_domain(sibling, cpu);
728 sg_span = sched_group_span(sg);
729 cpumask_or(covered, covered, sg_span);
731 init_overlap_sched_group(sd, sg);
745 free_sched_groups(first, 0);
752 * Package topology (also see the load-balance blurb in fair.c)
754 * The scheduler builds a tree structure to represent a number of important
755 * topology features. By default (default_topology[]) these include:
757 * - Simultaneous multithreading (SMT)
758 * - Multi-Core Cache (MC)
761 * Where the last one more or less denotes everything up to a NUMA node.
763 * The tree consists of 3 primary data structures:
765 * sched_domain -> sched_group -> sched_group_capacity
769 * The sched_domains are per-cpu and have a two way link (parent & child) and
770 * denote the ever growing mask of CPUs belonging to that level of topology.
772 * Each sched_domain has a circular (double) linked list of sched_group's, each
773 * denoting the domains of the level below (or individual CPUs in case of the
774 * first domain level). The sched_group linked by a sched_domain includes the
775 * CPU of that sched_domain [*].
777 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
779 * CPU 0 1 2 3 4 5 6 7
783 * SMT [ ] [ ] [ ] [ ]
787 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
788 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
789 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
791 * CPU 0 1 2 3 4 5 6 7
793 * One way to think about it is: sched_domain moves you up and down among these
794 * topology levels, while sched_group moves you sideways through it, at child
795 * domain granularity.
797 * sched_group_capacity ensures each unique sched_group has shared storage.
799 * There are two related construction problems, both require a CPU that
800 * uniquely identify each group (for a given domain):
802 * - The first is the balance_cpu (see should_we_balance() and the
803 * load-balance blub in fair.c); for each group we only want 1 CPU to
804 * continue balancing at a higher domain.
806 * - The second is the sched_group_capacity; we want all identical groups
807 * to share a single sched_group_capacity.
809 * Since these topologies are exclusive by construction. That is, its
810 * impossible for an SMT thread to belong to multiple cores, and cores to
811 * be part of multiple caches. There is a very clear and unique location
812 * for each CPU in the hierarchy.
814 * Therefore computing a unique CPU for each group is trivial (the iteration
815 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
816 * group), we can simply pick the first CPU in each group.
819 * [*] in other words, the first group of each domain is its child domain.
822 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
824 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
825 struct sched_domain *child = sd->child;
826 struct sched_group *sg;
829 cpu = cpumask_first(sched_domain_span(child));
831 sg = *per_cpu_ptr(sdd->sg, cpu);
832 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
834 /* For claim_allocations: */
835 atomic_inc(&sg->ref);
836 atomic_inc(&sg->sgc->ref);
839 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
840 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
842 cpumask_set_cpu(cpu, sched_group_span(sg));
843 cpumask_set_cpu(cpu, group_balance_mask(sg));
846 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
847 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
853 * build_sched_groups will build a circular linked list of the groups
854 * covered by the given span, and will set each group's ->cpumask correctly,
855 * and ->cpu_capacity to 0.
857 * Assumes the sched_domain tree is fully constructed
860 build_sched_groups(struct sched_domain *sd, int cpu)
862 struct sched_group *first = NULL, *last = NULL;
863 struct sd_data *sdd = sd->private;
864 const struct cpumask *span = sched_domain_span(sd);
865 struct cpumask *covered;
868 lockdep_assert_held(&sched_domains_mutex);
869 covered = sched_domains_tmpmask;
871 cpumask_clear(covered);
873 for_each_cpu_wrap(i, span, cpu) {
874 struct sched_group *sg;
876 if (cpumask_test_cpu(i, covered))
879 sg = get_group(i, sdd);
881 cpumask_or(covered, covered, sched_group_span(sg));
896 * Initialize sched groups cpu_capacity.
898 * cpu_capacity indicates the capacity of sched group, which is used while
899 * distributing the load between different sched groups in a sched domain.
900 * Typically cpu_capacity for all the groups in a sched domain will be same
901 * unless there are asymmetries in the topology. If there are asymmetries,
902 * group having more cpu_capacity will pickup more load compared to the
903 * group having less cpu_capacity.
905 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
907 struct sched_group *sg = sd->groups;
912 int cpu, max_cpu = -1;
914 sg->group_weight = cpumask_weight(sched_group_span(sg));
916 if (!(sd->flags & SD_ASYM_PACKING))
919 for_each_cpu(cpu, sched_group_span(sg)) {
922 else if (sched_asym_prefer(cpu, max_cpu))
925 sg->asym_prefer_cpu = max_cpu;
929 } while (sg != sd->groups);
931 if (cpu != group_balance_cpu(sg))
934 update_group_capacity(sd, cpu);
938 * Initializers for schedule domains
939 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
942 static int default_relax_domain_level = -1;
943 int sched_domain_level_max;
945 static int __init setup_relax_domain_level(char *str)
947 if (kstrtoint(str, 0, &default_relax_domain_level))
948 pr_warn("Unable to set relax_domain_level\n");
952 __setup("relax_domain_level=", setup_relax_domain_level);
954 static void set_domain_attribute(struct sched_domain *sd,
955 struct sched_domain_attr *attr)
959 if (!attr || attr->relax_domain_level < 0) {
960 if (default_relax_domain_level < 0)
963 request = default_relax_domain_level;
965 request = attr->relax_domain_level;
966 if (request < sd->level) {
967 /* Turn off idle balance on this domain: */
968 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
970 /* Turn on idle balance on this domain: */
971 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
975 static void __sdt_free(const struct cpumask *cpu_map);
976 static int __sdt_alloc(const struct cpumask *cpu_map);
978 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
979 const struct cpumask *cpu_map)
983 if (!atomic_read(&d->rd->refcount))
984 free_rootdomain(&d->rd->rcu);
998 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1000 memset(d, 0, sizeof(*d));
1002 if (__sdt_alloc(cpu_map))
1003 return sa_sd_storage;
1004 d->sd = alloc_percpu(struct sched_domain *);
1006 return sa_sd_storage;
1007 d->rd = alloc_rootdomain();
1010 return sa_rootdomain;
1014 * NULL the sd_data elements we've used to build the sched_domain and
1015 * sched_group structure so that the subsequent __free_domain_allocs()
1016 * will not free the data we're using.
1018 static void claim_allocations(int cpu, struct sched_domain *sd)
1020 struct sd_data *sdd = sd->private;
1022 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1023 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1025 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1026 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1028 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1029 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1031 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1032 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1036 static int sched_domains_numa_levels;
1037 enum numa_topology_type sched_numa_topology_type;
1038 static int *sched_domains_numa_distance;
1039 int sched_max_numa_distance;
1040 static struct cpumask ***sched_domains_numa_masks;
1041 static int sched_domains_curr_level;
1045 * SD_flags allowed in topology descriptions.
1047 * These flags are purely descriptive of the topology and do not prescribe
1048 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1051 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1052 * SD_SHARE_PKG_RESOURCES - describes shared caches
1053 * SD_NUMA - describes NUMA topologies
1054 * SD_SHARE_POWERDOMAIN - describes shared power domain
1055 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
1057 * Odd one out, which beside describing the topology has a quirk also
1058 * prescribes the desired behaviour that goes along with it:
1060 * SD_ASYM_PACKING - describes SMT quirks
1062 #define TOPOLOGY_SD_FLAGS \
1063 (SD_SHARE_CPUCAPACITY | \
1064 SD_SHARE_PKG_RESOURCES | \
1067 SD_ASYM_CPUCAPACITY | \
1068 SD_SHARE_POWERDOMAIN)
1070 static struct sched_domain *
1071 sd_init(struct sched_domain_topology_level *tl,
1072 const struct cpumask *cpu_map,
1073 struct sched_domain *child, int cpu)
1075 struct sd_data *sdd = &tl->data;
1076 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1077 int sd_id, sd_weight, sd_flags = 0;
1081 * Ugly hack to pass state to sd_numa_mask()...
1083 sched_domains_curr_level = tl->numa_level;
1086 sd_weight = cpumask_weight(tl->mask(cpu));
1089 sd_flags = (*tl->sd_flags)();
1090 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1091 "wrong sd_flags in topology description\n"))
1092 sd_flags &= ~TOPOLOGY_SD_FLAGS;
1094 *sd = (struct sched_domain){
1095 .min_interval = sd_weight,
1096 .max_interval = 2*sd_weight,
1098 .imbalance_pct = 125,
1100 .cache_nice_tries = 0,
1107 .flags = 1*SD_LOAD_BALANCE
1108 | 1*SD_BALANCE_NEWIDLE
1113 | 0*SD_SHARE_CPUCAPACITY
1114 | 0*SD_SHARE_PKG_RESOURCES
1116 | 0*SD_PREFER_SIBLING
1121 .last_balance = jiffies,
1122 .balance_interval = sd_weight,
1124 .max_newidle_lb_cost = 0,
1125 .next_decay_max_lb_cost = jiffies,
1127 #ifdef CONFIG_SCHED_DEBUG
1132 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1133 sd_id = cpumask_first(sched_domain_span(sd));
1136 * Convert topological properties into behaviour.
1139 if (sd->flags & SD_ASYM_CPUCAPACITY) {
1140 struct sched_domain *t = sd;
1142 for_each_lower_domain(t)
1143 t->flags |= SD_BALANCE_WAKE;
1146 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1147 sd->flags |= SD_PREFER_SIBLING;
1148 sd->imbalance_pct = 110;
1149 sd->smt_gain = 1178; /* ~15% */
1151 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1152 sd->flags |= SD_PREFER_SIBLING;
1153 sd->imbalance_pct = 117;
1154 sd->cache_nice_tries = 1;
1158 } else if (sd->flags & SD_NUMA) {
1159 sd->cache_nice_tries = 2;
1163 sd->flags |= SD_SERIALIZE;
1164 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
1165 sd->flags &= ~(SD_BALANCE_EXEC |
1172 sd->flags |= SD_PREFER_SIBLING;
1173 sd->cache_nice_tries = 1;
1179 * For all levels sharing cache; connect a sched_domain_shared
1182 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1183 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1184 atomic_inc(&sd->shared->ref);
1185 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1194 * Topology list, bottom-up.
1196 static struct sched_domain_topology_level default_topology[] = {
1197 #ifdef CONFIG_SCHED_SMT
1198 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1200 #ifdef CONFIG_SCHED_MC
1201 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1203 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1207 static struct sched_domain_topology_level *sched_domain_topology =
1210 #define for_each_sd_topology(tl) \
1211 for (tl = sched_domain_topology; tl->mask; tl++)
1213 void set_sched_topology(struct sched_domain_topology_level *tl)
1215 if (WARN_ON_ONCE(sched_smp_initialized))
1218 sched_domain_topology = tl;
1223 static const struct cpumask *sd_numa_mask(int cpu)
1225 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1228 static void sched_numa_warn(const char *str)
1230 static int done = false;
1238 printk(KERN_WARNING "ERROR: %s\n\n", str);
1240 for (i = 0; i < nr_node_ids; i++) {
1241 printk(KERN_WARNING " ");
1242 for (j = 0; j < nr_node_ids; j++)
1243 printk(KERN_CONT "%02d ", node_distance(i,j));
1244 printk(KERN_CONT "\n");
1246 printk(KERN_WARNING "\n");
1249 bool find_numa_distance(int distance)
1253 if (distance == node_distance(0, 0))
1256 for (i = 0; i < sched_domains_numa_levels; i++) {
1257 if (sched_domains_numa_distance[i] == distance)
1265 * A system can have three types of NUMA topology:
1266 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1267 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1268 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1270 * The difference between a glueless mesh topology and a backplane
1271 * topology lies in whether communication between not directly
1272 * connected nodes goes through intermediary nodes (where programs
1273 * could run), or through backplane controllers. This affects
1274 * placement of programs.
1276 * The type of topology can be discerned with the following tests:
1277 * - If the maximum distance between any nodes is 1 hop, the system
1278 * is directly connected.
1279 * - If for two nodes A and B, located N > 1 hops away from each other,
1280 * there is an intermediary node C, which is < N hops away from both
1281 * nodes A and B, the system is a glueless mesh.
1283 static void init_numa_topology_type(void)
1287 n = sched_max_numa_distance;
1289 if (sched_domains_numa_levels <= 1) {
1290 sched_numa_topology_type = NUMA_DIRECT;
1294 for_each_online_node(a) {
1295 for_each_online_node(b) {
1296 /* Find two nodes furthest removed from each other. */
1297 if (node_distance(a, b) < n)
1300 /* Is there an intermediary node between a and b? */
1301 for_each_online_node(c) {
1302 if (node_distance(a, c) < n &&
1303 node_distance(b, c) < n) {
1304 sched_numa_topology_type =
1310 sched_numa_topology_type = NUMA_BACKPLANE;
1316 void sched_init_numa(void)
1318 int next_distance, curr_distance = node_distance(0, 0);
1319 struct sched_domain_topology_level *tl;
1323 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
1324 if (!sched_domains_numa_distance)
1327 /* Includes NUMA identity node at level 0. */
1328 sched_domains_numa_distance[level++] = curr_distance;
1329 sched_domains_numa_levels = level;
1332 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1333 * unique distances in the node_distance() table.
1335 * Assumes node_distance(0,j) includes all distances in
1336 * node_distance(i,j) in order to avoid cubic time.
1338 next_distance = curr_distance;
1339 for (i = 0; i < nr_node_ids; i++) {
1340 for (j = 0; j < nr_node_ids; j++) {
1341 for (k = 0; k < nr_node_ids; k++) {
1342 int distance = node_distance(i, k);
1344 if (distance > curr_distance &&
1345 (distance < next_distance ||
1346 next_distance == curr_distance))
1347 next_distance = distance;
1350 * While not a strong assumption it would be nice to know
1351 * about cases where if node A is connected to B, B is not
1352 * equally connected to A.
1354 if (sched_debug() && node_distance(k, i) != distance)
1355 sched_numa_warn("Node-distance not symmetric");
1357 if (sched_debug() && i && !find_numa_distance(distance))
1358 sched_numa_warn("Node-0 not representative");
1360 if (next_distance != curr_distance) {
1361 sched_domains_numa_distance[level++] = next_distance;
1362 sched_domains_numa_levels = level;
1363 curr_distance = next_distance;
1368 * In case of sched_debug() we verify the above assumption.
1378 * 'level' contains the number of unique distances
1380 * The sched_domains_numa_distance[] array includes the actual distance
1385 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1386 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1387 * the array will contain less then 'level' members. This could be
1388 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1389 * in other functions.
1391 * We reset it to 'level' at the end of this function.
1393 sched_domains_numa_levels = 0;
1395 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1396 if (!sched_domains_numa_masks)
1400 * Now for each level, construct a mask per node which contains all
1401 * CPUs of nodes that are that many hops away from us.
1403 for (i = 0; i < level; i++) {
1404 sched_domains_numa_masks[i] =
1405 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1406 if (!sched_domains_numa_masks[i])
1409 for (j = 0; j < nr_node_ids; j++) {
1410 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1414 sched_domains_numa_masks[i][j] = mask;
1417 if (node_distance(j, k) > sched_domains_numa_distance[i])
1420 cpumask_or(mask, mask, cpumask_of_node(k));
1425 /* Compute default topology size */
1426 for (i = 0; sched_domain_topology[i].mask; i++);
1428 tl = kzalloc((i + level + 1) *
1429 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1434 * Copy the default topology bits..
1436 for (i = 0; sched_domain_topology[i].mask; i++)
1437 tl[i] = sched_domain_topology[i];
1440 * Add the NUMA identity distance, aka single NODE.
1442 tl[i++] = (struct sched_domain_topology_level){
1443 .mask = sd_numa_mask,
1449 * .. and append 'j' levels of NUMA goodness.
1451 for (j = 1; j < level; i++, j++) {
1452 tl[i] = (struct sched_domain_topology_level){
1453 .mask = sd_numa_mask,
1454 .sd_flags = cpu_numa_flags,
1455 .flags = SDTL_OVERLAP,
1461 sched_domain_topology = tl;
1463 sched_domains_numa_levels = level;
1464 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1466 init_numa_topology_type();
1469 void sched_domains_numa_masks_set(unsigned int cpu)
1471 int node = cpu_to_node(cpu);
1474 for (i = 0; i < sched_domains_numa_levels; i++) {
1475 for (j = 0; j < nr_node_ids; j++) {
1476 if (node_distance(j, node) <= sched_domains_numa_distance[i])
1477 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1482 void sched_domains_numa_masks_clear(unsigned int cpu)
1486 for (i = 0; i < sched_domains_numa_levels; i++) {
1487 for (j = 0; j < nr_node_ids; j++)
1488 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1492 #endif /* CONFIG_NUMA */
1494 static int __sdt_alloc(const struct cpumask *cpu_map)
1496 struct sched_domain_topology_level *tl;
1499 for_each_sd_topology(tl) {
1500 struct sd_data *sdd = &tl->data;
1502 sdd->sd = alloc_percpu(struct sched_domain *);
1506 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1510 sdd->sg = alloc_percpu(struct sched_group *);
1514 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1518 for_each_cpu(j, cpu_map) {
1519 struct sched_domain *sd;
1520 struct sched_domain_shared *sds;
1521 struct sched_group *sg;
1522 struct sched_group_capacity *sgc;
1524 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1525 GFP_KERNEL, cpu_to_node(j));
1529 *per_cpu_ptr(sdd->sd, j) = sd;
1531 sds = kzalloc_node(sizeof(struct sched_domain_shared),
1532 GFP_KERNEL, cpu_to_node(j));
1536 *per_cpu_ptr(sdd->sds, j) = sds;
1538 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1539 GFP_KERNEL, cpu_to_node(j));
1545 *per_cpu_ptr(sdd->sg, j) = sg;
1547 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1548 GFP_KERNEL, cpu_to_node(j));
1552 #ifdef CONFIG_SCHED_DEBUG
1556 *per_cpu_ptr(sdd->sgc, j) = sgc;
1563 static void __sdt_free(const struct cpumask *cpu_map)
1565 struct sched_domain_topology_level *tl;
1568 for_each_sd_topology(tl) {
1569 struct sd_data *sdd = &tl->data;
1571 for_each_cpu(j, cpu_map) {
1572 struct sched_domain *sd;
1575 sd = *per_cpu_ptr(sdd->sd, j);
1576 if (sd && (sd->flags & SD_OVERLAP))
1577 free_sched_groups(sd->groups, 0);
1578 kfree(*per_cpu_ptr(sdd->sd, j));
1582 kfree(*per_cpu_ptr(sdd->sds, j));
1584 kfree(*per_cpu_ptr(sdd->sg, j));
1586 kfree(*per_cpu_ptr(sdd->sgc, j));
1588 free_percpu(sdd->sd);
1590 free_percpu(sdd->sds);
1592 free_percpu(sdd->sg);
1594 free_percpu(sdd->sgc);
1599 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1600 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1601 struct sched_domain *child, int cpu)
1603 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
1606 sd->level = child->level + 1;
1607 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1610 if (!cpumask_subset(sched_domain_span(child),
1611 sched_domain_span(sd))) {
1612 pr_err("BUG: arch topology borken\n");
1613 #ifdef CONFIG_SCHED_DEBUG
1614 pr_err(" the %s domain not a subset of the %s domain\n",
1615 child->name, sd->name);
1617 /* Fixup, ensure @sd has at least @child cpus. */
1618 cpumask_or(sched_domain_span(sd),
1619 sched_domain_span(sd),
1620 sched_domain_span(child));
1624 set_domain_attribute(sd, attr);
1630 * Build sched domains for a given set of CPUs and attach the sched domains
1631 * to the individual CPUs
1634 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1636 enum s_alloc alloc_state;
1637 struct sched_domain *sd;
1639 struct rq *rq = NULL;
1640 int i, ret = -ENOMEM;
1642 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1643 if (alloc_state != sa_rootdomain)
1646 /* Set up domains for CPUs specified by the cpu_map: */
1647 for_each_cpu(i, cpu_map) {
1648 struct sched_domain_topology_level *tl;
1651 for_each_sd_topology(tl) {
1652 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
1653 if (tl == sched_domain_topology)
1654 *per_cpu_ptr(d.sd, i) = sd;
1655 if (tl->flags & SDTL_OVERLAP)
1656 sd->flags |= SD_OVERLAP;
1657 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
1662 /* Build the groups for the domains */
1663 for_each_cpu(i, cpu_map) {
1664 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1665 sd->span_weight = cpumask_weight(sched_domain_span(sd));
1666 if (sd->flags & SD_OVERLAP) {
1667 if (build_overlap_sched_groups(sd, i))
1670 if (build_sched_groups(sd, i))
1676 /* Calculate CPU capacity for physical packages and nodes */
1677 for (i = nr_cpumask_bits-1; i >= 0; i--) {
1678 if (!cpumask_test_cpu(i, cpu_map))
1681 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1682 claim_allocations(i, sd);
1683 init_sched_groups_capacity(i, sd);
1687 /* Attach the domains */
1689 for_each_cpu(i, cpu_map) {
1691 sd = *per_cpu_ptr(d.sd, i);
1693 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
1694 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
1695 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
1697 cpu_attach_domain(sd, d.rd, i);
1701 if (rq && sched_debug_enabled) {
1702 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
1703 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
1708 __free_domain_allocs(&d, alloc_state, cpu_map);
1712 /* Current sched domains: */
1713 static cpumask_var_t *doms_cur;
1715 /* Number of sched domains in 'doms_cur': */
1716 static int ndoms_cur;
1718 /* Attribues of custom domains in 'doms_cur' */
1719 static struct sched_domain_attr *dattr_cur;
1722 * Special case: If a kmalloc() of a doms_cur partition (array of
1723 * cpumask) fails, then fallback to a single sched domain,
1724 * as determined by the single cpumask fallback_doms.
1726 static cpumask_var_t fallback_doms;
1729 * arch_update_cpu_topology lets virtualized architectures update the
1730 * CPU core maps. It is supposed to return 1 if the topology changed
1731 * or 0 if it stayed the same.
1733 int __weak arch_update_cpu_topology(void)
1738 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
1741 cpumask_var_t *doms;
1743 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
1746 for (i = 0; i < ndoms; i++) {
1747 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
1748 free_sched_domains(doms, i);
1755 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
1758 for (i = 0; i < ndoms; i++)
1759 free_cpumask_var(doms[i]);
1764 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
1765 * For now this just excludes isolated CPUs, but could be used to
1766 * exclude other special cases in the future.
1768 int sched_init_domains(const struct cpumask *cpu_map)
1772 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
1773 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
1774 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
1776 arch_update_cpu_topology();
1778 doms_cur = alloc_sched_domains(ndoms_cur);
1780 doms_cur = &fallback_doms;
1781 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
1782 err = build_sched_domains(doms_cur[0], NULL);
1783 register_sched_domain_sysctl();
1789 * Detach sched domains from a group of CPUs specified in cpu_map
1790 * These CPUs will now be attached to the NULL domain
1792 static void detach_destroy_domains(const struct cpumask *cpu_map)
1797 for_each_cpu(i, cpu_map)
1798 cpu_attach_domain(NULL, &def_root_domain, i);
1802 /* handle null as "default" */
1803 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
1804 struct sched_domain_attr *new, int idx_new)
1806 struct sched_domain_attr tmp;
1813 return !memcmp(cur ? (cur + idx_cur) : &tmp,
1814 new ? (new + idx_new) : &tmp,
1815 sizeof(struct sched_domain_attr));
1819 * Partition sched domains as specified by the 'ndoms_new'
1820 * cpumasks in the array doms_new[] of cpumasks. This compares
1821 * doms_new[] to the current sched domain partitioning, doms_cur[].
1822 * It destroys each deleted domain and builds each new domain.
1824 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
1825 * The masks don't intersect (don't overlap.) We should setup one
1826 * sched domain for each mask. CPUs not in any of the cpumasks will
1827 * not be load balanced. If the same cpumask appears both in the
1828 * current 'doms_cur' domains and in the new 'doms_new', we can leave
1831 * The passed in 'doms_new' should be allocated using
1832 * alloc_sched_domains. This routine takes ownership of it and will
1833 * free_sched_domains it when done with it. If the caller failed the
1834 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
1835 * and partition_sched_domains() will fallback to the single partition
1836 * 'fallback_doms', it also forces the domains to be rebuilt.
1838 * If doms_new == NULL it will be replaced with cpu_online_mask.
1839 * ndoms_new == 0 is a special case for destroying existing domains,
1840 * and it will not create the default domain.
1842 * Call with hotplug lock held
1844 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
1845 struct sched_domain_attr *dattr_new)
1850 mutex_lock(&sched_domains_mutex);
1852 /* Always unregister in case we don't destroy any domains: */
1853 unregister_sched_domain_sysctl();
1855 /* Let the architecture update CPU core mappings: */
1856 new_topology = arch_update_cpu_topology();
1859 WARN_ON_ONCE(dattr_new);
1861 doms_new = alloc_sched_domains(1);
1864 cpumask_and(doms_new[0], cpu_active_mask,
1865 housekeeping_cpumask(HK_FLAG_DOMAIN));
1871 /* Destroy deleted domains: */
1872 for (i = 0; i < ndoms_cur; i++) {
1873 for (j = 0; j < n && !new_topology; j++) {
1874 if (cpumask_equal(doms_cur[i], doms_new[j])
1875 && dattrs_equal(dattr_cur, i, dattr_new, j))
1878 /* No match - a current sched domain not in new doms_new[] */
1879 detach_destroy_domains(doms_cur[i]);
1887 doms_new = &fallback_doms;
1888 cpumask_and(doms_new[0], cpu_active_mask,
1889 housekeeping_cpumask(HK_FLAG_DOMAIN));
1892 /* Build new domains: */
1893 for (i = 0; i < ndoms_new; i++) {
1894 for (j = 0; j < n && !new_topology; j++) {
1895 if (cpumask_equal(doms_new[i], doms_cur[j])
1896 && dattrs_equal(dattr_new, i, dattr_cur, j))
1899 /* No match - add a new doms_new */
1900 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
1905 /* Remember the new sched domains: */
1906 if (doms_cur != &fallback_doms)
1907 free_sched_domains(doms_cur, ndoms_cur);
1910 doms_cur = doms_new;
1911 dattr_cur = dattr_new;
1912 ndoms_cur = ndoms_new;
1914 register_sched_domain_sysctl();
1916 mutex_unlock(&sched_domains_mutex);