1 // SPDX-License-Identifier: GPL-2.0
3 * Scheduler topology setup/handling methods
7 DEFINE_MUTEX(sched_domains_mutex);
9 /* Protected by sched_domains_mutex: */
10 cpumask_var_t sched_domains_tmpmask;
11 cpumask_var_t sched_domains_tmpmask2;
13 #ifdef CONFIG_SCHED_DEBUG
15 static int __init sched_debug_setup(char *str)
17 sched_debug_enabled = true;
21 early_param("sched_debug", sched_debug_setup);
23 static inline bool sched_debug(void)
25 return sched_debug_enabled;
28 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
29 struct cpumask *groupmask)
31 struct sched_group *group = sd->groups;
33 cpumask_clear(groupmask);
35 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
37 if (!(sd->flags & SD_LOAD_BALANCE)) {
38 printk("does not load-balance\n");
40 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
44 printk(KERN_CONT "span=%*pbl level=%s\n",
45 cpumask_pr_args(sched_domain_span(sd)), sd->name);
47 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
48 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
50 if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
51 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
54 printk(KERN_DEBUG "%*s groups:", level + 1, "");
58 printk(KERN_ERR "ERROR: group is NULL\n");
62 if (!cpumask_weight(sched_group_span(group))) {
63 printk(KERN_CONT "\n");
64 printk(KERN_ERR "ERROR: empty group\n");
68 if (!(sd->flags & SD_OVERLAP) &&
69 cpumask_intersects(groupmask, sched_group_span(group))) {
70 printk(KERN_CONT "\n");
71 printk(KERN_ERR "ERROR: repeated CPUs\n");
75 cpumask_or(groupmask, groupmask, sched_group_span(group));
77 printk(KERN_CONT " %d:{ span=%*pbl",
79 cpumask_pr_args(sched_group_span(group)));
81 if ((sd->flags & SD_OVERLAP) &&
82 !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
83 printk(KERN_CONT " mask=%*pbl",
84 cpumask_pr_args(group_balance_mask(group)));
87 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
88 printk(KERN_CONT " cap=%lu", group->sgc->capacity);
90 if (group == sd->groups && sd->child &&
91 !cpumask_equal(sched_domain_span(sd->child),
92 sched_group_span(group))) {
93 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
96 printk(KERN_CONT " }");
100 if (group != sd->groups)
101 printk(KERN_CONT ",");
103 } while (group != sd->groups);
104 printk(KERN_CONT "\n");
106 if (!cpumask_equal(sched_domain_span(sd), groupmask))
107 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
110 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
111 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
115 static void sched_domain_debug(struct sched_domain *sd, int cpu)
119 if (!sched_debug_enabled)
123 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
127 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
130 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
138 #else /* !CONFIG_SCHED_DEBUG */
140 # define sched_debug_enabled 0
141 # define sched_domain_debug(sd, cpu) do { } while (0)
142 static inline bool sched_debug(void)
146 #endif /* CONFIG_SCHED_DEBUG */
148 static int sd_degenerate(struct sched_domain *sd)
150 if (cpumask_weight(sched_domain_span(sd)) == 1)
153 /* Following flags need at least 2 groups */
154 if (sd->flags & (SD_LOAD_BALANCE |
158 SD_SHARE_CPUCAPACITY |
159 SD_ASYM_CPUCAPACITY |
160 SD_SHARE_PKG_RESOURCES |
161 SD_SHARE_POWERDOMAIN)) {
162 if (sd->groups != sd->groups->next)
166 /* Following flags don't use groups */
167 if (sd->flags & (SD_WAKE_AFFINE))
174 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
176 unsigned long cflags = sd->flags, pflags = parent->flags;
178 if (sd_degenerate(parent))
181 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
184 /* Flags needing groups don't count if only 1 group in parent */
185 if (parent->groups == parent->groups->next) {
186 pflags &= ~(SD_LOAD_BALANCE |
190 SD_ASYM_CPUCAPACITY |
191 SD_SHARE_CPUCAPACITY |
192 SD_SHARE_PKG_RESOURCES |
194 SD_SHARE_POWERDOMAIN);
195 if (nr_node_ids == 1)
196 pflags &= ~SD_SERIALIZE;
198 if (~cflags & pflags)
204 static void free_rootdomain(struct rcu_head *rcu)
206 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
208 cpupri_cleanup(&rd->cpupri);
209 cpudl_cleanup(&rd->cpudl);
210 free_cpumask_var(rd->dlo_mask);
211 free_cpumask_var(rd->rto_mask);
212 free_cpumask_var(rd->online);
213 free_cpumask_var(rd->span);
217 void rq_attach_root(struct rq *rq, struct root_domain *rd)
219 struct root_domain *old_rd = NULL;
222 raw_spin_lock_irqsave(&rq->lock, flags);
227 if (cpumask_test_cpu(rq->cpu, old_rd->online))
230 cpumask_clear_cpu(rq->cpu, old_rd->span);
233 * If we dont want to free the old_rd yet then
234 * set old_rd to NULL to skip the freeing later
237 if (!atomic_dec_and_test(&old_rd->refcount))
241 atomic_inc(&rd->refcount);
244 cpumask_set_cpu(rq->cpu, rd->span);
245 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
248 raw_spin_unlock_irqrestore(&rq->lock, flags);
251 call_rcu_sched(&old_rd->rcu, free_rootdomain);
254 void sched_get_rd(struct root_domain *rd)
256 atomic_inc(&rd->refcount);
259 void sched_put_rd(struct root_domain *rd)
261 if (!atomic_dec_and_test(&rd->refcount))
264 call_rcu_sched(&rd->rcu, free_rootdomain);
267 static int init_rootdomain(struct root_domain *rd)
269 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
271 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
273 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
275 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
278 #ifdef HAVE_RT_PUSH_IPI
280 raw_spin_lock_init(&rd->rto_lock);
281 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
284 init_dl_bw(&rd->dl_bw);
285 if (cpudl_init(&rd->cpudl) != 0)
288 if (cpupri_init(&rd->cpupri) != 0)
293 cpudl_cleanup(&rd->cpudl);
295 free_cpumask_var(rd->rto_mask);
297 free_cpumask_var(rd->dlo_mask);
299 free_cpumask_var(rd->online);
301 free_cpumask_var(rd->span);
307 * By default the system creates a single root-domain with all CPUs as
308 * members (mimicking the global state we have today).
310 struct root_domain def_root_domain;
312 void init_defrootdomain(void)
314 init_rootdomain(&def_root_domain);
316 atomic_set(&def_root_domain.refcount, 1);
319 static struct root_domain *alloc_rootdomain(void)
321 struct root_domain *rd;
323 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
327 if (init_rootdomain(rd) != 0) {
335 static void free_sched_groups(struct sched_group *sg, int free_sgc)
337 struct sched_group *tmp, *first;
346 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
349 if (atomic_dec_and_test(&sg->ref))
352 } while (sg != first);
355 static void destroy_sched_domain(struct sched_domain *sd)
358 * A normal sched domain may have multiple group references, an
359 * overlapping domain, having private groups, only one. Iterate,
360 * dropping group/capacity references, freeing where none remain.
362 free_sched_groups(sd->groups, 1);
364 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
369 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
371 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
374 struct sched_domain *parent = sd->parent;
375 destroy_sched_domain(sd);
380 static void destroy_sched_domains(struct sched_domain *sd)
383 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
387 * Keep a special pointer to the highest sched_domain that has
388 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
389 * allows us to avoid some pointer chasing select_idle_sibling().
391 * Also keep a unique ID per domain (we use the first CPU number in
392 * the cpumask of the domain), this allows us to quickly tell if
393 * two CPUs are in the same cache domain, see cpus_share_cache().
395 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
396 DEFINE_PER_CPU(int, sd_llc_size);
397 DEFINE_PER_CPU(int, sd_llc_id);
398 DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
399 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
400 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
402 static void update_top_cache_domain(int cpu)
404 struct sched_domain_shared *sds = NULL;
405 struct sched_domain *sd;
409 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
411 id = cpumask_first(sched_domain_span(sd));
412 size = cpumask_weight(sched_domain_span(sd));
416 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
417 per_cpu(sd_llc_size, cpu) = size;
418 per_cpu(sd_llc_id, cpu) = id;
419 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
421 sd = lowest_flag_domain(cpu, SD_NUMA);
422 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
424 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
425 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
429 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
430 * hold the hotplug lock.
433 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
435 struct rq *rq = cpu_rq(cpu);
436 struct sched_domain *tmp;
438 /* Remove the sched domains which do not contribute to scheduling. */
439 for (tmp = sd; tmp; ) {
440 struct sched_domain *parent = tmp->parent;
444 if (sd_parent_degenerate(tmp, parent)) {
445 tmp->parent = parent->parent;
447 parent->parent->child = tmp;
449 * Transfer SD_PREFER_SIBLING down in case of a
450 * degenerate parent; the spans match for this
451 * so the property transfers.
453 if (parent->flags & SD_PREFER_SIBLING)
454 tmp->flags |= SD_PREFER_SIBLING;
455 destroy_sched_domain(parent);
460 if (sd && sd_degenerate(sd)) {
463 destroy_sched_domain(tmp);
468 sched_domain_debug(sd, cpu);
470 rq_attach_root(rq, rd);
472 rcu_assign_pointer(rq->sd, sd);
473 dirty_sched_domain_sysctl(cpu);
474 destroy_sched_domains(tmp);
476 update_top_cache_domain(cpu);
480 struct sched_domain ** __percpu sd;
481 struct root_domain *rd;
492 * Return the canonical balance CPU for this group, this is the first CPU
493 * of this group that's also in the balance mask.
495 * The balance mask are all those CPUs that could actually end up at this
496 * group. See build_balance_mask().
498 * Also see should_we_balance().
500 int group_balance_cpu(struct sched_group *sg)
502 return cpumask_first(group_balance_mask(sg));
507 * NUMA topology (first read the regular topology blurb below)
509 * Given a node-distance table, for example:
517 * which represents a 4 node ring topology like:
525 * We want to construct domains and groups to represent this. The way we go
526 * about doing this is to build the domains on 'hops'. For each NUMA level we
527 * construct the mask of all nodes reachable in @level hops.
529 * For the above NUMA topology that gives 3 levels:
531 * NUMA-2 0-3 0-3 0-3 0-3
532 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
534 * NUMA-1 0-1,3 0-2 1-3 0,2-3
535 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
540 * As can be seen; things don't nicely line up as with the regular topology.
541 * When we iterate a domain in child domain chunks some nodes can be
542 * represented multiple times -- hence the "overlap" naming for this part of
545 * In order to minimize this overlap, we only build enough groups to cover the
546 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
550 * - the first group of each domain is its child domain; this
551 * gets us the first 0-1,3
552 * - the only uncovered node is 2, who's child domain is 1-3.
554 * However, because of the overlap, computing a unique CPU for each group is
555 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
556 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
557 * end up at those groups (they would end up in group: 0-1,3).
559 * To correct this we have to introduce the group balance mask. This mask
560 * will contain those CPUs in the group that can reach this group given the
561 * (child) domain tree.
563 * With this we can once again compute balance_cpu and sched_group_capacity
566 * XXX include words on how balance_cpu is unique and therefore can be
567 * used for sched_group_capacity links.
570 * Another 'interesting' topology is:
578 * Which looks a little like:
586 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
589 * This leads to a few particularly weird cases where the sched_domain's are
590 * not of the same number for each CPU. Consider:
593 * groups: {0-2},{1-3} {1-3},{0-2}
595 * NUMA-1 0-2 0-3 0-3 1-3
603 * Build the balance mask; it contains only those CPUs that can arrive at this
604 * group and should be considered to continue balancing.
606 * We do this during the group creation pass, therefore the group information
607 * isn't complete yet, however since each group represents a (child) domain we
608 * can fully construct this using the sched_domain bits (which are already
612 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
614 const struct cpumask *sg_span = sched_group_span(sg);
615 struct sd_data *sdd = sd->private;
616 struct sched_domain *sibling;
621 for_each_cpu(i, sg_span) {
622 sibling = *per_cpu_ptr(sdd->sd, i);
625 * Can happen in the asymmetric case, where these siblings are
626 * unused. The mask will not be empty because those CPUs that
627 * do have the top domain _should_ span the domain.
632 /* If we would not end up here, we can't continue from here */
633 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
636 cpumask_set_cpu(i, mask);
639 /* We must not have empty masks here */
640 WARN_ON_ONCE(cpumask_empty(mask));
644 * XXX: This creates per-node group entries; since the load-balancer will
645 * immediately access remote memory to construct this group's load-balance
646 * statistics having the groups node local is of dubious benefit.
648 static struct sched_group *
649 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
651 struct sched_group *sg;
652 struct cpumask *sg_span;
654 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
655 GFP_KERNEL, cpu_to_node(cpu));
660 sg_span = sched_group_span(sg);
662 cpumask_copy(sg_span, sched_domain_span(sd->child));
664 cpumask_copy(sg_span, sched_domain_span(sd));
666 atomic_inc(&sg->ref);
670 static void init_overlap_sched_group(struct sched_domain *sd,
671 struct sched_group *sg)
673 struct cpumask *mask = sched_domains_tmpmask2;
674 struct sd_data *sdd = sd->private;
675 struct cpumask *sg_span;
678 build_balance_mask(sd, sg, mask);
679 cpu = cpumask_first_and(sched_group_span(sg), mask);
681 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
682 if (atomic_inc_return(&sg->sgc->ref) == 1)
683 cpumask_copy(group_balance_mask(sg), mask);
685 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
688 * Initialize sgc->capacity such that even if we mess up the
689 * domains and no possible iteration will get us here, we won't
692 sg_span = sched_group_span(sg);
693 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
694 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
698 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
700 struct sched_group *first = NULL, *last = NULL, *sg;
701 const struct cpumask *span = sched_domain_span(sd);
702 struct cpumask *covered = sched_domains_tmpmask;
703 struct sd_data *sdd = sd->private;
704 struct sched_domain *sibling;
707 cpumask_clear(covered);
709 for_each_cpu_wrap(i, span, cpu) {
710 struct cpumask *sg_span;
712 if (cpumask_test_cpu(i, covered))
715 sibling = *per_cpu_ptr(sdd->sd, i);
718 * Asymmetric node setups can result in situations where the
719 * domain tree is of unequal depth, make sure to skip domains
720 * that already cover the entire range.
722 * In that case build_sched_domains() will have terminated the
723 * iteration early and our sibling sd spans will be empty.
724 * Domains should always include the CPU they're built on, so
727 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
730 sg = build_group_from_child_sched_domain(sibling, cpu);
734 sg_span = sched_group_span(sg);
735 cpumask_or(covered, covered, sg_span);
737 init_overlap_sched_group(sd, sg);
751 free_sched_groups(first, 0);
758 * Package topology (also see the load-balance blurb in fair.c)
760 * The scheduler builds a tree structure to represent a number of important
761 * topology features. By default (default_topology[]) these include:
763 * - Simultaneous multithreading (SMT)
764 * - Multi-Core Cache (MC)
767 * Where the last one more or less denotes everything up to a NUMA node.
769 * The tree consists of 3 primary data structures:
771 * sched_domain -> sched_group -> sched_group_capacity
775 * The sched_domains are per-CPU and have a two way link (parent & child) and
776 * denote the ever growing mask of CPUs belonging to that level of topology.
778 * Each sched_domain has a circular (double) linked list of sched_group's, each
779 * denoting the domains of the level below (or individual CPUs in case of the
780 * first domain level). The sched_group linked by a sched_domain includes the
781 * CPU of that sched_domain [*].
783 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
785 * CPU 0 1 2 3 4 5 6 7
789 * SMT [ ] [ ] [ ] [ ]
793 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
794 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
795 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
797 * CPU 0 1 2 3 4 5 6 7
799 * One way to think about it is: sched_domain moves you up and down among these
800 * topology levels, while sched_group moves you sideways through it, at child
801 * domain granularity.
803 * sched_group_capacity ensures each unique sched_group has shared storage.
805 * There are two related construction problems, both require a CPU that
806 * uniquely identify each group (for a given domain):
808 * - The first is the balance_cpu (see should_we_balance() and the
809 * load-balance blub in fair.c); for each group we only want 1 CPU to
810 * continue balancing at a higher domain.
812 * - The second is the sched_group_capacity; we want all identical groups
813 * to share a single sched_group_capacity.
815 * Since these topologies are exclusive by construction. That is, its
816 * impossible for an SMT thread to belong to multiple cores, and cores to
817 * be part of multiple caches. There is a very clear and unique location
818 * for each CPU in the hierarchy.
820 * Therefore computing a unique CPU for each group is trivial (the iteration
821 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
822 * group), we can simply pick the first CPU in each group.
825 * [*] in other words, the first group of each domain is its child domain.
828 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
830 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
831 struct sched_domain *child = sd->child;
832 struct sched_group *sg;
835 cpu = cpumask_first(sched_domain_span(child));
837 sg = *per_cpu_ptr(sdd->sg, cpu);
838 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
840 /* For claim_allocations: */
841 atomic_inc(&sg->ref);
842 atomic_inc(&sg->sgc->ref);
845 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
846 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
848 cpumask_set_cpu(cpu, sched_group_span(sg));
849 cpumask_set_cpu(cpu, group_balance_mask(sg));
852 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
853 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
859 * build_sched_groups will build a circular linked list of the groups
860 * covered by the given span, and will set each group's ->cpumask correctly,
861 * and ->cpu_capacity to 0.
863 * Assumes the sched_domain tree is fully constructed
866 build_sched_groups(struct sched_domain *sd, int cpu)
868 struct sched_group *first = NULL, *last = NULL;
869 struct sd_data *sdd = sd->private;
870 const struct cpumask *span = sched_domain_span(sd);
871 struct cpumask *covered;
874 lockdep_assert_held(&sched_domains_mutex);
875 covered = sched_domains_tmpmask;
877 cpumask_clear(covered);
879 for_each_cpu_wrap(i, span, cpu) {
880 struct sched_group *sg;
882 if (cpumask_test_cpu(i, covered))
885 sg = get_group(i, sdd);
887 cpumask_or(covered, covered, sched_group_span(sg));
902 * Initialize sched groups cpu_capacity.
904 * cpu_capacity indicates the capacity of sched group, which is used while
905 * distributing the load between different sched groups in a sched domain.
906 * Typically cpu_capacity for all the groups in a sched domain will be same
907 * unless there are asymmetries in the topology. If there are asymmetries,
908 * group having more cpu_capacity will pickup more load compared to the
909 * group having less cpu_capacity.
911 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
913 struct sched_group *sg = sd->groups;
918 int cpu, max_cpu = -1;
920 sg->group_weight = cpumask_weight(sched_group_span(sg));
922 if (!(sd->flags & SD_ASYM_PACKING))
925 for_each_cpu(cpu, sched_group_span(sg)) {
928 else if (sched_asym_prefer(cpu, max_cpu))
931 sg->asym_prefer_cpu = max_cpu;
935 } while (sg != sd->groups);
937 if (cpu != group_balance_cpu(sg))
940 update_group_capacity(sd, cpu);
944 * Initializers for schedule domains
945 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
948 static int default_relax_domain_level = -1;
949 int sched_domain_level_max;
951 static int __init setup_relax_domain_level(char *str)
953 if (kstrtoint(str, 0, &default_relax_domain_level))
954 pr_warn("Unable to set relax_domain_level\n");
958 __setup("relax_domain_level=", setup_relax_domain_level);
960 static void set_domain_attribute(struct sched_domain *sd,
961 struct sched_domain_attr *attr)
965 if (!attr || attr->relax_domain_level < 0) {
966 if (default_relax_domain_level < 0)
969 request = default_relax_domain_level;
971 request = attr->relax_domain_level;
972 if (request < sd->level) {
973 /* Turn off idle balance on this domain: */
974 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
976 /* Turn on idle balance on this domain: */
977 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
981 static void __sdt_free(const struct cpumask *cpu_map);
982 static int __sdt_alloc(const struct cpumask *cpu_map);
984 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
985 const struct cpumask *cpu_map)
989 if (!atomic_read(&d->rd->refcount))
990 free_rootdomain(&d->rd->rcu);
1004 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1006 memset(d, 0, sizeof(*d));
1008 if (__sdt_alloc(cpu_map))
1009 return sa_sd_storage;
1010 d->sd = alloc_percpu(struct sched_domain *);
1012 return sa_sd_storage;
1013 d->rd = alloc_rootdomain();
1017 return sa_rootdomain;
1021 * NULL the sd_data elements we've used to build the sched_domain and
1022 * sched_group structure so that the subsequent __free_domain_allocs()
1023 * will not free the data we're using.
1025 static void claim_allocations(int cpu, struct sched_domain *sd)
1027 struct sd_data *sdd = sd->private;
1029 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1030 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1032 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1033 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1035 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1036 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1038 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1039 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1043 enum numa_topology_type sched_numa_topology_type;
1045 static int sched_domains_numa_levels;
1046 static int sched_domains_curr_level;
1048 int sched_max_numa_distance;
1049 static int *sched_domains_numa_distance;
1050 static struct cpumask ***sched_domains_numa_masks;
1054 * SD_flags allowed in topology descriptions.
1056 * These flags are purely descriptive of the topology and do not prescribe
1057 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1060 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1061 * SD_SHARE_PKG_RESOURCES - describes shared caches
1062 * SD_NUMA - describes NUMA topologies
1063 * SD_SHARE_POWERDOMAIN - describes shared power domain
1064 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
1066 * Odd one out, which beside describing the topology has a quirk also
1067 * prescribes the desired behaviour that goes along with it:
1069 * SD_ASYM_PACKING - describes SMT quirks
1071 #define TOPOLOGY_SD_FLAGS \
1072 (SD_SHARE_CPUCAPACITY | \
1073 SD_SHARE_PKG_RESOURCES | \
1076 SD_ASYM_CPUCAPACITY | \
1077 SD_SHARE_POWERDOMAIN)
1079 static struct sched_domain *
1080 sd_init(struct sched_domain_topology_level *tl,
1081 const struct cpumask *cpu_map,
1082 struct sched_domain *child, int cpu)
1084 struct sd_data *sdd = &tl->data;
1085 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1086 int sd_id, sd_weight, sd_flags = 0;
1090 * Ugly hack to pass state to sd_numa_mask()...
1092 sched_domains_curr_level = tl->numa_level;
1095 sd_weight = cpumask_weight(tl->mask(cpu));
1098 sd_flags = (*tl->sd_flags)();
1099 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1100 "wrong sd_flags in topology description\n"))
1101 sd_flags &= ~TOPOLOGY_SD_FLAGS;
1103 *sd = (struct sched_domain){
1104 .min_interval = sd_weight,
1105 .max_interval = 2*sd_weight,
1107 .imbalance_pct = 125,
1109 .cache_nice_tries = 0,
1116 .flags = 1*SD_LOAD_BALANCE
1117 | 1*SD_BALANCE_NEWIDLE
1122 | 0*SD_SHARE_CPUCAPACITY
1123 | 0*SD_SHARE_PKG_RESOURCES
1125 | 0*SD_PREFER_SIBLING
1130 .last_balance = jiffies,
1131 .balance_interval = sd_weight,
1133 .max_newidle_lb_cost = 0,
1134 .next_decay_max_lb_cost = jiffies,
1136 #ifdef CONFIG_SCHED_DEBUG
1141 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1142 sd_id = cpumask_first(sched_domain_span(sd));
1145 * Convert topological properties into behaviour.
1148 if (sd->flags & SD_ASYM_CPUCAPACITY) {
1149 struct sched_domain *t = sd;
1151 for_each_lower_domain(t)
1152 t->flags |= SD_BALANCE_WAKE;
1155 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1156 sd->flags |= SD_PREFER_SIBLING;
1157 sd->imbalance_pct = 110;
1158 sd->smt_gain = 1178; /* ~15% */
1160 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1161 sd->flags |= SD_PREFER_SIBLING;
1162 sd->imbalance_pct = 117;
1163 sd->cache_nice_tries = 1;
1167 } else if (sd->flags & SD_NUMA) {
1168 sd->cache_nice_tries = 2;
1172 sd->flags |= SD_SERIALIZE;
1173 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
1174 sd->flags &= ~(SD_BALANCE_EXEC |
1181 sd->flags |= SD_PREFER_SIBLING;
1182 sd->cache_nice_tries = 1;
1188 * For all levels sharing cache; connect a sched_domain_shared
1191 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1192 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1193 atomic_inc(&sd->shared->ref);
1194 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1203 * Topology list, bottom-up.
1205 static struct sched_domain_topology_level default_topology[] = {
1206 #ifdef CONFIG_SCHED_SMT
1207 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1209 #ifdef CONFIG_SCHED_MC
1210 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1212 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1216 static struct sched_domain_topology_level *sched_domain_topology =
1219 #define for_each_sd_topology(tl) \
1220 for (tl = sched_domain_topology; tl->mask; tl++)
1222 void set_sched_topology(struct sched_domain_topology_level *tl)
1224 if (WARN_ON_ONCE(sched_smp_initialized))
1227 sched_domain_topology = tl;
1232 static const struct cpumask *sd_numa_mask(int cpu)
1234 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1237 static void sched_numa_warn(const char *str)
1239 static int done = false;
1247 printk(KERN_WARNING "ERROR: %s\n\n", str);
1249 for (i = 0; i < nr_node_ids; i++) {
1250 printk(KERN_WARNING " ");
1251 for (j = 0; j < nr_node_ids; j++)
1252 printk(KERN_CONT "%02d ", node_distance(i,j));
1253 printk(KERN_CONT "\n");
1255 printk(KERN_WARNING "\n");
1258 bool find_numa_distance(int distance)
1262 if (distance == node_distance(0, 0))
1265 for (i = 0; i < sched_domains_numa_levels; i++) {
1266 if (sched_domains_numa_distance[i] == distance)
1274 * A system can have three types of NUMA topology:
1275 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1276 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1277 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1279 * The difference between a glueless mesh topology and a backplane
1280 * topology lies in whether communication between not directly
1281 * connected nodes goes through intermediary nodes (where programs
1282 * could run), or through backplane controllers. This affects
1283 * placement of programs.
1285 * The type of topology can be discerned with the following tests:
1286 * - If the maximum distance between any nodes is 1 hop, the system
1287 * is directly connected.
1288 * - If for two nodes A and B, located N > 1 hops away from each other,
1289 * there is an intermediary node C, which is < N hops away from both
1290 * nodes A and B, the system is a glueless mesh.
1292 static void init_numa_topology_type(void)
1296 n = sched_max_numa_distance;
1298 if (sched_domains_numa_levels <= 2) {
1299 sched_numa_topology_type = NUMA_DIRECT;
1303 for_each_online_node(a) {
1304 for_each_online_node(b) {
1305 /* Find two nodes furthest removed from each other. */
1306 if (node_distance(a, b) < n)
1309 /* Is there an intermediary node between a and b? */
1310 for_each_online_node(c) {
1311 if (node_distance(a, c) < n &&
1312 node_distance(b, c) < n) {
1313 sched_numa_topology_type =
1319 sched_numa_topology_type = NUMA_BACKPLANE;
1325 void sched_init_numa(void)
1327 int next_distance, curr_distance = node_distance(0, 0);
1328 struct sched_domain_topology_level *tl;
1332 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
1333 if (!sched_domains_numa_distance)
1336 /* Includes NUMA identity node at level 0. */
1337 sched_domains_numa_distance[level++] = curr_distance;
1338 sched_domains_numa_levels = level;
1341 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1342 * unique distances in the node_distance() table.
1344 * Assumes node_distance(0,j) includes all distances in
1345 * node_distance(i,j) in order to avoid cubic time.
1347 next_distance = curr_distance;
1348 for (i = 0; i < nr_node_ids; i++) {
1349 for (j = 0; j < nr_node_ids; j++) {
1350 for (k = 0; k < nr_node_ids; k++) {
1351 int distance = node_distance(i, k);
1353 if (distance > curr_distance &&
1354 (distance < next_distance ||
1355 next_distance == curr_distance))
1356 next_distance = distance;
1359 * While not a strong assumption it would be nice to know
1360 * about cases where if node A is connected to B, B is not
1361 * equally connected to A.
1363 if (sched_debug() && node_distance(k, i) != distance)
1364 sched_numa_warn("Node-distance not symmetric");
1366 if (sched_debug() && i && !find_numa_distance(distance))
1367 sched_numa_warn("Node-0 not representative");
1369 if (next_distance != curr_distance) {
1370 sched_domains_numa_distance[level++] = next_distance;
1371 sched_domains_numa_levels = level;
1372 curr_distance = next_distance;
1377 * In case of sched_debug() we verify the above assumption.
1384 * 'level' contains the number of unique distances
1386 * The sched_domains_numa_distance[] array includes the actual distance
1391 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1392 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1393 * the array will contain less then 'level' members. This could be
1394 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1395 * in other functions.
1397 * We reset it to 'level' at the end of this function.
1399 sched_domains_numa_levels = 0;
1401 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1402 if (!sched_domains_numa_masks)
1406 * Now for each level, construct a mask per node which contains all
1407 * CPUs of nodes that are that many hops away from us.
1409 for (i = 0; i < level; i++) {
1410 sched_domains_numa_masks[i] =
1411 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1412 if (!sched_domains_numa_masks[i])
1415 for (j = 0; j < nr_node_ids; j++) {
1416 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1420 sched_domains_numa_masks[i][j] = mask;
1423 if (node_distance(j, k) > sched_domains_numa_distance[i])
1426 cpumask_or(mask, mask, cpumask_of_node(k));
1431 /* Compute default topology size */
1432 for (i = 0; sched_domain_topology[i].mask; i++);
1434 tl = kzalloc((i + level + 1) *
1435 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1440 * Copy the default topology bits..
1442 for (i = 0; sched_domain_topology[i].mask; i++)
1443 tl[i] = sched_domain_topology[i];
1446 * Add the NUMA identity distance, aka single NODE.
1448 tl[i++] = (struct sched_domain_topology_level){
1449 .mask = sd_numa_mask,
1455 * .. and append 'j' levels of NUMA goodness.
1457 for (j = 1; j < level; i++, j++) {
1458 tl[i] = (struct sched_domain_topology_level){
1459 .mask = sd_numa_mask,
1460 .sd_flags = cpu_numa_flags,
1461 .flags = SDTL_OVERLAP,
1467 sched_domain_topology = tl;
1469 sched_domains_numa_levels = level;
1470 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1472 init_numa_topology_type();
1475 void sched_domains_numa_masks_set(unsigned int cpu)
1477 int node = cpu_to_node(cpu);
1480 for (i = 0; i < sched_domains_numa_levels; i++) {
1481 for (j = 0; j < nr_node_ids; j++) {
1482 if (node_distance(j, node) <= sched_domains_numa_distance[i])
1483 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1488 void sched_domains_numa_masks_clear(unsigned int cpu)
1492 for (i = 0; i < sched_domains_numa_levels; i++) {
1493 for (j = 0; j < nr_node_ids; j++)
1494 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1498 #endif /* CONFIG_NUMA */
1500 static int __sdt_alloc(const struct cpumask *cpu_map)
1502 struct sched_domain_topology_level *tl;
1505 for_each_sd_topology(tl) {
1506 struct sd_data *sdd = &tl->data;
1508 sdd->sd = alloc_percpu(struct sched_domain *);
1512 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1516 sdd->sg = alloc_percpu(struct sched_group *);
1520 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1524 for_each_cpu(j, cpu_map) {
1525 struct sched_domain *sd;
1526 struct sched_domain_shared *sds;
1527 struct sched_group *sg;
1528 struct sched_group_capacity *sgc;
1530 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1531 GFP_KERNEL, cpu_to_node(j));
1535 *per_cpu_ptr(sdd->sd, j) = sd;
1537 sds = kzalloc_node(sizeof(struct sched_domain_shared),
1538 GFP_KERNEL, cpu_to_node(j));
1542 *per_cpu_ptr(sdd->sds, j) = sds;
1544 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1545 GFP_KERNEL, cpu_to_node(j));
1551 *per_cpu_ptr(sdd->sg, j) = sg;
1553 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1554 GFP_KERNEL, cpu_to_node(j));
1558 #ifdef CONFIG_SCHED_DEBUG
1562 *per_cpu_ptr(sdd->sgc, j) = sgc;
1569 static void __sdt_free(const struct cpumask *cpu_map)
1571 struct sched_domain_topology_level *tl;
1574 for_each_sd_topology(tl) {
1575 struct sd_data *sdd = &tl->data;
1577 for_each_cpu(j, cpu_map) {
1578 struct sched_domain *sd;
1581 sd = *per_cpu_ptr(sdd->sd, j);
1582 if (sd && (sd->flags & SD_OVERLAP))
1583 free_sched_groups(sd->groups, 0);
1584 kfree(*per_cpu_ptr(sdd->sd, j));
1588 kfree(*per_cpu_ptr(sdd->sds, j));
1590 kfree(*per_cpu_ptr(sdd->sg, j));
1592 kfree(*per_cpu_ptr(sdd->sgc, j));
1594 free_percpu(sdd->sd);
1596 free_percpu(sdd->sds);
1598 free_percpu(sdd->sg);
1600 free_percpu(sdd->sgc);
1605 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1606 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1607 struct sched_domain *child, int cpu)
1609 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
1612 sd->level = child->level + 1;
1613 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1616 if (!cpumask_subset(sched_domain_span(child),
1617 sched_domain_span(sd))) {
1618 pr_err("BUG: arch topology borken\n");
1619 #ifdef CONFIG_SCHED_DEBUG
1620 pr_err(" the %s domain not a subset of the %s domain\n",
1621 child->name, sd->name);
1623 /* Fixup, ensure @sd has at least @child CPUs. */
1624 cpumask_or(sched_domain_span(sd),
1625 sched_domain_span(sd),
1626 sched_domain_span(child));
1630 set_domain_attribute(sd, attr);
1636 * Build sched domains for a given set of CPUs and attach the sched domains
1637 * to the individual CPUs
1640 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1642 enum s_alloc alloc_state;
1643 struct sched_domain *sd;
1645 struct rq *rq = NULL;
1646 int i, ret = -ENOMEM;
1648 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1649 if (alloc_state != sa_rootdomain)
1652 /* Set up domains for CPUs specified by the cpu_map: */
1653 for_each_cpu(i, cpu_map) {
1654 struct sched_domain_topology_level *tl;
1657 for_each_sd_topology(tl) {
1658 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
1659 if (tl == sched_domain_topology)
1660 *per_cpu_ptr(d.sd, i) = sd;
1661 if (tl->flags & SDTL_OVERLAP)
1662 sd->flags |= SD_OVERLAP;
1663 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
1668 /* Build the groups for the domains */
1669 for_each_cpu(i, cpu_map) {
1670 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1671 sd->span_weight = cpumask_weight(sched_domain_span(sd));
1672 if (sd->flags & SD_OVERLAP) {
1673 if (build_overlap_sched_groups(sd, i))
1676 if (build_sched_groups(sd, i))
1682 /* Calculate CPU capacity for physical packages and nodes */
1683 for (i = nr_cpumask_bits-1; i >= 0; i--) {
1684 if (!cpumask_test_cpu(i, cpu_map))
1687 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1688 claim_allocations(i, sd);
1689 init_sched_groups_capacity(i, sd);
1693 /* Attach the domains */
1695 for_each_cpu(i, cpu_map) {
1697 sd = *per_cpu_ptr(d.sd, i);
1699 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
1700 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
1701 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
1703 cpu_attach_domain(sd, d.rd, i);
1707 if (rq && sched_debug_enabled) {
1708 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
1709 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
1714 __free_domain_allocs(&d, alloc_state, cpu_map);
1719 /* Current sched domains: */
1720 static cpumask_var_t *doms_cur;
1722 /* Number of sched domains in 'doms_cur': */
1723 static int ndoms_cur;
1725 /* Attribues of custom domains in 'doms_cur' */
1726 static struct sched_domain_attr *dattr_cur;
1729 * Special case: If a kmalloc() of a doms_cur partition (array of
1730 * cpumask) fails, then fallback to a single sched domain,
1731 * as determined by the single cpumask fallback_doms.
1733 static cpumask_var_t fallback_doms;
1736 * arch_update_cpu_topology lets virtualized architectures update the
1737 * CPU core maps. It is supposed to return 1 if the topology changed
1738 * or 0 if it stayed the same.
1740 int __weak arch_update_cpu_topology(void)
1745 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
1748 cpumask_var_t *doms;
1750 doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
1753 for (i = 0; i < ndoms; i++) {
1754 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
1755 free_sched_domains(doms, i);
1762 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
1765 for (i = 0; i < ndoms; i++)
1766 free_cpumask_var(doms[i]);
1771 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
1772 * For now this just excludes isolated CPUs, but could be used to
1773 * exclude other special cases in the future.
1775 int sched_init_domains(const struct cpumask *cpu_map)
1779 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
1780 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
1781 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
1783 arch_update_cpu_topology();
1785 doms_cur = alloc_sched_domains(ndoms_cur);
1787 doms_cur = &fallback_doms;
1788 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
1789 err = build_sched_domains(doms_cur[0], NULL);
1790 register_sched_domain_sysctl();
1796 * Detach sched domains from a group of CPUs specified in cpu_map
1797 * These CPUs will now be attached to the NULL domain
1799 static void detach_destroy_domains(const struct cpumask *cpu_map)
1804 for_each_cpu(i, cpu_map)
1805 cpu_attach_domain(NULL, &def_root_domain, i);
1809 /* handle null as "default" */
1810 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
1811 struct sched_domain_attr *new, int idx_new)
1813 struct sched_domain_attr tmp;
1821 return !memcmp(cur ? (cur + idx_cur) : &tmp,
1822 new ? (new + idx_new) : &tmp,
1823 sizeof(struct sched_domain_attr));
1827 * Partition sched domains as specified by the 'ndoms_new'
1828 * cpumasks in the array doms_new[] of cpumasks. This compares
1829 * doms_new[] to the current sched domain partitioning, doms_cur[].
1830 * It destroys each deleted domain and builds each new domain.
1832 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
1833 * The masks don't intersect (don't overlap.) We should setup one
1834 * sched domain for each mask. CPUs not in any of the cpumasks will
1835 * not be load balanced. If the same cpumask appears both in the
1836 * current 'doms_cur' domains and in the new 'doms_new', we can leave
1839 * The passed in 'doms_new' should be allocated using
1840 * alloc_sched_domains. This routine takes ownership of it and will
1841 * free_sched_domains it when done with it. If the caller failed the
1842 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
1843 * and partition_sched_domains() will fallback to the single partition
1844 * 'fallback_doms', it also forces the domains to be rebuilt.
1846 * If doms_new == NULL it will be replaced with cpu_online_mask.
1847 * ndoms_new == 0 is a special case for destroying existing domains,
1848 * and it will not create the default domain.
1850 * Call with hotplug lock held
1852 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
1853 struct sched_domain_attr *dattr_new)
1858 mutex_lock(&sched_domains_mutex);
1860 /* Always unregister in case we don't destroy any domains: */
1861 unregister_sched_domain_sysctl();
1863 /* Let the architecture update CPU core mappings: */
1864 new_topology = arch_update_cpu_topology();
1867 WARN_ON_ONCE(dattr_new);
1869 doms_new = alloc_sched_domains(1);
1872 cpumask_and(doms_new[0], cpu_active_mask,
1873 housekeeping_cpumask(HK_FLAG_DOMAIN));
1879 /* Destroy deleted domains: */
1880 for (i = 0; i < ndoms_cur; i++) {
1881 for (j = 0; j < n && !new_topology; j++) {
1882 if (cpumask_equal(doms_cur[i], doms_new[j])
1883 && dattrs_equal(dattr_cur, i, dattr_new, j))
1886 /* No match - a current sched domain not in new doms_new[] */
1887 detach_destroy_domains(doms_cur[i]);
1895 doms_new = &fallback_doms;
1896 cpumask_and(doms_new[0], cpu_active_mask,
1897 housekeeping_cpumask(HK_FLAG_DOMAIN));
1900 /* Build new domains: */
1901 for (i = 0; i < ndoms_new; i++) {
1902 for (j = 0; j < n && !new_topology; j++) {
1903 if (cpumask_equal(doms_new[i], doms_cur[j])
1904 && dattrs_equal(dattr_new, i, dattr_cur, j))
1907 /* No match - add a new doms_new */
1908 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
1913 /* Remember the new sched domains: */
1914 if (doms_cur != &fallback_doms)
1915 free_sched_domains(doms_cur, ndoms_cur);
1918 doms_cur = doms_new;
1919 dattr_cur = dattr_new;
1920 ndoms_cur = ndoms_new;
1922 register_sched_domain_sysctl();
1924 mutex_unlock(&sched_domains_mutex);