1 // SPDX-License-Identifier: GPL-2.0
3 * Scheduler topology setup/handling methods
5 #include <linux/sched.h>
6 #include <linux/mutex.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 void sched_get_rd(struct root_domain *rd)
263 atomic_inc(&rd->refcount);
266 void sched_put_rd(struct root_domain *rd)
268 if (!atomic_dec_and_test(&rd->refcount))
271 call_rcu_sched(&rd->rcu, free_rootdomain);
274 static int init_rootdomain(struct root_domain *rd)
276 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
278 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
280 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
282 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
285 #ifdef HAVE_RT_PUSH_IPI
287 raw_spin_lock_init(&rd->rto_lock);
288 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
291 init_dl_bw(&rd->dl_bw);
292 if (cpudl_init(&rd->cpudl) != 0)
295 if (cpupri_init(&rd->cpupri) != 0)
300 cpudl_cleanup(&rd->cpudl);
302 free_cpumask_var(rd->rto_mask);
304 free_cpumask_var(rd->dlo_mask);
306 free_cpumask_var(rd->online);
308 free_cpumask_var(rd->span);
314 * By default the system creates a single root-domain with all CPUs as
315 * members (mimicking the global state we have today).
317 struct root_domain def_root_domain;
319 void init_defrootdomain(void)
321 init_rootdomain(&def_root_domain);
323 atomic_set(&def_root_domain.refcount, 1);
326 static struct root_domain *alloc_rootdomain(void)
328 struct root_domain *rd;
330 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
334 if (init_rootdomain(rd) != 0) {
342 static void free_sched_groups(struct sched_group *sg, int free_sgc)
344 struct sched_group *tmp, *first;
353 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
356 if (atomic_dec_and_test(&sg->ref))
359 } while (sg != first);
362 static void destroy_sched_domain(struct sched_domain *sd)
365 * A normal sched domain may have multiple group references, an
366 * overlapping domain, having private groups, only one. Iterate,
367 * dropping group/capacity references, freeing where none remain.
369 free_sched_groups(sd->groups, 1);
371 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
376 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
378 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
381 struct sched_domain *parent = sd->parent;
382 destroy_sched_domain(sd);
387 static void destroy_sched_domains(struct sched_domain *sd)
390 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
394 * Keep a special pointer to the highest sched_domain that has
395 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
396 * allows us to avoid some pointer chasing select_idle_sibling().
398 * Also keep a unique ID per domain (we use the first CPU number in
399 * the cpumask of the domain), this allows us to quickly tell if
400 * two CPUs are in the same cache domain, see cpus_share_cache().
402 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
403 DEFINE_PER_CPU(int, sd_llc_size);
404 DEFINE_PER_CPU(int, sd_llc_id);
405 DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
406 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
407 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
409 static void update_top_cache_domain(int cpu)
411 struct sched_domain_shared *sds = NULL;
412 struct sched_domain *sd;
416 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
418 id = cpumask_first(sched_domain_span(sd));
419 size = cpumask_weight(sched_domain_span(sd));
423 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
424 per_cpu(sd_llc_size, cpu) = size;
425 per_cpu(sd_llc_id, cpu) = id;
426 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
428 sd = lowest_flag_domain(cpu, SD_NUMA);
429 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
431 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
432 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
436 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
437 * hold the hotplug lock.
440 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
442 struct rq *rq = cpu_rq(cpu);
443 struct sched_domain *tmp;
445 /* Remove the sched domains which do not contribute to scheduling. */
446 for (tmp = sd; tmp; ) {
447 struct sched_domain *parent = tmp->parent;
451 if (sd_parent_degenerate(tmp, parent)) {
452 tmp->parent = parent->parent;
454 parent->parent->child = tmp;
456 * Transfer SD_PREFER_SIBLING down in case of a
457 * degenerate parent; the spans match for this
458 * so the property transfers.
460 if (parent->flags & SD_PREFER_SIBLING)
461 tmp->flags |= SD_PREFER_SIBLING;
462 destroy_sched_domain(parent);
467 if (sd && sd_degenerate(sd)) {
470 destroy_sched_domain(tmp);
475 sched_domain_debug(sd, cpu);
477 rq_attach_root(rq, rd);
479 rcu_assign_pointer(rq->sd, sd);
480 dirty_sched_domain_sysctl(cpu);
481 destroy_sched_domains(tmp);
483 update_top_cache_domain(cpu);
486 /* Setup the mask of CPUs configured for isolated domains */
487 static int __init isolated_cpu_setup(char *str)
491 alloc_bootmem_cpumask_var(&cpu_isolated_map);
492 ret = cpulist_parse(str, cpu_isolated_map);
494 pr_err("sched: Error, all isolcpus= values must be between 0 and %u\n", nr_cpu_ids);
499 __setup("isolcpus=", isolated_cpu_setup);
502 struct sched_domain * __percpu *sd;
503 struct root_domain *rd;
514 * Return the canonical balance CPU for this group, this is the first CPU
515 * of this group that's also in the balance mask.
517 * The balance mask are all those CPUs that could actually end up at this
518 * group. See build_balance_mask().
520 * Also see should_we_balance().
522 int group_balance_cpu(struct sched_group *sg)
524 return cpumask_first(group_balance_mask(sg));
529 * NUMA topology (first read the regular topology blurb below)
531 * Given a node-distance table, for example:
539 * which represents a 4 node ring topology like:
547 * We want to construct domains and groups to represent this. The way we go
548 * about doing this is to build the domains on 'hops'. For each NUMA level we
549 * construct the mask of all nodes reachable in @level hops.
551 * For the above NUMA topology that gives 3 levels:
553 * NUMA-2 0-3 0-3 0-3 0-3
554 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
556 * NUMA-1 0-1,3 0-2 1-3 0,2-3
557 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
562 * As can be seen; things don't nicely line up as with the regular topology.
563 * When we iterate a domain in child domain chunks some nodes can be
564 * represented multiple times -- hence the "overlap" naming for this part of
567 * In order to minimize this overlap, we only build enough groups to cover the
568 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
572 * - the first group of each domain is its child domain; this
573 * gets us the first 0-1,3
574 * - the only uncovered node is 2, who's child domain is 1-3.
576 * However, because of the overlap, computing a unique CPU for each group is
577 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
578 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
579 * end up at those groups (they would end up in group: 0-1,3).
581 * To correct this we have to introduce the group balance mask. This mask
582 * will contain those CPUs in the group that can reach this group given the
583 * (child) domain tree.
585 * With this we can once again compute balance_cpu and sched_group_capacity
588 * XXX include words on how balance_cpu is unique and therefore can be
589 * used for sched_group_capacity links.
592 * Another 'interesting' topology is:
600 * Which looks a little like:
608 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
611 * This leads to a few particularly weird cases where the sched_domain's are
612 * not of the same number for each cpu. Consider:
615 * groups: {0-2},{1-3} {1-3},{0-2}
617 * NUMA-1 0-2 0-3 0-3 1-3
625 * Build the balance mask; it contains only those CPUs that can arrive at this
626 * group and should be considered to continue balancing.
628 * We do this during the group creation pass, therefore the group information
629 * isn't complete yet, however since each group represents a (child) domain we
630 * can fully construct this using the sched_domain bits (which are already
634 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
636 const struct cpumask *sg_span = sched_group_span(sg);
637 struct sd_data *sdd = sd->private;
638 struct sched_domain *sibling;
643 for_each_cpu(i, sg_span) {
644 sibling = *per_cpu_ptr(sdd->sd, i);
647 * Can happen in the asymmetric case, where these siblings are
648 * unused. The mask will not be empty because those CPUs that
649 * do have the top domain _should_ span the domain.
654 /* If we would not end up here, we can't continue from here */
655 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
658 cpumask_set_cpu(i, mask);
661 /* We must not have empty masks here */
662 WARN_ON_ONCE(cpumask_empty(mask));
666 * XXX: This creates per-node group entries; since the load-balancer will
667 * immediately access remote memory to construct this group's load-balance
668 * statistics having the groups node local is of dubious benefit.
670 static struct sched_group *
671 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
673 struct sched_group *sg;
674 struct cpumask *sg_span;
676 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
677 GFP_KERNEL, cpu_to_node(cpu));
682 sg_span = sched_group_span(sg);
684 cpumask_copy(sg_span, sched_domain_span(sd->child));
686 cpumask_copy(sg_span, sched_domain_span(sd));
688 atomic_inc(&sg->ref);
692 static void init_overlap_sched_group(struct sched_domain *sd,
693 struct sched_group *sg)
695 struct cpumask *mask = sched_domains_tmpmask2;
696 struct sd_data *sdd = sd->private;
697 struct cpumask *sg_span;
700 build_balance_mask(sd, sg, mask);
701 cpu = cpumask_first_and(sched_group_span(sg), mask);
703 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
704 if (atomic_inc_return(&sg->sgc->ref) == 1)
705 cpumask_copy(group_balance_mask(sg), mask);
707 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
710 * Initialize sgc->capacity such that even if we mess up the
711 * domains and no possible iteration will get us here, we won't
714 sg_span = sched_group_span(sg);
715 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
716 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
720 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
722 struct sched_group *first = NULL, *last = NULL, *sg;
723 const struct cpumask *span = sched_domain_span(sd);
724 struct cpumask *covered = sched_domains_tmpmask;
725 struct sd_data *sdd = sd->private;
726 struct sched_domain *sibling;
729 cpumask_clear(covered);
731 for_each_cpu_wrap(i, span, cpu) {
732 struct cpumask *sg_span;
734 if (cpumask_test_cpu(i, covered))
737 sibling = *per_cpu_ptr(sdd->sd, i);
740 * Asymmetric node setups can result in situations where the
741 * domain tree is of unequal depth, make sure to skip domains
742 * that already cover the entire range.
744 * In that case build_sched_domains() will have terminated the
745 * iteration early and our sibling sd spans will be empty.
746 * Domains should always include the CPU they're built on, so
749 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
752 sg = build_group_from_child_sched_domain(sibling, cpu);
756 sg_span = sched_group_span(sg);
757 cpumask_or(covered, covered, sg_span);
759 init_overlap_sched_group(sd, sg);
773 free_sched_groups(first, 0);
780 * Package topology (also see the load-balance blurb in fair.c)
782 * The scheduler builds a tree structure to represent a number of important
783 * topology features. By default (default_topology[]) these include:
785 * - Simultaneous multithreading (SMT)
786 * - Multi-Core Cache (MC)
789 * Where the last one more or less denotes everything up to a NUMA node.
791 * The tree consists of 3 primary data structures:
793 * sched_domain -> sched_group -> sched_group_capacity
797 * The sched_domains are per-cpu and have a two way link (parent & child) and
798 * denote the ever growing mask of CPUs belonging to that level of topology.
800 * Each sched_domain has a circular (double) linked list of sched_group's, each
801 * denoting the domains of the level below (or individual CPUs in case of the
802 * first domain level). The sched_group linked by a sched_domain includes the
803 * CPU of that sched_domain [*].
805 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
807 * CPU 0 1 2 3 4 5 6 7
811 * SMT [ ] [ ] [ ] [ ]
815 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
816 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
817 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
819 * CPU 0 1 2 3 4 5 6 7
821 * One way to think about it is: sched_domain moves you up and down among these
822 * topology levels, while sched_group moves you sideways through it, at child
823 * domain granularity.
825 * sched_group_capacity ensures each unique sched_group has shared storage.
827 * There are two related construction problems, both require a CPU that
828 * uniquely identify each group (for a given domain):
830 * - The first is the balance_cpu (see should_we_balance() and the
831 * load-balance blub in fair.c); for each group we only want 1 CPU to
832 * continue balancing at a higher domain.
834 * - The second is the sched_group_capacity; we want all identical groups
835 * to share a single sched_group_capacity.
837 * Since these topologies are exclusive by construction. That is, its
838 * impossible for an SMT thread to belong to multiple cores, and cores to
839 * be part of multiple caches. There is a very clear and unique location
840 * for each CPU in the hierarchy.
842 * Therefore computing a unique CPU for each group is trivial (the iteration
843 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
844 * group), we can simply pick the first CPU in each group.
847 * [*] in other words, the first group of each domain is its child domain.
850 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
852 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
853 struct sched_domain *child = sd->child;
854 struct sched_group *sg;
857 cpu = cpumask_first(sched_domain_span(child));
859 sg = *per_cpu_ptr(sdd->sg, cpu);
860 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
862 /* For claim_allocations: */
863 atomic_inc(&sg->ref);
864 atomic_inc(&sg->sgc->ref);
867 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
868 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
870 cpumask_set_cpu(cpu, sched_group_span(sg));
871 cpumask_set_cpu(cpu, group_balance_mask(sg));
874 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
875 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
881 * build_sched_groups will build a circular linked list of the groups
882 * covered by the given span, and will set each group's ->cpumask correctly,
883 * and ->cpu_capacity to 0.
885 * Assumes the sched_domain tree is fully constructed
888 build_sched_groups(struct sched_domain *sd, int cpu)
890 struct sched_group *first = NULL, *last = NULL;
891 struct sd_data *sdd = sd->private;
892 const struct cpumask *span = sched_domain_span(sd);
893 struct cpumask *covered;
896 lockdep_assert_held(&sched_domains_mutex);
897 covered = sched_domains_tmpmask;
899 cpumask_clear(covered);
901 for_each_cpu_wrap(i, span, cpu) {
902 struct sched_group *sg;
904 if (cpumask_test_cpu(i, covered))
907 sg = get_group(i, sdd);
909 cpumask_or(covered, covered, sched_group_span(sg));
924 * Initialize sched groups cpu_capacity.
926 * cpu_capacity indicates the capacity of sched group, which is used while
927 * distributing the load between different sched groups in a sched domain.
928 * Typically cpu_capacity for all the groups in a sched domain will be same
929 * unless there are asymmetries in the topology. If there are asymmetries,
930 * group having more cpu_capacity will pickup more load compared to the
931 * group having less cpu_capacity.
933 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
935 struct sched_group *sg = sd->groups;
940 int cpu, max_cpu = -1;
942 sg->group_weight = cpumask_weight(sched_group_span(sg));
944 if (!(sd->flags & SD_ASYM_PACKING))
947 for_each_cpu(cpu, sched_group_span(sg)) {
950 else if (sched_asym_prefer(cpu, max_cpu))
953 sg->asym_prefer_cpu = max_cpu;
957 } while (sg != sd->groups);
959 if (cpu != group_balance_cpu(sg))
962 update_group_capacity(sd, cpu);
966 * Initializers for schedule domains
967 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
970 static int default_relax_domain_level = -1;
971 int sched_domain_level_max;
973 static int __init setup_relax_domain_level(char *str)
975 if (kstrtoint(str, 0, &default_relax_domain_level))
976 pr_warn("Unable to set relax_domain_level\n");
980 __setup("relax_domain_level=", setup_relax_domain_level);
982 static void set_domain_attribute(struct sched_domain *sd,
983 struct sched_domain_attr *attr)
987 if (!attr || attr->relax_domain_level < 0) {
988 if (default_relax_domain_level < 0)
991 request = default_relax_domain_level;
993 request = attr->relax_domain_level;
994 if (request < sd->level) {
995 /* Turn off idle balance on this domain: */
996 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
998 /* Turn on idle balance on this domain: */
999 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1003 static void __sdt_free(const struct cpumask *cpu_map);
1004 static int __sdt_alloc(const struct cpumask *cpu_map);
1006 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1007 const struct cpumask *cpu_map)
1011 if (!atomic_read(&d->rd->refcount))
1012 free_rootdomain(&d->rd->rcu);
1018 __sdt_free(cpu_map);
1026 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1028 memset(d, 0, sizeof(*d));
1030 if (__sdt_alloc(cpu_map))
1031 return sa_sd_storage;
1032 d->sd = alloc_percpu(struct sched_domain *);
1034 return sa_sd_storage;
1035 d->rd = alloc_rootdomain();
1038 return sa_rootdomain;
1042 * NULL the sd_data elements we've used to build the sched_domain and
1043 * sched_group structure so that the subsequent __free_domain_allocs()
1044 * will not free the data we're using.
1046 static void claim_allocations(int cpu, struct sched_domain *sd)
1048 struct sd_data *sdd = sd->private;
1050 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1051 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1053 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1054 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1056 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1057 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1059 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1060 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1064 static int sched_domains_numa_levels;
1065 enum numa_topology_type sched_numa_topology_type;
1066 static int *sched_domains_numa_distance;
1067 int sched_max_numa_distance;
1068 static struct cpumask ***sched_domains_numa_masks;
1069 static int sched_domains_curr_level;
1073 * SD_flags allowed in topology descriptions.
1075 * These flags are purely descriptive of the topology and do not prescribe
1076 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1079 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1080 * SD_SHARE_PKG_RESOURCES - describes shared caches
1081 * SD_NUMA - describes NUMA topologies
1082 * SD_SHARE_POWERDOMAIN - describes shared power domain
1083 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
1085 * Odd one out, which beside describing the topology has a quirk also
1086 * prescribes the desired behaviour that goes along with it:
1088 * SD_ASYM_PACKING - describes SMT quirks
1090 #define TOPOLOGY_SD_FLAGS \
1091 (SD_SHARE_CPUCAPACITY | \
1092 SD_SHARE_PKG_RESOURCES | \
1095 SD_ASYM_CPUCAPACITY | \
1096 SD_SHARE_POWERDOMAIN)
1098 static struct sched_domain *
1099 sd_init(struct sched_domain_topology_level *tl,
1100 const struct cpumask *cpu_map,
1101 struct sched_domain *child, int cpu)
1103 struct sd_data *sdd = &tl->data;
1104 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1105 int sd_id, sd_weight, sd_flags = 0;
1109 * Ugly hack to pass state to sd_numa_mask()...
1111 sched_domains_curr_level = tl->numa_level;
1114 sd_weight = cpumask_weight(tl->mask(cpu));
1117 sd_flags = (*tl->sd_flags)();
1118 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1119 "wrong sd_flags in topology description\n"))
1120 sd_flags &= TOPOLOGY_SD_FLAGS;
1122 *sd = (struct sched_domain){
1123 .min_interval = sd_weight,
1124 .max_interval = 2*sd_weight,
1126 .imbalance_pct = 125,
1128 .cache_nice_tries = 0,
1135 .flags = 1*SD_LOAD_BALANCE
1136 | 1*SD_BALANCE_NEWIDLE
1141 | 0*SD_SHARE_CPUCAPACITY
1142 | 0*SD_SHARE_PKG_RESOURCES
1144 | 0*SD_PREFER_SIBLING
1149 .last_balance = jiffies,
1150 .balance_interval = sd_weight,
1152 .max_newidle_lb_cost = 0,
1153 .next_decay_max_lb_cost = jiffies,
1155 #ifdef CONFIG_SCHED_DEBUG
1160 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1161 sd_id = cpumask_first(sched_domain_span(sd));
1164 * Convert topological properties into behaviour.
1167 if (sd->flags & SD_ASYM_CPUCAPACITY) {
1168 struct sched_domain *t = sd;
1170 for_each_lower_domain(t)
1171 t->flags |= SD_BALANCE_WAKE;
1174 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1175 sd->flags |= SD_PREFER_SIBLING;
1176 sd->imbalance_pct = 110;
1177 sd->smt_gain = 1178; /* ~15% */
1179 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1180 sd->imbalance_pct = 117;
1181 sd->cache_nice_tries = 1;
1185 } else if (sd->flags & SD_NUMA) {
1186 sd->cache_nice_tries = 2;
1190 sd->flags |= SD_SERIALIZE;
1191 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
1192 sd->flags &= ~(SD_BALANCE_EXEC |
1199 sd->flags |= SD_PREFER_SIBLING;
1200 sd->cache_nice_tries = 1;
1206 * For all levels sharing cache; connect a sched_domain_shared
1209 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1210 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1211 atomic_inc(&sd->shared->ref);
1212 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1221 * Topology list, bottom-up.
1223 static struct sched_domain_topology_level default_topology[] = {
1224 #ifdef CONFIG_SCHED_SMT
1225 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1227 #ifdef CONFIG_SCHED_MC
1228 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1230 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1234 static struct sched_domain_topology_level *sched_domain_topology =
1237 #define for_each_sd_topology(tl) \
1238 for (tl = sched_domain_topology; tl->mask; tl++)
1240 void set_sched_topology(struct sched_domain_topology_level *tl)
1242 if (WARN_ON_ONCE(sched_smp_initialized))
1245 sched_domain_topology = tl;
1250 static const struct cpumask *sd_numa_mask(int cpu)
1252 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1255 static void sched_numa_warn(const char *str)
1257 static int done = false;
1265 printk(KERN_WARNING "ERROR: %s\n\n", str);
1267 for (i = 0; i < nr_node_ids; i++) {
1268 printk(KERN_WARNING " ");
1269 for (j = 0; j < nr_node_ids; j++)
1270 printk(KERN_CONT "%02d ", node_distance(i,j));
1271 printk(KERN_CONT "\n");
1273 printk(KERN_WARNING "\n");
1276 bool find_numa_distance(int distance)
1280 if (distance == node_distance(0, 0))
1283 for (i = 0; i < sched_domains_numa_levels; i++) {
1284 if (sched_domains_numa_distance[i] == distance)
1292 * A system can have three types of NUMA topology:
1293 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1294 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1295 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1297 * The difference between a glueless mesh topology and a backplane
1298 * topology lies in whether communication between not directly
1299 * connected nodes goes through intermediary nodes (where programs
1300 * could run), or through backplane controllers. This affects
1301 * placement of programs.
1303 * The type of topology can be discerned with the following tests:
1304 * - If the maximum distance between any nodes is 1 hop, the system
1305 * is directly connected.
1306 * - If for two nodes A and B, located N > 1 hops away from each other,
1307 * there is an intermediary node C, which is < N hops away from both
1308 * nodes A and B, the system is a glueless mesh.
1310 static void init_numa_topology_type(void)
1314 n = sched_max_numa_distance;
1316 if (sched_domains_numa_levels <= 1) {
1317 sched_numa_topology_type = NUMA_DIRECT;
1321 for_each_online_node(a) {
1322 for_each_online_node(b) {
1323 /* Find two nodes furthest removed from each other. */
1324 if (node_distance(a, b) < n)
1327 /* Is there an intermediary node between a and b? */
1328 for_each_online_node(c) {
1329 if (node_distance(a, c) < n &&
1330 node_distance(b, c) < n) {
1331 sched_numa_topology_type =
1337 sched_numa_topology_type = NUMA_BACKPLANE;
1343 void sched_init_numa(void)
1345 int next_distance, curr_distance = node_distance(0, 0);
1346 struct sched_domain_topology_level *tl;
1350 sched_domains_numa_distance = kzalloc(sizeof(int) * (nr_node_ids + 1), GFP_KERNEL);
1351 if (!sched_domains_numa_distance)
1355 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1356 * unique distances in the node_distance() table.
1358 * Assumes node_distance(0,j) includes all distances in
1359 * node_distance(i,j) in order to avoid cubic time.
1361 next_distance = curr_distance;
1362 for (i = 0; i < nr_node_ids; i++) {
1363 for (j = 0; j < nr_node_ids; j++) {
1364 for (k = 0; k < nr_node_ids; k++) {
1365 int distance = node_distance(i, k);
1367 if (distance > curr_distance &&
1368 (distance < next_distance ||
1369 next_distance == curr_distance))
1370 next_distance = distance;
1373 * While not a strong assumption it would be nice to know
1374 * about cases where if node A is connected to B, B is not
1375 * equally connected to A.
1377 if (sched_debug() && node_distance(k, i) != distance)
1378 sched_numa_warn("Node-distance not symmetric");
1380 if (sched_debug() && i && !find_numa_distance(distance))
1381 sched_numa_warn("Node-0 not representative");
1383 if (next_distance != curr_distance) {
1384 sched_domains_numa_distance[level++] = next_distance;
1385 sched_domains_numa_levels = level;
1386 curr_distance = next_distance;
1391 * In case of sched_debug() we verify the above assumption.
1401 * 'level' contains the number of unique distances, excluding the
1402 * identity distance node_distance(i,i).
1404 * The sched_domains_numa_distance[] array includes the actual distance
1409 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1410 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1411 * the array will contain less then 'level' members. This could be
1412 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1413 * in other functions.
1415 * We reset it to 'level' at the end of this function.
1417 sched_domains_numa_levels = 0;
1419 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1420 if (!sched_domains_numa_masks)
1424 * Now for each level, construct a mask per node which contains all
1425 * CPUs of nodes that are that many hops away from us.
1427 for (i = 0; i < level; i++) {
1428 sched_domains_numa_masks[i] =
1429 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1430 if (!sched_domains_numa_masks[i])
1433 for (j = 0; j < nr_node_ids; j++) {
1434 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1438 sched_domains_numa_masks[i][j] = mask;
1441 if (node_distance(j, k) > sched_domains_numa_distance[i])
1444 cpumask_or(mask, mask, cpumask_of_node(k));
1449 /* Compute default topology size */
1450 for (i = 0; sched_domain_topology[i].mask; i++);
1452 tl = kzalloc((i + level + 1) *
1453 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1458 * Copy the default topology bits..
1460 for (i = 0; sched_domain_topology[i].mask; i++)
1461 tl[i] = sched_domain_topology[i];
1464 * .. and append 'j' levels of NUMA goodness.
1466 for (j = 0; j < level; i++, j++) {
1467 tl[i] = (struct sched_domain_topology_level){
1468 .mask = sd_numa_mask,
1469 .sd_flags = cpu_numa_flags,
1470 .flags = SDTL_OVERLAP,
1476 sched_domain_topology = tl;
1478 sched_domains_numa_levels = level;
1479 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1481 init_numa_topology_type();
1484 void sched_domains_numa_masks_set(unsigned int cpu)
1486 int node = cpu_to_node(cpu);
1489 for (i = 0; i < sched_domains_numa_levels; i++) {
1490 for (j = 0; j < nr_node_ids; j++) {
1491 if (node_distance(j, node) <= sched_domains_numa_distance[i])
1492 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1497 void sched_domains_numa_masks_clear(unsigned int cpu)
1501 for (i = 0; i < sched_domains_numa_levels; i++) {
1502 for (j = 0; j < nr_node_ids; j++)
1503 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1507 #endif /* CONFIG_NUMA */
1509 static int __sdt_alloc(const struct cpumask *cpu_map)
1511 struct sched_domain_topology_level *tl;
1514 for_each_sd_topology(tl) {
1515 struct sd_data *sdd = &tl->data;
1517 sdd->sd = alloc_percpu(struct sched_domain *);
1521 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1525 sdd->sg = alloc_percpu(struct sched_group *);
1529 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1533 for_each_cpu(j, cpu_map) {
1534 struct sched_domain *sd;
1535 struct sched_domain_shared *sds;
1536 struct sched_group *sg;
1537 struct sched_group_capacity *sgc;
1539 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1540 GFP_KERNEL, cpu_to_node(j));
1544 *per_cpu_ptr(sdd->sd, j) = sd;
1546 sds = kzalloc_node(sizeof(struct sched_domain_shared),
1547 GFP_KERNEL, cpu_to_node(j));
1551 *per_cpu_ptr(sdd->sds, j) = sds;
1553 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1554 GFP_KERNEL, cpu_to_node(j));
1560 *per_cpu_ptr(sdd->sg, j) = sg;
1562 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1563 GFP_KERNEL, cpu_to_node(j));
1567 #ifdef CONFIG_SCHED_DEBUG
1571 *per_cpu_ptr(sdd->sgc, j) = sgc;
1578 static void __sdt_free(const struct cpumask *cpu_map)
1580 struct sched_domain_topology_level *tl;
1583 for_each_sd_topology(tl) {
1584 struct sd_data *sdd = &tl->data;
1586 for_each_cpu(j, cpu_map) {
1587 struct sched_domain *sd;
1590 sd = *per_cpu_ptr(sdd->sd, j);
1591 if (sd && (sd->flags & SD_OVERLAP))
1592 free_sched_groups(sd->groups, 0);
1593 kfree(*per_cpu_ptr(sdd->sd, j));
1597 kfree(*per_cpu_ptr(sdd->sds, j));
1599 kfree(*per_cpu_ptr(sdd->sg, j));
1601 kfree(*per_cpu_ptr(sdd->sgc, j));
1603 free_percpu(sdd->sd);
1605 free_percpu(sdd->sds);
1607 free_percpu(sdd->sg);
1609 free_percpu(sdd->sgc);
1614 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1615 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1616 struct sched_domain *child, int cpu)
1618 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
1621 sd->level = child->level + 1;
1622 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1625 if (!cpumask_subset(sched_domain_span(child),
1626 sched_domain_span(sd))) {
1627 pr_err("BUG: arch topology borken\n");
1628 #ifdef CONFIG_SCHED_DEBUG
1629 pr_err(" the %s domain not a subset of the %s domain\n",
1630 child->name, sd->name);
1632 /* Fixup, ensure @sd has at least @child cpus. */
1633 cpumask_or(sched_domain_span(sd),
1634 sched_domain_span(sd),
1635 sched_domain_span(child));
1639 set_domain_attribute(sd, attr);
1645 * Build sched domains for a given set of CPUs and attach the sched domains
1646 * to the individual CPUs
1649 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1651 enum s_alloc alloc_state;
1652 struct sched_domain *sd;
1654 struct rq *rq = NULL;
1655 int i, ret = -ENOMEM;
1657 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1658 if (alloc_state != sa_rootdomain)
1661 /* Set up domains for CPUs specified by the cpu_map: */
1662 for_each_cpu(i, cpu_map) {
1663 struct sched_domain_topology_level *tl;
1666 for_each_sd_topology(tl) {
1667 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
1668 if (tl == sched_domain_topology)
1669 *per_cpu_ptr(d.sd, i) = sd;
1670 if (tl->flags & SDTL_OVERLAP)
1671 sd->flags |= SD_OVERLAP;
1672 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
1677 /* Build the groups for the domains */
1678 for_each_cpu(i, cpu_map) {
1679 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1680 sd->span_weight = cpumask_weight(sched_domain_span(sd));
1681 if (sd->flags & SD_OVERLAP) {
1682 if (build_overlap_sched_groups(sd, i))
1685 if (build_sched_groups(sd, i))
1691 /* Calculate CPU capacity for physical packages and nodes */
1692 for (i = nr_cpumask_bits-1; i >= 0; i--) {
1693 if (!cpumask_test_cpu(i, cpu_map))
1696 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1697 claim_allocations(i, sd);
1698 init_sched_groups_capacity(i, sd);
1702 /* Attach the domains */
1704 for_each_cpu(i, cpu_map) {
1706 sd = *per_cpu_ptr(d.sd, i);
1708 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
1709 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
1710 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
1712 cpu_attach_domain(sd, d.rd, i);
1716 if (rq && sched_debug_enabled) {
1717 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
1718 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
1723 __free_domain_allocs(&d, alloc_state, cpu_map);
1727 /* Current sched domains: */
1728 static cpumask_var_t *doms_cur;
1730 /* Number of sched domains in 'doms_cur': */
1731 static int ndoms_cur;
1733 /* Attribues of custom domains in 'doms_cur' */
1734 static struct sched_domain_attr *dattr_cur;
1737 * Special case: If a kmalloc() of a doms_cur partition (array of
1738 * cpumask) fails, then fallback to a single sched domain,
1739 * as determined by the single cpumask fallback_doms.
1741 static cpumask_var_t fallback_doms;
1744 * arch_update_cpu_topology lets virtualized architectures update the
1745 * CPU core maps. It is supposed to return 1 if the topology changed
1746 * or 0 if it stayed the same.
1748 int __weak arch_update_cpu_topology(void)
1753 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
1756 cpumask_var_t *doms;
1758 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
1761 for (i = 0; i < ndoms; i++) {
1762 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
1763 free_sched_domains(doms, i);
1770 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
1773 for (i = 0; i < ndoms; i++)
1774 free_cpumask_var(doms[i]);
1779 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
1780 * For now this just excludes isolated CPUs, but could be used to
1781 * exclude other special cases in the future.
1783 int sched_init_domains(const struct cpumask *cpu_map)
1787 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
1788 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
1789 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
1791 arch_update_cpu_topology();
1793 doms_cur = alloc_sched_domains(ndoms_cur);
1795 doms_cur = &fallback_doms;
1796 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
1797 err = build_sched_domains(doms_cur[0], NULL);
1798 register_sched_domain_sysctl();
1804 * Detach sched domains from a group of CPUs specified in cpu_map
1805 * These CPUs will now be attached to the NULL domain
1807 static void detach_destroy_domains(const struct cpumask *cpu_map)
1812 for_each_cpu(i, cpu_map)
1813 cpu_attach_domain(NULL, &def_root_domain, i);
1817 /* handle null as "default" */
1818 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
1819 struct sched_domain_attr *new, int idx_new)
1821 struct sched_domain_attr tmp;
1828 return !memcmp(cur ? (cur + idx_cur) : &tmp,
1829 new ? (new + idx_new) : &tmp,
1830 sizeof(struct sched_domain_attr));
1834 * Partition sched domains as specified by the 'ndoms_new'
1835 * cpumasks in the array doms_new[] of cpumasks. This compares
1836 * doms_new[] to the current sched domain partitioning, doms_cur[].
1837 * It destroys each deleted domain and builds each new domain.
1839 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
1840 * The masks don't intersect (don't overlap.) We should setup one
1841 * sched domain for each mask. CPUs not in any of the cpumasks will
1842 * not be load balanced. If the same cpumask appears both in the
1843 * current 'doms_cur' domains and in the new 'doms_new', we can leave
1846 * The passed in 'doms_new' should be allocated using
1847 * alloc_sched_domains. This routine takes ownership of it and will
1848 * free_sched_domains it when done with it. If the caller failed the
1849 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
1850 * and partition_sched_domains() will fallback to the single partition
1851 * 'fallback_doms', it also forces the domains to be rebuilt.
1853 * If doms_new == NULL it will be replaced with cpu_online_mask.
1854 * ndoms_new == 0 is a special case for destroying existing domains,
1855 * and it will not create the default domain.
1857 * Call with hotplug lock held
1859 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
1860 struct sched_domain_attr *dattr_new)
1865 mutex_lock(&sched_domains_mutex);
1867 /* Always unregister in case we don't destroy any domains: */
1868 unregister_sched_domain_sysctl();
1870 /* Let the architecture update CPU core mappings: */
1871 new_topology = arch_update_cpu_topology();
1874 WARN_ON_ONCE(dattr_new);
1876 doms_new = alloc_sched_domains(1);
1879 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
1885 /* Destroy deleted domains: */
1886 for (i = 0; i < ndoms_cur; i++) {
1887 for (j = 0; j < n && !new_topology; j++) {
1888 if (cpumask_equal(doms_cur[i], doms_new[j])
1889 && dattrs_equal(dattr_cur, i, dattr_new, j))
1892 /* No match - a current sched domain not in new doms_new[] */
1893 detach_destroy_domains(doms_cur[i]);
1901 doms_new = &fallback_doms;
1902 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
1905 /* Build new domains: */
1906 for (i = 0; i < ndoms_new; i++) {
1907 for (j = 0; j < n && !new_topology; j++) {
1908 if (cpumask_equal(doms_new[i], doms_cur[j])
1909 && dattrs_equal(dattr_new, i, dattr_cur, j))
1912 /* No match - add a new doms_new */
1913 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
1918 /* Remember the new sched domains: */
1919 if (doms_cur != &fallback_doms)
1920 free_sched_domains(doms_cur, ndoms_cur);
1923 doms_cur = doms_new;
1924 dattr_cur = dattr_new;
1925 ndoms_cur = ndoms_new;
1927 register_sched_domain_sysctl();
1929 mutex_unlock(&sched_domains_mutex);