1 // SPDX-License-Identifier: GPL-2.0
3 * Scheduler topology setup/handling methods
6 #include <linux/bsearch.h>
8 DEFINE_MUTEX(sched_domains_mutex);
10 /* Protected by sched_domains_mutex: */
11 static cpumask_var_t sched_domains_tmpmask;
12 static cpumask_var_t sched_domains_tmpmask2;
14 #ifdef CONFIG_SCHED_DEBUG
16 static int __init sched_debug_setup(char *str)
18 sched_debug_verbose = true;
22 early_param("sched_verbose", sched_debug_setup);
24 static inline bool sched_debug(void)
26 return sched_debug_verbose;
29 #define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
30 const struct sd_flag_debug sd_flag_debug[] = {
31 #include <linux/sched/sd_flags.h>
35 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
36 struct cpumask *groupmask)
38 struct sched_group *group = sd->groups;
39 unsigned long flags = sd->flags;
42 cpumask_clear(groupmask);
44 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
45 printk(KERN_CONT "span=%*pbl level=%s\n",
46 cpumask_pr_args(sched_domain_span(sd)), sd->name);
48 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
49 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
51 if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
52 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
55 for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
56 unsigned int flag = BIT(idx);
57 unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
59 if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
60 !(sd->child->flags & flag))
61 printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
62 sd_flag_debug[idx].name);
64 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
65 !(sd->parent->flags & flag))
66 printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
67 sd_flag_debug[idx].name);
70 printk(KERN_DEBUG "%*s groups:", level + 1, "");
74 printk(KERN_ERR "ERROR: group is NULL\n");
78 if (cpumask_empty(sched_group_span(group))) {
79 printk(KERN_CONT "\n");
80 printk(KERN_ERR "ERROR: empty group\n");
84 if (!(sd->flags & SD_OVERLAP) &&
85 cpumask_intersects(groupmask, sched_group_span(group))) {
86 printk(KERN_CONT "\n");
87 printk(KERN_ERR "ERROR: repeated CPUs\n");
91 cpumask_or(groupmask, groupmask, sched_group_span(group));
93 printk(KERN_CONT " %d:{ span=%*pbl",
95 cpumask_pr_args(sched_group_span(group)));
97 if ((sd->flags & SD_OVERLAP) &&
98 !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
99 printk(KERN_CONT " mask=%*pbl",
100 cpumask_pr_args(group_balance_mask(group)));
103 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
104 printk(KERN_CONT " cap=%lu", group->sgc->capacity);
106 if (group == sd->groups && sd->child &&
107 !cpumask_equal(sched_domain_span(sd->child),
108 sched_group_span(group))) {
109 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
112 printk(KERN_CONT " }");
116 if (group != sd->groups)
117 printk(KERN_CONT ",");
119 } while (group != sd->groups);
120 printk(KERN_CONT "\n");
122 if (!cpumask_equal(sched_domain_span(sd), groupmask))
123 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
126 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
127 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
131 static void sched_domain_debug(struct sched_domain *sd, int cpu)
135 if (!sched_debug_verbose)
139 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
143 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
146 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
154 #else /* !CONFIG_SCHED_DEBUG */
156 # define sched_debug_verbose 0
157 # define sched_domain_debug(sd, cpu) do { } while (0)
158 static inline bool sched_debug(void)
162 #endif /* CONFIG_SCHED_DEBUG */
164 /* Generate a mask of SD flags with the SDF_NEEDS_GROUPS metaflag */
165 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
166 static const unsigned int SD_DEGENERATE_GROUPS_MASK =
167 #include <linux/sched/sd_flags.h>
171 static int sd_degenerate(struct sched_domain *sd)
173 if (cpumask_weight(sched_domain_span(sd)) == 1)
176 /* Following flags need at least 2 groups */
177 if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
178 (sd->groups != sd->groups->next))
181 /* Following flags don't use groups */
182 if (sd->flags & (SD_WAKE_AFFINE))
189 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
191 unsigned long cflags = sd->flags, pflags = parent->flags;
193 if (sd_degenerate(parent))
196 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
199 /* Flags needing groups don't count if only 1 group in parent */
200 if (parent->groups == parent->groups->next)
201 pflags &= ~SD_DEGENERATE_GROUPS_MASK;
203 if (~cflags & pflags)
209 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
210 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
211 static unsigned int sysctl_sched_energy_aware = 1;
212 static DEFINE_MUTEX(sched_energy_mutex);
213 static bool sched_energy_update;
215 static bool sched_is_eas_possible(const struct cpumask *cpu_mask)
217 bool any_asym_capacity = false;
218 struct cpufreq_policy *policy;
219 struct cpufreq_governor *gov;
222 /* EAS is enabled for asymmetric CPU capacity topologies. */
223 for_each_cpu(i, cpu_mask) {
224 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, i))) {
225 any_asym_capacity = true;
229 if (!any_asym_capacity) {
231 pr_info("rd %*pbl: Checking EAS, CPUs do not have asymmetric capacities\n",
232 cpumask_pr_args(cpu_mask));
237 /* EAS definitely does *not* handle SMT */
238 if (sched_smt_active()) {
240 pr_info("rd %*pbl: Checking EAS, SMT is not supported\n",
241 cpumask_pr_args(cpu_mask));
246 if (!arch_scale_freq_invariant()) {
248 pr_info("rd %*pbl: Checking EAS: frequency-invariant load tracking not yet supported",
249 cpumask_pr_args(cpu_mask));
254 /* Do not attempt EAS if schedutil is not being used. */
255 for_each_cpu(i, cpu_mask) {
256 policy = cpufreq_cpu_get(i);
259 pr_info("rd %*pbl: Checking EAS, cpufreq policy not set for CPU: %d",
260 cpumask_pr_args(cpu_mask), i);
264 gov = policy->governor;
265 cpufreq_cpu_put(policy);
266 if (gov != &schedutil_gov) {
268 pr_info("rd %*pbl: Checking EAS, schedutil is mandatory\n",
269 cpumask_pr_args(cpu_mask));
278 void rebuild_sched_domains_energy(void)
280 mutex_lock(&sched_energy_mutex);
281 sched_energy_update = true;
282 rebuild_sched_domains();
283 sched_energy_update = false;
284 mutex_unlock(&sched_energy_mutex);
287 #ifdef CONFIG_PROC_SYSCTL
288 static int sched_energy_aware_handler(struct ctl_table *table, int write,
289 void *buffer, size_t *lenp, loff_t *ppos)
293 if (write && !capable(CAP_SYS_ADMIN))
296 if (!sched_is_eas_possible(cpu_active_mask)) {
305 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
307 state = static_branch_unlikely(&sched_energy_present);
308 if (state != sysctl_sched_energy_aware)
309 rebuild_sched_domains_energy();
315 static struct ctl_table sched_energy_aware_sysctls[] = {
317 .procname = "sched_energy_aware",
318 .data = &sysctl_sched_energy_aware,
319 .maxlen = sizeof(unsigned int),
321 .proc_handler = sched_energy_aware_handler,
322 .extra1 = SYSCTL_ZERO,
323 .extra2 = SYSCTL_ONE,
327 static int __init sched_energy_aware_sysctl_init(void)
329 register_sysctl_init("kernel", sched_energy_aware_sysctls);
333 late_initcall(sched_energy_aware_sysctl_init);
336 static void free_pd(struct perf_domain *pd)
338 struct perf_domain *tmp;
347 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
350 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
358 static struct perf_domain *pd_init(int cpu)
360 struct em_perf_domain *obj = em_cpu_get(cpu);
361 struct perf_domain *pd;
365 pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
369 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
377 static void perf_domain_debug(const struct cpumask *cpu_map,
378 struct perf_domain *pd)
380 if (!sched_debug() || !pd)
383 printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
386 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
387 cpumask_first(perf_domain_span(pd)),
388 cpumask_pr_args(perf_domain_span(pd)),
389 em_pd_nr_perf_states(pd->em_pd));
393 printk(KERN_CONT "\n");
396 static void destroy_perf_domain_rcu(struct rcu_head *rp)
398 struct perf_domain *pd;
400 pd = container_of(rp, struct perf_domain, rcu);
404 static void sched_energy_set(bool has_eas)
406 if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
408 pr_info("%s: stopping EAS\n", __func__);
409 static_branch_disable_cpuslocked(&sched_energy_present);
410 } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
412 pr_info("%s: starting EAS\n", __func__);
413 static_branch_enable_cpuslocked(&sched_energy_present);
418 * EAS can be used on a root domain if it meets all the following conditions:
419 * 1. an Energy Model (EM) is available;
420 * 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
421 * 3. no SMT is detected.
422 * 4. schedutil is driving the frequency of all CPUs of the rd;
423 * 5. frequency invariance support is present;
425 static bool build_perf_domains(const struct cpumask *cpu_map)
428 struct perf_domain *pd = NULL, *tmp;
429 int cpu = cpumask_first(cpu_map);
430 struct root_domain *rd = cpu_rq(cpu)->rd;
432 if (!sysctl_sched_energy_aware)
435 if (!sched_is_eas_possible(cpu_map))
438 for_each_cpu(i, cpu_map) {
439 /* Skip already covered CPUs. */
443 /* Create the new pd and add it to the local list. */
451 perf_domain_debug(cpu_map, pd);
453 /* Attach the new list of performance domains to the root domain. */
455 rcu_assign_pointer(rd->pd, pd);
457 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
464 rcu_assign_pointer(rd->pd, NULL);
466 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
471 static void free_pd(struct perf_domain *pd) { }
472 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
474 static void free_rootdomain(struct rcu_head *rcu)
476 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
478 cpupri_cleanup(&rd->cpupri);
479 cpudl_cleanup(&rd->cpudl);
480 free_cpumask_var(rd->dlo_mask);
481 free_cpumask_var(rd->rto_mask);
482 free_cpumask_var(rd->online);
483 free_cpumask_var(rd->span);
488 void rq_attach_root(struct rq *rq, struct root_domain *rd)
490 struct root_domain *old_rd = NULL;
493 rq_lock_irqsave(rq, &rf);
498 if (cpumask_test_cpu(rq->cpu, old_rd->online))
501 cpumask_clear_cpu(rq->cpu, old_rd->span);
504 * If we dont want to free the old_rd yet then
505 * set old_rd to NULL to skip the freeing later
508 if (!atomic_dec_and_test(&old_rd->refcount))
512 atomic_inc(&rd->refcount);
515 cpumask_set_cpu(rq->cpu, rd->span);
516 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
519 rq_unlock_irqrestore(rq, &rf);
522 call_rcu(&old_rd->rcu, free_rootdomain);
525 void sched_get_rd(struct root_domain *rd)
527 atomic_inc(&rd->refcount);
530 void sched_put_rd(struct root_domain *rd)
532 if (!atomic_dec_and_test(&rd->refcount))
535 call_rcu(&rd->rcu, free_rootdomain);
538 static int init_rootdomain(struct root_domain *rd)
540 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
542 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
544 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
546 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
549 #ifdef HAVE_RT_PUSH_IPI
551 raw_spin_lock_init(&rd->rto_lock);
552 rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
556 init_dl_bw(&rd->dl_bw);
557 if (cpudl_init(&rd->cpudl) != 0)
560 if (cpupri_init(&rd->cpupri) != 0)
565 cpudl_cleanup(&rd->cpudl);
567 free_cpumask_var(rd->rto_mask);
569 free_cpumask_var(rd->dlo_mask);
571 free_cpumask_var(rd->online);
573 free_cpumask_var(rd->span);
579 * By default the system creates a single root-domain with all CPUs as
580 * members (mimicking the global state we have today).
582 struct root_domain def_root_domain;
584 void __init init_defrootdomain(void)
586 init_rootdomain(&def_root_domain);
588 atomic_set(&def_root_domain.refcount, 1);
591 static struct root_domain *alloc_rootdomain(void)
593 struct root_domain *rd;
595 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
599 if (init_rootdomain(rd) != 0) {
607 static void free_sched_groups(struct sched_group *sg, int free_sgc)
609 struct sched_group *tmp, *first;
618 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
621 if (atomic_dec_and_test(&sg->ref))
624 } while (sg != first);
627 static void destroy_sched_domain(struct sched_domain *sd)
630 * A normal sched domain may have multiple group references, an
631 * overlapping domain, having private groups, only one. Iterate,
632 * dropping group/capacity references, freeing where none remain.
634 free_sched_groups(sd->groups, 1);
636 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
641 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
643 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
646 struct sched_domain *parent = sd->parent;
647 destroy_sched_domain(sd);
652 static void destroy_sched_domains(struct sched_domain *sd)
655 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
659 * Keep a special pointer to the highest sched_domain that has SD_SHARE_LLC set
660 * (Last Level Cache Domain) for this allows us to avoid some pointer chasing
661 * select_idle_sibling().
663 * Also keep a unique ID per domain (we use the first CPU number in the cpumask
664 * of the domain), this allows us to quickly tell if two CPUs are in the same
665 * cache domain, see cpus_share_cache().
667 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
668 DEFINE_PER_CPU(int, sd_llc_size);
669 DEFINE_PER_CPU(int, sd_llc_id);
670 DEFINE_PER_CPU(int, sd_share_id);
671 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
672 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
673 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
674 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
676 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
677 DEFINE_STATIC_KEY_FALSE(sched_cluster_active);
679 static void update_top_cache_domain(int cpu)
681 struct sched_domain_shared *sds = NULL;
682 struct sched_domain *sd;
686 sd = highest_flag_domain(cpu, SD_SHARE_LLC);
688 id = cpumask_first(sched_domain_span(sd));
689 size = cpumask_weight(sched_domain_span(sd));
693 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
694 per_cpu(sd_llc_size, cpu) = size;
695 per_cpu(sd_llc_id, cpu) = id;
696 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
698 sd = lowest_flag_domain(cpu, SD_CLUSTER);
700 id = cpumask_first(sched_domain_span(sd));
703 * This assignment should be placed after the sd_llc_id as
704 * we want this id equals to cluster id on cluster machines
705 * but equals to LLC id on non-Cluster machines.
707 per_cpu(sd_share_id, cpu) = id;
709 sd = lowest_flag_domain(cpu, SD_NUMA);
710 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
712 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
713 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
715 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
716 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
720 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
721 * hold the hotplug lock.
724 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
726 struct rq *rq = cpu_rq(cpu);
727 struct sched_domain *tmp;
729 /* Remove the sched domains which do not contribute to scheduling. */
730 for (tmp = sd; tmp; ) {
731 struct sched_domain *parent = tmp->parent;
735 if (sd_parent_degenerate(tmp, parent)) {
736 tmp->parent = parent->parent;
738 if (parent->parent) {
739 parent->parent->child = tmp;
740 parent->parent->groups->flags = tmp->flags;
744 * Transfer SD_PREFER_SIBLING down in case of a
745 * degenerate parent; the spans match for this
746 * so the property transfers.
748 if (parent->flags & SD_PREFER_SIBLING)
749 tmp->flags |= SD_PREFER_SIBLING;
750 destroy_sched_domain(parent);
755 if (sd && sd_degenerate(sd)) {
758 destroy_sched_domain(tmp);
760 struct sched_group *sg = sd->groups;
763 * sched groups hold the flags of the child sched
764 * domain for convenience. Clear such flags since
765 * the child is being destroyed.
769 } while (sg != sd->groups);
775 sched_domain_debug(sd, cpu);
777 rq_attach_root(rq, rd);
779 rcu_assign_pointer(rq->sd, sd);
780 dirty_sched_domain_sysctl(cpu);
781 destroy_sched_domains(tmp);
783 update_top_cache_domain(cpu);
787 struct sched_domain * __percpu *sd;
788 struct root_domain *rd;
799 * Return the canonical balance CPU for this group, this is the first CPU
800 * of this group that's also in the balance mask.
802 * The balance mask are all those CPUs that could actually end up at this
803 * group. See build_balance_mask().
805 * Also see should_we_balance().
807 int group_balance_cpu(struct sched_group *sg)
809 return cpumask_first(group_balance_mask(sg));
814 * NUMA topology (first read the regular topology blurb below)
816 * Given a node-distance table, for example:
824 * which represents a 4 node ring topology like:
832 * We want to construct domains and groups to represent this. The way we go
833 * about doing this is to build the domains on 'hops'. For each NUMA level we
834 * construct the mask of all nodes reachable in @level hops.
836 * For the above NUMA topology that gives 3 levels:
838 * NUMA-2 0-3 0-3 0-3 0-3
839 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
841 * NUMA-1 0-1,3 0-2 1-3 0,2-3
842 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
847 * As can be seen; things don't nicely line up as with the regular topology.
848 * When we iterate a domain in child domain chunks some nodes can be
849 * represented multiple times -- hence the "overlap" naming for this part of
852 * In order to minimize this overlap, we only build enough groups to cover the
853 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
857 * - the first group of each domain is its child domain; this
858 * gets us the first 0-1,3
859 * - the only uncovered node is 2, who's child domain is 1-3.
861 * However, because of the overlap, computing a unique CPU for each group is
862 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
863 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
864 * end up at those groups (they would end up in group: 0-1,3).
866 * To correct this we have to introduce the group balance mask. This mask
867 * will contain those CPUs in the group that can reach this group given the
868 * (child) domain tree.
870 * With this we can once again compute balance_cpu and sched_group_capacity
873 * XXX include words on how balance_cpu is unique and therefore can be
874 * used for sched_group_capacity links.
877 * Another 'interesting' topology is:
885 * Which looks a little like:
893 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
896 * This leads to a few particularly weird cases where the sched_domain's are
897 * not of the same number for each CPU. Consider:
900 * groups: {0-2},{1-3} {1-3},{0-2}
902 * NUMA-1 0-2 0-3 0-3 1-3
910 * Build the balance mask; it contains only those CPUs that can arrive at this
911 * group and should be considered to continue balancing.
913 * We do this during the group creation pass, therefore the group information
914 * isn't complete yet, however since each group represents a (child) domain we
915 * can fully construct this using the sched_domain bits (which are already
919 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
921 const struct cpumask *sg_span = sched_group_span(sg);
922 struct sd_data *sdd = sd->private;
923 struct sched_domain *sibling;
928 for_each_cpu(i, sg_span) {
929 sibling = *per_cpu_ptr(sdd->sd, i);
932 * Can happen in the asymmetric case, where these siblings are
933 * unused. The mask will not be empty because those CPUs that
934 * do have the top domain _should_ span the domain.
939 /* If we would not end up here, we can't continue from here */
940 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
943 cpumask_set_cpu(i, mask);
946 /* We must not have empty masks here */
947 WARN_ON_ONCE(cpumask_empty(mask));
951 * XXX: This creates per-node group entries; since the load-balancer will
952 * immediately access remote memory to construct this group's load-balance
953 * statistics having the groups node local is of dubious benefit.
955 static struct sched_group *
956 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
958 struct sched_group *sg;
959 struct cpumask *sg_span;
961 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
962 GFP_KERNEL, cpu_to_node(cpu));
967 sg_span = sched_group_span(sg);
969 cpumask_copy(sg_span, sched_domain_span(sd->child));
970 sg->flags = sd->child->flags;
972 cpumask_copy(sg_span, sched_domain_span(sd));
975 atomic_inc(&sg->ref);
979 static void init_overlap_sched_group(struct sched_domain *sd,
980 struct sched_group *sg)
982 struct cpumask *mask = sched_domains_tmpmask2;
983 struct sd_data *sdd = sd->private;
984 struct cpumask *sg_span;
987 build_balance_mask(sd, sg, mask);
988 cpu = cpumask_first(mask);
990 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
991 if (atomic_inc_return(&sg->sgc->ref) == 1)
992 cpumask_copy(group_balance_mask(sg), mask);
994 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
997 * Initialize sgc->capacity such that even if we mess up the
998 * domains and no possible iteration will get us here, we won't
1001 sg_span = sched_group_span(sg);
1002 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
1003 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1004 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1007 static struct sched_domain *
1008 find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
1011 * The proper descendant would be the one whose child won't span out
1014 while (sibling->child &&
1015 !cpumask_subset(sched_domain_span(sibling->child),
1016 sched_domain_span(sd)))
1017 sibling = sibling->child;
1020 * As we are referencing sgc across different topology level, we need
1021 * to go down to skip those sched_domains which don't contribute to
1022 * scheduling because they will be degenerated in cpu_attach_domain
1024 while (sibling->child &&
1025 cpumask_equal(sched_domain_span(sibling->child),
1026 sched_domain_span(sibling)))
1027 sibling = sibling->child;
1033 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
1035 struct sched_group *first = NULL, *last = NULL, *sg;
1036 const struct cpumask *span = sched_domain_span(sd);
1037 struct cpumask *covered = sched_domains_tmpmask;
1038 struct sd_data *sdd = sd->private;
1039 struct sched_domain *sibling;
1042 cpumask_clear(covered);
1044 for_each_cpu_wrap(i, span, cpu) {
1045 struct cpumask *sg_span;
1047 if (cpumask_test_cpu(i, covered))
1050 sibling = *per_cpu_ptr(sdd->sd, i);
1053 * Asymmetric node setups can result in situations where the
1054 * domain tree is of unequal depth, make sure to skip domains
1055 * that already cover the entire range.
1057 * In that case build_sched_domains() will have terminated the
1058 * iteration early and our sibling sd spans will be empty.
1059 * Domains should always include the CPU they're built on, so
1062 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1066 * Usually we build sched_group by sibling's child sched_domain
1067 * But for machines whose NUMA diameter are 3 or above, we move
1068 * to build sched_group by sibling's proper descendant's child
1069 * domain because sibling's child sched_domain will span out of
1070 * the sched_domain being built as below.
1072 * Smallest diameter=3 topology is:
1080 * 0 --- 1 --- 2 --- 3
1082 * NUMA-3 0-3 N/A N/A 0-3
1083 * groups: {0-2},{1-3} {1-3},{0-2}
1085 * NUMA-2 0-2 0-3 0-3 1-3
1086 * groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2}
1088 * NUMA-1 0-1 0-2 1-3 2-3
1089 * groups: {0},{1} {1},{2},{0} {2},{3},{1} {3},{2}
1093 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1094 * group span isn't a subset of the domain span.
1096 if (sibling->child &&
1097 !cpumask_subset(sched_domain_span(sibling->child), span))
1098 sibling = find_descended_sibling(sd, sibling);
1100 sg = build_group_from_child_sched_domain(sibling, cpu);
1104 sg_span = sched_group_span(sg);
1105 cpumask_or(covered, covered, sg_span);
1107 init_overlap_sched_group(sibling, sg);
1121 free_sched_groups(first, 0);
1128 * Package topology (also see the load-balance blurb in fair.c)
1130 * The scheduler builds a tree structure to represent a number of important
1131 * topology features. By default (default_topology[]) these include:
1133 * - Simultaneous multithreading (SMT)
1134 * - Multi-Core Cache (MC)
1137 * Where the last one more or less denotes everything up to a NUMA node.
1139 * The tree consists of 3 primary data structures:
1141 * sched_domain -> sched_group -> sched_group_capacity
1145 * The sched_domains are per-CPU and have a two way link (parent & child) and
1146 * denote the ever growing mask of CPUs belonging to that level of topology.
1148 * Each sched_domain has a circular (double) linked list of sched_group's, each
1149 * denoting the domains of the level below (or individual CPUs in case of the
1150 * first domain level). The sched_group linked by a sched_domain includes the
1151 * CPU of that sched_domain [*].
1153 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1155 * CPU 0 1 2 3 4 5 6 7
1159 * SMT [ ] [ ] [ ] [ ]
1163 * PKG 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1164 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1165 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1167 * CPU 0 1 2 3 4 5 6 7
1169 * One way to think about it is: sched_domain moves you up and down among these
1170 * topology levels, while sched_group moves you sideways through it, at child
1171 * domain granularity.
1173 * sched_group_capacity ensures each unique sched_group has shared storage.
1175 * There are two related construction problems, both require a CPU that
1176 * uniquely identify each group (for a given domain):
1178 * - The first is the balance_cpu (see should_we_balance() and the
1179 * load-balance blub in fair.c); for each group we only want 1 CPU to
1180 * continue balancing at a higher domain.
1182 * - The second is the sched_group_capacity; we want all identical groups
1183 * to share a single sched_group_capacity.
1185 * Since these topologies are exclusive by construction. That is, its
1186 * impossible for an SMT thread to belong to multiple cores, and cores to
1187 * be part of multiple caches. There is a very clear and unique location
1188 * for each CPU in the hierarchy.
1190 * Therefore computing a unique CPU for each group is trivial (the iteration
1191 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1192 * group), we can simply pick the first CPU in each group.
1195 * [*] in other words, the first group of each domain is its child domain.
1198 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1200 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1201 struct sched_domain *child = sd->child;
1202 struct sched_group *sg;
1203 bool already_visited;
1206 cpu = cpumask_first(sched_domain_span(child));
1208 sg = *per_cpu_ptr(sdd->sg, cpu);
1209 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1211 /* Increase refcounts for claim_allocations: */
1212 already_visited = atomic_inc_return(&sg->ref) > 1;
1213 /* sgc visits should follow a similar trend as sg */
1214 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1216 /* If we have already visited that group, it's already initialized. */
1217 if (already_visited)
1221 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1222 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1223 sg->flags = child->flags;
1225 cpumask_set_cpu(cpu, sched_group_span(sg));
1226 cpumask_set_cpu(cpu, group_balance_mask(sg));
1229 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1230 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1231 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1237 * build_sched_groups will build a circular linked list of the groups
1238 * covered by the given span, will set each group's ->cpumask correctly,
1239 * and will initialize their ->sgc.
1241 * Assumes the sched_domain tree is fully constructed
1244 build_sched_groups(struct sched_domain *sd, int cpu)
1246 struct sched_group *first = NULL, *last = NULL;
1247 struct sd_data *sdd = sd->private;
1248 const struct cpumask *span = sched_domain_span(sd);
1249 struct cpumask *covered;
1252 lockdep_assert_held(&sched_domains_mutex);
1253 covered = sched_domains_tmpmask;
1255 cpumask_clear(covered);
1257 for_each_cpu_wrap(i, span, cpu) {
1258 struct sched_group *sg;
1260 if (cpumask_test_cpu(i, covered))
1263 sg = get_group(i, sdd);
1265 cpumask_or(covered, covered, sched_group_span(sg));
1280 * Initialize sched groups cpu_capacity.
1282 * cpu_capacity indicates the capacity of sched group, which is used while
1283 * distributing the load between different sched groups in a sched domain.
1284 * Typically cpu_capacity for all the groups in a sched domain will be same
1285 * unless there are asymmetries in the topology. If there are asymmetries,
1286 * group having more cpu_capacity will pickup more load compared to the
1287 * group having less cpu_capacity.
1289 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1291 struct sched_group *sg = sd->groups;
1292 struct cpumask *mask = sched_domains_tmpmask2;
1297 int cpu, cores = 0, max_cpu = -1;
1299 sg->group_weight = cpumask_weight(sched_group_span(sg));
1301 cpumask_copy(mask, sched_group_span(sg));
1302 for_each_cpu(cpu, mask) {
1304 #ifdef CONFIG_SCHED_SMT
1305 cpumask_andnot(mask, mask, cpu_smt_mask(cpu));
1310 if (!(sd->flags & SD_ASYM_PACKING))
1313 for_each_cpu(cpu, sched_group_span(sg)) {
1316 else if (sched_asym_prefer(cpu, max_cpu))
1319 sg->asym_prefer_cpu = max_cpu;
1323 } while (sg != sd->groups);
1325 if (cpu != group_balance_cpu(sg))
1328 update_group_capacity(sd, cpu);
1332 * Set of available CPUs grouped by their corresponding capacities
1333 * Each list entry contains a CPU mask reflecting CPUs that share the same
1335 * The lifespan of data is unlimited.
1337 LIST_HEAD(asym_cap_list);
1340 * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1341 * Provides sd_flags reflecting the asymmetry scope.
1344 asym_cpu_capacity_classify(const struct cpumask *sd_span,
1345 const struct cpumask *cpu_map)
1347 struct asym_cap_data *entry;
1348 int count = 0, miss = 0;
1351 * Count how many unique CPU capacities this domain spans across
1352 * (compare sched_domain CPUs mask with ones representing available
1353 * CPUs capacities). Take into account CPUs that might be offline:
1356 list_for_each_entry(entry, &asym_cap_list, link) {
1357 if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1359 else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1363 WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1365 /* No asymmetry detected */
1368 /* Some of the available CPU capacity values have not been detected */
1370 return SD_ASYM_CPUCAPACITY;
1372 /* Full asymmetry */
1373 return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1377 static void free_asym_cap_entry(struct rcu_head *head)
1379 struct asym_cap_data *entry = container_of(head, struct asym_cap_data, rcu);
1383 static inline void asym_cpu_capacity_update_data(int cpu)
1385 unsigned long capacity = arch_scale_cpu_capacity(cpu);
1386 struct asym_cap_data *insert_entry = NULL;
1387 struct asym_cap_data *entry;
1390 * Search if capacity already exits. If not, track which the entry
1391 * where we should insert to keep the list ordered descendingly.
1393 list_for_each_entry(entry, &asym_cap_list, link) {
1394 if (capacity == entry->capacity)
1396 else if (!insert_entry && capacity > entry->capacity)
1397 insert_entry = list_prev_entry(entry, link);
1400 entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1401 if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1403 entry->capacity = capacity;
1405 /* If NULL then the new capacity is the smallest, add last. */
1407 list_add_tail_rcu(&entry->link, &asym_cap_list);
1409 list_add_rcu(&entry->link, &insert_entry->link);
1411 __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1415 * Build-up/update list of CPUs grouped by their capacities
1416 * An update requires explicit request to rebuild sched domains
1417 * with state indicating CPU topology changes.
1419 static void asym_cpu_capacity_scan(void)
1421 struct asym_cap_data *entry, *next;
1424 list_for_each_entry(entry, &asym_cap_list, link)
1425 cpumask_clear(cpu_capacity_span(entry));
1427 for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1428 asym_cpu_capacity_update_data(cpu);
1430 list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1431 if (cpumask_empty(cpu_capacity_span(entry))) {
1432 list_del_rcu(&entry->link);
1433 call_rcu(&entry->rcu, free_asym_cap_entry);
1438 * Only one capacity value has been detected i.e. this system is symmetric.
1439 * No need to keep this data around.
1441 if (list_is_singular(&asym_cap_list)) {
1442 entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1443 list_del_rcu(&entry->link);
1444 call_rcu(&entry->rcu, free_asym_cap_entry);
1449 * Initializers for schedule domains
1450 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1453 static int default_relax_domain_level = -1;
1454 int sched_domain_level_max;
1456 static int __init setup_relax_domain_level(char *str)
1458 if (kstrtoint(str, 0, &default_relax_domain_level))
1459 pr_warn("Unable to set relax_domain_level\n");
1463 __setup("relax_domain_level=", setup_relax_domain_level);
1465 static void set_domain_attribute(struct sched_domain *sd,
1466 struct sched_domain_attr *attr)
1470 if (!attr || attr->relax_domain_level < 0) {
1471 if (default_relax_domain_level < 0)
1473 request = default_relax_domain_level;
1475 request = attr->relax_domain_level;
1477 if (sd->level >= request) {
1478 /* Turn off idle balance on this domain: */
1479 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1483 static void __sdt_free(const struct cpumask *cpu_map);
1484 static int __sdt_alloc(const struct cpumask *cpu_map);
1486 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1487 const struct cpumask *cpu_map)
1491 if (!atomic_read(&d->rd->refcount))
1492 free_rootdomain(&d->rd->rcu);
1498 __sdt_free(cpu_map);
1506 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1508 memset(d, 0, sizeof(*d));
1510 if (__sdt_alloc(cpu_map))
1511 return sa_sd_storage;
1512 d->sd = alloc_percpu(struct sched_domain *);
1514 return sa_sd_storage;
1515 d->rd = alloc_rootdomain();
1519 return sa_rootdomain;
1523 * NULL the sd_data elements we've used to build the sched_domain and
1524 * sched_group structure so that the subsequent __free_domain_allocs()
1525 * will not free the data we're using.
1527 static void claim_allocations(int cpu, struct sched_domain *sd)
1529 struct sd_data *sdd = sd->private;
1531 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1532 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1534 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1535 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1537 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1538 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1540 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1541 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1545 enum numa_topology_type sched_numa_topology_type;
1547 static int sched_domains_numa_levels;
1548 static int sched_domains_curr_level;
1550 int sched_max_numa_distance;
1551 static int *sched_domains_numa_distance;
1552 static struct cpumask ***sched_domains_numa_masks;
1556 * SD_flags allowed in topology descriptions.
1558 * These flags are purely descriptive of the topology and do not prescribe
1559 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1560 * function. For details, see include/linux/sched/sd_flags.h.
1562 * SD_SHARE_CPUCAPACITY
1567 * Odd one out, which beside describing the topology has a quirk also
1568 * prescribes the desired behaviour that goes along with it:
1570 * SD_ASYM_PACKING - describes SMT quirks
1572 #define TOPOLOGY_SD_FLAGS \
1573 (SD_SHARE_CPUCAPACITY | \
1579 static struct sched_domain *
1580 sd_init(struct sched_domain_topology_level *tl,
1581 const struct cpumask *cpu_map,
1582 struct sched_domain *child, int cpu)
1584 struct sd_data *sdd = &tl->data;
1585 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1586 int sd_id, sd_weight, sd_flags = 0;
1587 struct cpumask *sd_span;
1591 * Ugly hack to pass state to sd_numa_mask()...
1593 sched_domains_curr_level = tl->numa_level;
1596 sd_weight = cpumask_weight(tl->mask(cpu));
1599 sd_flags = (*tl->sd_flags)();
1600 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1601 "wrong sd_flags in topology description\n"))
1602 sd_flags &= TOPOLOGY_SD_FLAGS;
1604 *sd = (struct sched_domain){
1605 .min_interval = sd_weight,
1606 .max_interval = 2*sd_weight,
1608 .imbalance_pct = 117,
1610 .cache_nice_tries = 0,
1612 .flags = 1*SD_BALANCE_NEWIDLE
1617 | 0*SD_SHARE_CPUCAPACITY
1620 | 1*SD_PREFER_SIBLING
1625 .last_balance = jiffies,
1626 .balance_interval = sd_weight,
1627 .max_newidle_lb_cost = 0,
1628 .last_decay_max_lb_cost = jiffies,
1630 #ifdef CONFIG_SCHED_DEBUG
1635 sd_span = sched_domain_span(sd);
1636 cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1637 sd_id = cpumask_first(sd_span);
1639 sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1641 WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1642 (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1643 "CPU capacity asymmetry not supported on SMT\n");
1646 * Convert topological properties into behaviour.
1648 /* Don't attempt to spread across CPUs of different capacities. */
1649 if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1650 sd->child->flags &= ~SD_PREFER_SIBLING;
1652 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1653 sd->imbalance_pct = 110;
1655 } else if (sd->flags & SD_SHARE_LLC) {
1656 sd->imbalance_pct = 117;
1657 sd->cache_nice_tries = 1;
1660 } else if (sd->flags & SD_NUMA) {
1661 sd->cache_nice_tries = 2;
1663 sd->flags &= ~SD_PREFER_SIBLING;
1664 sd->flags |= SD_SERIALIZE;
1665 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1666 sd->flags &= ~(SD_BALANCE_EXEC |
1673 sd->cache_nice_tries = 1;
1677 * For all levels sharing cache; connect a sched_domain_shared
1680 if (sd->flags & SD_SHARE_LLC) {
1681 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1682 atomic_inc(&sd->shared->ref);
1683 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1692 * Topology list, bottom-up.
1694 static struct sched_domain_topology_level default_topology[] = {
1695 #ifdef CONFIG_SCHED_SMT
1696 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1699 #ifdef CONFIG_SCHED_CLUSTER
1700 { cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1703 #ifdef CONFIG_SCHED_MC
1704 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1706 { cpu_cpu_mask, SD_INIT_NAME(PKG) },
1710 static struct sched_domain_topology_level *sched_domain_topology =
1712 static struct sched_domain_topology_level *sched_domain_topology_saved;
1714 #define for_each_sd_topology(tl) \
1715 for (tl = sched_domain_topology; tl->mask; tl++)
1717 void __init set_sched_topology(struct sched_domain_topology_level *tl)
1719 if (WARN_ON_ONCE(sched_smp_initialized))
1722 sched_domain_topology = tl;
1723 sched_domain_topology_saved = NULL;
1728 static const struct cpumask *sd_numa_mask(int cpu)
1730 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1733 static void sched_numa_warn(const char *str)
1735 static int done = false;
1743 printk(KERN_WARNING "ERROR: %s\n\n", str);
1745 for (i = 0; i < nr_node_ids; i++) {
1746 printk(KERN_WARNING " ");
1747 for (j = 0; j < nr_node_ids; j++) {
1748 if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
1749 printk(KERN_CONT "(%02d) ", node_distance(i,j));
1751 printk(KERN_CONT " %02d ", node_distance(i,j));
1753 printk(KERN_CONT "\n");
1755 printk(KERN_WARNING "\n");
1758 bool find_numa_distance(int distance)
1763 if (distance == node_distance(0, 0))
1767 distances = rcu_dereference(sched_domains_numa_distance);
1770 for (i = 0; i < sched_domains_numa_levels; i++) {
1771 if (distances[i] == distance) {
1782 #define for_each_cpu_node_but(n, nbut) \
1783 for_each_node_state(n, N_CPU) \
1789 * A system can have three types of NUMA topology:
1790 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1791 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1792 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1794 * The difference between a glueless mesh topology and a backplane
1795 * topology lies in whether communication between not directly
1796 * connected nodes goes through intermediary nodes (where programs
1797 * could run), or through backplane controllers. This affects
1798 * placement of programs.
1800 * The type of topology can be discerned with the following tests:
1801 * - If the maximum distance between any nodes is 1 hop, the system
1802 * is directly connected.
1803 * - If for two nodes A and B, located N > 1 hops away from each other,
1804 * there is an intermediary node C, which is < N hops away from both
1805 * nodes A and B, the system is a glueless mesh.
1807 static void init_numa_topology_type(int offline_node)
1811 n = sched_max_numa_distance;
1813 if (sched_domains_numa_levels <= 2) {
1814 sched_numa_topology_type = NUMA_DIRECT;
1818 for_each_cpu_node_but(a, offline_node) {
1819 for_each_cpu_node_but(b, offline_node) {
1820 /* Find two nodes furthest removed from each other. */
1821 if (node_distance(a, b) < n)
1824 /* Is there an intermediary node between a and b? */
1825 for_each_cpu_node_but(c, offline_node) {
1826 if (node_distance(a, c) < n &&
1827 node_distance(b, c) < n) {
1828 sched_numa_topology_type =
1834 sched_numa_topology_type = NUMA_BACKPLANE;
1839 pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1840 sched_numa_topology_type = NUMA_DIRECT;
1844 #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1846 void sched_init_numa(int offline_node)
1848 struct sched_domain_topology_level *tl;
1849 unsigned long *distance_map;
1853 struct cpumask ***masks;
1856 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1857 * unique distances in the node_distance() table.
1859 distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1863 bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1864 for_each_cpu_node_but(i, offline_node) {
1865 for_each_cpu_node_but(j, offline_node) {
1866 int distance = node_distance(i, j);
1868 if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1869 sched_numa_warn("Invalid distance value range");
1870 bitmap_free(distance_map);
1874 bitmap_set(distance_map, distance, 1);
1878 * We can now figure out how many unique distance values there are and
1879 * allocate memory accordingly.
1881 nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1883 distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1885 bitmap_free(distance_map);
1889 for (i = 0, j = 0; i < nr_levels; i++, j++) {
1890 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1893 rcu_assign_pointer(sched_domains_numa_distance, distances);
1895 bitmap_free(distance_map);
1898 * 'nr_levels' contains the number of unique distances
1900 * The sched_domains_numa_distance[] array includes the actual distance
1905 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1906 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1907 * the array will contain less then 'nr_levels' members. This could be
1908 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1909 * in other functions.
1911 * We reset it to 'nr_levels' at the end of this function.
1913 sched_domains_numa_levels = 0;
1915 masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1920 * Now for each level, construct a mask per node which contains all
1921 * CPUs of nodes that are that many hops away from us.
1923 for (i = 0; i < nr_levels; i++) {
1924 masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1928 for_each_cpu_node_but(j, offline_node) {
1929 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1937 for_each_cpu_node_but(k, offline_node) {
1938 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1939 sched_numa_warn("Node-distance not symmetric");
1941 if (node_distance(j, k) > sched_domains_numa_distance[i])
1944 cpumask_or(mask, mask, cpumask_of_node(k));
1948 rcu_assign_pointer(sched_domains_numa_masks, masks);
1950 /* Compute default topology size */
1951 for (i = 0; sched_domain_topology[i].mask; i++);
1953 tl = kzalloc((i + nr_levels + 1) *
1954 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1959 * Copy the default topology bits..
1961 for (i = 0; sched_domain_topology[i].mask; i++)
1962 tl[i] = sched_domain_topology[i];
1965 * Add the NUMA identity distance, aka single NODE.
1967 tl[i++] = (struct sched_domain_topology_level){
1968 .mask = sd_numa_mask,
1974 * .. and append 'j' levels of NUMA goodness.
1976 for (j = 1; j < nr_levels; i++, j++) {
1977 tl[i] = (struct sched_domain_topology_level){
1978 .mask = sd_numa_mask,
1979 .sd_flags = cpu_numa_flags,
1980 .flags = SDTL_OVERLAP,
1986 sched_domain_topology_saved = sched_domain_topology;
1987 sched_domain_topology = tl;
1989 sched_domains_numa_levels = nr_levels;
1990 WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]);
1992 init_numa_topology_type(offline_node);
1996 static void sched_reset_numa(void)
1998 int nr_levels, *distances;
1999 struct cpumask ***masks;
2001 nr_levels = sched_domains_numa_levels;
2002 sched_domains_numa_levels = 0;
2003 sched_max_numa_distance = 0;
2004 sched_numa_topology_type = NUMA_DIRECT;
2005 distances = sched_domains_numa_distance;
2006 rcu_assign_pointer(sched_domains_numa_distance, NULL);
2007 masks = sched_domains_numa_masks;
2008 rcu_assign_pointer(sched_domains_numa_masks, NULL);
2009 if (distances || masks) {
2014 for (i = 0; i < nr_levels && masks; i++) {
2023 if (sched_domain_topology_saved) {
2024 kfree(sched_domain_topology);
2025 sched_domain_topology = sched_domain_topology_saved;
2026 sched_domain_topology_saved = NULL;
2031 * Call with hotplug lock held
2033 void sched_update_numa(int cpu, bool online)
2037 node = cpu_to_node(cpu);
2039 * Scheduler NUMA topology is updated when the first CPU of a
2040 * node is onlined or the last CPU of a node is offlined.
2042 if (cpumask_weight(cpumask_of_node(node)) != 1)
2046 sched_init_numa(online ? NUMA_NO_NODE : node);
2049 void sched_domains_numa_masks_set(unsigned int cpu)
2051 int node = cpu_to_node(cpu);
2054 for (i = 0; i < sched_domains_numa_levels; i++) {
2055 for (j = 0; j < nr_node_ids; j++) {
2056 if (!node_state(j, N_CPU))
2059 /* Set ourselves in the remote node's masks */
2060 if (node_distance(j, node) <= sched_domains_numa_distance[i])
2061 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
2066 void sched_domains_numa_masks_clear(unsigned int cpu)
2070 for (i = 0; i < sched_domains_numa_levels; i++) {
2071 for (j = 0; j < nr_node_ids; j++) {
2072 if (sched_domains_numa_masks[i][j])
2073 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
2079 * sched_numa_find_closest() - given the NUMA topology, find the cpu
2080 * closest to @cpu from @cpumask.
2081 * cpumask: cpumask to find a cpu from
2082 * cpu: cpu to be close to
2084 * returns: cpu, or nr_cpu_ids when nothing found.
2086 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2088 int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2089 struct cpumask ***masks;
2092 masks = rcu_dereference(sched_domains_numa_masks);
2095 for (i = 0; i < sched_domains_numa_levels; i++) {
2098 cpu = cpumask_any_and(cpus, masks[i][j]);
2099 if (cpu < nr_cpu_ids) {
2111 const struct cpumask *cpus;
2112 struct cpumask ***masks;
2118 static int hop_cmp(const void *a, const void *b)
2120 struct cpumask **prev_hop, **cur_hop = *(struct cpumask ***)b;
2121 struct __cmp_key *k = (struct __cmp_key *)a;
2123 if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu)
2126 if (b == k->masks) {
2131 prev_hop = *((struct cpumask ***)b - 1);
2132 k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]);
2140 * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth closest CPU
2141 * from @cpus to @cpu, taking into account distance
2142 * from a given @node.
2143 * @cpus: cpumask to find a cpu from
2144 * @cpu: CPU to start searching
2145 * @node: NUMA node to order CPUs by distance
2147 * Return: cpu, or nr_cpu_ids when nothing found.
2149 int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
2151 struct __cmp_key k = { .cpus = cpus, .cpu = cpu };
2152 struct cpumask ***hop_masks;
2153 int hop, ret = nr_cpu_ids;
2155 if (node == NUMA_NO_NODE)
2156 return cpumask_nth_and(cpu, cpus, cpu_online_mask);
2160 /* CPU-less node entries are uninitialized in sched_domains_numa_masks */
2161 node = numa_nearest_node(node, N_CPU);
2164 k.masks = rcu_dereference(sched_domains_numa_masks);
2168 hop_masks = bsearch(&k, k.masks, sched_domains_numa_levels, sizeof(k.masks[0]), hop_cmp);
2169 hop = hop_masks - k.masks;
2172 cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) :
2173 cpumask_nth_and(cpu, cpus, k.masks[0][node]);
2178 EXPORT_SYMBOL_GPL(sched_numa_find_nth_cpu);
2181 * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2183 * @node: The node to count hops from.
2184 * @hops: Include CPUs up to that many hops away. 0 means local node.
2186 * Return: On success, a pointer to a cpumask of CPUs at most @hops away from
2187 * @node, an error value otherwise.
2189 * Requires rcu_lock to be held. Returned cpumask is only valid within that
2190 * read-side section, copy it if required beyond that.
2192 * Note that not all hops are equal in distance; see sched_init_numa() for how
2193 * distances and masks are handled.
2194 * Also note that this is a reflection of sched_domains_numa_masks, which may change
2195 * during the lifetime of the system (offline nodes are taken out of the masks).
2197 const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops)
2199 struct cpumask ***masks;
2201 if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
2202 return ERR_PTR(-EINVAL);
2204 masks = rcu_dereference(sched_domains_numa_masks);
2206 return ERR_PTR(-EBUSY);
2208 return masks[hops][node];
2210 EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
2212 #endif /* CONFIG_NUMA */
2214 static int __sdt_alloc(const struct cpumask *cpu_map)
2216 struct sched_domain_topology_level *tl;
2219 for_each_sd_topology(tl) {
2220 struct sd_data *sdd = &tl->data;
2222 sdd->sd = alloc_percpu(struct sched_domain *);
2226 sdd->sds = alloc_percpu(struct sched_domain_shared *);
2230 sdd->sg = alloc_percpu(struct sched_group *);
2234 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2238 for_each_cpu(j, cpu_map) {
2239 struct sched_domain *sd;
2240 struct sched_domain_shared *sds;
2241 struct sched_group *sg;
2242 struct sched_group_capacity *sgc;
2244 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2245 GFP_KERNEL, cpu_to_node(j));
2249 *per_cpu_ptr(sdd->sd, j) = sd;
2251 sds = kzalloc_node(sizeof(struct sched_domain_shared),
2252 GFP_KERNEL, cpu_to_node(j));
2256 *per_cpu_ptr(sdd->sds, j) = sds;
2258 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2259 GFP_KERNEL, cpu_to_node(j));
2265 *per_cpu_ptr(sdd->sg, j) = sg;
2267 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2268 GFP_KERNEL, cpu_to_node(j));
2272 #ifdef CONFIG_SCHED_DEBUG
2276 *per_cpu_ptr(sdd->sgc, j) = sgc;
2283 static void __sdt_free(const struct cpumask *cpu_map)
2285 struct sched_domain_topology_level *tl;
2288 for_each_sd_topology(tl) {
2289 struct sd_data *sdd = &tl->data;
2291 for_each_cpu(j, cpu_map) {
2292 struct sched_domain *sd;
2295 sd = *per_cpu_ptr(sdd->sd, j);
2296 if (sd && (sd->flags & SD_OVERLAP))
2297 free_sched_groups(sd->groups, 0);
2298 kfree(*per_cpu_ptr(sdd->sd, j));
2302 kfree(*per_cpu_ptr(sdd->sds, j));
2304 kfree(*per_cpu_ptr(sdd->sg, j));
2306 kfree(*per_cpu_ptr(sdd->sgc, j));
2308 free_percpu(sdd->sd);
2310 free_percpu(sdd->sds);
2312 free_percpu(sdd->sg);
2314 free_percpu(sdd->sgc);
2319 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2320 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2321 struct sched_domain *child, int cpu)
2323 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2326 sd->level = child->level + 1;
2327 sched_domain_level_max = max(sched_domain_level_max, sd->level);
2330 if (!cpumask_subset(sched_domain_span(child),
2331 sched_domain_span(sd))) {
2332 pr_err("BUG: arch topology borken\n");
2333 #ifdef CONFIG_SCHED_DEBUG
2334 pr_err(" the %s domain not a subset of the %s domain\n",
2335 child->name, sd->name);
2337 /* Fixup, ensure @sd has at least @child CPUs. */
2338 cpumask_or(sched_domain_span(sd),
2339 sched_domain_span(sd),
2340 sched_domain_span(child));
2344 set_domain_attribute(sd, attr);
2350 * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2351 * any two given CPUs at this (non-NUMA) topology level.
2353 static bool topology_span_sane(struct sched_domain_topology_level *tl,
2354 const struct cpumask *cpu_map, int cpu)
2358 /* NUMA levels are allowed to overlap */
2359 if (tl->flags & SDTL_OVERLAP)
2363 * Non-NUMA levels cannot partially overlap - they must be either
2364 * completely equal or completely disjoint. Otherwise we can end up
2365 * breaking the sched_group lists - i.e. a later get_group() pass
2366 * breaks the linking done for an earlier span.
2368 for_each_cpu_from(i, cpu_map) {
2370 * We should 'and' all those masks with 'cpu_map' to exactly
2371 * match the topology we're about to build, but that can only
2372 * remove CPUs, which only lessens our ability to detect
2375 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2376 cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2384 * Build sched domains for a given set of CPUs and attach the sched domains
2385 * to the individual CPUs
2388 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2390 enum s_alloc alloc_state = sa_none;
2391 struct sched_domain *sd;
2393 struct rq *rq = NULL;
2394 int i, ret = -ENOMEM;
2395 bool has_asym = false;
2396 bool has_cluster = false;
2398 if (WARN_ON(cpumask_empty(cpu_map)))
2401 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2402 if (alloc_state != sa_rootdomain)
2405 /* Set up domains for CPUs specified by the cpu_map: */
2406 for_each_cpu(i, cpu_map) {
2407 struct sched_domain_topology_level *tl;
2410 for_each_sd_topology(tl) {
2412 if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2415 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2417 has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2419 if (tl == sched_domain_topology)
2420 *per_cpu_ptr(d.sd, i) = sd;
2421 if (tl->flags & SDTL_OVERLAP)
2422 sd->flags |= SD_OVERLAP;
2423 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2428 /* Build the groups for the domains */
2429 for_each_cpu(i, cpu_map) {
2430 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2431 sd->span_weight = cpumask_weight(sched_domain_span(sd));
2432 if (sd->flags & SD_OVERLAP) {
2433 if (build_overlap_sched_groups(sd, i))
2436 if (build_sched_groups(sd, i))
2443 * Calculate an allowed NUMA imbalance such that LLCs do not get
2446 for_each_cpu(i, cpu_map) {
2447 unsigned int imb = 0;
2448 unsigned int imb_span = 1;
2450 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2451 struct sched_domain *child = sd->child;
2453 if (!(sd->flags & SD_SHARE_LLC) && child &&
2454 (child->flags & SD_SHARE_LLC)) {
2455 struct sched_domain __rcu *top_p;
2456 unsigned int nr_llcs;
2459 * For a single LLC per node, allow an
2460 * imbalance up to 12.5% of the node. This is
2461 * arbitrary cutoff based two factors -- SMT and
2462 * memory channels. For SMT-2, the intent is to
2463 * avoid premature sharing of HT resources but
2464 * SMT-4 or SMT-8 *may* benefit from a different
2465 * cutoff. For memory channels, this is a very
2466 * rough estimate of how many channels may be
2467 * active and is based on recent CPUs with
2470 * For multiple LLCs, allow an imbalance
2471 * until multiple tasks would share an LLC
2472 * on one node while LLCs on another node
2473 * remain idle. This assumes that there are
2474 * enough logical CPUs per LLC to avoid SMT
2475 * factors and that there is a correlation
2476 * between LLCs and memory channels.
2478 nr_llcs = sd->span_weight / child->span_weight;
2480 imb = sd->span_weight >> 3;
2484 sd->imb_numa_nr = imb;
2486 /* Set span based on the first NUMA domain. */
2488 while (top_p && !(top_p->flags & SD_NUMA)) {
2489 top_p = top_p->parent;
2491 imb_span = top_p ? top_p->span_weight : sd->span_weight;
2493 int factor = max(1U, (sd->span_weight / imb_span));
2495 sd->imb_numa_nr = imb * factor;
2500 /* Calculate CPU capacity for physical packages and nodes */
2501 for (i = nr_cpumask_bits-1; i >= 0; i--) {
2502 if (!cpumask_test_cpu(i, cpu_map))
2505 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2506 claim_allocations(i, sd);
2507 init_sched_groups_capacity(i, sd);
2511 /* Attach the domains */
2513 for_each_cpu(i, cpu_map) {
2515 sd = *per_cpu_ptr(d.sd, i);
2517 cpu_attach_domain(sd, d.rd, i);
2519 if (lowest_flag_domain(i, SD_CLUSTER))
2525 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2528 static_branch_inc_cpuslocked(&sched_cluster_active);
2530 if (rq && sched_debug_verbose)
2531 pr_info("root domain span: %*pbl\n", cpumask_pr_args(cpu_map));
2535 __free_domain_allocs(&d, alloc_state, cpu_map);
2540 /* Current sched domains: */
2541 static cpumask_var_t *doms_cur;
2543 /* Number of sched domains in 'doms_cur': */
2544 static int ndoms_cur;
2546 /* Attributes of custom domains in 'doms_cur' */
2547 static struct sched_domain_attr *dattr_cur;
2550 * Special case: If a kmalloc() of a doms_cur partition (array of
2551 * cpumask) fails, then fallback to a single sched domain,
2552 * as determined by the single cpumask fallback_doms.
2554 static cpumask_var_t fallback_doms;
2557 * arch_update_cpu_topology lets virtualized architectures update the
2558 * CPU core maps. It is supposed to return 1 if the topology changed
2559 * or 0 if it stayed the same.
2561 int __weak arch_update_cpu_topology(void)
2566 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2569 cpumask_var_t *doms;
2571 doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2574 for (i = 0; i < ndoms; i++) {
2575 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2576 free_sched_domains(doms, i);
2583 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2586 for (i = 0; i < ndoms; i++)
2587 free_cpumask_var(doms[i]);
2592 * Set up scheduler domains and groups. For now this just excludes isolated
2593 * CPUs, but could be used to exclude other special cases in the future.
2595 int __init sched_init_domains(const struct cpumask *cpu_map)
2599 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2600 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2601 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2603 arch_update_cpu_topology();
2604 asym_cpu_capacity_scan();
2606 doms_cur = alloc_sched_domains(ndoms_cur);
2608 doms_cur = &fallback_doms;
2609 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
2610 err = build_sched_domains(doms_cur[0], NULL);
2616 * Detach sched domains from a group of CPUs specified in cpu_map
2617 * These CPUs will now be attached to the NULL domain
2619 static void detach_destroy_domains(const struct cpumask *cpu_map)
2621 unsigned int cpu = cpumask_any(cpu_map);
2624 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2625 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2627 if (static_branch_unlikely(&sched_cluster_active))
2628 static_branch_dec_cpuslocked(&sched_cluster_active);
2631 for_each_cpu(i, cpu_map)
2632 cpu_attach_domain(NULL, &def_root_domain, i);
2636 /* handle null as "default" */
2637 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2638 struct sched_domain_attr *new, int idx_new)
2640 struct sched_domain_attr tmp;
2648 return !memcmp(cur ? (cur + idx_cur) : &tmp,
2649 new ? (new + idx_new) : &tmp,
2650 sizeof(struct sched_domain_attr));
2654 * Partition sched domains as specified by the 'ndoms_new'
2655 * cpumasks in the array doms_new[] of cpumasks. This compares
2656 * doms_new[] to the current sched domain partitioning, doms_cur[].
2657 * It destroys each deleted domain and builds each new domain.
2659 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2660 * The masks don't intersect (don't overlap.) We should setup one
2661 * sched domain for each mask. CPUs not in any of the cpumasks will
2662 * not be load balanced. If the same cpumask appears both in the
2663 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2666 * The passed in 'doms_new' should be allocated using
2667 * alloc_sched_domains. This routine takes ownership of it and will
2668 * free_sched_domains it when done with it. If the caller failed the
2669 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2670 * and partition_sched_domains() will fallback to the single partition
2671 * 'fallback_doms', it also forces the domains to be rebuilt.
2673 * If doms_new == NULL it will be replaced with cpu_online_mask.
2674 * ndoms_new == 0 is a special case for destroying existing domains,
2675 * and it will not create the default domain.
2677 * Call with hotplug lock and sched_domains_mutex held
2679 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2680 struct sched_domain_attr *dattr_new)
2682 bool __maybe_unused has_eas = false;
2686 lockdep_assert_held(&sched_domains_mutex);
2688 /* Let the architecture update CPU core mappings: */
2689 new_topology = arch_update_cpu_topology();
2690 /* Trigger rebuilding CPU capacity asymmetry data */
2692 asym_cpu_capacity_scan();
2695 WARN_ON_ONCE(dattr_new);
2697 doms_new = alloc_sched_domains(1);
2700 cpumask_and(doms_new[0], cpu_active_mask,
2701 housekeeping_cpumask(HK_TYPE_DOMAIN));
2707 /* Destroy deleted domains: */
2708 for (i = 0; i < ndoms_cur; i++) {
2709 for (j = 0; j < n && !new_topology; j++) {
2710 if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2711 dattrs_equal(dattr_cur, i, dattr_new, j)) {
2712 struct root_domain *rd;
2715 * This domain won't be destroyed and as such
2716 * its dl_bw->total_bw needs to be cleared. It
2717 * will be recomputed in function
2718 * update_tasks_root_domain().
2720 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2721 dl_clear_root_domain(rd);
2725 /* No match - a current sched domain not in new doms_new[] */
2726 detach_destroy_domains(doms_cur[i]);
2734 doms_new = &fallback_doms;
2735 cpumask_and(doms_new[0], cpu_active_mask,
2736 housekeeping_cpumask(HK_TYPE_DOMAIN));
2739 /* Build new domains: */
2740 for (i = 0; i < ndoms_new; i++) {
2741 for (j = 0; j < n && !new_topology; j++) {
2742 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2743 dattrs_equal(dattr_new, i, dattr_cur, j))
2746 /* No match - add a new doms_new */
2747 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2752 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2753 /* Build perf. domains: */
2754 for (i = 0; i < ndoms_new; i++) {
2755 for (j = 0; j < n && !sched_energy_update; j++) {
2756 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2757 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2762 /* No match - add perf. domains for a new rd */
2763 has_eas |= build_perf_domains(doms_new[i]);
2767 sched_energy_set(has_eas);
2770 /* Remember the new sched domains: */
2771 if (doms_cur != &fallback_doms)
2772 free_sched_domains(doms_cur, ndoms_cur);
2775 doms_cur = doms_new;
2776 dattr_cur = dattr_new;
2777 ndoms_cur = ndoms_new;
2779 update_sched_domain_debugfs();
2783 * Call with hotplug lock held
2785 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2786 struct sched_domain_attr *dattr_new)
2788 mutex_lock(&sched_domains_mutex);
2789 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2790 mutex_unlock(&sched_domains_mutex);