1 // SPDX-License-Identifier: GPL-2.0
3 * Scheduler topology setup/handling methods
7 DEFINE_MUTEX(sched_domains_mutex);
9 /* Protected by sched_domains_mutex: */
10 static cpumask_var_t sched_domains_tmpmask;
11 static cpumask_var_t sched_domains_tmpmask2;
13 #ifdef CONFIG_SCHED_DEBUG
15 static int __init sched_debug_setup(char *str)
17 sched_debug_enabled = true;
21 early_param("sched_debug", sched_debug_setup);
23 static inline bool sched_debug(void)
25 return sched_debug_enabled;
28 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
29 struct cpumask *groupmask)
31 struct sched_group *group = sd->groups;
33 cpumask_clear(groupmask);
35 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
37 if (!(sd->flags & SD_LOAD_BALANCE)) {
38 printk("does not load-balance\n");
40 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
44 printk(KERN_CONT "span=%*pbl level=%s\n",
45 cpumask_pr_args(sched_domain_span(sd)), sd->name);
47 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
48 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
50 if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
51 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
54 printk(KERN_DEBUG "%*s groups:", level + 1, "");
58 printk(KERN_ERR "ERROR: group is NULL\n");
62 if (!cpumask_weight(sched_group_span(group))) {
63 printk(KERN_CONT "\n");
64 printk(KERN_ERR "ERROR: empty group\n");
68 if (!(sd->flags & SD_OVERLAP) &&
69 cpumask_intersects(groupmask, sched_group_span(group))) {
70 printk(KERN_CONT "\n");
71 printk(KERN_ERR "ERROR: repeated CPUs\n");
75 cpumask_or(groupmask, groupmask, sched_group_span(group));
77 printk(KERN_CONT " %d:{ span=%*pbl",
79 cpumask_pr_args(sched_group_span(group)));
81 if ((sd->flags & SD_OVERLAP) &&
82 !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
83 printk(KERN_CONT " mask=%*pbl",
84 cpumask_pr_args(group_balance_mask(group)));
87 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
88 printk(KERN_CONT " cap=%lu", group->sgc->capacity);
90 if (group == sd->groups && sd->child &&
91 !cpumask_equal(sched_domain_span(sd->child),
92 sched_group_span(group))) {
93 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
96 printk(KERN_CONT " }");
100 if (group != sd->groups)
101 printk(KERN_CONT ",");
103 } while (group != sd->groups);
104 printk(KERN_CONT "\n");
106 if (!cpumask_equal(sched_domain_span(sd), groupmask))
107 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
110 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
111 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
115 static void sched_domain_debug(struct sched_domain *sd, int cpu)
119 if (!sched_debug_enabled)
123 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
127 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
130 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
138 #else /* !CONFIG_SCHED_DEBUG */
140 # define sched_debug_enabled 0
141 # define sched_domain_debug(sd, cpu) do { } while (0)
142 static inline bool sched_debug(void)
146 #endif /* CONFIG_SCHED_DEBUG */
148 static int sd_degenerate(struct sched_domain *sd)
150 if (cpumask_weight(sched_domain_span(sd)) == 1)
153 /* Following flags need at least 2 groups */
154 if (sd->flags & (SD_LOAD_BALANCE |
158 SD_SHARE_CPUCAPACITY |
159 SD_ASYM_CPUCAPACITY |
160 SD_SHARE_PKG_RESOURCES |
161 SD_SHARE_POWERDOMAIN)) {
162 if (sd->groups != sd->groups->next)
166 /* Following flags don't use groups */
167 if (sd->flags & (SD_WAKE_AFFINE))
174 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
176 unsigned long cflags = sd->flags, pflags = parent->flags;
178 if (sd_degenerate(parent))
181 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
184 /* Flags needing groups don't count if only 1 group in parent */
185 if (parent->groups == parent->groups->next) {
186 pflags &= ~(SD_LOAD_BALANCE |
190 SD_ASYM_CPUCAPACITY |
191 SD_SHARE_CPUCAPACITY |
192 SD_SHARE_PKG_RESOURCES |
194 SD_SHARE_POWERDOMAIN);
195 if (nr_node_ids == 1)
196 pflags &= ~SD_SERIALIZE;
198 if (~cflags & pflags)
204 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
205 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
206 unsigned int sysctl_sched_energy_aware = 1;
207 DEFINE_MUTEX(sched_energy_mutex);
208 bool sched_energy_update;
210 #ifdef CONFIG_PROC_SYSCTL
211 int sched_energy_aware_handler(struct ctl_table *table, int write,
212 void __user *buffer, size_t *lenp, loff_t *ppos)
216 if (write && !capable(CAP_SYS_ADMIN))
219 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
221 state = static_branch_unlikely(&sched_energy_present);
222 if (state != sysctl_sched_energy_aware) {
223 mutex_lock(&sched_energy_mutex);
224 sched_energy_update = 1;
225 rebuild_sched_domains();
226 sched_energy_update = 0;
227 mutex_unlock(&sched_energy_mutex);
235 static void free_pd(struct perf_domain *pd)
237 struct perf_domain *tmp;
246 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
249 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
257 static struct perf_domain *pd_init(int cpu)
259 struct em_perf_domain *obj = em_cpu_get(cpu);
260 struct perf_domain *pd;
264 pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
268 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
276 static void perf_domain_debug(const struct cpumask *cpu_map,
277 struct perf_domain *pd)
279 if (!sched_debug() || !pd)
282 printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
285 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_cstate=%d }",
286 cpumask_first(perf_domain_span(pd)),
287 cpumask_pr_args(perf_domain_span(pd)),
288 em_pd_nr_cap_states(pd->em_pd));
292 printk(KERN_CONT "\n");
295 static void destroy_perf_domain_rcu(struct rcu_head *rp)
297 struct perf_domain *pd;
299 pd = container_of(rp, struct perf_domain, rcu);
303 static void sched_energy_set(bool has_eas)
305 if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
307 pr_info("%s: stopping EAS\n", __func__);
308 static_branch_disable_cpuslocked(&sched_energy_present);
309 } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
311 pr_info("%s: starting EAS\n", __func__);
312 static_branch_enable_cpuslocked(&sched_energy_present);
317 * EAS can be used on a root domain if it meets all the following conditions:
318 * 1. an Energy Model (EM) is available;
319 * 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
320 * 3. the EM complexity is low enough to keep scheduling overheads low;
321 * 4. schedutil is driving the frequency of all CPUs of the rd;
323 * The complexity of the Energy Model is defined as:
325 * C = nr_pd * (nr_cpus + nr_cs)
327 * with parameters defined as:
328 * - nr_pd: the number of performance domains
329 * - nr_cpus: the number of CPUs
330 * - nr_cs: the sum of the number of capacity states of all performance
331 * domains (for example, on a system with 2 performance domains,
332 * with 10 capacity states each, nr_cs = 2 * 10 = 20).
334 * It is generally not a good idea to use such a model in the wake-up path on
335 * very complex platforms because of the associated scheduling overheads. The
336 * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
337 * with per-CPU DVFS and less than 8 capacity states each, for example.
339 #define EM_MAX_COMPLEXITY 2048
341 extern struct cpufreq_governor schedutil_gov;
342 static bool build_perf_domains(const struct cpumask *cpu_map)
344 int i, nr_pd = 0, nr_cs = 0, nr_cpus = cpumask_weight(cpu_map);
345 struct perf_domain *pd = NULL, *tmp;
346 int cpu = cpumask_first(cpu_map);
347 struct root_domain *rd = cpu_rq(cpu)->rd;
348 struct cpufreq_policy *policy;
349 struct cpufreq_governor *gov;
351 if (!sysctl_sched_energy_aware)
354 /* EAS is enabled for asymmetric CPU capacity topologies. */
355 if (!per_cpu(sd_asym_cpucapacity, cpu)) {
357 pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
358 cpumask_pr_args(cpu_map));
363 for_each_cpu(i, cpu_map) {
364 /* Skip already covered CPUs. */
368 /* Do not attempt EAS if schedutil is not being used. */
369 policy = cpufreq_cpu_get(i);
372 gov = policy->governor;
373 cpufreq_cpu_put(policy);
374 if (gov != &schedutil_gov) {
376 pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
377 cpumask_pr_args(cpu_map));
381 /* Create the new pd and add it to the local list. */
389 * Count performance domains and capacity states for the
393 nr_cs += em_pd_nr_cap_states(pd->em_pd);
396 /* Bail out if the Energy Model complexity is too high. */
397 if (nr_pd * (nr_cs + nr_cpus) > EM_MAX_COMPLEXITY) {
398 WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
399 cpumask_pr_args(cpu_map));
403 perf_domain_debug(cpu_map, pd);
405 /* Attach the new list of performance domains to the root domain. */
407 rcu_assign_pointer(rd->pd, pd);
409 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
416 rcu_assign_pointer(rd->pd, NULL);
418 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
423 static void free_pd(struct perf_domain *pd) { }
424 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
426 static void free_rootdomain(struct rcu_head *rcu)
428 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
430 cpupri_cleanup(&rd->cpupri);
431 cpudl_cleanup(&rd->cpudl);
432 free_cpumask_var(rd->dlo_mask);
433 free_cpumask_var(rd->rto_mask);
434 free_cpumask_var(rd->online);
435 free_cpumask_var(rd->span);
440 void rq_attach_root(struct rq *rq, struct root_domain *rd)
442 struct root_domain *old_rd = NULL;
445 raw_spin_lock_irqsave(&rq->lock, flags);
450 if (cpumask_test_cpu(rq->cpu, old_rd->online))
453 cpumask_clear_cpu(rq->cpu, old_rd->span);
456 * If we dont want to free the old_rd yet then
457 * set old_rd to NULL to skip the freeing later
460 if (!atomic_dec_and_test(&old_rd->refcount))
464 atomic_inc(&rd->refcount);
467 cpumask_set_cpu(rq->cpu, rd->span);
468 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
471 raw_spin_unlock_irqrestore(&rq->lock, flags);
474 call_rcu(&old_rd->rcu, free_rootdomain);
477 void sched_get_rd(struct root_domain *rd)
479 atomic_inc(&rd->refcount);
482 void sched_put_rd(struct root_domain *rd)
484 if (!atomic_dec_and_test(&rd->refcount))
487 call_rcu(&rd->rcu, free_rootdomain);
490 static int init_rootdomain(struct root_domain *rd)
492 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
494 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
496 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
498 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
501 #ifdef HAVE_RT_PUSH_IPI
503 raw_spin_lock_init(&rd->rto_lock);
504 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
507 init_dl_bw(&rd->dl_bw);
508 if (cpudl_init(&rd->cpudl) != 0)
511 if (cpupri_init(&rd->cpupri) != 0)
516 cpudl_cleanup(&rd->cpudl);
518 free_cpumask_var(rd->rto_mask);
520 free_cpumask_var(rd->dlo_mask);
522 free_cpumask_var(rd->online);
524 free_cpumask_var(rd->span);
530 * By default the system creates a single root-domain with all CPUs as
531 * members (mimicking the global state we have today).
533 struct root_domain def_root_domain;
535 void init_defrootdomain(void)
537 init_rootdomain(&def_root_domain);
539 atomic_set(&def_root_domain.refcount, 1);
542 static struct root_domain *alloc_rootdomain(void)
544 struct root_domain *rd;
546 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
550 if (init_rootdomain(rd) != 0) {
558 static void free_sched_groups(struct sched_group *sg, int free_sgc)
560 struct sched_group *tmp, *first;
569 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
572 if (atomic_dec_and_test(&sg->ref))
575 } while (sg != first);
578 static void destroy_sched_domain(struct sched_domain *sd)
581 * A normal sched domain may have multiple group references, an
582 * overlapping domain, having private groups, only one. Iterate,
583 * dropping group/capacity references, freeing where none remain.
585 free_sched_groups(sd->groups, 1);
587 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
592 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
594 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
597 struct sched_domain *parent = sd->parent;
598 destroy_sched_domain(sd);
603 static void destroy_sched_domains(struct sched_domain *sd)
606 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
610 * Keep a special pointer to the highest sched_domain that has
611 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
612 * allows us to avoid some pointer chasing select_idle_sibling().
614 * Also keep a unique ID per domain (we use the first CPU number in
615 * the cpumask of the domain), this allows us to quickly tell if
616 * two CPUs are in the same cache domain, see cpus_share_cache().
618 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
619 DEFINE_PER_CPU(int, sd_llc_size);
620 DEFINE_PER_CPU(int, sd_llc_id);
621 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
622 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
623 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
624 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
625 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
627 static void update_top_cache_domain(int cpu)
629 struct sched_domain_shared *sds = NULL;
630 struct sched_domain *sd;
634 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
636 id = cpumask_first(sched_domain_span(sd));
637 size = cpumask_weight(sched_domain_span(sd));
641 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
642 per_cpu(sd_llc_size, cpu) = size;
643 per_cpu(sd_llc_id, cpu) = id;
644 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
646 sd = lowest_flag_domain(cpu, SD_NUMA);
647 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
649 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
650 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
652 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY);
653 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
657 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
658 * hold the hotplug lock.
661 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
663 struct rq *rq = cpu_rq(cpu);
664 struct sched_domain *tmp;
666 /* Remove the sched domains which do not contribute to scheduling. */
667 for (tmp = sd; tmp; ) {
668 struct sched_domain *parent = tmp->parent;
672 if (sd_parent_degenerate(tmp, parent)) {
673 tmp->parent = parent->parent;
675 parent->parent->child = tmp;
677 * Transfer SD_PREFER_SIBLING down in case of a
678 * degenerate parent; the spans match for this
679 * so the property transfers.
681 if (parent->flags & SD_PREFER_SIBLING)
682 tmp->flags |= SD_PREFER_SIBLING;
683 destroy_sched_domain(parent);
688 if (sd && sd_degenerate(sd)) {
691 destroy_sched_domain(tmp);
696 sched_domain_debug(sd, cpu);
698 rq_attach_root(rq, rd);
700 rcu_assign_pointer(rq->sd, sd);
701 dirty_sched_domain_sysctl(cpu);
702 destroy_sched_domains(tmp);
704 update_top_cache_domain(cpu);
708 struct sched_domain * __percpu *sd;
709 struct root_domain *rd;
720 * Return the canonical balance CPU for this group, this is the first CPU
721 * of this group that's also in the balance mask.
723 * The balance mask are all those CPUs that could actually end up at this
724 * group. See build_balance_mask().
726 * Also see should_we_balance().
728 int group_balance_cpu(struct sched_group *sg)
730 return cpumask_first(group_balance_mask(sg));
735 * NUMA topology (first read the regular topology blurb below)
737 * Given a node-distance table, for example:
745 * which represents a 4 node ring topology like:
753 * We want to construct domains and groups to represent this. The way we go
754 * about doing this is to build the domains on 'hops'. For each NUMA level we
755 * construct the mask of all nodes reachable in @level hops.
757 * For the above NUMA topology that gives 3 levels:
759 * NUMA-2 0-3 0-3 0-3 0-3
760 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
762 * NUMA-1 0-1,3 0-2 1-3 0,2-3
763 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
768 * As can be seen; things don't nicely line up as with the regular topology.
769 * When we iterate a domain in child domain chunks some nodes can be
770 * represented multiple times -- hence the "overlap" naming for this part of
773 * In order to minimize this overlap, we only build enough groups to cover the
774 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
778 * - the first group of each domain is its child domain; this
779 * gets us the first 0-1,3
780 * - the only uncovered node is 2, who's child domain is 1-3.
782 * However, because of the overlap, computing a unique CPU for each group is
783 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
784 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
785 * end up at those groups (they would end up in group: 0-1,3).
787 * To correct this we have to introduce the group balance mask. This mask
788 * will contain those CPUs in the group that can reach this group given the
789 * (child) domain tree.
791 * With this we can once again compute balance_cpu and sched_group_capacity
794 * XXX include words on how balance_cpu is unique and therefore can be
795 * used for sched_group_capacity links.
798 * Another 'interesting' topology is:
806 * Which looks a little like:
814 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
817 * This leads to a few particularly weird cases where the sched_domain's are
818 * not of the same number for each CPU. Consider:
821 * groups: {0-2},{1-3} {1-3},{0-2}
823 * NUMA-1 0-2 0-3 0-3 1-3
831 * Build the balance mask; it contains only those CPUs that can arrive at this
832 * group and should be considered to continue balancing.
834 * We do this during the group creation pass, therefore the group information
835 * isn't complete yet, however since each group represents a (child) domain we
836 * can fully construct this using the sched_domain bits (which are already
840 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
842 const struct cpumask *sg_span = sched_group_span(sg);
843 struct sd_data *sdd = sd->private;
844 struct sched_domain *sibling;
849 for_each_cpu(i, sg_span) {
850 sibling = *per_cpu_ptr(sdd->sd, i);
853 * Can happen in the asymmetric case, where these siblings are
854 * unused. The mask will not be empty because those CPUs that
855 * do have the top domain _should_ span the domain.
860 /* If we would not end up here, we can't continue from here */
861 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
864 cpumask_set_cpu(i, mask);
867 /* We must not have empty masks here */
868 WARN_ON_ONCE(cpumask_empty(mask));
872 * XXX: This creates per-node group entries; since the load-balancer will
873 * immediately access remote memory to construct this group's load-balance
874 * statistics having the groups node local is of dubious benefit.
876 static struct sched_group *
877 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
879 struct sched_group *sg;
880 struct cpumask *sg_span;
882 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
883 GFP_KERNEL, cpu_to_node(cpu));
888 sg_span = sched_group_span(sg);
890 cpumask_copy(sg_span, sched_domain_span(sd->child));
892 cpumask_copy(sg_span, sched_domain_span(sd));
894 atomic_inc(&sg->ref);
898 static void init_overlap_sched_group(struct sched_domain *sd,
899 struct sched_group *sg)
901 struct cpumask *mask = sched_domains_tmpmask2;
902 struct sd_data *sdd = sd->private;
903 struct cpumask *sg_span;
906 build_balance_mask(sd, sg, mask);
907 cpu = cpumask_first_and(sched_group_span(sg), mask);
909 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
910 if (atomic_inc_return(&sg->sgc->ref) == 1)
911 cpumask_copy(group_balance_mask(sg), mask);
913 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
916 * Initialize sgc->capacity such that even if we mess up the
917 * domains and no possible iteration will get us here, we won't
920 sg_span = sched_group_span(sg);
921 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
922 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
923 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
927 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
929 struct sched_group *first = NULL, *last = NULL, *sg;
930 const struct cpumask *span = sched_domain_span(sd);
931 struct cpumask *covered = sched_domains_tmpmask;
932 struct sd_data *sdd = sd->private;
933 struct sched_domain *sibling;
936 cpumask_clear(covered);
938 for_each_cpu_wrap(i, span, cpu) {
939 struct cpumask *sg_span;
941 if (cpumask_test_cpu(i, covered))
944 sibling = *per_cpu_ptr(sdd->sd, i);
947 * Asymmetric node setups can result in situations where the
948 * domain tree is of unequal depth, make sure to skip domains
949 * that already cover the entire range.
951 * In that case build_sched_domains() will have terminated the
952 * iteration early and our sibling sd spans will be empty.
953 * Domains should always include the CPU they're built on, so
956 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
959 sg = build_group_from_child_sched_domain(sibling, cpu);
963 sg_span = sched_group_span(sg);
964 cpumask_or(covered, covered, sg_span);
966 init_overlap_sched_group(sd, sg);
980 free_sched_groups(first, 0);
987 * Package topology (also see the load-balance blurb in fair.c)
989 * The scheduler builds a tree structure to represent a number of important
990 * topology features. By default (default_topology[]) these include:
992 * - Simultaneous multithreading (SMT)
993 * - Multi-Core Cache (MC)
996 * Where the last one more or less denotes everything up to a NUMA node.
998 * The tree consists of 3 primary data structures:
1000 * sched_domain -> sched_group -> sched_group_capacity
1004 * The sched_domains are per-CPU and have a two way link (parent & child) and
1005 * denote the ever growing mask of CPUs belonging to that level of topology.
1007 * Each sched_domain has a circular (double) linked list of sched_group's, each
1008 * denoting the domains of the level below (or individual CPUs in case of the
1009 * first domain level). The sched_group linked by a sched_domain includes the
1010 * CPU of that sched_domain [*].
1012 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1014 * CPU 0 1 2 3 4 5 6 7
1018 * SMT [ ] [ ] [ ] [ ]
1022 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1023 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1024 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1026 * CPU 0 1 2 3 4 5 6 7
1028 * One way to think about it is: sched_domain moves you up and down among these
1029 * topology levels, while sched_group moves you sideways through it, at child
1030 * domain granularity.
1032 * sched_group_capacity ensures each unique sched_group has shared storage.
1034 * There are two related construction problems, both require a CPU that
1035 * uniquely identify each group (for a given domain):
1037 * - The first is the balance_cpu (see should_we_balance() and the
1038 * load-balance blub in fair.c); for each group we only want 1 CPU to
1039 * continue balancing at a higher domain.
1041 * - The second is the sched_group_capacity; we want all identical groups
1042 * to share a single sched_group_capacity.
1044 * Since these topologies are exclusive by construction. That is, its
1045 * impossible for an SMT thread to belong to multiple cores, and cores to
1046 * be part of multiple caches. There is a very clear and unique location
1047 * for each CPU in the hierarchy.
1049 * Therefore computing a unique CPU for each group is trivial (the iteration
1050 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1051 * group), we can simply pick the first CPU in each group.
1054 * [*] in other words, the first group of each domain is its child domain.
1057 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1059 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1060 struct sched_domain *child = sd->child;
1061 struct sched_group *sg;
1062 bool already_visited;
1065 cpu = cpumask_first(sched_domain_span(child));
1067 sg = *per_cpu_ptr(sdd->sg, cpu);
1068 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1070 /* Increase refcounts for claim_allocations: */
1071 already_visited = atomic_inc_return(&sg->ref) > 1;
1072 /* sgc visits should follow a similar trend as sg */
1073 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1075 /* If we have already visited that group, it's already initialized. */
1076 if (already_visited)
1080 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1081 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1083 cpumask_set_cpu(cpu, sched_group_span(sg));
1084 cpumask_set_cpu(cpu, group_balance_mask(sg));
1087 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1088 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1089 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1095 * build_sched_groups will build a circular linked list of the groups
1096 * covered by the given span, will set each group's ->cpumask correctly,
1097 * and will initialize their ->sgc.
1099 * Assumes the sched_domain tree is fully constructed
1102 build_sched_groups(struct sched_domain *sd, int cpu)
1104 struct sched_group *first = NULL, *last = NULL;
1105 struct sd_data *sdd = sd->private;
1106 const struct cpumask *span = sched_domain_span(sd);
1107 struct cpumask *covered;
1110 lockdep_assert_held(&sched_domains_mutex);
1111 covered = sched_domains_tmpmask;
1113 cpumask_clear(covered);
1115 for_each_cpu_wrap(i, span, cpu) {
1116 struct sched_group *sg;
1118 if (cpumask_test_cpu(i, covered))
1121 sg = get_group(i, sdd);
1123 cpumask_or(covered, covered, sched_group_span(sg));
1138 * Initialize sched groups cpu_capacity.
1140 * cpu_capacity indicates the capacity of sched group, which is used while
1141 * distributing the load between different sched groups in a sched domain.
1142 * Typically cpu_capacity for all the groups in a sched domain will be same
1143 * unless there are asymmetries in the topology. If there are asymmetries,
1144 * group having more cpu_capacity will pickup more load compared to the
1145 * group having less cpu_capacity.
1147 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1149 struct sched_group *sg = sd->groups;
1154 int cpu, max_cpu = -1;
1156 sg->group_weight = cpumask_weight(sched_group_span(sg));
1158 if (!(sd->flags & SD_ASYM_PACKING))
1161 for_each_cpu(cpu, sched_group_span(sg)) {
1164 else if (sched_asym_prefer(cpu, max_cpu))
1167 sg->asym_prefer_cpu = max_cpu;
1171 } while (sg != sd->groups);
1173 if (cpu != group_balance_cpu(sg))
1176 update_group_capacity(sd, cpu);
1180 * Initializers for schedule domains
1181 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1184 static int default_relax_domain_level = -1;
1185 int sched_domain_level_max;
1187 static int __init setup_relax_domain_level(char *str)
1189 if (kstrtoint(str, 0, &default_relax_domain_level))
1190 pr_warn("Unable to set relax_domain_level\n");
1194 __setup("relax_domain_level=", setup_relax_domain_level);
1196 static void set_domain_attribute(struct sched_domain *sd,
1197 struct sched_domain_attr *attr)
1201 if (!attr || attr->relax_domain_level < 0) {
1202 if (default_relax_domain_level < 0)
1205 request = default_relax_domain_level;
1207 request = attr->relax_domain_level;
1208 if (request < sd->level) {
1209 /* Turn off idle balance on this domain: */
1210 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1212 /* Turn on idle balance on this domain: */
1213 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1217 static void __sdt_free(const struct cpumask *cpu_map);
1218 static int __sdt_alloc(const struct cpumask *cpu_map);
1220 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1221 const struct cpumask *cpu_map)
1225 if (!atomic_read(&d->rd->refcount))
1226 free_rootdomain(&d->rd->rcu);
1232 __sdt_free(cpu_map);
1240 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1242 memset(d, 0, sizeof(*d));
1244 if (__sdt_alloc(cpu_map))
1245 return sa_sd_storage;
1246 d->sd = alloc_percpu(struct sched_domain *);
1248 return sa_sd_storage;
1249 d->rd = alloc_rootdomain();
1253 return sa_rootdomain;
1257 * NULL the sd_data elements we've used to build the sched_domain and
1258 * sched_group structure so that the subsequent __free_domain_allocs()
1259 * will not free the data we're using.
1261 static void claim_allocations(int cpu, struct sched_domain *sd)
1263 struct sd_data *sdd = sd->private;
1265 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1266 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1268 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1269 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1271 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1272 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1274 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1275 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1279 enum numa_topology_type sched_numa_topology_type;
1281 static int sched_domains_numa_levels;
1282 static int sched_domains_curr_level;
1284 int sched_max_numa_distance;
1285 static int *sched_domains_numa_distance;
1286 static struct cpumask ***sched_domains_numa_masks;
1290 * SD_flags allowed in topology descriptions.
1292 * These flags are purely descriptive of the topology and do not prescribe
1293 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1296 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1297 * SD_SHARE_PKG_RESOURCES - describes shared caches
1298 * SD_NUMA - describes NUMA topologies
1299 * SD_SHARE_POWERDOMAIN - describes shared power domain
1301 * Odd one out, which beside describing the topology has a quirk also
1302 * prescribes the desired behaviour that goes along with it:
1304 * SD_ASYM_PACKING - describes SMT quirks
1306 #define TOPOLOGY_SD_FLAGS \
1307 (SD_SHARE_CPUCAPACITY | \
1308 SD_SHARE_PKG_RESOURCES | \
1311 SD_SHARE_POWERDOMAIN)
1313 static struct sched_domain *
1314 sd_init(struct sched_domain_topology_level *tl,
1315 const struct cpumask *cpu_map,
1316 struct sched_domain *child, int dflags, int cpu)
1318 struct sd_data *sdd = &tl->data;
1319 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1320 int sd_id, sd_weight, sd_flags = 0;
1324 * Ugly hack to pass state to sd_numa_mask()...
1326 sched_domains_curr_level = tl->numa_level;
1329 sd_weight = cpumask_weight(tl->mask(cpu));
1332 sd_flags = (*tl->sd_flags)();
1333 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1334 "wrong sd_flags in topology description\n"))
1335 sd_flags &= ~TOPOLOGY_SD_FLAGS;
1337 /* Apply detected topology flags */
1340 *sd = (struct sched_domain){
1341 .min_interval = sd_weight,
1342 .max_interval = 2*sd_weight,
1344 .imbalance_pct = 125,
1346 .cache_nice_tries = 0,
1348 .flags = 1*SD_LOAD_BALANCE
1349 | 1*SD_BALANCE_NEWIDLE
1354 | 0*SD_SHARE_CPUCAPACITY
1355 | 0*SD_SHARE_PKG_RESOURCES
1357 | 1*SD_PREFER_SIBLING
1362 .last_balance = jiffies,
1363 .balance_interval = sd_weight,
1364 .max_newidle_lb_cost = 0,
1365 .next_decay_max_lb_cost = jiffies,
1367 #ifdef CONFIG_SCHED_DEBUG
1372 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1373 sd_id = cpumask_first(sched_domain_span(sd));
1376 * Convert topological properties into behaviour.
1379 if (sd->flags & SD_ASYM_CPUCAPACITY) {
1380 struct sched_domain *t = sd;
1383 * Don't attempt to spread across CPUs of different capacities.
1386 sd->child->flags &= ~SD_PREFER_SIBLING;
1388 for_each_lower_domain(t)
1389 t->flags |= SD_BALANCE_WAKE;
1392 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1393 sd->imbalance_pct = 110;
1395 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1396 sd->imbalance_pct = 117;
1397 sd->cache_nice_tries = 1;
1400 } else if (sd->flags & SD_NUMA) {
1401 sd->cache_nice_tries = 2;
1403 sd->flags &= ~SD_PREFER_SIBLING;
1404 sd->flags |= SD_SERIALIZE;
1405 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
1406 sd->flags &= ~(SD_BALANCE_EXEC |
1413 sd->cache_nice_tries = 1;
1417 * For all levels sharing cache; connect a sched_domain_shared
1420 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1421 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1422 atomic_inc(&sd->shared->ref);
1423 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1432 * Topology list, bottom-up.
1434 static struct sched_domain_topology_level default_topology[] = {
1435 #ifdef CONFIG_SCHED_SMT
1436 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1438 #ifdef CONFIG_SCHED_MC
1439 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1441 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1445 static struct sched_domain_topology_level *sched_domain_topology =
1448 #define for_each_sd_topology(tl) \
1449 for (tl = sched_domain_topology; tl->mask; tl++)
1451 void set_sched_topology(struct sched_domain_topology_level *tl)
1453 if (WARN_ON_ONCE(sched_smp_initialized))
1456 sched_domain_topology = tl;
1461 static const struct cpumask *sd_numa_mask(int cpu)
1463 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1466 static void sched_numa_warn(const char *str)
1468 static int done = false;
1476 printk(KERN_WARNING "ERROR: %s\n\n", str);
1478 for (i = 0; i < nr_node_ids; i++) {
1479 printk(KERN_WARNING " ");
1480 for (j = 0; j < nr_node_ids; j++)
1481 printk(KERN_CONT "%02d ", node_distance(i,j));
1482 printk(KERN_CONT "\n");
1484 printk(KERN_WARNING "\n");
1487 bool find_numa_distance(int distance)
1491 if (distance == node_distance(0, 0))
1494 for (i = 0; i < sched_domains_numa_levels; i++) {
1495 if (sched_domains_numa_distance[i] == distance)
1503 * A system can have three types of NUMA topology:
1504 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1505 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1506 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1508 * The difference between a glueless mesh topology and a backplane
1509 * topology lies in whether communication between not directly
1510 * connected nodes goes through intermediary nodes (where programs
1511 * could run), or through backplane controllers. This affects
1512 * placement of programs.
1514 * The type of topology can be discerned with the following tests:
1515 * - If the maximum distance between any nodes is 1 hop, the system
1516 * is directly connected.
1517 * - If for two nodes A and B, located N > 1 hops away from each other,
1518 * there is an intermediary node C, which is < N hops away from both
1519 * nodes A and B, the system is a glueless mesh.
1521 static void init_numa_topology_type(void)
1525 n = sched_max_numa_distance;
1527 if (sched_domains_numa_levels <= 2) {
1528 sched_numa_topology_type = NUMA_DIRECT;
1532 for_each_online_node(a) {
1533 for_each_online_node(b) {
1534 /* Find two nodes furthest removed from each other. */
1535 if (node_distance(a, b) < n)
1538 /* Is there an intermediary node between a and b? */
1539 for_each_online_node(c) {
1540 if (node_distance(a, c) < n &&
1541 node_distance(b, c) < n) {
1542 sched_numa_topology_type =
1548 sched_numa_topology_type = NUMA_BACKPLANE;
1554 void sched_init_numa(void)
1556 int next_distance, curr_distance = node_distance(0, 0);
1557 struct sched_domain_topology_level *tl;
1561 sched_domains_numa_distance = kzalloc(sizeof(int) * (nr_node_ids + 1), GFP_KERNEL);
1562 if (!sched_domains_numa_distance)
1565 /* Includes NUMA identity node at level 0. */
1566 sched_domains_numa_distance[level++] = curr_distance;
1567 sched_domains_numa_levels = level;
1570 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1571 * unique distances in the node_distance() table.
1573 * Assumes node_distance(0,j) includes all distances in
1574 * node_distance(i,j) in order to avoid cubic time.
1576 next_distance = curr_distance;
1577 for (i = 0; i < nr_node_ids; i++) {
1578 for (j = 0; j < nr_node_ids; j++) {
1579 for (k = 0; k < nr_node_ids; k++) {
1580 int distance = node_distance(i, k);
1582 if (distance > curr_distance &&
1583 (distance < next_distance ||
1584 next_distance == curr_distance))
1585 next_distance = distance;
1588 * While not a strong assumption it would be nice to know
1589 * about cases where if node A is connected to B, B is not
1590 * equally connected to A.
1592 if (sched_debug() && node_distance(k, i) != distance)
1593 sched_numa_warn("Node-distance not symmetric");
1595 if (sched_debug() && i && !find_numa_distance(distance))
1596 sched_numa_warn("Node-0 not representative");
1598 if (next_distance != curr_distance) {
1599 sched_domains_numa_distance[level++] = next_distance;
1600 sched_domains_numa_levels = level;
1601 curr_distance = next_distance;
1606 * In case of sched_debug() we verify the above assumption.
1613 * 'level' contains the number of unique distances
1615 * The sched_domains_numa_distance[] array includes the actual distance
1620 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1621 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1622 * the array will contain less then 'level' members. This could be
1623 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1624 * in other functions.
1626 * We reset it to 'level' at the end of this function.
1628 sched_domains_numa_levels = 0;
1630 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1631 if (!sched_domains_numa_masks)
1635 * Now for each level, construct a mask per node which contains all
1636 * CPUs of nodes that are that many hops away from us.
1638 for (i = 0; i < level; i++) {
1639 sched_domains_numa_masks[i] =
1640 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1641 if (!sched_domains_numa_masks[i])
1644 for (j = 0; j < nr_node_ids; j++) {
1645 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1649 sched_domains_numa_masks[i][j] = mask;
1652 if (node_distance(j, k) > sched_domains_numa_distance[i])
1655 cpumask_or(mask, mask, cpumask_of_node(k));
1660 /* Compute default topology size */
1661 for (i = 0; sched_domain_topology[i].mask; i++);
1663 tl = kzalloc((i + level + 1) *
1664 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1669 * Copy the default topology bits..
1671 for (i = 0; sched_domain_topology[i].mask; i++)
1672 tl[i] = sched_domain_topology[i];
1675 * Add the NUMA identity distance, aka single NODE.
1677 tl[i++] = (struct sched_domain_topology_level){
1678 .mask = sd_numa_mask,
1684 * .. and append 'j' levels of NUMA goodness.
1686 for (j = 1; j < level; i++, j++) {
1687 tl[i] = (struct sched_domain_topology_level){
1688 .mask = sd_numa_mask,
1689 .sd_flags = cpu_numa_flags,
1690 .flags = SDTL_OVERLAP,
1696 sched_domain_topology = tl;
1698 sched_domains_numa_levels = level;
1699 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1701 init_numa_topology_type();
1704 void sched_domains_numa_masks_set(unsigned int cpu)
1706 int node = cpu_to_node(cpu);
1709 for (i = 0; i < sched_domains_numa_levels; i++) {
1710 for (j = 0; j < nr_node_ids; j++) {
1711 if (node_distance(j, node) <= sched_domains_numa_distance[i])
1712 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1717 void sched_domains_numa_masks_clear(unsigned int cpu)
1721 for (i = 0; i < sched_domains_numa_levels; i++) {
1722 for (j = 0; j < nr_node_ids; j++)
1723 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1728 * sched_numa_find_closest() - given the NUMA topology, find the cpu
1729 * closest to @cpu from @cpumask.
1730 * cpumask: cpumask to find a cpu from
1731 * cpu: cpu to be close to
1733 * returns: cpu, or nr_cpu_ids when nothing found.
1735 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1737 int i, j = cpu_to_node(cpu);
1739 for (i = 0; i < sched_domains_numa_levels; i++) {
1740 cpu = cpumask_any_and(cpus, sched_domains_numa_masks[i][j]);
1741 if (cpu < nr_cpu_ids)
1747 #endif /* CONFIG_NUMA */
1749 static int __sdt_alloc(const struct cpumask *cpu_map)
1751 struct sched_domain_topology_level *tl;
1754 for_each_sd_topology(tl) {
1755 struct sd_data *sdd = &tl->data;
1757 sdd->sd = alloc_percpu(struct sched_domain *);
1761 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1765 sdd->sg = alloc_percpu(struct sched_group *);
1769 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1773 for_each_cpu(j, cpu_map) {
1774 struct sched_domain *sd;
1775 struct sched_domain_shared *sds;
1776 struct sched_group *sg;
1777 struct sched_group_capacity *sgc;
1779 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1780 GFP_KERNEL, cpu_to_node(j));
1784 *per_cpu_ptr(sdd->sd, j) = sd;
1786 sds = kzalloc_node(sizeof(struct sched_domain_shared),
1787 GFP_KERNEL, cpu_to_node(j));
1791 *per_cpu_ptr(sdd->sds, j) = sds;
1793 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1794 GFP_KERNEL, cpu_to_node(j));
1800 *per_cpu_ptr(sdd->sg, j) = sg;
1802 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1803 GFP_KERNEL, cpu_to_node(j));
1807 #ifdef CONFIG_SCHED_DEBUG
1811 *per_cpu_ptr(sdd->sgc, j) = sgc;
1818 static void __sdt_free(const struct cpumask *cpu_map)
1820 struct sched_domain_topology_level *tl;
1823 for_each_sd_topology(tl) {
1824 struct sd_data *sdd = &tl->data;
1826 for_each_cpu(j, cpu_map) {
1827 struct sched_domain *sd;
1830 sd = *per_cpu_ptr(sdd->sd, j);
1831 if (sd && (sd->flags & SD_OVERLAP))
1832 free_sched_groups(sd->groups, 0);
1833 kfree(*per_cpu_ptr(sdd->sd, j));
1837 kfree(*per_cpu_ptr(sdd->sds, j));
1839 kfree(*per_cpu_ptr(sdd->sg, j));
1841 kfree(*per_cpu_ptr(sdd->sgc, j));
1843 free_percpu(sdd->sd);
1845 free_percpu(sdd->sds);
1847 free_percpu(sdd->sg);
1849 free_percpu(sdd->sgc);
1854 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1855 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1856 struct sched_domain *child, int dflags, int cpu)
1858 struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu);
1861 sd->level = child->level + 1;
1862 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1865 if (!cpumask_subset(sched_domain_span(child),
1866 sched_domain_span(sd))) {
1867 pr_err("BUG: arch topology borken\n");
1868 #ifdef CONFIG_SCHED_DEBUG
1869 pr_err(" the %s domain not a subset of the %s domain\n",
1870 child->name, sd->name);
1872 /* Fixup, ensure @sd has at least @child CPUs. */
1873 cpumask_or(sched_domain_span(sd),
1874 sched_domain_span(sd),
1875 sched_domain_span(child));
1879 set_domain_attribute(sd, attr);
1885 * Find the sched_domain_topology_level where all CPU capacities are visible
1888 static struct sched_domain_topology_level
1889 *asym_cpu_capacity_level(const struct cpumask *cpu_map)
1891 int i, j, asym_level = 0;
1893 struct sched_domain_topology_level *tl, *asym_tl = NULL;
1896 /* Is there any asymmetry? */
1897 cap = arch_scale_cpu_capacity(cpumask_first(cpu_map));
1899 for_each_cpu(i, cpu_map) {
1900 if (arch_scale_cpu_capacity(i) != cap) {
1910 * Examine topology from all CPU's point of views to detect the lowest
1911 * sched_domain_topology_level where a highest capacity CPU is visible
1914 for_each_cpu(i, cpu_map) {
1915 unsigned long max_capacity = arch_scale_cpu_capacity(i);
1918 for_each_sd_topology(tl) {
1919 if (tl_id < asym_level)
1922 for_each_cpu_and(j, tl->mask(i), cpu_map) {
1923 unsigned long capacity;
1925 capacity = arch_scale_cpu_capacity(j);
1927 if (capacity <= max_capacity)
1930 max_capacity = capacity;
1944 * Build sched domains for a given set of CPUs and attach the sched domains
1945 * to the individual CPUs
1948 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1950 enum s_alloc alloc_state;
1951 struct sched_domain *sd;
1953 struct rq *rq = NULL;
1954 int i, ret = -ENOMEM;
1955 struct sched_domain_topology_level *tl_asym;
1956 bool has_asym = false;
1958 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1959 if (alloc_state != sa_rootdomain)
1962 tl_asym = asym_cpu_capacity_level(cpu_map);
1964 /* Set up domains for CPUs specified by the cpu_map: */
1965 for_each_cpu(i, cpu_map) {
1966 struct sched_domain_topology_level *tl;
1969 for_each_sd_topology(tl) {
1972 if (tl == tl_asym) {
1973 dflags |= SD_ASYM_CPUCAPACITY;
1977 sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i);
1979 if (tl == sched_domain_topology)
1980 *per_cpu_ptr(d.sd, i) = sd;
1981 if (tl->flags & SDTL_OVERLAP)
1982 sd->flags |= SD_OVERLAP;
1983 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
1988 /* Build the groups for the domains */
1989 for_each_cpu(i, cpu_map) {
1990 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1991 sd->span_weight = cpumask_weight(sched_domain_span(sd));
1992 if (sd->flags & SD_OVERLAP) {
1993 if (build_overlap_sched_groups(sd, i))
1996 if (build_sched_groups(sd, i))
2002 /* Calculate CPU capacity for physical packages and nodes */
2003 for (i = nr_cpumask_bits-1; i >= 0; i--) {
2004 if (!cpumask_test_cpu(i, cpu_map))
2007 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2008 claim_allocations(i, sd);
2009 init_sched_groups_capacity(i, sd);
2013 /* Attach the domains */
2015 for_each_cpu(i, cpu_map) {
2017 sd = *per_cpu_ptr(d.sd, i);
2019 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2020 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2021 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2023 cpu_attach_domain(sd, d.rd, i);
2028 static_branch_enable_cpuslocked(&sched_asym_cpucapacity);
2030 if (rq && sched_debug_enabled) {
2031 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2032 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2037 __free_domain_allocs(&d, alloc_state, cpu_map);
2042 /* Current sched domains: */
2043 static cpumask_var_t *doms_cur;
2045 /* Number of sched domains in 'doms_cur': */
2046 static int ndoms_cur;
2048 /* Attribues of custom domains in 'doms_cur' */
2049 static struct sched_domain_attr *dattr_cur;
2052 * Special case: If a kmalloc() of a doms_cur partition (array of
2053 * cpumask) fails, then fallback to a single sched domain,
2054 * as determined by the single cpumask fallback_doms.
2056 static cpumask_var_t fallback_doms;
2059 * arch_update_cpu_topology lets virtualized architectures update the
2060 * CPU core maps. It is supposed to return 1 if the topology changed
2061 * or 0 if it stayed the same.
2063 int __weak arch_update_cpu_topology(void)
2068 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2071 cpumask_var_t *doms;
2073 doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2076 for (i = 0; i < ndoms; i++) {
2077 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2078 free_sched_domains(doms, i);
2085 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2088 for (i = 0; i < ndoms; i++)
2089 free_cpumask_var(doms[i]);
2094 * Set up scheduler domains and groups. For now this just excludes isolated
2095 * CPUs, but could be used to exclude other special cases in the future.
2097 int sched_init_domains(const struct cpumask *cpu_map)
2101 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2102 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2103 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2105 arch_update_cpu_topology();
2107 doms_cur = alloc_sched_domains(ndoms_cur);
2109 doms_cur = &fallback_doms;
2110 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
2111 err = build_sched_domains(doms_cur[0], NULL);
2112 register_sched_domain_sysctl();
2118 * Detach sched domains from a group of CPUs specified in cpu_map
2119 * These CPUs will now be attached to the NULL domain
2121 static void detach_destroy_domains(const struct cpumask *cpu_map)
2126 for_each_cpu(i, cpu_map)
2127 cpu_attach_domain(NULL, &def_root_domain, i);
2131 /* handle null as "default" */
2132 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2133 struct sched_domain_attr *new, int idx_new)
2135 struct sched_domain_attr tmp;
2143 return !memcmp(cur ? (cur + idx_cur) : &tmp,
2144 new ? (new + idx_new) : &tmp,
2145 sizeof(struct sched_domain_attr));
2149 * Partition sched domains as specified by the 'ndoms_new'
2150 * cpumasks in the array doms_new[] of cpumasks. This compares
2151 * doms_new[] to the current sched domain partitioning, doms_cur[].
2152 * It destroys each deleted domain and builds each new domain.
2154 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2155 * The masks don't intersect (don't overlap.) We should setup one
2156 * sched domain for each mask. CPUs not in any of the cpumasks will
2157 * not be load balanced. If the same cpumask appears both in the
2158 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2161 * The passed in 'doms_new' should be allocated using
2162 * alloc_sched_domains. This routine takes ownership of it and will
2163 * free_sched_domains it when done with it. If the caller failed the
2164 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2165 * and partition_sched_domains() will fallback to the single partition
2166 * 'fallback_doms', it also forces the domains to be rebuilt.
2168 * If doms_new == NULL it will be replaced with cpu_online_mask.
2169 * ndoms_new == 0 is a special case for destroying existing domains,
2170 * and it will not create the default domain.
2172 * Call with hotplug lock and sched_domains_mutex held
2174 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2175 struct sched_domain_attr *dattr_new)
2177 bool __maybe_unused has_eas = false;
2181 lockdep_assert_held(&sched_domains_mutex);
2183 /* Always unregister in case we don't destroy any domains: */
2184 unregister_sched_domain_sysctl();
2186 /* Let the architecture update CPU core mappings: */
2187 new_topology = arch_update_cpu_topology();
2190 WARN_ON_ONCE(dattr_new);
2192 doms_new = alloc_sched_domains(1);
2195 cpumask_and(doms_new[0], cpu_active_mask,
2196 housekeeping_cpumask(HK_FLAG_DOMAIN));
2202 /* Destroy deleted domains: */
2203 for (i = 0; i < ndoms_cur; i++) {
2204 for (j = 0; j < n && !new_topology; j++) {
2205 if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2206 dattrs_equal(dattr_cur, i, dattr_new, j))
2209 /* No match - a current sched domain not in new doms_new[] */
2210 detach_destroy_domains(doms_cur[i]);
2218 doms_new = &fallback_doms;
2219 cpumask_and(doms_new[0], cpu_active_mask,
2220 housekeeping_cpumask(HK_FLAG_DOMAIN));
2223 /* Build new domains: */
2224 for (i = 0; i < ndoms_new; i++) {
2225 for (j = 0; j < n && !new_topology; j++) {
2226 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2227 dattrs_equal(dattr_new, i, dattr_cur, j))
2230 /* No match - add a new doms_new */
2231 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2236 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2237 /* Build perf. domains: */
2238 for (i = 0; i < ndoms_new; i++) {
2239 for (j = 0; j < n && !sched_energy_update; j++) {
2240 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2241 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2246 /* No match - add perf. domains for a new rd */
2247 has_eas |= build_perf_domains(doms_new[i]);
2251 sched_energy_set(has_eas);
2254 /* Remember the new sched domains: */
2255 if (doms_cur != &fallback_doms)
2256 free_sched_domains(doms_cur, ndoms_cur);
2259 doms_cur = doms_new;
2260 dattr_cur = dattr_new;
2261 ndoms_cur = ndoms_new;
2263 register_sched_domain_sysctl();
2267 * Call with hotplug lock held
2269 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2270 struct sched_domain_attr *dattr_new)
2272 mutex_lock(&sched_domains_mutex);
2273 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2274 mutex_unlock(&sched_domains_mutex);