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_verbose = true;
21 early_param("sched_verbose", sched_debug_setup);
23 static inline bool sched_debug(void)
25 return sched_debug_verbose;
28 #define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
29 const struct sd_flag_debug sd_flag_debug[] = {
30 #include <linux/sched/sd_flags.h>
34 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
35 struct cpumask *groupmask)
37 struct sched_group *group = sd->groups;
38 unsigned long flags = sd->flags;
41 cpumask_clear(groupmask);
43 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
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 for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
55 unsigned int flag = BIT(idx);
56 unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
58 if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
59 !(sd->child->flags & flag))
60 printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
61 sd_flag_debug[idx].name);
63 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
64 !(sd->parent->flags & flag))
65 printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
66 sd_flag_debug[idx].name);
69 printk(KERN_DEBUG "%*s groups:", level + 1, "");
73 printk(KERN_ERR "ERROR: group is NULL\n");
77 if (!cpumask_weight(sched_group_span(group))) {
78 printk(KERN_CONT "\n");
79 printk(KERN_ERR "ERROR: empty group\n");
83 if (!(sd->flags & SD_OVERLAP) &&
84 cpumask_intersects(groupmask, sched_group_span(group))) {
85 printk(KERN_CONT "\n");
86 printk(KERN_ERR "ERROR: repeated CPUs\n");
90 cpumask_or(groupmask, groupmask, sched_group_span(group));
92 printk(KERN_CONT " %d:{ span=%*pbl",
94 cpumask_pr_args(sched_group_span(group)));
96 if ((sd->flags & SD_OVERLAP) &&
97 !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
98 printk(KERN_CONT " mask=%*pbl",
99 cpumask_pr_args(group_balance_mask(group)));
102 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
103 printk(KERN_CONT " cap=%lu", group->sgc->capacity);
105 if (group == sd->groups && sd->child &&
106 !cpumask_equal(sched_domain_span(sd->child),
107 sched_group_span(group))) {
108 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
111 printk(KERN_CONT " }");
115 if (group != sd->groups)
116 printk(KERN_CONT ",");
118 } while (group != sd->groups);
119 printk(KERN_CONT "\n");
121 if (!cpumask_equal(sched_domain_span(sd), groupmask))
122 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
125 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
126 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
130 static void sched_domain_debug(struct sched_domain *sd, int cpu)
134 if (!sched_debug_verbose)
138 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
142 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
145 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
153 #else /* !CONFIG_SCHED_DEBUG */
155 # define sched_debug_verbose 0
156 # define sched_domain_debug(sd, cpu) do { } while (0)
157 static inline bool sched_debug(void)
161 #endif /* CONFIG_SCHED_DEBUG */
163 /* Generate a mask of SD flags with the SDF_NEEDS_GROUPS metaflag */
164 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
165 static const unsigned int SD_DEGENERATE_GROUPS_MASK =
166 #include <linux/sched/sd_flags.h>
170 static int sd_degenerate(struct sched_domain *sd)
172 if (cpumask_weight(sched_domain_span(sd)) == 1)
175 /* Following flags need at least 2 groups */
176 if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
177 (sd->groups != sd->groups->next))
180 /* Following flags don't use groups */
181 if (sd->flags & (SD_WAKE_AFFINE))
188 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
190 unsigned long cflags = sd->flags, pflags = parent->flags;
192 if (sd_degenerate(parent))
195 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
198 /* Flags needing groups don't count if only 1 group in parent */
199 if (parent->groups == parent->groups->next)
200 pflags &= ~SD_DEGENERATE_GROUPS_MASK;
202 if (~cflags & pflags)
208 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
209 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
210 unsigned int sysctl_sched_energy_aware = 1;
211 DEFINE_MUTEX(sched_energy_mutex);
212 bool sched_energy_update;
214 void rebuild_sched_domains_energy(void)
216 mutex_lock(&sched_energy_mutex);
217 sched_energy_update = true;
218 rebuild_sched_domains();
219 sched_energy_update = false;
220 mutex_unlock(&sched_energy_mutex);
223 #ifdef CONFIG_PROC_SYSCTL
224 int sched_energy_aware_handler(struct ctl_table *table, int write,
225 void *buffer, size_t *lenp, loff_t *ppos)
229 if (write && !capable(CAP_SYS_ADMIN))
232 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
234 state = static_branch_unlikely(&sched_energy_present);
235 if (state != sysctl_sched_energy_aware)
236 rebuild_sched_domains_energy();
243 static void free_pd(struct perf_domain *pd)
245 struct perf_domain *tmp;
254 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
257 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
265 static struct perf_domain *pd_init(int cpu)
267 struct em_perf_domain *obj = em_cpu_get(cpu);
268 struct perf_domain *pd;
272 pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
276 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
284 static void perf_domain_debug(const struct cpumask *cpu_map,
285 struct perf_domain *pd)
287 if (!sched_debug() || !pd)
290 printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
293 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
294 cpumask_first(perf_domain_span(pd)),
295 cpumask_pr_args(perf_domain_span(pd)),
296 em_pd_nr_perf_states(pd->em_pd));
300 printk(KERN_CONT "\n");
303 static void destroy_perf_domain_rcu(struct rcu_head *rp)
305 struct perf_domain *pd;
307 pd = container_of(rp, struct perf_domain, rcu);
311 static void sched_energy_set(bool has_eas)
313 if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
315 pr_info("%s: stopping EAS\n", __func__);
316 static_branch_disable_cpuslocked(&sched_energy_present);
317 } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
319 pr_info("%s: starting EAS\n", __func__);
320 static_branch_enable_cpuslocked(&sched_energy_present);
325 * EAS can be used on a root domain if it meets all the following conditions:
326 * 1. an Energy Model (EM) is available;
327 * 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
328 * 3. no SMT is detected.
329 * 4. the EM complexity is low enough to keep scheduling overheads low;
330 * 5. schedutil is driving the frequency of all CPUs of the rd;
331 * 6. frequency invariance support is present;
333 * The complexity of the Energy Model is defined as:
335 * C = nr_pd * (nr_cpus + nr_ps)
337 * with parameters defined as:
338 * - nr_pd: the number of performance domains
339 * - nr_cpus: the number of CPUs
340 * - nr_ps: the sum of the number of performance states of all performance
341 * domains (for example, on a system with 2 performance domains,
342 * with 10 performance states each, nr_ps = 2 * 10 = 20).
344 * It is generally not a good idea to use such a model in the wake-up path on
345 * very complex platforms because of the associated scheduling overheads. The
346 * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
347 * with per-CPU DVFS and less than 8 performance states each, for example.
349 #define EM_MAX_COMPLEXITY 2048
351 extern struct cpufreq_governor schedutil_gov;
352 static bool build_perf_domains(const struct cpumask *cpu_map)
354 int i, nr_pd = 0, nr_ps = 0, nr_cpus = cpumask_weight(cpu_map);
355 struct perf_domain *pd = NULL, *tmp;
356 int cpu = cpumask_first(cpu_map);
357 struct root_domain *rd = cpu_rq(cpu)->rd;
358 struct cpufreq_policy *policy;
359 struct cpufreq_governor *gov;
361 if (!sysctl_sched_energy_aware)
364 /* EAS is enabled for asymmetric CPU capacity topologies. */
365 if (!per_cpu(sd_asym_cpucapacity, cpu)) {
367 pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
368 cpumask_pr_args(cpu_map));
373 /* EAS definitely does *not* handle SMT */
374 if (sched_smt_active()) {
375 pr_warn("rd %*pbl: Disabling EAS, SMT is not supported\n",
376 cpumask_pr_args(cpu_map));
380 if (!arch_scale_freq_invariant()) {
382 pr_warn("rd %*pbl: Disabling EAS: frequency-invariant load tracking not yet supported",
383 cpumask_pr_args(cpu_map));
388 for_each_cpu(i, cpu_map) {
389 /* Skip already covered CPUs. */
393 /* Do not attempt EAS if schedutil is not being used. */
394 policy = cpufreq_cpu_get(i);
397 gov = policy->governor;
398 cpufreq_cpu_put(policy);
399 if (gov != &schedutil_gov) {
401 pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
402 cpumask_pr_args(cpu_map));
406 /* Create the new pd and add it to the local list. */
414 * Count performance domains and performance states for the
418 nr_ps += em_pd_nr_perf_states(pd->em_pd);
421 /* Bail out if the Energy Model complexity is too high. */
422 if (nr_pd * (nr_ps + nr_cpus) > EM_MAX_COMPLEXITY) {
423 WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
424 cpumask_pr_args(cpu_map));
428 perf_domain_debug(cpu_map, pd);
430 /* Attach the new list of performance domains to the root domain. */
432 rcu_assign_pointer(rd->pd, pd);
434 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
441 rcu_assign_pointer(rd->pd, NULL);
443 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
448 static void free_pd(struct perf_domain *pd) { }
449 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
451 static void free_rootdomain(struct rcu_head *rcu)
453 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
455 cpupri_cleanup(&rd->cpupri);
456 cpudl_cleanup(&rd->cpudl);
457 free_cpumask_var(rd->dlo_mask);
458 free_cpumask_var(rd->rto_mask);
459 free_cpumask_var(rd->online);
460 free_cpumask_var(rd->span);
465 void rq_attach_root(struct rq *rq, struct root_domain *rd)
467 struct root_domain *old_rd = NULL;
470 raw_spin_rq_lock_irqsave(rq, flags);
475 if (cpumask_test_cpu(rq->cpu, old_rd->online))
478 cpumask_clear_cpu(rq->cpu, old_rd->span);
481 * If we dont want to free the old_rd yet then
482 * set old_rd to NULL to skip the freeing later
485 if (!atomic_dec_and_test(&old_rd->refcount))
489 atomic_inc(&rd->refcount);
492 cpumask_set_cpu(rq->cpu, rd->span);
493 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
496 raw_spin_rq_unlock_irqrestore(rq, flags);
499 call_rcu(&old_rd->rcu, free_rootdomain);
502 void sched_get_rd(struct root_domain *rd)
504 atomic_inc(&rd->refcount);
507 void sched_put_rd(struct root_domain *rd)
509 if (!atomic_dec_and_test(&rd->refcount))
512 call_rcu(&rd->rcu, free_rootdomain);
515 static int init_rootdomain(struct root_domain *rd)
517 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
519 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
521 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
523 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
526 #ifdef HAVE_RT_PUSH_IPI
528 raw_spin_lock_init(&rd->rto_lock);
529 rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
533 init_dl_bw(&rd->dl_bw);
534 if (cpudl_init(&rd->cpudl) != 0)
537 if (cpupri_init(&rd->cpupri) != 0)
542 cpudl_cleanup(&rd->cpudl);
544 free_cpumask_var(rd->rto_mask);
546 free_cpumask_var(rd->dlo_mask);
548 free_cpumask_var(rd->online);
550 free_cpumask_var(rd->span);
556 * By default the system creates a single root-domain with all CPUs as
557 * members (mimicking the global state we have today).
559 struct root_domain def_root_domain;
561 void init_defrootdomain(void)
563 init_rootdomain(&def_root_domain);
565 atomic_set(&def_root_domain.refcount, 1);
568 static struct root_domain *alloc_rootdomain(void)
570 struct root_domain *rd;
572 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
576 if (init_rootdomain(rd) != 0) {
584 static void free_sched_groups(struct sched_group *sg, int free_sgc)
586 struct sched_group *tmp, *first;
595 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
598 if (atomic_dec_and_test(&sg->ref))
601 } while (sg != first);
604 static void destroy_sched_domain(struct sched_domain *sd)
607 * A normal sched domain may have multiple group references, an
608 * overlapping domain, having private groups, only one. Iterate,
609 * dropping group/capacity references, freeing where none remain.
611 free_sched_groups(sd->groups, 1);
613 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
618 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
620 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
623 struct sched_domain *parent = sd->parent;
624 destroy_sched_domain(sd);
629 static void destroy_sched_domains(struct sched_domain *sd)
632 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
636 * Keep a special pointer to the highest sched_domain that has
637 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
638 * allows us to avoid some pointer chasing select_idle_sibling().
640 * Also keep a unique ID per domain (we use the first CPU number in
641 * the cpumask of the domain), this allows us to quickly tell if
642 * two CPUs are in the same cache domain, see cpus_share_cache().
644 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
645 DEFINE_PER_CPU(int, sd_llc_size);
646 DEFINE_PER_CPU(int, sd_llc_id);
647 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
648 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
649 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
650 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
651 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
653 static void update_top_cache_domain(int cpu)
655 struct sched_domain_shared *sds = NULL;
656 struct sched_domain *sd;
660 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
662 id = cpumask_first(sched_domain_span(sd));
663 size = cpumask_weight(sched_domain_span(sd));
667 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
668 per_cpu(sd_llc_size, cpu) = size;
669 per_cpu(sd_llc_id, cpu) = id;
670 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
672 sd = lowest_flag_domain(cpu, SD_NUMA);
673 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
675 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
676 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
678 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
679 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
683 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
684 * hold the hotplug lock.
687 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
689 struct rq *rq = cpu_rq(cpu);
690 struct sched_domain *tmp;
692 /* Remove the sched domains which do not contribute to scheduling. */
693 for (tmp = sd; tmp; ) {
694 struct sched_domain *parent = tmp->parent;
698 if (sd_parent_degenerate(tmp, parent)) {
699 tmp->parent = parent->parent;
701 parent->parent->child = tmp;
703 * Transfer SD_PREFER_SIBLING down in case of a
704 * degenerate parent; the spans match for this
705 * so the property transfers.
707 if (parent->flags & SD_PREFER_SIBLING)
708 tmp->flags |= SD_PREFER_SIBLING;
709 destroy_sched_domain(parent);
714 if (sd && sd_degenerate(sd)) {
717 destroy_sched_domain(tmp);
719 struct sched_group *sg = sd->groups;
722 * sched groups hold the flags of the child sched
723 * domain for convenience. Clear such flags since
724 * the child is being destroyed.
728 } while (sg != sd->groups);
734 sched_domain_debug(sd, cpu);
736 rq_attach_root(rq, rd);
738 rcu_assign_pointer(rq->sd, sd);
739 dirty_sched_domain_sysctl(cpu);
740 destroy_sched_domains(tmp);
742 update_top_cache_domain(cpu);
746 struct sched_domain * __percpu *sd;
747 struct root_domain *rd;
758 * Return the canonical balance CPU for this group, this is the first CPU
759 * of this group that's also in the balance mask.
761 * The balance mask are all those CPUs that could actually end up at this
762 * group. See build_balance_mask().
764 * Also see should_we_balance().
766 int group_balance_cpu(struct sched_group *sg)
768 return cpumask_first(group_balance_mask(sg));
773 * NUMA topology (first read the regular topology blurb below)
775 * Given a node-distance table, for example:
783 * which represents a 4 node ring topology like:
791 * We want to construct domains and groups to represent this. The way we go
792 * about doing this is to build the domains on 'hops'. For each NUMA level we
793 * construct the mask of all nodes reachable in @level hops.
795 * For the above NUMA topology that gives 3 levels:
797 * NUMA-2 0-3 0-3 0-3 0-3
798 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
800 * NUMA-1 0-1,3 0-2 1-3 0,2-3
801 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
806 * As can be seen; things don't nicely line up as with the regular topology.
807 * When we iterate a domain in child domain chunks some nodes can be
808 * represented multiple times -- hence the "overlap" naming for this part of
811 * In order to minimize this overlap, we only build enough groups to cover the
812 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
816 * - the first group of each domain is its child domain; this
817 * gets us the first 0-1,3
818 * - the only uncovered node is 2, who's child domain is 1-3.
820 * However, because of the overlap, computing a unique CPU for each group is
821 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
822 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
823 * end up at those groups (they would end up in group: 0-1,3).
825 * To correct this we have to introduce the group balance mask. This mask
826 * will contain those CPUs in the group that can reach this group given the
827 * (child) domain tree.
829 * With this we can once again compute balance_cpu and sched_group_capacity
832 * XXX include words on how balance_cpu is unique and therefore can be
833 * used for sched_group_capacity links.
836 * Another 'interesting' topology is:
844 * Which looks a little like:
852 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
855 * This leads to a few particularly weird cases where the sched_domain's are
856 * not of the same number for each CPU. Consider:
859 * groups: {0-2},{1-3} {1-3},{0-2}
861 * NUMA-1 0-2 0-3 0-3 1-3
869 * Build the balance mask; it contains only those CPUs that can arrive at this
870 * group and should be considered to continue balancing.
872 * We do this during the group creation pass, therefore the group information
873 * isn't complete yet, however since each group represents a (child) domain we
874 * can fully construct this using the sched_domain bits (which are already
878 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
880 const struct cpumask *sg_span = sched_group_span(sg);
881 struct sd_data *sdd = sd->private;
882 struct sched_domain *sibling;
887 for_each_cpu(i, sg_span) {
888 sibling = *per_cpu_ptr(sdd->sd, i);
891 * Can happen in the asymmetric case, where these siblings are
892 * unused. The mask will not be empty because those CPUs that
893 * do have the top domain _should_ span the domain.
898 /* If we would not end up here, we can't continue from here */
899 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
902 cpumask_set_cpu(i, mask);
905 /* We must not have empty masks here */
906 WARN_ON_ONCE(cpumask_empty(mask));
910 * XXX: This creates per-node group entries; since the load-balancer will
911 * immediately access remote memory to construct this group's load-balance
912 * statistics having the groups node local is of dubious benefit.
914 static struct sched_group *
915 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
917 struct sched_group *sg;
918 struct cpumask *sg_span;
920 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
921 GFP_KERNEL, cpu_to_node(cpu));
926 sg_span = sched_group_span(sg);
928 cpumask_copy(sg_span, sched_domain_span(sd->child));
929 sg->flags = sd->child->flags;
931 cpumask_copy(sg_span, sched_domain_span(sd));
934 atomic_inc(&sg->ref);
938 static void init_overlap_sched_group(struct sched_domain *sd,
939 struct sched_group *sg)
941 struct cpumask *mask = sched_domains_tmpmask2;
942 struct sd_data *sdd = sd->private;
943 struct cpumask *sg_span;
946 build_balance_mask(sd, sg, mask);
947 cpu = cpumask_first(mask);
949 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
950 if (atomic_inc_return(&sg->sgc->ref) == 1)
951 cpumask_copy(group_balance_mask(sg), mask);
953 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
956 * Initialize sgc->capacity such that even if we mess up the
957 * domains and no possible iteration will get us here, we won't
960 sg_span = sched_group_span(sg);
961 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
962 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
963 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
966 static struct sched_domain *
967 find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
970 * The proper descendant would be the one whose child won't span out
973 while (sibling->child &&
974 !cpumask_subset(sched_domain_span(sibling->child),
975 sched_domain_span(sd)))
976 sibling = sibling->child;
979 * As we are referencing sgc across different topology level, we need
980 * to go down to skip those sched_domains which don't contribute to
981 * scheduling because they will be degenerated in cpu_attach_domain
983 while (sibling->child &&
984 cpumask_equal(sched_domain_span(sibling->child),
985 sched_domain_span(sibling)))
986 sibling = sibling->child;
992 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
994 struct sched_group *first = NULL, *last = NULL, *sg;
995 const struct cpumask *span = sched_domain_span(sd);
996 struct cpumask *covered = sched_domains_tmpmask;
997 struct sd_data *sdd = sd->private;
998 struct sched_domain *sibling;
1001 cpumask_clear(covered);
1003 for_each_cpu_wrap(i, span, cpu) {
1004 struct cpumask *sg_span;
1006 if (cpumask_test_cpu(i, covered))
1009 sibling = *per_cpu_ptr(sdd->sd, i);
1012 * Asymmetric node setups can result in situations where the
1013 * domain tree is of unequal depth, make sure to skip domains
1014 * that already cover the entire range.
1016 * In that case build_sched_domains() will have terminated the
1017 * iteration early and our sibling sd spans will be empty.
1018 * Domains should always include the CPU they're built on, so
1021 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1025 * Usually we build sched_group by sibling's child sched_domain
1026 * But for machines whose NUMA diameter are 3 or above, we move
1027 * to build sched_group by sibling's proper descendant's child
1028 * domain because sibling's child sched_domain will span out of
1029 * the sched_domain being built as below.
1031 * Smallest diameter=3 topology is:
1039 * 0 --- 1 --- 2 --- 3
1041 * NUMA-3 0-3 N/A N/A 0-3
1042 * groups: {0-2},{1-3} {1-3},{0-2}
1044 * NUMA-2 0-2 0-3 0-3 1-3
1045 * groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2}
1047 * NUMA-1 0-1 0-2 1-3 2-3
1048 * groups: {0},{1} {1},{2},{0} {2},{3},{1} {3},{2}
1052 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1053 * group span isn't a subset of the domain span.
1055 if (sibling->child &&
1056 !cpumask_subset(sched_domain_span(sibling->child), span))
1057 sibling = find_descended_sibling(sd, sibling);
1059 sg = build_group_from_child_sched_domain(sibling, cpu);
1063 sg_span = sched_group_span(sg);
1064 cpumask_or(covered, covered, sg_span);
1066 init_overlap_sched_group(sibling, sg);
1080 free_sched_groups(first, 0);
1087 * Package topology (also see the load-balance blurb in fair.c)
1089 * The scheduler builds a tree structure to represent a number of important
1090 * topology features. By default (default_topology[]) these include:
1092 * - Simultaneous multithreading (SMT)
1093 * - Multi-Core Cache (MC)
1096 * Where the last one more or less denotes everything up to a NUMA node.
1098 * The tree consists of 3 primary data structures:
1100 * sched_domain -> sched_group -> sched_group_capacity
1104 * The sched_domains are per-CPU and have a two way link (parent & child) and
1105 * denote the ever growing mask of CPUs belonging to that level of topology.
1107 * Each sched_domain has a circular (double) linked list of sched_group's, each
1108 * denoting the domains of the level below (or individual CPUs in case of the
1109 * first domain level). The sched_group linked by a sched_domain includes the
1110 * CPU of that sched_domain [*].
1112 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1114 * CPU 0 1 2 3 4 5 6 7
1118 * SMT [ ] [ ] [ ] [ ]
1122 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1123 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1124 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1126 * CPU 0 1 2 3 4 5 6 7
1128 * One way to think about it is: sched_domain moves you up and down among these
1129 * topology levels, while sched_group moves you sideways through it, at child
1130 * domain granularity.
1132 * sched_group_capacity ensures each unique sched_group has shared storage.
1134 * There are two related construction problems, both require a CPU that
1135 * uniquely identify each group (for a given domain):
1137 * - The first is the balance_cpu (see should_we_balance() and the
1138 * load-balance blub in fair.c); for each group we only want 1 CPU to
1139 * continue balancing at a higher domain.
1141 * - The second is the sched_group_capacity; we want all identical groups
1142 * to share a single sched_group_capacity.
1144 * Since these topologies are exclusive by construction. That is, its
1145 * impossible for an SMT thread to belong to multiple cores, and cores to
1146 * be part of multiple caches. There is a very clear and unique location
1147 * for each CPU in the hierarchy.
1149 * Therefore computing a unique CPU for each group is trivial (the iteration
1150 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1151 * group), we can simply pick the first CPU in each group.
1154 * [*] in other words, the first group of each domain is its child domain.
1157 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1159 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1160 struct sched_domain *child = sd->child;
1161 struct sched_group *sg;
1162 bool already_visited;
1165 cpu = cpumask_first(sched_domain_span(child));
1167 sg = *per_cpu_ptr(sdd->sg, cpu);
1168 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1170 /* Increase refcounts for claim_allocations: */
1171 already_visited = atomic_inc_return(&sg->ref) > 1;
1172 /* sgc visits should follow a similar trend as sg */
1173 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1175 /* If we have already visited that group, it's already initialized. */
1176 if (already_visited)
1180 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1181 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1182 sg->flags = child->flags;
1184 cpumask_set_cpu(cpu, sched_group_span(sg));
1185 cpumask_set_cpu(cpu, group_balance_mask(sg));
1188 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1189 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1190 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1196 * build_sched_groups will build a circular linked list of the groups
1197 * covered by the given span, will set each group's ->cpumask correctly,
1198 * and will initialize their ->sgc.
1200 * Assumes the sched_domain tree is fully constructed
1203 build_sched_groups(struct sched_domain *sd, int cpu)
1205 struct sched_group *first = NULL, *last = NULL;
1206 struct sd_data *sdd = sd->private;
1207 const struct cpumask *span = sched_domain_span(sd);
1208 struct cpumask *covered;
1211 lockdep_assert_held(&sched_domains_mutex);
1212 covered = sched_domains_tmpmask;
1214 cpumask_clear(covered);
1216 for_each_cpu_wrap(i, span, cpu) {
1217 struct sched_group *sg;
1219 if (cpumask_test_cpu(i, covered))
1222 sg = get_group(i, sdd);
1224 cpumask_or(covered, covered, sched_group_span(sg));
1239 * Initialize sched groups cpu_capacity.
1241 * cpu_capacity indicates the capacity of sched group, which is used while
1242 * distributing the load between different sched groups in a sched domain.
1243 * Typically cpu_capacity for all the groups in a sched domain will be same
1244 * unless there are asymmetries in the topology. If there are asymmetries,
1245 * group having more cpu_capacity will pickup more load compared to the
1246 * group having less cpu_capacity.
1248 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1250 struct sched_group *sg = sd->groups;
1255 int cpu, max_cpu = -1;
1257 sg->group_weight = cpumask_weight(sched_group_span(sg));
1259 if (!(sd->flags & SD_ASYM_PACKING))
1262 for_each_cpu(cpu, sched_group_span(sg)) {
1265 else if (sched_asym_prefer(cpu, max_cpu))
1268 sg->asym_prefer_cpu = max_cpu;
1272 } while (sg != sd->groups);
1274 if (cpu != group_balance_cpu(sg))
1277 update_group_capacity(sd, cpu);
1281 * Asymmetric CPU capacity bits
1283 struct asym_cap_data {
1284 struct list_head link;
1285 unsigned long capacity;
1286 unsigned long cpus[];
1290 * Set of available CPUs grouped by their corresponding capacities
1291 * Each list entry contains a CPU mask reflecting CPUs that share the same
1293 * The lifespan of data is unlimited.
1295 static LIST_HEAD(asym_cap_list);
1297 #define cpu_capacity_span(asym_data) to_cpumask((asym_data)->cpus)
1300 * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1301 * Provides sd_flags reflecting the asymmetry scope.
1304 asym_cpu_capacity_classify(const struct cpumask *sd_span,
1305 const struct cpumask *cpu_map)
1307 struct asym_cap_data *entry;
1308 int count = 0, miss = 0;
1311 * Count how many unique CPU capacities this domain spans across
1312 * (compare sched_domain CPUs mask with ones representing available
1313 * CPUs capacities). Take into account CPUs that might be offline:
1316 list_for_each_entry(entry, &asym_cap_list, link) {
1317 if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1319 else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1323 WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1325 /* No asymmetry detected */
1328 /* Some of the available CPU capacity values have not been detected */
1330 return SD_ASYM_CPUCAPACITY;
1332 /* Full asymmetry */
1333 return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1337 static inline void asym_cpu_capacity_update_data(int cpu)
1339 unsigned long capacity = arch_scale_cpu_capacity(cpu);
1340 struct asym_cap_data *entry = NULL;
1342 list_for_each_entry(entry, &asym_cap_list, link) {
1343 if (capacity == entry->capacity)
1347 entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1348 if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1350 entry->capacity = capacity;
1351 list_add(&entry->link, &asym_cap_list);
1353 __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1357 * Build-up/update list of CPUs grouped by their capacities
1358 * An update requires explicit request to rebuild sched domains
1359 * with state indicating CPU topology changes.
1361 static void asym_cpu_capacity_scan(void)
1363 struct asym_cap_data *entry, *next;
1366 list_for_each_entry(entry, &asym_cap_list, link)
1367 cpumask_clear(cpu_capacity_span(entry));
1369 for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_FLAG_DOMAIN))
1370 asym_cpu_capacity_update_data(cpu);
1372 list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1373 if (cpumask_empty(cpu_capacity_span(entry))) {
1374 list_del(&entry->link);
1380 * Only one capacity value has been detected i.e. this system is symmetric.
1381 * No need to keep this data around.
1383 if (list_is_singular(&asym_cap_list)) {
1384 entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1385 list_del(&entry->link);
1391 * Initializers for schedule domains
1392 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1395 static int default_relax_domain_level = -1;
1396 int sched_domain_level_max;
1398 static int __init setup_relax_domain_level(char *str)
1400 if (kstrtoint(str, 0, &default_relax_domain_level))
1401 pr_warn("Unable to set relax_domain_level\n");
1405 __setup("relax_domain_level=", setup_relax_domain_level);
1407 static void set_domain_attribute(struct sched_domain *sd,
1408 struct sched_domain_attr *attr)
1412 if (!attr || attr->relax_domain_level < 0) {
1413 if (default_relax_domain_level < 0)
1415 request = default_relax_domain_level;
1417 request = attr->relax_domain_level;
1419 if (sd->level > request) {
1420 /* Turn off idle balance on this domain: */
1421 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1425 static void __sdt_free(const struct cpumask *cpu_map);
1426 static int __sdt_alloc(const struct cpumask *cpu_map);
1428 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1429 const struct cpumask *cpu_map)
1433 if (!atomic_read(&d->rd->refcount))
1434 free_rootdomain(&d->rd->rcu);
1440 __sdt_free(cpu_map);
1448 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1450 memset(d, 0, sizeof(*d));
1452 if (__sdt_alloc(cpu_map))
1453 return sa_sd_storage;
1454 d->sd = alloc_percpu(struct sched_domain *);
1456 return sa_sd_storage;
1457 d->rd = alloc_rootdomain();
1461 return sa_rootdomain;
1465 * NULL the sd_data elements we've used to build the sched_domain and
1466 * sched_group structure so that the subsequent __free_domain_allocs()
1467 * will not free the data we're using.
1469 static void claim_allocations(int cpu, struct sched_domain *sd)
1471 struct sd_data *sdd = sd->private;
1473 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1474 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1476 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1477 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1479 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1480 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1482 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1483 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1487 enum numa_topology_type sched_numa_topology_type;
1489 static int sched_domains_numa_levels;
1490 static int sched_domains_curr_level;
1492 int sched_max_numa_distance;
1493 static int *sched_domains_numa_distance;
1494 static struct cpumask ***sched_domains_numa_masks;
1496 static unsigned long __read_mostly *sched_numa_onlined_nodes;
1500 * SD_flags allowed in topology descriptions.
1502 * These flags are purely descriptive of the topology and do not prescribe
1503 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1506 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1507 * SD_SHARE_PKG_RESOURCES - describes shared caches
1508 * SD_NUMA - describes NUMA topologies
1510 * Odd one out, which beside describing the topology has a quirk also
1511 * prescribes the desired behaviour that goes along with it:
1513 * SD_ASYM_PACKING - describes SMT quirks
1515 #define TOPOLOGY_SD_FLAGS \
1516 (SD_SHARE_CPUCAPACITY | \
1517 SD_SHARE_PKG_RESOURCES | \
1521 static struct sched_domain *
1522 sd_init(struct sched_domain_topology_level *tl,
1523 const struct cpumask *cpu_map,
1524 struct sched_domain *child, int cpu)
1526 struct sd_data *sdd = &tl->data;
1527 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1528 int sd_id, sd_weight, sd_flags = 0;
1529 struct cpumask *sd_span;
1533 * Ugly hack to pass state to sd_numa_mask()...
1535 sched_domains_curr_level = tl->numa_level;
1538 sd_weight = cpumask_weight(tl->mask(cpu));
1541 sd_flags = (*tl->sd_flags)();
1542 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1543 "wrong sd_flags in topology description\n"))
1544 sd_flags &= TOPOLOGY_SD_FLAGS;
1546 *sd = (struct sched_domain){
1547 .min_interval = sd_weight,
1548 .max_interval = 2*sd_weight,
1550 .imbalance_pct = 117,
1552 .cache_nice_tries = 0,
1554 .flags = 1*SD_BALANCE_NEWIDLE
1559 | 0*SD_SHARE_CPUCAPACITY
1560 | 0*SD_SHARE_PKG_RESOURCES
1562 | 1*SD_PREFER_SIBLING
1567 .last_balance = jiffies,
1568 .balance_interval = sd_weight,
1569 .max_newidle_lb_cost = 0,
1570 .last_decay_max_lb_cost = jiffies,
1572 #ifdef CONFIG_SCHED_DEBUG
1577 sd_span = sched_domain_span(sd);
1578 cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1579 sd_id = cpumask_first(sd_span);
1581 sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1583 WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1584 (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1585 "CPU capacity asymmetry not supported on SMT\n");
1588 * Convert topological properties into behaviour.
1590 /* Don't attempt to spread across CPUs of different capacities. */
1591 if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1592 sd->child->flags &= ~SD_PREFER_SIBLING;
1594 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1595 sd->imbalance_pct = 110;
1597 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1598 sd->imbalance_pct = 117;
1599 sd->cache_nice_tries = 1;
1602 } else if (sd->flags & SD_NUMA) {
1603 sd->cache_nice_tries = 2;
1605 sd->flags &= ~SD_PREFER_SIBLING;
1606 sd->flags |= SD_SERIALIZE;
1607 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1608 sd->flags &= ~(SD_BALANCE_EXEC |
1615 sd->cache_nice_tries = 1;
1619 * For all levels sharing cache; connect a sched_domain_shared
1622 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1623 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1624 atomic_inc(&sd->shared->ref);
1625 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1634 * Topology list, bottom-up.
1636 static struct sched_domain_topology_level default_topology[] = {
1637 #ifdef CONFIG_SCHED_SMT
1638 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1641 #ifdef CONFIG_SCHED_CLUSTER
1642 { cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1645 #ifdef CONFIG_SCHED_MC
1646 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1648 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1652 static struct sched_domain_topology_level *sched_domain_topology =
1655 #define for_each_sd_topology(tl) \
1656 for (tl = sched_domain_topology; tl->mask; tl++)
1658 void set_sched_topology(struct sched_domain_topology_level *tl)
1660 if (WARN_ON_ONCE(sched_smp_initialized))
1663 sched_domain_topology = tl;
1668 static const struct cpumask *sd_numa_mask(int cpu)
1670 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1673 static void sched_numa_warn(const char *str)
1675 static int done = false;
1683 printk(KERN_WARNING "ERROR: %s\n\n", str);
1685 for (i = 0; i < nr_node_ids; i++) {
1686 printk(KERN_WARNING " ");
1687 for (j = 0; j < nr_node_ids; j++)
1688 printk(KERN_CONT "%02d ", node_distance(i,j));
1689 printk(KERN_CONT "\n");
1691 printk(KERN_WARNING "\n");
1694 bool find_numa_distance(int distance)
1698 if (distance == node_distance(0, 0))
1701 for (i = 0; i < sched_domains_numa_levels; i++) {
1702 if (sched_domains_numa_distance[i] == distance)
1710 * A system can have three types of NUMA topology:
1711 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1712 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1713 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1715 * The difference between a glueless mesh topology and a backplane
1716 * topology lies in whether communication between not directly
1717 * connected nodes goes through intermediary nodes (where programs
1718 * could run), or through backplane controllers. This affects
1719 * placement of programs.
1721 * The type of topology can be discerned with the following tests:
1722 * - If the maximum distance between any nodes is 1 hop, the system
1723 * is directly connected.
1724 * - If for two nodes A and B, located N > 1 hops away from each other,
1725 * there is an intermediary node C, which is < N hops away from both
1726 * nodes A and B, the system is a glueless mesh.
1728 static void init_numa_topology_type(void)
1732 n = sched_max_numa_distance;
1734 if (sched_domains_numa_levels <= 2) {
1735 sched_numa_topology_type = NUMA_DIRECT;
1739 for_each_online_node(a) {
1740 for_each_online_node(b) {
1741 /* Find two nodes furthest removed from each other. */
1742 if (node_distance(a, b) < n)
1745 /* Is there an intermediary node between a and b? */
1746 for_each_online_node(c) {
1747 if (node_distance(a, c) < n &&
1748 node_distance(b, c) < n) {
1749 sched_numa_topology_type =
1755 sched_numa_topology_type = NUMA_BACKPLANE;
1762 #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1764 void sched_init_numa(void)
1766 struct sched_domain_topology_level *tl;
1767 unsigned long *distance_map;
1772 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1773 * unique distances in the node_distance() table.
1775 distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1779 bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1780 for (i = 0; i < nr_node_ids; i++) {
1781 for (j = 0; j < nr_node_ids; j++) {
1782 int distance = node_distance(i, j);
1784 if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1785 sched_numa_warn("Invalid distance value range");
1789 bitmap_set(distance_map, distance, 1);
1793 * We can now figure out how many unique distance values there are and
1794 * allocate memory accordingly.
1796 nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1798 sched_domains_numa_distance = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1799 if (!sched_domains_numa_distance) {
1800 bitmap_free(distance_map);
1804 for (i = 0, j = 0; i < nr_levels; i++, j++) {
1805 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1806 sched_domains_numa_distance[i] = j;
1809 bitmap_free(distance_map);
1812 * 'nr_levels' contains the number of unique distances
1814 * The sched_domains_numa_distance[] array includes the actual distance
1819 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1820 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1821 * the array will contain less then 'nr_levels' members. This could be
1822 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1823 * in other functions.
1825 * We reset it to 'nr_levels' at the end of this function.
1827 sched_domains_numa_levels = 0;
1829 sched_domains_numa_masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1830 if (!sched_domains_numa_masks)
1834 * Now for each level, construct a mask per node which contains all
1835 * CPUs of nodes that are that many hops away from us.
1837 for (i = 0; i < nr_levels; i++) {
1838 sched_domains_numa_masks[i] =
1839 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1840 if (!sched_domains_numa_masks[i])
1843 for (j = 0; j < nr_node_ids; j++) {
1844 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1850 sched_domains_numa_masks[i][j] = mask;
1854 * Distance information can be unreliable for
1855 * offline nodes, defer building the node
1856 * masks to its bringup.
1857 * This relies on all unique distance values
1858 * still being visible at init time.
1860 if (!node_online(j))
1863 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1864 sched_numa_warn("Node-distance not symmetric");
1866 if (node_distance(j, k) > sched_domains_numa_distance[i])
1869 cpumask_or(mask, mask, cpumask_of_node(k));
1874 /* Compute default topology size */
1875 for (i = 0; sched_domain_topology[i].mask; i++);
1877 tl = kzalloc((i + nr_levels + 1) *
1878 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1883 * Copy the default topology bits..
1885 for (i = 0; sched_domain_topology[i].mask; i++)
1886 tl[i] = sched_domain_topology[i];
1889 * Add the NUMA identity distance, aka single NODE.
1891 tl[i++] = (struct sched_domain_topology_level){
1892 .mask = sd_numa_mask,
1898 * .. and append 'j' levels of NUMA goodness.
1900 for (j = 1; j < nr_levels; i++, j++) {
1901 tl[i] = (struct sched_domain_topology_level){
1902 .mask = sd_numa_mask,
1903 .sd_flags = cpu_numa_flags,
1904 .flags = SDTL_OVERLAP,
1910 sched_domain_topology = tl;
1912 sched_domains_numa_levels = nr_levels;
1913 sched_max_numa_distance = sched_domains_numa_distance[nr_levels - 1];
1915 init_numa_topology_type();
1917 sched_numa_onlined_nodes = bitmap_alloc(nr_node_ids, GFP_KERNEL);
1918 if (!sched_numa_onlined_nodes)
1921 bitmap_zero(sched_numa_onlined_nodes, nr_node_ids);
1922 for_each_online_node(i)
1923 bitmap_set(sched_numa_onlined_nodes, i, 1);
1926 static void __sched_domains_numa_masks_set(unsigned int node)
1931 * NUMA masks are not built for offline nodes in sched_init_numa().
1932 * Thus, when a CPU of a never-onlined-before node gets plugged in,
1933 * adding that new CPU to the right NUMA masks is not sufficient: the
1934 * masks of that CPU's node must also be updated.
1936 if (test_bit(node, sched_numa_onlined_nodes))
1939 bitmap_set(sched_numa_onlined_nodes, node, 1);
1941 for (i = 0; i < sched_domains_numa_levels; i++) {
1942 for (j = 0; j < nr_node_ids; j++) {
1943 if (!node_online(j) || node == j)
1946 if (node_distance(j, node) > sched_domains_numa_distance[i])
1949 /* Add remote nodes in our masks */
1950 cpumask_or(sched_domains_numa_masks[i][node],
1951 sched_domains_numa_masks[i][node],
1952 sched_domains_numa_masks[0][j]);
1957 * A new node has been brought up, potentially changing the topology
1960 * Note that this is racy vs any use of sched_numa_topology_type :/
1962 init_numa_topology_type();
1965 void sched_domains_numa_masks_set(unsigned int cpu)
1967 int node = cpu_to_node(cpu);
1970 __sched_domains_numa_masks_set(node);
1972 for (i = 0; i < sched_domains_numa_levels; i++) {
1973 for (j = 0; j < nr_node_ids; j++) {
1974 if (!node_online(j))
1977 /* Set ourselves in the remote node's masks */
1978 if (node_distance(j, node) <= sched_domains_numa_distance[i])
1979 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1984 void sched_domains_numa_masks_clear(unsigned int cpu)
1988 for (i = 0; i < sched_domains_numa_levels; i++) {
1989 for (j = 0; j < nr_node_ids; j++)
1990 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1995 * sched_numa_find_closest() - given the NUMA topology, find the cpu
1996 * closest to @cpu from @cpumask.
1997 * cpumask: cpumask to find a cpu from
1998 * cpu: cpu to be close to
2000 * returns: cpu, or nr_cpu_ids when nothing found.
2002 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2004 int i, j = cpu_to_node(cpu);
2006 for (i = 0; i < sched_domains_numa_levels; i++) {
2007 cpu = cpumask_any_and(cpus, sched_domains_numa_masks[i][j]);
2008 if (cpu < nr_cpu_ids)
2014 #endif /* CONFIG_NUMA */
2016 static int __sdt_alloc(const struct cpumask *cpu_map)
2018 struct sched_domain_topology_level *tl;
2021 for_each_sd_topology(tl) {
2022 struct sd_data *sdd = &tl->data;
2024 sdd->sd = alloc_percpu(struct sched_domain *);
2028 sdd->sds = alloc_percpu(struct sched_domain_shared *);
2032 sdd->sg = alloc_percpu(struct sched_group *);
2036 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2040 for_each_cpu(j, cpu_map) {
2041 struct sched_domain *sd;
2042 struct sched_domain_shared *sds;
2043 struct sched_group *sg;
2044 struct sched_group_capacity *sgc;
2046 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2047 GFP_KERNEL, cpu_to_node(j));
2051 *per_cpu_ptr(sdd->sd, j) = sd;
2053 sds = kzalloc_node(sizeof(struct sched_domain_shared),
2054 GFP_KERNEL, cpu_to_node(j));
2058 *per_cpu_ptr(sdd->sds, j) = sds;
2060 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2061 GFP_KERNEL, cpu_to_node(j));
2067 *per_cpu_ptr(sdd->sg, j) = sg;
2069 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2070 GFP_KERNEL, cpu_to_node(j));
2074 #ifdef CONFIG_SCHED_DEBUG
2078 *per_cpu_ptr(sdd->sgc, j) = sgc;
2085 static void __sdt_free(const struct cpumask *cpu_map)
2087 struct sched_domain_topology_level *tl;
2090 for_each_sd_topology(tl) {
2091 struct sd_data *sdd = &tl->data;
2093 for_each_cpu(j, cpu_map) {
2094 struct sched_domain *sd;
2097 sd = *per_cpu_ptr(sdd->sd, j);
2098 if (sd && (sd->flags & SD_OVERLAP))
2099 free_sched_groups(sd->groups, 0);
2100 kfree(*per_cpu_ptr(sdd->sd, j));
2104 kfree(*per_cpu_ptr(sdd->sds, j));
2106 kfree(*per_cpu_ptr(sdd->sg, j));
2108 kfree(*per_cpu_ptr(sdd->sgc, j));
2110 free_percpu(sdd->sd);
2112 free_percpu(sdd->sds);
2114 free_percpu(sdd->sg);
2116 free_percpu(sdd->sgc);
2121 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2122 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2123 struct sched_domain *child, int cpu)
2125 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2128 sd->level = child->level + 1;
2129 sched_domain_level_max = max(sched_domain_level_max, sd->level);
2132 if (!cpumask_subset(sched_domain_span(child),
2133 sched_domain_span(sd))) {
2134 pr_err("BUG: arch topology borken\n");
2135 #ifdef CONFIG_SCHED_DEBUG
2136 pr_err(" the %s domain not a subset of the %s domain\n",
2137 child->name, sd->name);
2139 /* Fixup, ensure @sd has at least @child CPUs. */
2140 cpumask_or(sched_domain_span(sd),
2141 sched_domain_span(sd),
2142 sched_domain_span(child));
2146 set_domain_attribute(sd, attr);
2152 * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2153 * any two given CPUs at this (non-NUMA) topology level.
2155 static bool topology_span_sane(struct sched_domain_topology_level *tl,
2156 const struct cpumask *cpu_map, int cpu)
2160 /* NUMA levels are allowed to overlap */
2161 if (tl->flags & SDTL_OVERLAP)
2165 * Non-NUMA levels cannot partially overlap - they must be either
2166 * completely equal or completely disjoint. Otherwise we can end up
2167 * breaking the sched_group lists - i.e. a later get_group() pass
2168 * breaks the linking done for an earlier span.
2170 for_each_cpu(i, cpu_map) {
2174 * We should 'and' all those masks with 'cpu_map' to exactly
2175 * match the topology we're about to build, but that can only
2176 * remove CPUs, which only lessens our ability to detect
2179 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2180 cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2188 * Build sched domains for a given set of CPUs and attach the sched domains
2189 * to the individual CPUs
2192 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2194 enum s_alloc alloc_state = sa_none;
2195 struct sched_domain *sd;
2197 struct rq *rq = NULL;
2198 int i, ret = -ENOMEM;
2199 bool has_asym = false;
2201 if (WARN_ON(cpumask_empty(cpu_map)))
2204 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2205 if (alloc_state != sa_rootdomain)
2208 /* Set up domains for CPUs specified by the cpu_map: */
2209 for_each_cpu(i, cpu_map) {
2210 struct sched_domain_topology_level *tl;
2213 for_each_sd_topology(tl) {
2215 if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2218 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2220 has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2222 if (tl == sched_domain_topology)
2223 *per_cpu_ptr(d.sd, i) = sd;
2224 if (tl->flags & SDTL_OVERLAP)
2225 sd->flags |= SD_OVERLAP;
2226 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2231 /* Build the groups for the domains */
2232 for_each_cpu(i, cpu_map) {
2233 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2234 sd->span_weight = cpumask_weight(sched_domain_span(sd));
2235 if (sd->flags & SD_OVERLAP) {
2236 if (build_overlap_sched_groups(sd, i))
2239 if (build_sched_groups(sd, i))
2245 /* Calculate CPU capacity for physical packages and nodes */
2246 for (i = nr_cpumask_bits-1; i >= 0; i--) {
2247 if (!cpumask_test_cpu(i, cpu_map))
2250 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2251 claim_allocations(i, sd);
2252 init_sched_groups_capacity(i, sd);
2256 /* Attach the domains */
2258 for_each_cpu(i, cpu_map) {
2260 sd = *per_cpu_ptr(d.sd, i);
2262 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2263 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2264 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2266 cpu_attach_domain(sd, d.rd, i);
2271 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2273 if (rq && sched_debug_verbose) {
2274 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2275 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2280 __free_domain_allocs(&d, alloc_state, cpu_map);
2285 /* Current sched domains: */
2286 static cpumask_var_t *doms_cur;
2288 /* Number of sched domains in 'doms_cur': */
2289 static int ndoms_cur;
2291 /* Attributes of custom domains in 'doms_cur' */
2292 static struct sched_domain_attr *dattr_cur;
2295 * Special case: If a kmalloc() of a doms_cur partition (array of
2296 * cpumask) fails, then fallback to a single sched domain,
2297 * as determined by the single cpumask fallback_doms.
2299 static cpumask_var_t fallback_doms;
2302 * arch_update_cpu_topology lets virtualized architectures update the
2303 * CPU core maps. It is supposed to return 1 if the topology changed
2304 * or 0 if it stayed the same.
2306 int __weak arch_update_cpu_topology(void)
2311 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2314 cpumask_var_t *doms;
2316 doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2319 for (i = 0; i < ndoms; i++) {
2320 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2321 free_sched_domains(doms, i);
2328 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2331 for (i = 0; i < ndoms; i++)
2332 free_cpumask_var(doms[i]);
2337 * Set up scheduler domains and groups. For now this just excludes isolated
2338 * CPUs, but could be used to exclude other special cases in the future.
2340 int sched_init_domains(const struct cpumask *cpu_map)
2344 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2345 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2346 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2348 arch_update_cpu_topology();
2349 asym_cpu_capacity_scan();
2351 doms_cur = alloc_sched_domains(ndoms_cur);
2353 doms_cur = &fallback_doms;
2354 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
2355 err = build_sched_domains(doms_cur[0], NULL);
2361 * Detach sched domains from a group of CPUs specified in cpu_map
2362 * These CPUs will now be attached to the NULL domain
2364 static void detach_destroy_domains(const struct cpumask *cpu_map)
2366 unsigned int cpu = cpumask_any(cpu_map);
2369 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2370 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2373 for_each_cpu(i, cpu_map)
2374 cpu_attach_domain(NULL, &def_root_domain, i);
2378 /* handle null as "default" */
2379 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2380 struct sched_domain_attr *new, int idx_new)
2382 struct sched_domain_attr tmp;
2390 return !memcmp(cur ? (cur + idx_cur) : &tmp,
2391 new ? (new + idx_new) : &tmp,
2392 sizeof(struct sched_domain_attr));
2396 * Partition sched domains as specified by the 'ndoms_new'
2397 * cpumasks in the array doms_new[] of cpumasks. This compares
2398 * doms_new[] to the current sched domain partitioning, doms_cur[].
2399 * It destroys each deleted domain and builds each new domain.
2401 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2402 * The masks don't intersect (don't overlap.) We should setup one
2403 * sched domain for each mask. CPUs not in any of the cpumasks will
2404 * not be load balanced. If the same cpumask appears both in the
2405 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2408 * The passed in 'doms_new' should be allocated using
2409 * alloc_sched_domains. This routine takes ownership of it and will
2410 * free_sched_domains it when done with it. If the caller failed the
2411 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2412 * and partition_sched_domains() will fallback to the single partition
2413 * 'fallback_doms', it also forces the domains to be rebuilt.
2415 * If doms_new == NULL it will be replaced with cpu_online_mask.
2416 * ndoms_new == 0 is a special case for destroying existing domains,
2417 * and it will not create the default domain.
2419 * Call with hotplug lock and sched_domains_mutex held
2421 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2422 struct sched_domain_attr *dattr_new)
2424 bool __maybe_unused has_eas = false;
2428 lockdep_assert_held(&sched_domains_mutex);
2430 /* Let the architecture update CPU core mappings: */
2431 new_topology = arch_update_cpu_topology();
2432 /* Trigger rebuilding CPU capacity asymmetry data */
2434 asym_cpu_capacity_scan();
2437 WARN_ON_ONCE(dattr_new);
2439 doms_new = alloc_sched_domains(1);
2442 cpumask_and(doms_new[0], cpu_active_mask,
2443 housekeeping_cpumask(HK_FLAG_DOMAIN));
2449 /* Destroy deleted domains: */
2450 for (i = 0; i < ndoms_cur; i++) {
2451 for (j = 0; j < n && !new_topology; j++) {
2452 if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2453 dattrs_equal(dattr_cur, i, dattr_new, j)) {
2454 struct root_domain *rd;
2457 * This domain won't be destroyed and as such
2458 * its dl_bw->total_bw needs to be cleared. It
2459 * will be recomputed in function
2460 * update_tasks_root_domain().
2462 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2463 dl_clear_root_domain(rd);
2467 /* No match - a current sched domain not in new doms_new[] */
2468 detach_destroy_domains(doms_cur[i]);
2476 doms_new = &fallback_doms;
2477 cpumask_and(doms_new[0], cpu_active_mask,
2478 housekeeping_cpumask(HK_FLAG_DOMAIN));
2481 /* Build new domains: */
2482 for (i = 0; i < ndoms_new; i++) {
2483 for (j = 0; j < n && !new_topology; j++) {
2484 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2485 dattrs_equal(dattr_new, i, dattr_cur, j))
2488 /* No match - add a new doms_new */
2489 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2494 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2495 /* Build perf. domains: */
2496 for (i = 0; i < ndoms_new; i++) {
2497 for (j = 0; j < n && !sched_energy_update; j++) {
2498 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2499 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2504 /* No match - add perf. domains for a new rd */
2505 has_eas |= build_perf_domains(doms_new[i]);
2509 sched_energy_set(has_eas);
2512 /* Remember the new sched domains: */
2513 if (doms_cur != &fallback_doms)
2514 free_sched_domains(doms_cur, ndoms_cur);
2517 doms_cur = doms_new;
2518 dattr_cur = dattr_new;
2519 ndoms_cur = ndoms_new;
2521 update_sched_domain_debugfs();
2525 * Call with hotplug lock held
2527 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2528 struct sched_domain_attr *dattr_new)
2530 mutex_lock(&sched_domains_mutex);
2531 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2532 mutex_unlock(&sched_domains_mutex);