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_MUTEX(sched_energy_mutex);
206 bool sched_energy_update;
208 static void free_pd(struct perf_domain *pd)
210 struct perf_domain *tmp;
219 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
222 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
230 static struct perf_domain *pd_init(int cpu)
232 struct em_perf_domain *obj = em_cpu_get(cpu);
233 struct perf_domain *pd;
237 pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
241 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
249 static void perf_domain_debug(const struct cpumask *cpu_map,
250 struct perf_domain *pd)
252 if (!sched_debug() || !pd)
255 printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
258 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_cstate=%d }",
259 cpumask_first(perf_domain_span(pd)),
260 cpumask_pr_args(perf_domain_span(pd)),
261 em_pd_nr_cap_states(pd->em_pd));
265 printk(KERN_CONT "\n");
268 static void destroy_perf_domain_rcu(struct rcu_head *rp)
270 struct perf_domain *pd;
272 pd = container_of(rp, struct perf_domain, rcu);
277 * EAS can be used on a root domain if it meets all the following conditions:
278 * 1. an Energy Model (EM) is available;
279 * 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
280 * 3. the EM complexity is low enough to keep scheduling overheads low;
281 * 4. schedutil is driving the frequency of all CPUs of the rd;
283 * The complexity of the Energy Model is defined as:
285 * C = nr_pd * (nr_cpus + nr_cs)
287 * with parameters defined as:
288 * - nr_pd: the number of performance domains
289 * - nr_cpus: the number of CPUs
290 * - nr_cs: the sum of the number of capacity states of all performance
291 * domains (for example, on a system with 2 performance domains,
292 * with 10 capacity states each, nr_cs = 2 * 10 = 20).
294 * It is generally not a good idea to use such a model in the wake-up path on
295 * very complex platforms because of the associated scheduling overheads. The
296 * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
297 * with per-CPU DVFS and less than 8 capacity states each, for example.
299 #define EM_MAX_COMPLEXITY 2048
301 extern struct cpufreq_governor schedutil_gov;
302 static void build_perf_domains(const struct cpumask *cpu_map)
304 int i, nr_pd = 0, nr_cs = 0, nr_cpus = cpumask_weight(cpu_map);
305 struct perf_domain *pd = NULL, *tmp;
306 int cpu = cpumask_first(cpu_map);
307 struct root_domain *rd = cpu_rq(cpu)->rd;
308 struct cpufreq_policy *policy;
309 struct cpufreq_governor *gov;
311 /* EAS is enabled for asymmetric CPU capacity topologies. */
312 if (!per_cpu(sd_asym_cpucapacity, cpu)) {
314 pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
315 cpumask_pr_args(cpu_map));
320 for_each_cpu(i, cpu_map) {
321 /* Skip already covered CPUs. */
325 /* Do not attempt EAS if schedutil is not being used. */
326 policy = cpufreq_cpu_get(i);
329 gov = policy->governor;
330 cpufreq_cpu_put(policy);
331 if (gov != &schedutil_gov) {
333 pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
334 cpumask_pr_args(cpu_map));
338 /* Create the new pd and add it to the local list. */
346 * Count performance domains and capacity states for the
350 nr_cs += em_pd_nr_cap_states(pd->em_pd);
353 /* Bail out if the Energy Model complexity is too high. */
354 if (nr_pd * (nr_cs + nr_cpus) > EM_MAX_COMPLEXITY) {
355 WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
356 cpumask_pr_args(cpu_map));
360 perf_domain_debug(cpu_map, pd);
362 /* Attach the new list of performance domains to the root domain. */
364 rcu_assign_pointer(rd->pd, pd);
366 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
373 rcu_assign_pointer(rd->pd, NULL);
375 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
378 static void free_pd(struct perf_domain *pd) { }
379 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
381 static void free_rootdomain(struct rcu_head *rcu)
383 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
385 cpupri_cleanup(&rd->cpupri);
386 cpudl_cleanup(&rd->cpudl);
387 free_cpumask_var(rd->dlo_mask);
388 free_cpumask_var(rd->rto_mask);
389 free_cpumask_var(rd->online);
390 free_cpumask_var(rd->span);
395 void rq_attach_root(struct rq *rq, struct root_domain *rd)
397 struct root_domain *old_rd = NULL;
400 raw_spin_lock_irqsave(&rq->lock, flags);
405 if (cpumask_test_cpu(rq->cpu, old_rd->online))
408 cpumask_clear_cpu(rq->cpu, old_rd->span);
411 * If we dont want to free the old_rd yet then
412 * set old_rd to NULL to skip the freeing later
415 if (!atomic_dec_and_test(&old_rd->refcount))
419 atomic_inc(&rd->refcount);
422 cpumask_set_cpu(rq->cpu, rd->span);
423 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
426 raw_spin_unlock_irqrestore(&rq->lock, flags);
429 call_rcu_sched(&old_rd->rcu, free_rootdomain);
432 void sched_get_rd(struct root_domain *rd)
434 atomic_inc(&rd->refcount);
437 void sched_put_rd(struct root_domain *rd)
439 if (!atomic_dec_and_test(&rd->refcount))
442 call_rcu_sched(&rd->rcu, free_rootdomain);
445 static int init_rootdomain(struct root_domain *rd)
447 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
449 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
451 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
453 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
456 #ifdef HAVE_RT_PUSH_IPI
458 raw_spin_lock_init(&rd->rto_lock);
459 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
462 init_dl_bw(&rd->dl_bw);
463 if (cpudl_init(&rd->cpudl) != 0)
466 if (cpupri_init(&rd->cpupri) != 0)
471 cpudl_cleanup(&rd->cpudl);
473 free_cpumask_var(rd->rto_mask);
475 free_cpumask_var(rd->dlo_mask);
477 free_cpumask_var(rd->online);
479 free_cpumask_var(rd->span);
485 * By default the system creates a single root-domain with all CPUs as
486 * members (mimicking the global state we have today).
488 struct root_domain def_root_domain;
490 void init_defrootdomain(void)
492 init_rootdomain(&def_root_domain);
494 atomic_set(&def_root_domain.refcount, 1);
497 static struct root_domain *alloc_rootdomain(void)
499 struct root_domain *rd;
501 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
505 if (init_rootdomain(rd) != 0) {
513 static void free_sched_groups(struct sched_group *sg, int free_sgc)
515 struct sched_group *tmp, *first;
524 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
527 if (atomic_dec_and_test(&sg->ref))
530 } while (sg != first);
533 static void destroy_sched_domain(struct sched_domain *sd)
536 * A normal sched domain may have multiple group references, an
537 * overlapping domain, having private groups, only one. Iterate,
538 * dropping group/capacity references, freeing where none remain.
540 free_sched_groups(sd->groups, 1);
542 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
547 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
549 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
552 struct sched_domain *parent = sd->parent;
553 destroy_sched_domain(sd);
558 static void destroy_sched_domains(struct sched_domain *sd)
561 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
565 * Keep a special pointer to the highest sched_domain that has
566 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
567 * allows us to avoid some pointer chasing select_idle_sibling().
569 * Also keep a unique ID per domain (we use the first CPU number in
570 * the cpumask of the domain), this allows us to quickly tell if
571 * two CPUs are in the same cache domain, see cpus_share_cache().
573 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
574 DEFINE_PER_CPU(int, sd_llc_size);
575 DEFINE_PER_CPU(int, sd_llc_id);
576 DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
577 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
578 DEFINE_PER_CPU(struct sched_domain *, sd_asym_packing);
579 DEFINE_PER_CPU(struct sched_domain *, sd_asym_cpucapacity);
580 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
582 static void update_top_cache_domain(int cpu)
584 struct sched_domain_shared *sds = NULL;
585 struct sched_domain *sd;
589 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
591 id = cpumask_first(sched_domain_span(sd));
592 size = cpumask_weight(sched_domain_span(sd));
596 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
597 per_cpu(sd_llc_size, cpu) = size;
598 per_cpu(sd_llc_id, cpu) = id;
599 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
601 sd = lowest_flag_domain(cpu, SD_NUMA);
602 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
604 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
605 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
607 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY);
608 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
612 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
613 * hold the hotplug lock.
616 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
618 struct rq *rq = cpu_rq(cpu);
619 struct sched_domain *tmp;
621 /* Remove the sched domains which do not contribute to scheduling. */
622 for (tmp = sd; tmp; ) {
623 struct sched_domain *parent = tmp->parent;
627 if (sd_parent_degenerate(tmp, parent)) {
628 tmp->parent = parent->parent;
630 parent->parent->child = tmp;
632 * Transfer SD_PREFER_SIBLING down in case of a
633 * degenerate parent; the spans match for this
634 * so the property transfers.
636 if (parent->flags & SD_PREFER_SIBLING)
637 tmp->flags |= SD_PREFER_SIBLING;
638 destroy_sched_domain(parent);
643 if (sd && sd_degenerate(sd)) {
646 destroy_sched_domain(tmp);
651 sched_domain_debug(sd, cpu);
653 rq_attach_root(rq, rd);
655 rcu_assign_pointer(rq->sd, sd);
656 dirty_sched_domain_sysctl(cpu);
657 destroy_sched_domains(tmp);
659 update_top_cache_domain(cpu);
663 struct sched_domain ** __percpu sd;
664 struct root_domain *rd;
675 * Return the canonical balance CPU for this group, this is the first CPU
676 * of this group that's also in the balance mask.
678 * The balance mask are all those CPUs that could actually end up at this
679 * group. See build_balance_mask().
681 * Also see should_we_balance().
683 int group_balance_cpu(struct sched_group *sg)
685 return cpumask_first(group_balance_mask(sg));
690 * NUMA topology (first read the regular topology blurb below)
692 * Given a node-distance table, for example:
700 * which represents a 4 node ring topology like:
708 * We want to construct domains and groups to represent this. The way we go
709 * about doing this is to build the domains on 'hops'. For each NUMA level we
710 * construct the mask of all nodes reachable in @level hops.
712 * For the above NUMA topology that gives 3 levels:
714 * NUMA-2 0-3 0-3 0-3 0-3
715 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
717 * NUMA-1 0-1,3 0-2 1-3 0,2-3
718 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
723 * As can be seen; things don't nicely line up as with the regular topology.
724 * When we iterate a domain in child domain chunks some nodes can be
725 * represented multiple times -- hence the "overlap" naming for this part of
728 * In order to minimize this overlap, we only build enough groups to cover the
729 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
733 * - the first group of each domain is its child domain; this
734 * gets us the first 0-1,3
735 * - the only uncovered node is 2, who's child domain is 1-3.
737 * However, because of the overlap, computing a unique CPU for each group is
738 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
739 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
740 * end up at those groups (they would end up in group: 0-1,3).
742 * To correct this we have to introduce the group balance mask. This mask
743 * will contain those CPUs in the group that can reach this group given the
744 * (child) domain tree.
746 * With this we can once again compute balance_cpu and sched_group_capacity
749 * XXX include words on how balance_cpu is unique and therefore can be
750 * used for sched_group_capacity links.
753 * Another 'interesting' topology is:
761 * Which looks a little like:
769 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
772 * This leads to a few particularly weird cases where the sched_domain's are
773 * not of the same number for each CPU. Consider:
776 * groups: {0-2},{1-3} {1-3},{0-2}
778 * NUMA-1 0-2 0-3 0-3 1-3
786 * Build the balance mask; it contains only those CPUs that can arrive at this
787 * group and should be considered to continue balancing.
789 * We do this during the group creation pass, therefore the group information
790 * isn't complete yet, however since each group represents a (child) domain we
791 * can fully construct this using the sched_domain bits (which are already
795 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
797 const struct cpumask *sg_span = sched_group_span(sg);
798 struct sd_data *sdd = sd->private;
799 struct sched_domain *sibling;
804 for_each_cpu(i, sg_span) {
805 sibling = *per_cpu_ptr(sdd->sd, i);
808 * Can happen in the asymmetric case, where these siblings are
809 * unused. The mask will not be empty because those CPUs that
810 * do have the top domain _should_ span the domain.
815 /* If we would not end up here, we can't continue from here */
816 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
819 cpumask_set_cpu(i, mask);
822 /* We must not have empty masks here */
823 WARN_ON_ONCE(cpumask_empty(mask));
827 * XXX: This creates per-node group entries; since the load-balancer will
828 * immediately access remote memory to construct this group's load-balance
829 * statistics having the groups node local is of dubious benefit.
831 static struct sched_group *
832 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
834 struct sched_group *sg;
835 struct cpumask *sg_span;
837 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
838 GFP_KERNEL, cpu_to_node(cpu));
843 sg_span = sched_group_span(sg);
845 cpumask_copy(sg_span, sched_domain_span(sd->child));
847 cpumask_copy(sg_span, sched_domain_span(sd));
849 atomic_inc(&sg->ref);
853 static void init_overlap_sched_group(struct sched_domain *sd,
854 struct sched_group *sg)
856 struct cpumask *mask = sched_domains_tmpmask2;
857 struct sd_data *sdd = sd->private;
858 struct cpumask *sg_span;
861 build_balance_mask(sd, sg, mask);
862 cpu = cpumask_first_and(sched_group_span(sg), mask);
864 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
865 if (atomic_inc_return(&sg->sgc->ref) == 1)
866 cpumask_copy(group_balance_mask(sg), mask);
868 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
871 * Initialize sgc->capacity such that even if we mess up the
872 * domains and no possible iteration will get us here, we won't
875 sg_span = sched_group_span(sg);
876 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
877 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
878 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
882 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
884 struct sched_group *first = NULL, *last = NULL, *sg;
885 const struct cpumask *span = sched_domain_span(sd);
886 struct cpumask *covered = sched_domains_tmpmask;
887 struct sd_data *sdd = sd->private;
888 struct sched_domain *sibling;
891 cpumask_clear(covered);
893 for_each_cpu_wrap(i, span, cpu) {
894 struct cpumask *sg_span;
896 if (cpumask_test_cpu(i, covered))
899 sibling = *per_cpu_ptr(sdd->sd, i);
902 * Asymmetric node setups can result in situations where the
903 * domain tree is of unequal depth, make sure to skip domains
904 * that already cover the entire range.
906 * In that case build_sched_domains() will have terminated the
907 * iteration early and our sibling sd spans will be empty.
908 * Domains should always include the CPU they're built on, so
911 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
914 sg = build_group_from_child_sched_domain(sibling, cpu);
918 sg_span = sched_group_span(sg);
919 cpumask_or(covered, covered, sg_span);
921 init_overlap_sched_group(sd, sg);
935 free_sched_groups(first, 0);
942 * Package topology (also see the load-balance blurb in fair.c)
944 * The scheduler builds a tree structure to represent a number of important
945 * topology features. By default (default_topology[]) these include:
947 * - Simultaneous multithreading (SMT)
948 * - Multi-Core Cache (MC)
951 * Where the last one more or less denotes everything up to a NUMA node.
953 * The tree consists of 3 primary data structures:
955 * sched_domain -> sched_group -> sched_group_capacity
959 * The sched_domains are per-CPU and have a two way link (parent & child) and
960 * denote the ever growing mask of CPUs belonging to that level of topology.
962 * Each sched_domain has a circular (double) linked list of sched_group's, each
963 * denoting the domains of the level below (or individual CPUs in case of the
964 * first domain level). The sched_group linked by a sched_domain includes the
965 * CPU of that sched_domain [*].
967 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
969 * CPU 0 1 2 3 4 5 6 7
973 * SMT [ ] [ ] [ ] [ ]
977 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
978 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
979 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
981 * CPU 0 1 2 3 4 5 6 7
983 * One way to think about it is: sched_domain moves you up and down among these
984 * topology levels, while sched_group moves you sideways through it, at child
985 * domain granularity.
987 * sched_group_capacity ensures each unique sched_group has shared storage.
989 * There are two related construction problems, both require a CPU that
990 * uniquely identify each group (for a given domain):
992 * - The first is the balance_cpu (see should_we_balance() and the
993 * load-balance blub in fair.c); for each group we only want 1 CPU to
994 * continue balancing at a higher domain.
996 * - The second is the sched_group_capacity; we want all identical groups
997 * to share a single sched_group_capacity.
999 * Since these topologies are exclusive by construction. That is, its
1000 * impossible for an SMT thread to belong to multiple cores, and cores to
1001 * be part of multiple caches. There is a very clear and unique location
1002 * for each CPU in the hierarchy.
1004 * Therefore computing a unique CPU for each group is trivial (the iteration
1005 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1006 * group), we can simply pick the first CPU in each group.
1009 * [*] in other words, the first group of each domain is its child domain.
1012 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1014 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1015 struct sched_domain *child = sd->child;
1016 struct sched_group *sg;
1019 cpu = cpumask_first(sched_domain_span(child));
1021 sg = *per_cpu_ptr(sdd->sg, cpu);
1022 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1024 /* For claim_allocations: */
1025 atomic_inc(&sg->ref);
1026 atomic_inc(&sg->sgc->ref);
1029 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1030 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1032 cpumask_set_cpu(cpu, sched_group_span(sg));
1033 cpumask_set_cpu(cpu, group_balance_mask(sg));
1036 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1037 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1038 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1044 * build_sched_groups will build a circular linked list of the groups
1045 * covered by the given span, and will set each group's ->cpumask correctly,
1046 * and ->cpu_capacity to 0.
1048 * Assumes the sched_domain tree is fully constructed
1051 build_sched_groups(struct sched_domain *sd, int cpu)
1053 struct sched_group *first = NULL, *last = NULL;
1054 struct sd_data *sdd = sd->private;
1055 const struct cpumask *span = sched_domain_span(sd);
1056 struct cpumask *covered;
1059 lockdep_assert_held(&sched_domains_mutex);
1060 covered = sched_domains_tmpmask;
1062 cpumask_clear(covered);
1064 for_each_cpu_wrap(i, span, cpu) {
1065 struct sched_group *sg;
1067 if (cpumask_test_cpu(i, covered))
1070 sg = get_group(i, sdd);
1072 cpumask_or(covered, covered, sched_group_span(sg));
1087 * Initialize sched groups cpu_capacity.
1089 * cpu_capacity indicates the capacity of sched group, which is used while
1090 * distributing the load between different sched groups in a sched domain.
1091 * Typically cpu_capacity for all the groups in a sched domain will be same
1092 * unless there are asymmetries in the topology. If there are asymmetries,
1093 * group having more cpu_capacity will pickup more load compared to the
1094 * group having less cpu_capacity.
1096 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1098 struct sched_group *sg = sd->groups;
1103 int cpu, max_cpu = -1;
1105 sg->group_weight = cpumask_weight(sched_group_span(sg));
1107 if (!(sd->flags & SD_ASYM_PACKING))
1110 for_each_cpu(cpu, sched_group_span(sg)) {
1113 else if (sched_asym_prefer(cpu, max_cpu))
1116 sg->asym_prefer_cpu = max_cpu;
1120 } while (sg != sd->groups);
1122 if (cpu != group_balance_cpu(sg))
1125 update_group_capacity(sd, cpu);
1129 * Initializers for schedule domains
1130 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1133 static int default_relax_domain_level = -1;
1134 int sched_domain_level_max;
1136 static int __init setup_relax_domain_level(char *str)
1138 if (kstrtoint(str, 0, &default_relax_domain_level))
1139 pr_warn("Unable to set relax_domain_level\n");
1143 __setup("relax_domain_level=", setup_relax_domain_level);
1145 static void set_domain_attribute(struct sched_domain *sd,
1146 struct sched_domain_attr *attr)
1150 if (!attr || attr->relax_domain_level < 0) {
1151 if (default_relax_domain_level < 0)
1154 request = default_relax_domain_level;
1156 request = attr->relax_domain_level;
1157 if (request < sd->level) {
1158 /* Turn off idle balance on this domain: */
1159 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1161 /* Turn on idle balance on this domain: */
1162 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1166 static void __sdt_free(const struct cpumask *cpu_map);
1167 static int __sdt_alloc(const struct cpumask *cpu_map);
1169 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1170 const struct cpumask *cpu_map)
1174 if (!atomic_read(&d->rd->refcount))
1175 free_rootdomain(&d->rd->rcu);
1181 __sdt_free(cpu_map);
1189 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1191 memset(d, 0, sizeof(*d));
1193 if (__sdt_alloc(cpu_map))
1194 return sa_sd_storage;
1195 d->sd = alloc_percpu(struct sched_domain *);
1197 return sa_sd_storage;
1198 d->rd = alloc_rootdomain();
1202 return sa_rootdomain;
1206 * NULL the sd_data elements we've used to build the sched_domain and
1207 * sched_group structure so that the subsequent __free_domain_allocs()
1208 * will not free the data we're using.
1210 static void claim_allocations(int cpu, struct sched_domain *sd)
1212 struct sd_data *sdd = sd->private;
1214 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1215 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1217 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1218 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1220 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1221 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1223 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1224 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1228 enum numa_topology_type sched_numa_topology_type;
1230 static int sched_domains_numa_levels;
1231 static int sched_domains_curr_level;
1233 int sched_max_numa_distance;
1234 static int *sched_domains_numa_distance;
1235 static struct cpumask ***sched_domains_numa_masks;
1239 * SD_flags allowed in topology descriptions.
1241 * These flags are purely descriptive of the topology and do not prescribe
1242 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1245 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1246 * SD_SHARE_PKG_RESOURCES - describes shared caches
1247 * SD_NUMA - describes NUMA topologies
1248 * SD_SHARE_POWERDOMAIN - describes shared power domain
1250 * Odd one out, which beside describing the topology has a quirk also
1251 * prescribes the desired behaviour that goes along with it:
1253 * SD_ASYM_PACKING - describes SMT quirks
1255 #define TOPOLOGY_SD_FLAGS \
1256 (SD_SHARE_CPUCAPACITY | \
1257 SD_SHARE_PKG_RESOURCES | \
1260 SD_SHARE_POWERDOMAIN)
1262 static struct sched_domain *
1263 sd_init(struct sched_domain_topology_level *tl,
1264 const struct cpumask *cpu_map,
1265 struct sched_domain *child, int dflags, int cpu)
1267 struct sd_data *sdd = &tl->data;
1268 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1269 int sd_id, sd_weight, sd_flags = 0;
1273 * Ugly hack to pass state to sd_numa_mask()...
1275 sched_domains_curr_level = tl->numa_level;
1278 sd_weight = cpumask_weight(tl->mask(cpu));
1281 sd_flags = (*tl->sd_flags)();
1282 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1283 "wrong sd_flags in topology description\n"))
1284 sd_flags &= ~TOPOLOGY_SD_FLAGS;
1286 /* Apply detected topology flags */
1289 *sd = (struct sched_domain){
1290 .min_interval = sd_weight,
1291 .max_interval = 2*sd_weight,
1293 .imbalance_pct = 125,
1295 .cache_nice_tries = 0,
1302 .flags = 1*SD_LOAD_BALANCE
1303 | 1*SD_BALANCE_NEWIDLE
1308 | 0*SD_SHARE_CPUCAPACITY
1309 | 0*SD_SHARE_PKG_RESOURCES
1311 | 1*SD_PREFER_SIBLING
1316 .last_balance = jiffies,
1317 .balance_interval = sd_weight,
1318 .max_newidle_lb_cost = 0,
1319 .next_decay_max_lb_cost = jiffies,
1321 #ifdef CONFIG_SCHED_DEBUG
1326 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1327 sd_id = cpumask_first(sched_domain_span(sd));
1330 * Convert topological properties into behaviour.
1333 if (sd->flags & SD_ASYM_CPUCAPACITY) {
1334 struct sched_domain *t = sd;
1337 * Don't attempt to spread across CPUs of different capacities.
1340 sd->child->flags &= ~SD_PREFER_SIBLING;
1342 for_each_lower_domain(t)
1343 t->flags |= SD_BALANCE_WAKE;
1346 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1347 sd->imbalance_pct = 110;
1349 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1350 sd->imbalance_pct = 117;
1351 sd->cache_nice_tries = 1;
1355 } else if (sd->flags & SD_NUMA) {
1356 sd->cache_nice_tries = 2;
1360 sd->flags &= ~SD_PREFER_SIBLING;
1361 sd->flags |= SD_SERIALIZE;
1362 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
1363 sd->flags &= ~(SD_BALANCE_EXEC |
1370 sd->cache_nice_tries = 1;
1376 * For all levels sharing cache; connect a sched_domain_shared
1379 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1380 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1381 atomic_inc(&sd->shared->ref);
1382 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1391 * Topology list, bottom-up.
1393 static struct sched_domain_topology_level default_topology[] = {
1394 #ifdef CONFIG_SCHED_SMT
1395 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1397 #ifdef CONFIG_SCHED_MC
1398 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1400 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1404 static struct sched_domain_topology_level *sched_domain_topology =
1407 #define for_each_sd_topology(tl) \
1408 for (tl = sched_domain_topology; tl->mask; tl++)
1410 void set_sched_topology(struct sched_domain_topology_level *tl)
1412 if (WARN_ON_ONCE(sched_smp_initialized))
1415 sched_domain_topology = tl;
1420 static const struct cpumask *sd_numa_mask(int cpu)
1422 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1425 static void sched_numa_warn(const char *str)
1427 static int done = false;
1435 printk(KERN_WARNING "ERROR: %s\n\n", str);
1437 for (i = 0; i < nr_node_ids; i++) {
1438 printk(KERN_WARNING " ");
1439 for (j = 0; j < nr_node_ids; j++)
1440 printk(KERN_CONT "%02d ", node_distance(i,j));
1441 printk(KERN_CONT "\n");
1443 printk(KERN_WARNING "\n");
1446 bool find_numa_distance(int distance)
1450 if (distance == node_distance(0, 0))
1453 for (i = 0; i < sched_domains_numa_levels; i++) {
1454 if (sched_domains_numa_distance[i] == distance)
1462 * A system can have three types of NUMA topology:
1463 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1464 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1465 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1467 * The difference between a glueless mesh topology and a backplane
1468 * topology lies in whether communication between not directly
1469 * connected nodes goes through intermediary nodes (where programs
1470 * could run), or through backplane controllers. This affects
1471 * placement of programs.
1473 * The type of topology can be discerned with the following tests:
1474 * - If the maximum distance between any nodes is 1 hop, the system
1475 * is directly connected.
1476 * - If for two nodes A and B, located N > 1 hops away from each other,
1477 * there is an intermediary node C, which is < N hops away from both
1478 * nodes A and B, the system is a glueless mesh.
1480 static void init_numa_topology_type(void)
1484 n = sched_max_numa_distance;
1486 if (sched_domains_numa_levels <= 2) {
1487 sched_numa_topology_type = NUMA_DIRECT;
1491 for_each_online_node(a) {
1492 for_each_online_node(b) {
1493 /* Find two nodes furthest removed from each other. */
1494 if (node_distance(a, b) < n)
1497 /* Is there an intermediary node between a and b? */
1498 for_each_online_node(c) {
1499 if (node_distance(a, c) < n &&
1500 node_distance(b, c) < n) {
1501 sched_numa_topology_type =
1507 sched_numa_topology_type = NUMA_BACKPLANE;
1513 void sched_init_numa(void)
1515 int next_distance, curr_distance = node_distance(0, 0);
1516 struct sched_domain_topology_level *tl;
1520 sched_domains_numa_distance = kzalloc(sizeof(int) * (nr_node_ids + 1), GFP_KERNEL);
1521 if (!sched_domains_numa_distance)
1524 /* Includes NUMA identity node at level 0. */
1525 sched_domains_numa_distance[level++] = curr_distance;
1526 sched_domains_numa_levels = level;
1529 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1530 * unique distances in the node_distance() table.
1532 * Assumes node_distance(0,j) includes all distances in
1533 * node_distance(i,j) in order to avoid cubic time.
1535 next_distance = curr_distance;
1536 for (i = 0; i < nr_node_ids; i++) {
1537 for (j = 0; j < nr_node_ids; j++) {
1538 for (k = 0; k < nr_node_ids; k++) {
1539 int distance = node_distance(i, k);
1541 if (distance > curr_distance &&
1542 (distance < next_distance ||
1543 next_distance == curr_distance))
1544 next_distance = distance;
1547 * While not a strong assumption it would be nice to know
1548 * about cases where if node A is connected to B, B is not
1549 * equally connected to A.
1551 if (sched_debug() && node_distance(k, i) != distance)
1552 sched_numa_warn("Node-distance not symmetric");
1554 if (sched_debug() && i && !find_numa_distance(distance))
1555 sched_numa_warn("Node-0 not representative");
1557 if (next_distance != curr_distance) {
1558 sched_domains_numa_distance[level++] = next_distance;
1559 sched_domains_numa_levels = level;
1560 curr_distance = next_distance;
1565 * In case of sched_debug() we verify the above assumption.
1572 * 'level' contains the number of unique distances
1574 * The sched_domains_numa_distance[] array includes the actual distance
1579 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1580 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1581 * the array will contain less then 'level' members. This could be
1582 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1583 * in other functions.
1585 * We reset it to 'level' at the end of this function.
1587 sched_domains_numa_levels = 0;
1589 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1590 if (!sched_domains_numa_masks)
1594 * Now for each level, construct a mask per node which contains all
1595 * CPUs of nodes that are that many hops away from us.
1597 for (i = 0; i < level; i++) {
1598 sched_domains_numa_masks[i] =
1599 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1600 if (!sched_domains_numa_masks[i])
1603 for (j = 0; j < nr_node_ids; j++) {
1604 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1608 sched_domains_numa_masks[i][j] = mask;
1611 if (node_distance(j, k) > sched_domains_numa_distance[i])
1614 cpumask_or(mask, mask, cpumask_of_node(k));
1619 /* Compute default topology size */
1620 for (i = 0; sched_domain_topology[i].mask; i++);
1622 tl = kzalloc((i + level + 1) *
1623 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1628 * Copy the default topology bits..
1630 for (i = 0; sched_domain_topology[i].mask; i++)
1631 tl[i] = sched_domain_topology[i];
1634 * Add the NUMA identity distance, aka single NODE.
1636 tl[i++] = (struct sched_domain_topology_level){
1637 .mask = sd_numa_mask,
1643 * .. and append 'j' levels of NUMA goodness.
1645 for (j = 1; j < level; i++, j++) {
1646 tl[i] = (struct sched_domain_topology_level){
1647 .mask = sd_numa_mask,
1648 .sd_flags = cpu_numa_flags,
1649 .flags = SDTL_OVERLAP,
1655 sched_domain_topology = tl;
1657 sched_domains_numa_levels = level;
1658 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1660 init_numa_topology_type();
1663 void sched_domains_numa_masks_set(unsigned int cpu)
1665 int node = cpu_to_node(cpu);
1668 for (i = 0; i < sched_domains_numa_levels; i++) {
1669 for (j = 0; j < nr_node_ids; j++) {
1670 if (node_distance(j, node) <= sched_domains_numa_distance[i])
1671 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1676 void sched_domains_numa_masks_clear(unsigned int cpu)
1680 for (i = 0; i < sched_domains_numa_levels; i++) {
1681 for (j = 0; j < nr_node_ids; j++)
1682 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1686 #endif /* CONFIG_NUMA */
1688 static int __sdt_alloc(const struct cpumask *cpu_map)
1690 struct sched_domain_topology_level *tl;
1693 for_each_sd_topology(tl) {
1694 struct sd_data *sdd = &tl->data;
1696 sdd->sd = alloc_percpu(struct sched_domain *);
1700 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1704 sdd->sg = alloc_percpu(struct sched_group *);
1708 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1712 for_each_cpu(j, cpu_map) {
1713 struct sched_domain *sd;
1714 struct sched_domain_shared *sds;
1715 struct sched_group *sg;
1716 struct sched_group_capacity *sgc;
1718 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1719 GFP_KERNEL, cpu_to_node(j));
1723 *per_cpu_ptr(sdd->sd, j) = sd;
1725 sds = kzalloc_node(sizeof(struct sched_domain_shared),
1726 GFP_KERNEL, cpu_to_node(j));
1730 *per_cpu_ptr(sdd->sds, j) = sds;
1732 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1733 GFP_KERNEL, cpu_to_node(j));
1739 *per_cpu_ptr(sdd->sg, j) = sg;
1741 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1742 GFP_KERNEL, cpu_to_node(j));
1746 #ifdef CONFIG_SCHED_DEBUG
1750 *per_cpu_ptr(sdd->sgc, j) = sgc;
1757 static void __sdt_free(const struct cpumask *cpu_map)
1759 struct sched_domain_topology_level *tl;
1762 for_each_sd_topology(tl) {
1763 struct sd_data *sdd = &tl->data;
1765 for_each_cpu(j, cpu_map) {
1766 struct sched_domain *sd;
1769 sd = *per_cpu_ptr(sdd->sd, j);
1770 if (sd && (sd->flags & SD_OVERLAP))
1771 free_sched_groups(sd->groups, 0);
1772 kfree(*per_cpu_ptr(sdd->sd, j));
1776 kfree(*per_cpu_ptr(sdd->sds, j));
1778 kfree(*per_cpu_ptr(sdd->sg, j));
1780 kfree(*per_cpu_ptr(sdd->sgc, j));
1782 free_percpu(sdd->sd);
1784 free_percpu(sdd->sds);
1786 free_percpu(sdd->sg);
1788 free_percpu(sdd->sgc);
1793 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1794 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1795 struct sched_domain *child, int dflags, int cpu)
1797 struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu);
1800 sd->level = child->level + 1;
1801 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1804 if (!cpumask_subset(sched_domain_span(child),
1805 sched_domain_span(sd))) {
1806 pr_err("BUG: arch topology borken\n");
1807 #ifdef CONFIG_SCHED_DEBUG
1808 pr_err(" the %s domain not a subset of the %s domain\n",
1809 child->name, sd->name);
1811 /* Fixup, ensure @sd has at least @child CPUs. */
1812 cpumask_or(sched_domain_span(sd),
1813 sched_domain_span(sd),
1814 sched_domain_span(child));
1818 set_domain_attribute(sd, attr);
1824 * Find the sched_domain_topology_level where all CPU capacities are visible
1827 static struct sched_domain_topology_level
1828 *asym_cpu_capacity_level(const struct cpumask *cpu_map)
1830 int i, j, asym_level = 0;
1832 struct sched_domain_topology_level *tl, *asym_tl = NULL;
1835 /* Is there any asymmetry? */
1836 cap = arch_scale_cpu_capacity(NULL, cpumask_first(cpu_map));
1838 for_each_cpu(i, cpu_map) {
1839 if (arch_scale_cpu_capacity(NULL, i) != cap) {
1849 * Examine topology from all CPU's point of views to detect the lowest
1850 * sched_domain_topology_level where a highest capacity CPU is visible
1853 for_each_cpu(i, cpu_map) {
1854 unsigned long max_capacity = arch_scale_cpu_capacity(NULL, i);
1857 for_each_sd_topology(tl) {
1858 if (tl_id < asym_level)
1861 for_each_cpu_and(j, tl->mask(i), cpu_map) {
1862 unsigned long capacity;
1864 capacity = arch_scale_cpu_capacity(NULL, j);
1866 if (capacity <= max_capacity)
1869 max_capacity = capacity;
1883 * Build sched domains for a given set of CPUs and attach the sched domains
1884 * to the individual CPUs
1887 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1889 enum s_alloc alloc_state;
1890 struct sched_domain *sd;
1892 struct rq *rq = NULL;
1893 int i, ret = -ENOMEM;
1894 struct sched_domain_topology_level *tl_asym;
1895 bool has_asym = false;
1897 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1898 if (alloc_state != sa_rootdomain)
1901 tl_asym = asym_cpu_capacity_level(cpu_map);
1903 /* Set up domains for CPUs specified by the cpu_map: */
1904 for_each_cpu(i, cpu_map) {
1905 struct sched_domain_topology_level *tl;
1908 for_each_sd_topology(tl) {
1911 if (tl == tl_asym) {
1912 dflags |= SD_ASYM_CPUCAPACITY;
1916 sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i);
1918 if (tl == sched_domain_topology)
1919 *per_cpu_ptr(d.sd, i) = sd;
1920 if (tl->flags & SDTL_OVERLAP)
1921 sd->flags |= SD_OVERLAP;
1922 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
1927 /* Build the groups for the domains */
1928 for_each_cpu(i, cpu_map) {
1929 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1930 sd->span_weight = cpumask_weight(sched_domain_span(sd));
1931 if (sd->flags & SD_OVERLAP) {
1932 if (build_overlap_sched_groups(sd, i))
1935 if (build_sched_groups(sd, i))
1941 /* Calculate CPU capacity for physical packages and nodes */
1942 for (i = nr_cpumask_bits-1; i >= 0; i--) {
1943 if (!cpumask_test_cpu(i, cpu_map))
1946 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1947 claim_allocations(i, sd);
1948 init_sched_groups_capacity(i, sd);
1952 /* Attach the domains */
1954 for_each_cpu(i, cpu_map) {
1956 sd = *per_cpu_ptr(d.sd, i);
1958 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
1959 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
1960 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
1962 cpu_attach_domain(sd, d.rd, i);
1967 static_branch_enable_cpuslocked(&sched_asym_cpucapacity);
1969 if (rq && sched_debug_enabled) {
1970 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
1971 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
1976 __free_domain_allocs(&d, alloc_state, cpu_map);
1981 /* Current sched domains: */
1982 static cpumask_var_t *doms_cur;
1984 /* Number of sched domains in 'doms_cur': */
1985 static int ndoms_cur;
1987 /* Attribues of custom domains in 'doms_cur' */
1988 static struct sched_domain_attr *dattr_cur;
1991 * Special case: If a kmalloc() of a doms_cur partition (array of
1992 * cpumask) fails, then fallback to a single sched domain,
1993 * as determined by the single cpumask fallback_doms.
1995 static cpumask_var_t fallback_doms;
1998 * arch_update_cpu_topology lets virtualized architectures update the
1999 * CPU core maps. It is supposed to return 1 if the topology changed
2000 * or 0 if it stayed the same.
2002 int __weak arch_update_cpu_topology(void)
2007 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2010 cpumask_var_t *doms;
2012 doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2015 for (i = 0; i < ndoms; i++) {
2016 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2017 free_sched_domains(doms, i);
2024 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2027 for (i = 0; i < ndoms; i++)
2028 free_cpumask_var(doms[i]);
2033 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
2034 * For now this just excludes isolated CPUs, but could be used to
2035 * exclude other special cases in the future.
2037 int sched_init_domains(const struct cpumask *cpu_map)
2041 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2042 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2043 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2045 arch_update_cpu_topology();
2047 doms_cur = alloc_sched_domains(ndoms_cur);
2049 doms_cur = &fallback_doms;
2050 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
2051 err = build_sched_domains(doms_cur[0], NULL);
2052 register_sched_domain_sysctl();
2058 * Detach sched domains from a group of CPUs specified in cpu_map
2059 * These CPUs will now be attached to the NULL domain
2061 static void detach_destroy_domains(const struct cpumask *cpu_map)
2066 for_each_cpu(i, cpu_map)
2067 cpu_attach_domain(NULL, &def_root_domain, i);
2071 /* handle null as "default" */
2072 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2073 struct sched_domain_attr *new, int idx_new)
2075 struct sched_domain_attr tmp;
2083 return !memcmp(cur ? (cur + idx_cur) : &tmp,
2084 new ? (new + idx_new) : &tmp,
2085 sizeof(struct sched_domain_attr));
2089 * Partition sched domains as specified by the 'ndoms_new'
2090 * cpumasks in the array doms_new[] of cpumasks. This compares
2091 * doms_new[] to the current sched domain partitioning, doms_cur[].
2092 * It destroys each deleted domain and builds each new domain.
2094 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2095 * The masks don't intersect (don't overlap.) We should setup one
2096 * sched domain for each mask. CPUs not in any of the cpumasks will
2097 * not be load balanced. If the same cpumask appears both in the
2098 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2101 * The passed in 'doms_new' should be allocated using
2102 * alloc_sched_domains. This routine takes ownership of it and will
2103 * free_sched_domains it when done with it. If the caller failed the
2104 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2105 * and partition_sched_domains() will fallback to the single partition
2106 * 'fallback_doms', it also forces the domains to be rebuilt.
2108 * If doms_new == NULL it will be replaced with cpu_online_mask.
2109 * ndoms_new == 0 is a special case for destroying existing domains,
2110 * and it will not create the default domain.
2112 * Call with hotplug lock held
2114 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2115 struct sched_domain_attr *dattr_new)
2120 mutex_lock(&sched_domains_mutex);
2122 /* Always unregister in case we don't destroy any domains: */
2123 unregister_sched_domain_sysctl();
2125 /* Let the architecture update CPU core mappings: */
2126 new_topology = arch_update_cpu_topology();
2129 WARN_ON_ONCE(dattr_new);
2131 doms_new = alloc_sched_domains(1);
2134 cpumask_and(doms_new[0], cpu_active_mask,
2135 housekeeping_cpumask(HK_FLAG_DOMAIN));
2141 /* Destroy deleted domains: */
2142 for (i = 0; i < ndoms_cur; i++) {
2143 for (j = 0; j < n && !new_topology; j++) {
2144 if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2145 dattrs_equal(dattr_cur, i, dattr_new, j))
2148 /* No match - a current sched domain not in new doms_new[] */
2149 detach_destroy_domains(doms_cur[i]);
2157 doms_new = &fallback_doms;
2158 cpumask_and(doms_new[0], cpu_active_mask,
2159 housekeeping_cpumask(HK_FLAG_DOMAIN));
2162 /* Build new domains: */
2163 for (i = 0; i < ndoms_new; i++) {
2164 for (j = 0; j < n && !new_topology; j++) {
2165 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2166 dattrs_equal(dattr_new, i, dattr_cur, j))
2169 /* No match - add a new doms_new */
2170 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2175 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2176 /* Build perf. domains: */
2177 for (i = 0; i < ndoms_new; i++) {
2178 for (j = 0; j < n && !sched_energy_update; j++) {
2179 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2180 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd)
2183 /* No match - add perf. domains for a new rd */
2184 build_perf_domains(doms_new[i]);
2190 /* Remember the new sched domains: */
2191 if (doms_cur != &fallback_doms)
2192 free_sched_domains(doms_cur, ndoms_cur);
2195 doms_cur = doms_new;
2196 dattr_cur = dattr_new;
2197 ndoms_cur = ndoms_new;
2199 register_sched_domain_sysctl();
2201 mutex_unlock(&sched_domains_mutex);