1 // SPDX-License-Identifier: GPL-2.0
3 * Scheduler topology setup/handling methods
6 DEFINE_MUTEX(sched_domains_mutex);
8 /* Protected by sched_domains_mutex: */
9 static cpumask_var_t sched_domains_tmpmask;
10 static cpumask_var_t sched_domains_tmpmask2;
12 #ifdef CONFIG_SCHED_DEBUG
14 static int __init sched_debug_setup(char *str)
16 sched_debug_verbose = true;
20 early_param("sched_verbose", sched_debug_setup);
22 static inline bool sched_debug(void)
24 return sched_debug_verbose;
27 #define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
28 const struct sd_flag_debug sd_flag_debug[] = {
29 #include <linux/sched/sd_flags.h>
33 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
34 struct cpumask *groupmask)
36 struct sched_group *group = sd->groups;
37 unsigned long flags = sd->flags;
40 cpumask_clear(groupmask);
42 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
43 printk(KERN_CONT "span=%*pbl level=%s\n",
44 cpumask_pr_args(sched_domain_span(sd)), sd->name);
46 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
47 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
49 if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
50 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
53 for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
54 unsigned int flag = BIT(idx);
55 unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
57 if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
58 !(sd->child->flags & flag))
59 printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
60 sd_flag_debug[idx].name);
62 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
63 !(sd->parent->flags & flag))
64 printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
65 sd_flag_debug[idx].name);
68 printk(KERN_DEBUG "%*s groups:", level + 1, "");
72 printk(KERN_ERR "ERROR: group is NULL\n");
76 if (cpumask_empty(sched_group_span(group))) {
77 printk(KERN_CONT "\n");
78 printk(KERN_ERR "ERROR: empty group\n");
82 if (!(sd->flags & SD_OVERLAP) &&
83 cpumask_intersects(groupmask, sched_group_span(group))) {
84 printk(KERN_CONT "\n");
85 printk(KERN_ERR "ERROR: repeated CPUs\n");
89 cpumask_or(groupmask, groupmask, sched_group_span(group));
91 printk(KERN_CONT " %d:{ span=%*pbl",
93 cpumask_pr_args(sched_group_span(group)));
95 if ((sd->flags & SD_OVERLAP) &&
96 !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
97 printk(KERN_CONT " mask=%*pbl",
98 cpumask_pr_args(group_balance_mask(group)));
101 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
102 printk(KERN_CONT " cap=%lu", group->sgc->capacity);
104 if (group == sd->groups && sd->child &&
105 !cpumask_equal(sched_domain_span(sd->child),
106 sched_group_span(group))) {
107 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
110 printk(KERN_CONT " }");
114 if (group != sd->groups)
115 printk(KERN_CONT ",");
117 } while (group != sd->groups);
118 printk(KERN_CONT "\n");
120 if (!cpumask_equal(sched_domain_span(sd), groupmask))
121 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
124 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
125 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
129 static void sched_domain_debug(struct sched_domain *sd, int cpu)
133 if (!sched_debug_verbose)
137 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
141 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
144 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
152 #else /* !CONFIG_SCHED_DEBUG */
154 # define sched_debug_verbose 0
155 # define sched_domain_debug(sd, cpu) do { } while (0)
156 static inline bool sched_debug(void)
160 #endif /* CONFIG_SCHED_DEBUG */
162 /* Generate a mask of SD flags with the SDF_NEEDS_GROUPS metaflag */
163 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
164 static const unsigned int SD_DEGENERATE_GROUPS_MASK =
165 #include <linux/sched/sd_flags.h>
169 static int sd_degenerate(struct sched_domain *sd)
171 if (cpumask_weight(sched_domain_span(sd)) == 1)
174 /* Following flags need at least 2 groups */
175 if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
176 (sd->groups != sd->groups->next))
179 /* Following flags don't use groups */
180 if (sd->flags & (SD_WAKE_AFFINE))
187 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
189 unsigned long cflags = sd->flags, pflags = parent->flags;
191 if (sd_degenerate(parent))
194 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
197 /* Flags needing groups don't count if only 1 group in parent */
198 if (parent->groups == parent->groups->next)
199 pflags &= ~SD_DEGENERATE_GROUPS_MASK;
201 if (~cflags & pflags)
207 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
208 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
209 unsigned int sysctl_sched_energy_aware = 1;
210 DEFINE_MUTEX(sched_energy_mutex);
211 bool sched_energy_update;
213 void rebuild_sched_domains_energy(void)
215 mutex_lock(&sched_energy_mutex);
216 sched_energy_update = true;
217 rebuild_sched_domains();
218 sched_energy_update = false;
219 mutex_unlock(&sched_energy_mutex);
222 #ifdef CONFIG_PROC_SYSCTL
223 int sched_energy_aware_handler(struct ctl_table *table, int write,
224 void *buffer, size_t *lenp, loff_t *ppos)
228 if (write && !capable(CAP_SYS_ADMIN))
231 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
233 state = static_branch_unlikely(&sched_energy_present);
234 if (state != sysctl_sched_energy_aware)
235 rebuild_sched_domains_energy();
242 static void free_pd(struct perf_domain *pd)
244 struct perf_domain *tmp;
253 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
256 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
264 static struct perf_domain *pd_init(int cpu)
266 struct em_perf_domain *obj = em_cpu_get(cpu);
267 struct perf_domain *pd;
271 pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
275 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
283 static void perf_domain_debug(const struct cpumask *cpu_map,
284 struct perf_domain *pd)
286 if (!sched_debug() || !pd)
289 printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
292 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
293 cpumask_first(perf_domain_span(pd)),
294 cpumask_pr_args(perf_domain_span(pd)),
295 em_pd_nr_perf_states(pd->em_pd));
299 printk(KERN_CONT "\n");
302 static void destroy_perf_domain_rcu(struct rcu_head *rp)
304 struct perf_domain *pd;
306 pd = container_of(rp, struct perf_domain, rcu);
310 static void sched_energy_set(bool has_eas)
312 if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
314 pr_info("%s: stopping EAS\n", __func__);
315 static_branch_disable_cpuslocked(&sched_energy_present);
316 } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
318 pr_info("%s: starting EAS\n", __func__);
319 static_branch_enable_cpuslocked(&sched_energy_present);
324 * EAS can be used on a root domain if it meets all the following conditions:
325 * 1. an Energy Model (EM) is available;
326 * 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
327 * 3. no SMT is detected.
328 * 4. the EM complexity is low enough to keep scheduling overheads low;
329 * 5. schedutil is driving the frequency of all CPUs of the rd;
330 * 6. frequency invariance support is present;
332 * The complexity of the Energy Model is defined as:
334 * C = nr_pd * (nr_cpus + nr_ps)
336 * with parameters defined as:
337 * - nr_pd: the number of performance domains
338 * - nr_cpus: the number of CPUs
339 * - nr_ps: the sum of the number of performance states of all performance
340 * domains (for example, on a system with 2 performance domains,
341 * with 10 performance states each, nr_ps = 2 * 10 = 20).
343 * It is generally not a good idea to use such a model in the wake-up path on
344 * very complex platforms because of the associated scheduling overheads. The
345 * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
346 * with per-CPU DVFS and less than 8 performance states each, for example.
348 #define EM_MAX_COMPLEXITY 2048
350 extern struct cpufreq_governor schedutil_gov;
351 static bool build_perf_domains(const struct cpumask *cpu_map)
353 int i, nr_pd = 0, nr_ps = 0, nr_cpus = cpumask_weight(cpu_map);
354 struct perf_domain *pd = NULL, *tmp;
355 int cpu = cpumask_first(cpu_map);
356 struct root_domain *rd = cpu_rq(cpu)->rd;
357 struct cpufreq_policy *policy;
358 struct cpufreq_governor *gov;
360 if (!sysctl_sched_energy_aware)
363 /* EAS is enabled for asymmetric CPU capacity topologies. */
364 if (!per_cpu(sd_asym_cpucapacity, cpu)) {
366 pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
367 cpumask_pr_args(cpu_map));
372 /* EAS definitely does *not* handle SMT */
373 if (sched_smt_active()) {
374 pr_warn("rd %*pbl: Disabling EAS, SMT is not supported\n",
375 cpumask_pr_args(cpu_map));
379 if (!arch_scale_freq_invariant()) {
381 pr_warn("rd %*pbl: Disabling EAS: frequency-invariant load tracking not yet supported",
382 cpumask_pr_args(cpu_map));
387 for_each_cpu(i, cpu_map) {
388 /* Skip already covered CPUs. */
392 /* Do not attempt EAS if schedutil is not being used. */
393 policy = cpufreq_cpu_get(i);
396 gov = policy->governor;
397 cpufreq_cpu_put(policy);
398 if (gov != &schedutil_gov) {
400 pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
401 cpumask_pr_args(cpu_map));
405 /* Create the new pd and add it to the local list. */
413 * Count performance domains and performance states for the
417 nr_ps += em_pd_nr_perf_states(pd->em_pd);
420 /* Bail out if the Energy Model complexity is too high. */
421 if (nr_pd * (nr_ps + nr_cpus) > EM_MAX_COMPLEXITY) {
422 WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
423 cpumask_pr_args(cpu_map));
427 perf_domain_debug(cpu_map, pd);
429 /* Attach the new list of performance domains to the root domain. */
431 rcu_assign_pointer(rd->pd, pd);
433 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
440 rcu_assign_pointer(rd->pd, NULL);
442 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
447 static void free_pd(struct perf_domain *pd) { }
448 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
450 static void free_rootdomain(struct rcu_head *rcu)
452 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
454 cpupri_cleanup(&rd->cpupri);
455 cpudl_cleanup(&rd->cpudl);
456 free_cpumask_var(rd->dlo_mask);
457 free_cpumask_var(rd->rto_mask);
458 free_cpumask_var(rd->online);
459 free_cpumask_var(rd->span);
464 void rq_attach_root(struct rq *rq, struct root_domain *rd)
466 struct root_domain *old_rd = NULL;
469 raw_spin_rq_lock_irqsave(rq, flags);
474 if (cpumask_test_cpu(rq->cpu, old_rd->online))
477 cpumask_clear_cpu(rq->cpu, old_rd->span);
480 * If we dont want to free the old_rd yet then
481 * set old_rd to NULL to skip the freeing later
484 if (!atomic_dec_and_test(&old_rd->refcount))
488 atomic_inc(&rd->refcount);
491 cpumask_set_cpu(rq->cpu, rd->span);
492 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
495 raw_spin_rq_unlock_irqrestore(rq, flags);
498 call_rcu(&old_rd->rcu, free_rootdomain);
501 void sched_get_rd(struct root_domain *rd)
503 atomic_inc(&rd->refcount);
506 void sched_put_rd(struct root_domain *rd)
508 if (!atomic_dec_and_test(&rd->refcount))
511 call_rcu(&rd->rcu, free_rootdomain);
514 static int init_rootdomain(struct root_domain *rd)
516 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
518 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
520 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
522 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
525 #ifdef HAVE_RT_PUSH_IPI
527 raw_spin_lock_init(&rd->rto_lock);
528 rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
532 init_dl_bw(&rd->dl_bw);
533 if (cpudl_init(&rd->cpudl) != 0)
536 if (cpupri_init(&rd->cpupri) != 0)
541 cpudl_cleanup(&rd->cpudl);
543 free_cpumask_var(rd->rto_mask);
545 free_cpumask_var(rd->dlo_mask);
547 free_cpumask_var(rd->online);
549 free_cpumask_var(rd->span);
555 * By default the system creates a single root-domain with all CPUs as
556 * members (mimicking the global state we have today).
558 struct root_domain def_root_domain;
560 void init_defrootdomain(void)
562 init_rootdomain(&def_root_domain);
564 atomic_set(&def_root_domain.refcount, 1);
567 static struct root_domain *alloc_rootdomain(void)
569 struct root_domain *rd;
571 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
575 if (init_rootdomain(rd) != 0) {
583 static void free_sched_groups(struct sched_group *sg, int free_sgc)
585 struct sched_group *tmp, *first;
594 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
597 if (atomic_dec_and_test(&sg->ref))
600 } while (sg != first);
603 static void destroy_sched_domain(struct sched_domain *sd)
606 * A normal sched domain may have multiple group references, an
607 * overlapping domain, having private groups, only one. Iterate,
608 * dropping group/capacity references, freeing where none remain.
610 free_sched_groups(sd->groups, 1);
612 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
617 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
619 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
622 struct sched_domain *parent = sd->parent;
623 destroy_sched_domain(sd);
628 static void destroy_sched_domains(struct sched_domain *sd)
631 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
635 * Keep a special pointer to the highest sched_domain that has
636 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
637 * allows us to avoid some pointer chasing select_idle_sibling().
639 * Also keep a unique ID per domain (we use the first CPU number in
640 * the cpumask of the domain), this allows us to quickly tell if
641 * two CPUs are in the same cache domain, see cpus_share_cache().
643 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
644 DEFINE_PER_CPU(int, sd_llc_size);
645 DEFINE_PER_CPU(int, sd_llc_id);
646 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
647 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
648 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
649 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
650 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
652 static void update_top_cache_domain(int cpu)
654 struct sched_domain_shared *sds = NULL;
655 struct sched_domain *sd;
659 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
661 id = cpumask_first(sched_domain_span(sd));
662 size = cpumask_weight(sched_domain_span(sd));
666 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
667 per_cpu(sd_llc_size, cpu) = size;
668 per_cpu(sd_llc_id, cpu) = id;
669 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
671 sd = lowest_flag_domain(cpu, SD_NUMA);
672 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
674 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
675 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
677 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
678 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
682 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
683 * hold the hotplug lock.
686 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
688 struct rq *rq = cpu_rq(cpu);
689 struct sched_domain *tmp;
691 /* Remove the sched domains which do not contribute to scheduling. */
692 for (tmp = sd; tmp; ) {
693 struct sched_domain *parent = tmp->parent;
697 if (sd_parent_degenerate(tmp, parent)) {
698 tmp->parent = parent->parent;
700 parent->parent->child = tmp;
702 * Transfer SD_PREFER_SIBLING down in case of a
703 * degenerate parent; the spans match for this
704 * so the property transfers.
706 if (parent->flags & SD_PREFER_SIBLING)
707 tmp->flags |= SD_PREFER_SIBLING;
708 destroy_sched_domain(parent);
713 if (sd && sd_degenerate(sd)) {
716 destroy_sched_domain(tmp);
718 struct sched_group *sg = sd->groups;
721 * sched groups hold the flags of the child sched
722 * domain for convenience. Clear such flags since
723 * the child is being destroyed.
727 } while (sg != sd->groups);
733 sched_domain_debug(sd, cpu);
735 rq_attach_root(rq, rd);
737 rcu_assign_pointer(rq->sd, sd);
738 dirty_sched_domain_sysctl(cpu);
739 destroy_sched_domains(tmp);
741 update_top_cache_domain(cpu);
745 struct sched_domain * __percpu *sd;
746 struct root_domain *rd;
757 * Return the canonical balance CPU for this group, this is the first CPU
758 * of this group that's also in the balance mask.
760 * The balance mask are all those CPUs that could actually end up at this
761 * group. See build_balance_mask().
763 * Also see should_we_balance().
765 int group_balance_cpu(struct sched_group *sg)
767 return cpumask_first(group_balance_mask(sg));
772 * NUMA topology (first read the regular topology blurb below)
774 * Given a node-distance table, for example:
782 * which represents a 4 node ring topology like:
790 * We want to construct domains and groups to represent this. The way we go
791 * about doing this is to build the domains on 'hops'. For each NUMA level we
792 * construct the mask of all nodes reachable in @level hops.
794 * For the above NUMA topology that gives 3 levels:
796 * NUMA-2 0-3 0-3 0-3 0-3
797 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
799 * NUMA-1 0-1,3 0-2 1-3 0,2-3
800 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
805 * As can be seen; things don't nicely line up as with the regular topology.
806 * When we iterate a domain in child domain chunks some nodes can be
807 * represented multiple times -- hence the "overlap" naming for this part of
810 * In order to minimize this overlap, we only build enough groups to cover the
811 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
815 * - the first group of each domain is its child domain; this
816 * gets us the first 0-1,3
817 * - the only uncovered node is 2, who's child domain is 1-3.
819 * However, because of the overlap, computing a unique CPU for each group is
820 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
821 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
822 * end up at those groups (they would end up in group: 0-1,3).
824 * To correct this we have to introduce the group balance mask. This mask
825 * will contain those CPUs in the group that can reach this group given the
826 * (child) domain tree.
828 * With this we can once again compute balance_cpu and sched_group_capacity
831 * XXX include words on how balance_cpu is unique and therefore can be
832 * used for sched_group_capacity links.
835 * Another 'interesting' topology is:
843 * Which looks a little like:
851 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
854 * This leads to a few particularly weird cases where the sched_domain's are
855 * not of the same number for each CPU. Consider:
858 * groups: {0-2},{1-3} {1-3},{0-2}
860 * NUMA-1 0-2 0-3 0-3 1-3
868 * Build the balance mask; it contains only those CPUs that can arrive at this
869 * group and should be considered to continue balancing.
871 * We do this during the group creation pass, therefore the group information
872 * isn't complete yet, however since each group represents a (child) domain we
873 * can fully construct this using the sched_domain bits (which are already
877 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
879 const struct cpumask *sg_span = sched_group_span(sg);
880 struct sd_data *sdd = sd->private;
881 struct sched_domain *sibling;
886 for_each_cpu(i, sg_span) {
887 sibling = *per_cpu_ptr(sdd->sd, i);
890 * Can happen in the asymmetric case, where these siblings are
891 * unused. The mask will not be empty because those CPUs that
892 * do have the top domain _should_ span the domain.
897 /* If we would not end up here, we can't continue from here */
898 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
901 cpumask_set_cpu(i, mask);
904 /* We must not have empty masks here */
905 WARN_ON_ONCE(cpumask_empty(mask));
909 * XXX: This creates per-node group entries; since the load-balancer will
910 * immediately access remote memory to construct this group's load-balance
911 * statistics having the groups node local is of dubious benefit.
913 static struct sched_group *
914 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
916 struct sched_group *sg;
917 struct cpumask *sg_span;
919 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
920 GFP_KERNEL, cpu_to_node(cpu));
925 sg_span = sched_group_span(sg);
927 cpumask_copy(sg_span, sched_domain_span(sd->child));
928 sg->flags = sd->child->flags;
930 cpumask_copy(sg_span, sched_domain_span(sd));
933 atomic_inc(&sg->ref);
937 static void init_overlap_sched_group(struct sched_domain *sd,
938 struct sched_group *sg)
940 struct cpumask *mask = sched_domains_tmpmask2;
941 struct sd_data *sdd = sd->private;
942 struct cpumask *sg_span;
945 build_balance_mask(sd, sg, mask);
946 cpu = cpumask_first(mask);
948 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
949 if (atomic_inc_return(&sg->sgc->ref) == 1)
950 cpumask_copy(group_balance_mask(sg), mask);
952 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
955 * Initialize sgc->capacity such that even if we mess up the
956 * domains and no possible iteration will get us here, we won't
959 sg_span = sched_group_span(sg);
960 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
961 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
962 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
965 static struct sched_domain *
966 find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
969 * The proper descendant would be the one whose child won't span out
972 while (sibling->child &&
973 !cpumask_subset(sched_domain_span(sibling->child),
974 sched_domain_span(sd)))
975 sibling = sibling->child;
978 * As we are referencing sgc across different topology level, we need
979 * to go down to skip those sched_domains which don't contribute to
980 * scheduling because they will be degenerated in cpu_attach_domain
982 while (sibling->child &&
983 cpumask_equal(sched_domain_span(sibling->child),
984 sched_domain_span(sibling)))
985 sibling = sibling->child;
991 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
993 struct sched_group *first = NULL, *last = NULL, *sg;
994 const struct cpumask *span = sched_domain_span(sd);
995 struct cpumask *covered = sched_domains_tmpmask;
996 struct sd_data *sdd = sd->private;
997 struct sched_domain *sibling;
1000 cpumask_clear(covered);
1002 for_each_cpu_wrap(i, span, cpu) {
1003 struct cpumask *sg_span;
1005 if (cpumask_test_cpu(i, covered))
1008 sibling = *per_cpu_ptr(sdd->sd, i);
1011 * Asymmetric node setups can result in situations where the
1012 * domain tree is of unequal depth, make sure to skip domains
1013 * that already cover the entire range.
1015 * In that case build_sched_domains() will have terminated the
1016 * iteration early and our sibling sd spans will be empty.
1017 * Domains should always include the CPU they're built on, so
1020 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1024 * Usually we build sched_group by sibling's child sched_domain
1025 * But for machines whose NUMA diameter are 3 or above, we move
1026 * to build sched_group by sibling's proper descendant's child
1027 * domain because sibling's child sched_domain will span out of
1028 * the sched_domain being built as below.
1030 * Smallest diameter=3 topology is:
1038 * 0 --- 1 --- 2 --- 3
1040 * NUMA-3 0-3 N/A N/A 0-3
1041 * groups: {0-2},{1-3} {1-3},{0-2}
1043 * NUMA-2 0-2 0-3 0-3 1-3
1044 * groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2}
1046 * NUMA-1 0-1 0-2 1-3 2-3
1047 * groups: {0},{1} {1},{2},{0} {2},{3},{1} {3},{2}
1051 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1052 * group span isn't a subset of the domain span.
1054 if (sibling->child &&
1055 !cpumask_subset(sched_domain_span(sibling->child), span))
1056 sibling = find_descended_sibling(sd, sibling);
1058 sg = build_group_from_child_sched_domain(sibling, cpu);
1062 sg_span = sched_group_span(sg);
1063 cpumask_or(covered, covered, sg_span);
1065 init_overlap_sched_group(sibling, sg);
1079 free_sched_groups(first, 0);
1086 * Package topology (also see the load-balance blurb in fair.c)
1088 * The scheduler builds a tree structure to represent a number of important
1089 * topology features. By default (default_topology[]) these include:
1091 * - Simultaneous multithreading (SMT)
1092 * - Multi-Core Cache (MC)
1095 * Where the last one more or less denotes everything up to a NUMA node.
1097 * The tree consists of 3 primary data structures:
1099 * sched_domain -> sched_group -> sched_group_capacity
1103 * The sched_domains are per-CPU and have a two way link (parent & child) and
1104 * denote the ever growing mask of CPUs belonging to that level of topology.
1106 * Each sched_domain has a circular (double) linked list of sched_group's, each
1107 * denoting the domains of the level below (or individual CPUs in case of the
1108 * first domain level). The sched_group linked by a sched_domain includes the
1109 * CPU of that sched_domain [*].
1111 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1113 * CPU 0 1 2 3 4 5 6 7
1117 * SMT [ ] [ ] [ ] [ ]
1121 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1122 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1123 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1125 * CPU 0 1 2 3 4 5 6 7
1127 * One way to think about it is: sched_domain moves you up and down among these
1128 * topology levels, while sched_group moves you sideways through it, at child
1129 * domain granularity.
1131 * sched_group_capacity ensures each unique sched_group has shared storage.
1133 * There are two related construction problems, both require a CPU that
1134 * uniquely identify each group (for a given domain):
1136 * - The first is the balance_cpu (see should_we_balance() and the
1137 * load-balance blub in fair.c); for each group we only want 1 CPU to
1138 * continue balancing at a higher domain.
1140 * - The second is the sched_group_capacity; we want all identical groups
1141 * to share a single sched_group_capacity.
1143 * Since these topologies are exclusive by construction. That is, its
1144 * impossible for an SMT thread to belong to multiple cores, and cores to
1145 * be part of multiple caches. There is a very clear and unique location
1146 * for each CPU in the hierarchy.
1148 * Therefore computing a unique CPU for each group is trivial (the iteration
1149 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1150 * group), we can simply pick the first CPU in each group.
1153 * [*] in other words, the first group of each domain is its child domain.
1156 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1158 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1159 struct sched_domain *child = sd->child;
1160 struct sched_group *sg;
1161 bool already_visited;
1164 cpu = cpumask_first(sched_domain_span(child));
1166 sg = *per_cpu_ptr(sdd->sg, cpu);
1167 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1169 /* Increase refcounts for claim_allocations: */
1170 already_visited = atomic_inc_return(&sg->ref) > 1;
1171 /* sgc visits should follow a similar trend as sg */
1172 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1174 /* If we have already visited that group, it's already initialized. */
1175 if (already_visited)
1179 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1180 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1181 sg->flags = child->flags;
1183 cpumask_set_cpu(cpu, sched_group_span(sg));
1184 cpumask_set_cpu(cpu, group_balance_mask(sg));
1187 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1188 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1189 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1195 * build_sched_groups will build a circular linked list of the groups
1196 * covered by the given span, will set each group's ->cpumask correctly,
1197 * and will initialize their ->sgc.
1199 * Assumes the sched_domain tree is fully constructed
1202 build_sched_groups(struct sched_domain *sd, int cpu)
1204 struct sched_group *first = NULL, *last = NULL;
1205 struct sd_data *sdd = sd->private;
1206 const struct cpumask *span = sched_domain_span(sd);
1207 struct cpumask *covered;
1210 lockdep_assert_held(&sched_domains_mutex);
1211 covered = sched_domains_tmpmask;
1213 cpumask_clear(covered);
1215 for_each_cpu_wrap(i, span, cpu) {
1216 struct sched_group *sg;
1218 if (cpumask_test_cpu(i, covered))
1221 sg = get_group(i, sdd);
1223 cpumask_or(covered, covered, sched_group_span(sg));
1238 * Initialize sched groups cpu_capacity.
1240 * cpu_capacity indicates the capacity of sched group, which is used while
1241 * distributing the load between different sched groups in a sched domain.
1242 * Typically cpu_capacity for all the groups in a sched domain will be same
1243 * unless there are asymmetries in the topology. If there are asymmetries,
1244 * group having more cpu_capacity will pickup more load compared to the
1245 * group having less cpu_capacity.
1247 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1249 struct sched_group *sg = sd->groups;
1254 int cpu, max_cpu = -1;
1256 sg->group_weight = cpumask_weight(sched_group_span(sg));
1258 if (!(sd->flags & SD_ASYM_PACKING))
1261 for_each_cpu(cpu, sched_group_span(sg)) {
1264 else if (sched_asym_prefer(cpu, max_cpu))
1267 sg->asym_prefer_cpu = max_cpu;
1271 } while (sg != sd->groups);
1273 if (cpu != group_balance_cpu(sg))
1276 update_group_capacity(sd, cpu);
1280 * Asymmetric CPU capacity bits
1282 struct asym_cap_data {
1283 struct list_head link;
1284 unsigned long capacity;
1285 unsigned long cpus[];
1289 * Set of available CPUs grouped by their corresponding capacities
1290 * Each list entry contains a CPU mask reflecting CPUs that share the same
1292 * The lifespan of data is unlimited.
1294 static LIST_HEAD(asym_cap_list);
1296 #define cpu_capacity_span(asym_data) to_cpumask((asym_data)->cpus)
1299 * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1300 * Provides sd_flags reflecting the asymmetry scope.
1303 asym_cpu_capacity_classify(const struct cpumask *sd_span,
1304 const struct cpumask *cpu_map)
1306 struct asym_cap_data *entry;
1307 int count = 0, miss = 0;
1310 * Count how many unique CPU capacities this domain spans across
1311 * (compare sched_domain CPUs mask with ones representing available
1312 * CPUs capacities). Take into account CPUs that might be offline:
1315 list_for_each_entry(entry, &asym_cap_list, link) {
1316 if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1318 else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1322 WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1324 /* No asymmetry detected */
1327 /* Some of the available CPU capacity values have not been detected */
1329 return SD_ASYM_CPUCAPACITY;
1331 /* Full asymmetry */
1332 return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1336 static inline void asym_cpu_capacity_update_data(int cpu)
1338 unsigned long capacity = arch_scale_cpu_capacity(cpu);
1339 struct asym_cap_data *entry = NULL;
1341 list_for_each_entry(entry, &asym_cap_list, link) {
1342 if (capacity == entry->capacity)
1346 entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1347 if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1349 entry->capacity = capacity;
1350 list_add(&entry->link, &asym_cap_list);
1352 __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1356 * Build-up/update list of CPUs grouped by their capacities
1357 * An update requires explicit request to rebuild sched domains
1358 * with state indicating CPU topology changes.
1360 static void asym_cpu_capacity_scan(void)
1362 struct asym_cap_data *entry, *next;
1365 list_for_each_entry(entry, &asym_cap_list, link)
1366 cpumask_clear(cpu_capacity_span(entry));
1368 for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1369 asym_cpu_capacity_update_data(cpu);
1371 list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1372 if (cpumask_empty(cpu_capacity_span(entry))) {
1373 list_del(&entry->link);
1379 * Only one capacity value has been detected i.e. this system is symmetric.
1380 * No need to keep this data around.
1382 if (list_is_singular(&asym_cap_list)) {
1383 entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1384 list_del(&entry->link);
1390 * Initializers for schedule domains
1391 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1394 static int default_relax_domain_level = -1;
1395 int sched_domain_level_max;
1397 static int __init setup_relax_domain_level(char *str)
1399 if (kstrtoint(str, 0, &default_relax_domain_level))
1400 pr_warn("Unable to set relax_domain_level\n");
1404 __setup("relax_domain_level=", setup_relax_domain_level);
1406 static void set_domain_attribute(struct sched_domain *sd,
1407 struct sched_domain_attr *attr)
1411 if (!attr || attr->relax_domain_level < 0) {
1412 if (default_relax_domain_level < 0)
1414 request = default_relax_domain_level;
1416 request = attr->relax_domain_level;
1418 if (sd->level > request) {
1419 /* Turn off idle balance on this domain: */
1420 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1424 static void __sdt_free(const struct cpumask *cpu_map);
1425 static int __sdt_alloc(const struct cpumask *cpu_map);
1427 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1428 const struct cpumask *cpu_map)
1432 if (!atomic_read(&d->rd->refcount))
1433 free_rootdomain(&d->rd->rcu);
1439 __sdt_free(cpu_map);
1447 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1449 memset(d, 0, sizeof(*d));
1451 if (__sdt_alloc(cpu_map))
1452 return sa_sd_storage;
1453 d->sd = alloc_percpu(struct sched_domain *);
1455 return sa_sd_storage;
1456 d->rd = alloc_rootdomain();
1460 return sa_rootdomain;
1464 * NULL the sd_data elements we've used to build the sched_domain and
1465 * sched_group structure so that the subsequent __free_domain_allocs()
1466 * will not free the data we're using.
1468 static void claim_allocations(int cpu, struct sched_domain *sd)
1470 struct sd_data *sdd = sd->private;
1472 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1473 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1475 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1476 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1478 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1479 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1481 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1482 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1486 enum numa_topology_type sched_numa_topology_type;
1488 static int sched_domains_numa_levels;
1489 static int sched_domains_curr_level;
1491 int sched_max_numa_distance;
1492 static int *sched_domains_numa_distance;
1493 static struct cpumask ***sched_domains_numa_masks;
1497 * SD_flags allowed in topology descriptions.
1499 * These flags are purely descriptive of the topology and do not prescribe
1500 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1503 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1504 * SD_SHARE_PKG_RESOURCES - describes shared caches
1505 * SD_NUMA - describes NUMA topologies
1507 * Odd one out, which beside describing the topology has a quirk also
1508 * prescribes the desired behaviour that goes along with it:
1510 * SD_ASYM_PACKING - describes SMT quirks
1512 #define TOPOLOGY_SD_FLAGS \
1513 (SD_SHARE_CPUCAPACITY | \
1514 SD_SHARE_PKG_RESOURCES | \
1518 static struct sched_domain *
1519 sd_init(struct sched_domain_topology_level *tl,
1520 const struct cpumask *cpu_map,
1521 struct sched_domain *child, int cpu)
1523 struct sd_data *sdd = &tl->data;
1524 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1525 int sd_id, sd_weight, sd_flags = 0;
1526 struct cpumask *sd_span;
1530 * Ugly hack to pass state to sd_numa_mask()...
1532 sched_domains_curr_level = tl->numa_level;
1535 sd_weight = cpumask_weight(tl->mask(cpu));
1538 sd_flags = (*tl->sd_flags)();
1539 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1540 "wrong sd_flags in topology description\n"))
1541 sd_flags &= TOPOLOGY_SD_FLAGS;
1543 *sd = (struct sched_domain){
1544 .min_interval = sd_weight,
1545 .max_interval = 2*sd_weight,
1547 .imbalance_pct = 117,
1549 .cache_nice_tries = 0,
1551 .flags = 1*SD_BALANCE_NEWIDLE
1556 | 0*SD_SHARE_CPUCAPACITY
1557 | 0*SD_SHARE_PKG_RESOURCES
1559 | 1*SD_PREFER_SIBLING
1564 .last_balance = jiffies,
1565 .balance_interval = sd_weight,
1566 .max_newidle_lb_cost = 0,
1567 .last_decay_max_lb_cost = jiffies,
1569 #ifdef CONFIG_SCHED_DEBUG
1574 sd_span = sched_domain_span(sd);
1575 cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1576 sd_id = cpumask_first(sd_span);
1578 sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1580 WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1581 (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1582 "CPU capacity asymmetry not supported on SMT\n");
1585 * Convert topological properties into behaviour.
1587 /* Don't attempt to spread across CPUs of different capacities. */
1588 if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1589 sd->child->flags &= ~SD_PREFER_SIBLING;
1591 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1592 sd->imbalance_pct = 110;
1594 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1595 sd->imbalance_pct = 117;
1596 sd->cache_nice_tries = 1;
1599 } else if (sd->flags & SD_NUMA) {
1600 sd->cache_nice_tries = 2;
1602 sd->flags &= ~SD_PREFER_SIBLING;
1603 sd->flags |= SD_SERIALIZE;
1604 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1605 sd->flags &= ~(SD_BALANCE_EXEC |
1612 sd->cache_nice_tries = 1;
1616 * For all levels sharing cache; connect a sched_domain_shared
1619 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1620 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1621 atomic_inc(&sd->shared->ref);
1622 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1631 * Topology list, bottom-up.
1633 static struct sched_domain_topology_level default_topology[] = {
1634 #ifdef CONFIG_SCHED_SMT
1635 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1638 #ifdef CONFIG_SCHED_CLUSTER
1639 { cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1642 #ifdef CONFIG_SCHED_MC
1643 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1645 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1649 static struct sched_domain_topology_level *sched_domain_topology =
1651 static struct sched_domain_topology_level *sched_domain_topology_saved;
1653 #define for_each_sd_topology(tl) \
1654 for (tl = sched_domain_topology; tl->mask; tl++)
1656 void set_sched_topology(struct sched_domain_topology_level *tl)
1658 if (WARN_ON_ONCE(sched_smp_initialized))
1661 sched_domain_topology = tl;
1662 sched_domain_topology_saved = NULL;
1667 static const struct cpumask *sd_numa_mask(int cpu)
1669 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1672 static void sched_numa_warn(const char *str)
1674 static int done = false;
1682 printk(KERN_WARNING "ERROR: %s\n\n", str);
1684 for (i = 0; i < nr_node_ids; i++) {
1685 printk(KERN_WARNING " ");
1686 for (j = 0; j < nr_node_ids; j++) {
1687 if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
1688 printk(KERN_CONT "(%02d) ", node_distance(i,j));
1690 printk(KERN_CONT " %02d ", node_distance(i,j));
1692 printk(KERN_CONT "\n");
1694 printk(KERN_WARNING "\n");
1697 bool find_numa_distance(int distance)
1702 if (distance == node_distance(0, 0))
1706 distances = rcu_dereference(sched_domains_numa_distance);
1709 for (i = 0; i < sched_domains_numa_levels; i++) {
1710 if (distances[i] == distance) {
1721 #define for_each_cpu_node_but(n, nbut) \
1722 for_each_node_state(n, N_CPU) \
1728 * A system can have three types of NUMA topology:
1729 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1730 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1731 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1733 * The difference between a glueless mesh topology and a backplane
1734 * topology lies in whether communication between not directly
1735 * connected nodes goes through intermediary nodes (where programs
1736 * could run), or through backplane controllers. This affects
1737 * placement of programs.
1739 * The type of topology can be discerned with the following tests:
1740 * - If the maximum distance between any nodes is 1 hop, the system
1741 * is directly connected.
1742 * - If for two nodes A and B, located N > 1 hops away from each other,
1743 * there is an intermediary node C, which is < N hops away from both
1744 * nodes A and B, the system is a glueless mesh.
1746 static void init_numa_topology_type(int offline_node)
1750 n = sched_max_numa_distance;
1752 if (sched_domains_numa_levels <= 2) {
1753 sched_numa_topology_type = NUMA_DIRECT;
1757 for_each_cpu_node_but(a, offline_node) {
1758 for_each_cpu_node_but(b, offline_node) {
1759 /* Find two nodes furthest removed from each other. */
1760 if (node_distance(a, b) < n)
1763 /* Is there an intermediary node between a and b? */
1764 for_each_cpu_node_but(c, offline_node) {
1765 if (node_distance(a, c) < n &&
1766 node_distance(b, c) < n) {
1767 sched_numa_topology_type =
1773 sched_numa_topology_type = NUMA_BACKPLANE;
1778 pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1779 sched_numa_topology_type = NUMA_DIRECT;
1783 #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1785 void sched_init_numa(int offline_node)
1787 struct sched_domain_topology_level *tl;
1788 unsigned long *distance_map;
1792 struct cpumask ***masks;
1795 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1796 * unique distances in the node_distance() table.
1798 distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1802 bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1803 for_each_cpu_node_but(i, offline_node) {
1804 for_each_cpu_node_but(j, offline_node) {
1805 int distance = node_distance(i, j);
1807 if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1808 sched_numa_warn("Invalid distance value range");
1809 bitmap_free(distance_map);
1813 bitmap_set(distance_map, distance, 1);
1817 * We can now figure out how many unique distance values there are and
1818 * allocate memory accordingly.
1820 nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1822 distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1824 bitmap_free(distance_map);
1828 for (i = 0, j = 0; i < nr_levels; i++, j++) {
1829 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1832 rcu_assign_pointer(sched_domains_numa_distance, distances);
1834 bitmap_free(distance_map);
1837 * 'nr_levels' contains the number of unique distances
1839 * The sched_domains_numa_distance[] array includes the actual distance
1844 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1845 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1846 * the array will contain less then 'nr_levels' members. This could be
1847 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1848 * in other functions.
1850 * We reset it to 'nr_levels' at the end of this function.
1852 sched_domains_numa_levels = 0;
1854 masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1859 * Now for each level, construct a mask per node which contains all
1860 * CPUs of nodes that are that many hops away from us.
1862 for (i = 0; i < nr_levels; i++) {
1863 masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1867 for_each_cpu_node_but(j, offline_node) {
1868 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1876 for_each_cpu_node_but(k, offline_node) {
1877 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1878 sched_numa_warn("Node-distance not symmetric");
1880 if (node_distance(j, k) > sched_domains_numa_distance[i])
1883 cpumask_or(mask, mask, cpumask_of_node(k));
1887 rcu_assign_pointer(sched_domains_numa_masks, masks);
1889 /* Compute default topology size */
1890 for (i = 0; sched_domain_topology[i].mask; i++);
1892 tl = kzalloc((i + nr_levels + 1) *
1893 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1898 * Copy the default topology bits..
1900 for (i = 0; sched_domain_topology[i].mask; i++)
1901 tl[i] = sched_domain_topology[i];
1904 * Add the NUMA identity distance, aka single NODE.
1906 tl[i++] = (struct sched_domain_topology_level){
1907 .mask = sd_numa_mask,
1913 * .. and append 'j' levels of NUMA goodness.
1915 for (j = 1; j < nr_levels; i++, j++) {
1916 tl[i] = (struct sched_domain_topology_level){
1917 .mask = sd_numa_mask,
1918 .sd_flags = cpu_numa_flags,
1919 .flags = SDTL_OVERLAP,
1925 sched_domain_topology_saved = sched_domain_topology;
1926 sched_domain_topology = tl;
1928 sched_domains_numa_levels = nr_levels;
1929 WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]);
1931 init_numa_topology_type(offline_node);
1935 static void sched_reset_numa(void)
1937 int nr_levels, *distances;
1938 struct cpumask ***masks;
1940 nr_levels = sched_domains_numa_levels;
1941 sched_domains_numa_levels = 0;
1942 sched_max_numa_distance = 0;
1943 sched_numa_topology_type = NUMA_DIRECT;
1944 distances = sched_domains_numa_distance;
1945 rcu_assign_pointer(sched_domains_numa_distance, NULL);
1946 masks = sched_domains_numa_masks;
1947 rcu_assign_pointer(sched_domains_numa_masks, NULL);
1948 if (distances || masks) {
1953 for (i = 0; i < nr_levels && masks; i++) {
1962 if (sched_domain_topology_saved) {
1963 kfree(sched_domain_topology);
1964 sched_domain_topology = sched_domain_topology_saved;
1965 sched_domain_topology_saved = NULL;
1970 * Call with hotplug lock held
1972 void sched_update_numa(int cpu, bool online)
1976 node = cpu_to_node(cpu);
1978 * Scheduler NUMA topology is updated when the first CPU of a
1979 * node is onlined or the last CPU of a node is offlined.
1981 if (cpumask_weight(cpumask_of_node(node)) != 1)
1985 sched_init_numa(online ? NUMA_NO_NODE : node);
1988 void sched_domains_numa_masks_set(unsigned int cpu)
1990 int node = cpu_to_node(cpu);
1993 for (i = 0; i < sched_domains_numa_levels; i++) {
1994 for (j = 0; j < nr_node_ids; j++) {
1995 if (!node_state(j, N_CPU))
1998 /* Set ourselves in the remote node's masks */
1999 if (node_distance(j, node) <= sched_domains_numa_distance[i])
2000 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
2005 void sched_domains_numa_masks_clear(unsigned int cpu)
2009 for (i = 0; i < sched_domains_numa_levels; i++) {
2010 for (j = 0; j < nr_node_ids; j++) {
2011 if (sched_domains_numa_masks[i][j])
2012 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
2018 * sched_numa_find_closest() - given the NUMA topology, find the cpu
2019 * closest to @cpu from @cpumask.
2020 * cpumask: cpumask to find a cpu from
2021 * cpu: cpu to be close to
2023 * returns: cpu, or nr_cpu_ids when nothing found.
2025 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2027 int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2028 struct cpumask ***masks;
2031 masks = rcu_dereference(sched_domains_numa_masks);
2034 for (i = 0; i < sched_domains_numa_levels; i++) {
2037 cpu = cpumask_any_and(cpus, masks[i][j]);
2038 if (cpu < nr_cpu_ids) {
2049 #endif /* CONFIG_NUMA */
2051 static int __sdt_alloc(const struct cpumask *cpu_map)
2053 struct sched_domain_topology_level *tl;
2056 for_each_sd_topology(tl) {
2057 struct sd_data *sdd = &tl->data;
2059 sdd->sd = alloc_percpu(struct sched_domain *);
2063 sdd->sds = alloc_percpu(struct sched_domain_shared *);
2067 sdd->sg = alloc_percpu(struct sched_group *);
2071 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2075 for_each_cpu(j, cpu_map) {
2076 struct sched_domain *sd;
2077 struct sched_domain_shared *sds;
2078 struct sched_group *sg;
2079 struct sched_group_capacity *sgc;
2081 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2082 GFP_KERNEL, cpu_to_node(j));
2086 *per_cpu_ptr(sdd->sd, j) = sd;
2088 sds = kzalloc_node(sizeof(struct sched_domain_shared),
2089 GFP_KERNEL, cpu_to_node(j));
2093 *per_cpu_ptr(sdd->sds, j) = sds;
2095 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2096 GFP_KERNEL, cpu_to_node(j));
2102 *per_cpu_ptr(sdd->sg, j) = sg;
2104 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2105 GFP_KERNEL, cpu_to_node(j));
2109 #ifdef CONFIG_SCHED_DEBUG
2113 *per_cpu_ptr(sdd->sgc, j) = sgc;
2120 static void __sdt_free(const struct cpumask *cpu_map)
2122 struct sched_domain_topology_level *tl;
2125 for_each_sd_topology(tl) {
2126 struct sd_data *sdd = &tl->data;
2128 for_each_cpu(j, cpu_map) {
2129 struct sched_domain *sd;
2132 sd = *per_cpu_ptr(sdd->sd, j);
2133 if (sd && (sd->flags & SD_OVERLAP))
2134 free_sched_groups(sd->groups, 0);
2135 kfree(*per_cpu_ptr(sdd->sd, j));
2139 kfree(*per_cpu_ptr(sdd->sds, j));
2141 kfree(*per_cpu_ptr(sdd->sg, j));
2143 kfree(*per_cpu_ptr(sdd->sgc, j));
2145 free_percpu(sdd->sd);
2147 free_percpu(sdd->sds);
2149 free_percpu(sdd->sg);
2151 free_percpu(sdd->sgc);
2156 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2157 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2158 struct sched_domain *child, int cpu)
2160 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2163 sd->level = child->level + 1;
2164 sched_domain_level_max = max(sched_domain_level_max, sd->level);
2167 if (!cpumask_subset(sched_domain_span(child),
2168 sched_domain_span(sd))) {
2169 pr_err("BUG: arch topology borken\n");
2170 #ifdef CONFIG_SCHED_DEBUG
2171 pr_err(" the %s domain not a subset of the %s domain\n",
2172 child->name, sd->name);
2174 /* Fixup, ensure @sd has at least @child CPUs. */
2175 cpumask_or(sched_domain_span(sd),
2176 sched_domain_span(sd),
2177 sched_domain_span(child));
2181 set_domain_attribute(sd, attr);
2187 * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2188 * any two given CPUs at this (non-NUMA) topology level.
2190 static bool topology_span_sane(struct sched_domain_topology_level *tl,
2191 const struct cpumask *cpu_map, int cpu)
2195 /* NUMA levels are allowed to overlap */
2196 if (tl->flags & SDTL_OVERLAP)
2200 * Non-NUMA levels cannot partially overlap - they must be either
2201 * completely equal or completely disjoint. Otherwise we can end up
2202 * breaking the sched_group lists - i.e. a later get_group() pass
2203 * breaks the linking done for an earlier span.
2205 for_each_cpu(i, cpu_map) {
2209 * We should 'and' all those masks with 'cpu_map' to exactly
2210 * match the topology we're about to build, but that can only
2211 * remove CPUs, which only lessens our ability to detect
2214 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2215 cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2223 * Build sched domains for a given set of CPUs and attach the sched domains
2224 * to the individual CPUs
2227 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2229 enum s_alloc alloc_state = sa_none;
2230 struct sched_domain *sd;
2232 struct rq *rq = NULL;
2233 int i, ret = -ENOMEM;
2234 bool has_asym = false;
2236 if (WARN_ON(cpumask_empty(cpu_map)))
2239 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2240 if (alloc_state != sa_rootdomain)
2243 /* Set up domains for CPUs specified by the cpu_map: */
2244 for_each_cpu(i, cpu_map) {
2245 struct sched_domain_topology_level *tl;
2248 for_each_sd_topology(tl) {
2250 if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2253 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2255 has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2257 if (tl == sched_domain_topology)
2258 *per_cpu_ptr(d.sd, i) = sd;
2259 if (tl->flags & SDTL_OVERLAP)
2260 sd->flags |= SD_OVERLAP;
2261 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2266 /* Build the groups for the domains */
2267 for_each_cpu(i, cpu_map) {
2268 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2269 sd->span_weight = cpumask_weight(sched_domain_span(sd));
2270 if (sd->flags & SD_OVERLAP) {
2271 if (build_overlap_sched_groups(sd, i))
2274 if (build_sched_groups(sd, i))
2281 * Calculate an allowed NUMA imbalance such that LLCs do not get
2284 for_each_cpu(i, cpu_map) {
2285 unsigned int imb = 0;
2286 unsigned int imb_span = 1;
2288 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2289 struct sched_domain *child = sd->child;
2291 if (!(sd->flags & SD_SHARE_PKG_RESOURCES) && child &&
2292 (child->flags & SD_SHARE_PKG_RESOURCES)) {
2293 struct sched_domain __rcu *top_p;
2294 unsigned int nr_llcs;
2297 * For a single LLC per node, allow an
2298 * imbalance up to 25% of the node. This is an
2299 * arbitrary cutoff based on SMT-2 to balance
2300 * between memory bandwidth and avoiding
2301 * premature sharing of HT resources and SMT-4
2302 * or SMT-8 *may* benefit from a different
2305 * For multiple LLCs, allow an imbalance
2306 * until multiple tasks would share an LLC
2307 * on one node while LLCs on another node
2310 nr_llcs = sd->span_weight / child->span_weight;
2312 imb = sd->span_weight >> 2;
2315 sd->imb_numa_nr = imb;
2317 /* Set span based on the first NUMA domain. */
2319 while (top_p && !(top_p->flags & SD_NUMA)) {
2320 top_p = top_p->parent;
2322 imb_span = top_p ? top_p->span_weight : sd->span_weight;
2324 int factor = max(1U, (sd->span_weight / imb_span));
2326 sd->imb_numa_nr = imb * factor;
2331 /* Calculate CPU capacity for physical packages and nodes */
2332 for (i = nr_cpumask_bits-1; i >= 0; i--) {
2333 if (!cpumask_test_cpu(i, cpu_map))
2336 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2337 claim_allocations(i, sd);
2338 init_sched_groups_capacity(i, sd);
2342 /* Attach the domains */
2344 for_each_cpu(i, cpu_map) {
2346 sd = *per_cpu_ptr(d.sd, i);
2348 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2349 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2350 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2352 cpu_attach_domain(sd, d.rd, i);
2357 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2359 if (rq && sched_debug_verbose) {
2360 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2361 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2366 __free_domain_allocs(&d, alloc_state, cpu_map);
2371 /* Current sched domains: */
2372 static cpumask_var_t *doms_cur;
2374 /* Number of sched domains in 'doms_cur': */
2375 static int ndoms_cur;
2377 /* Attributes of custom domains in 'doms_cur' */
2378 static struct sched_domain_attr *dattr_cur;
2381 * Special case: If a kmalloc() of a doms_cur partition (array of
2382 * cpumask) fails, then fallback to a single sched domain,
2383 * as determined by the single cpumask fallback_doms.
2385 static cpumask_var_t fallback_doms;
2388 * arch_update_cpu_topology lets virtualized architectures update the
2389 * CPU core maps. It is supposed to return 1 if the topology changed
2390 * or 0 if it stayed the same.
2392 int __weak arch_update_cpu_topology(void)
2397 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2400 cpumask_var_t *doms;
2402 doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2405 for (i = 0; i < ndoms; i++) {
2406 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2407 free_sched_domains(doms, i);
2414 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2417 for (i = 0; i < ndoms; i++)
2418 free_cpumask_var(doms[i]);
2423 * Set up scheduler domains and groups. For now this just excludes isolated
2424 * CPUs, but could be used to exclude other special cases in the future.
2426 int sched_init_domains(const struct cpumask *cpu_map)
2430 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2431 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2432 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2434 arch_update_cpu_topology();
2435 asym_cpu_capacity_scan();
2437 doms_cur = alloc_sched_domains(ndoms_cur);
2439 doms_cur = &fallback_doms;
2440 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
2441 err = build_sched_domains(doms_cur[0], NULL);
2447 * Detach sched domains from a group of CPUs specified in cpu_map
2448 * These CPUs will now be attached to the NULL domain
2450 static void detach_destroy_domains(const struct cpumask *cpu_map)
2452 unsigned int cpu = cpumask_any(cpu_map);
2455 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2456 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2459 for_each_cpu(i, cpu_map)
2460 cpu_attach_domain(NULL, &def_root_domain, i);
2464 /* handle null as "default" */
2465 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2466 struct sched_domain_attr *new, int idx_new)
2468 struct sched_domain_attr tmp;
2476 return !memcmp(cur ? (cur + idx_cur) : &tmp,
2477 new ? (new + idx_new) : &tmp,
2478 sizeof(struct sched_domain_attr));
2482 * Partition sched domains as specified by the 'ndoms_new'
2483 * cpumasks in the array doms_new[] of cpumasks. This compares
2484 * doms_new[] to the current sched domain partitioning, doms_cur[].
2485 * It destroys each deleted domain and builds each new domain.
2487 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2488 * The masks don't intersect (don't overlap.) We should setup one
2489 * sched domain for each mask. CPUs not in any of the cpumasks will
2490 * not be load balanced. If the same cpumask appears both in the
2491 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2494 * The passed in 'doms_new' should be allocated using
2495 * alloc_sched_domains. This routine takes ownership of it and will
2496 * free_sched_domains it when done with it. If the caller failed the
2497 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2498 * and partition_sched_domains() will fallback to the single partition
2499 * 'fallback_doms', it also forces the domains to be rebuilt.
2501 * If doms_new == NULL it will be replaced with cpu_online_mask.
2502 * ndoms_new == 0 is a special case for destroying existing domains,
2503 * and it will not create the default domain.
2505 * Call with hotplug lock and sched_domains_mutex held
2507 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2508 struct sched_domain_attr *dattr_new)
2510 bool __maybe_unused has_eas = false;
2514 lockdep_assert_held(&sched_domains_mutex);
2516 /* Let the architecture update CPU core mappings: */
2517 new_topology = arch_update_cpu_topology();
2518 /* Trigger rebuilding CPU capacity asymmetry data */
2520 asym_cpu_capacity_scan();
2523 WARN_ON_ONCE(dattr_new);
2525 doms_new = alloc_sched_domains(1);
2528 cpumask_and(doms_new[0], cpu_active_mask,
2529 housekeeping_cpumask(HK_TYPE_DOMAIN));
2535 /* Destroy deleted domains: */
2536 for (i = 0; i < ndoms_cur; i++) {
2537 for (j = 0; j < n && !new_topology; j++) {
2538 if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2539 dattrs_equal(dattr_cur, i, dattr_new, j)) {
2540 struct root_domain *rd;
2543 * This domain won't be destroyed and as such
2544 * its dl_bw->total_bw needs to be cleared. It
2545 * will be recomputed in function
2546 * update_tasks_root_domain().
2548 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2549 dl_clear_root_domain(rd);
2553 /* No match - a current sched domain not in new doms_new[] */
2554 detach_destroy_domains(doms_cur[i]);
2562 doms_new = &fallback_doms;
2563 cpumask_and(doms_new[0], cpu_active_mask,
2564 housekeeping_cpumask(HK_TYPE_DOMAIN));
2567 /* Build new domains: */
2568 for (i = 0; i < ndoms_new; i++) {
2569 for (j = 0; j < n && !new_topology; j++) {
2570 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2571 dattrs_equal(dattr_new, i, dattr_cur, j))
2574 /* No match - add a new doms_new */
2575 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2580 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2581 /* Build perf. domains: */
2582 for (i = 0; i < ndoms_new; i++) {
2583 for (j = 0; j < n && !sched_energy_update; j++) {
2584 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2585 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2590 /* No match - add perf. domains for a new rd */
2591 has_eas |= build_perf_domains(doms_new[i]);
2595 sched_energy_set(has_eas);
2598 /* Remember the new sched domains: */
2599 if (doms_cur != &fallback_doms)
2600 free_sched_domains(doms_cur, ndoms_cur);
2603 doms_cur = doms_new;
2604 dattr_cur = dattr_new;
2605 ndoms_cur = ndoms_new;
2607 update_sched_domain_debugfs();
2611 * Call with hotplug lock held
2613 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2614 struct sched_domain_attr *dattr_new)
2616 mutex_lock(&sched_domains_mutex);
2617 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2618 mutex_unlock(&sched_domains_mutex);