Lines Matching +full:cpu +full:- +full:3
1 // SPDX-License-Identifier: GPL-2.0
34 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, in sched_domain_debug_one() argument
37 struct sched_group *group = sd->groups; in sched_domain_debug_one()
38 unsigned long flags = sd->flags; in sched_domain_debug_one()
43 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level); in sched_domain_debug_one()
45 cpumask_pr_args(sched_domain_span(sd)), sd->name); in sched_domain_debug_one()
47 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { in sched_domain_debug_one()
48 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); in sched_domain_debug_one()
50 if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) { in sched_domain_debug_one()
51 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); in sched_domain_debug_one()
58 if ((meta_flags & SDF_SHARED_CHILD) && sd->child && in sched_domain_debug_one()
59 !(sd->child->flags & flag)) in sched_domain_debug_one()
63 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent && in sched_domain_debug_one()
64 !(sd->parent->flags & flag)) in sched_domain_debug_one()
83 if (!(sd->flags & SD_OVERLAP) && in sched_domain_debug_one()
93 group->sgc->id, in sched_domain_debug_one()
96 if ((sd->flags & SD_OVERLAP) && in sched_domain_debug_one()
102 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) in sched_domain_debug_one()
103 printk(KERN_CONT " cap=%lu", group->sgc->capacity); in sched_domain_debug_one()
105 if (group == sd->groups && sd->child && in sched_domain_debug_one()
106 !cpumask_equal(sched_domain_span(sd->child), in sched_domain_debug_one()
108 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n"); in sched_domain_debug_one()
113 group = group->next; in sched_domain_debug_one()
115 if (group != sd->groups) in sched_domain_debug_one()
118 } while (group != sd->groups); in sched_domain_debug_one()
122 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); in sched_domain_debug_one()
124 if (sd->parent && in sched_domain_debug_one()
125 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) in sched_domain_debug_one()
126 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); in sched_domain_debug_one()
130 static void sched_domain_debug(struct sched_domain *sd, int cpu) in sched_domain_debug() argument
138 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); in sched_domain_debug()
142 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu); in sched_domain_debug()
145 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) in sched_domain_debug()
148 sd = sd->parent; in sched_domain_debug()
156 # define sched_domain_debug(sd, cpu) do { } while (0) argument
176 if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) && in sd_degenerate()
177 (sd->groups != sd->groups->next)) in sd_degenerate()
181 if (sd->flags & (SD_WAKE_AFFINE)) in sd_degenerate()
190 unsigned long cflags = sd->flags, pflags = parent->flags; in sd_parent_degenerate()
199 if (parent->groups == parent->groups->next) in sd_parent_degenerate()
221 return -EPERM; in sched_energy_aware_handler()
244 tmp = pd->next; in free_pd()
250 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu) in find_pd() argument
253 if (cpumask_test_cpu(cpu, perf_domain_span(pd))) in find_pd()
255 pd = pd->next; in find_pd()
261 static struct perf_domain *pd_init(int cpu) in pd_init() argument
263 struct em_perf_domain *obj = em_cpu_get(cpu); in pd_init()
268 pr_info("%s: no EM found for CPU%d\n", __func__, cpu); in pd_init()
275 pd->em_pd = obj; in pd_init()
292 em_pd_nr_perf_states(pd->em_pd)); in perf_domain_debug()
293 pd = pd->next; in perf_domain_debug()
324 * 3. no SMT is detected.
333 * - nr_pd: the number of performance domains
334 * - nr_cpus: the number of CPUs
335 * - nr_ps: the sum of the number of performance states of all performance
339 * It is generally not a good idea to use such a model in the wake-up path on
342 * with per-CPU DVFS and less than 8 performance states each, for example.
351 int cpu = cpumask_first(cpu_map); in build_perf_domains() local
352 struct root_domain *rd = cpu_rq(cpu)->rd; in build_perf_domains()
359 /* EAS is enabled for asymmetric CPU capacity topologies. */ in build_perf_domains()
360 if (!per_cpu(sd_asym_cpucapacity, cpu)) { in build_perf_domains()
384 gov = policy->governor; in build_perf_domains()
387 if (rd->pd) in build_perf_domains()
397 tmp->next = pd; in build_perf_domains()
405 nr_ps += em_pd_nr_perf_states(pd->em_pd); in build_perf_domains()
418 tmp = rd->pd; in build_perf_domains()
419 rcu_assign_pointer(rd->pd, pd); in build_perf_domains()
421 call_rcu(&tmp->rcu, destroy_perf_domain_rcu); in build_perf_domains()
427 tmp = rd->pd; in build_perf_domains()
428 rcu_assign_pointer(rd->pd, NULL); in build_perf_domains()
430 call_rcu(&tmp->rcu, destroy_perf_domain_rcu); in build_perf_domains()
442 cpupri_cleanup(&rd->cpupri); in free_rootdomain()
443 cpudl_cleanup(&rd->cpudl); in free_rootdomain()
444 free_cpumask_var(rd->dlo_mask); in free_rootdomain()
445 free_cpumask_var(rd->rto_mask); in free_rootdomain()
446 free_cpumask_var(rd->online); in free_rootdomain()
447 free_cpumask_var(rd->span); in free_rootdomain()
448 free_pd(rd->pd); in free_rootdomain()
457 raw_spin_lock_irqsave(&rq->lock, flags); in rq_attach_root()
459 if (rq->rd) { in rq_attach_root()
460 old_rd = rq->rd; in rq_attach_root()
462 if (cpumask_test_cpu(rq->cpu, old_rd->online)) in rq_attach_root()
465 cpumask_clear_cpu(rq->cpu, old_rd->span); in rq_attach_root()
472 if (!atomic_dec_and_test(&old_rd->refcount)) in rq_attach_root()
476 atomic_inc(&rd->refcount); in rq_attach_root()
477 rq->rd = rd; in rq_attach_root()
479 cpumask_set_cpu(rq->cpu, rd->span); in rq_attach_root()
480 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) in rq_attach_root()
483 raw_spin_unlock_irqrestore(&rq->lock, flags); in rq_attach_root()
486 call_rcu(&old_rd->rcu, free_rootdomain); in rq_attach_root()
491 atomic_inc(&rd->refcount); in sched_get_rd()
496 if (!atomic_dec_and_test(&rd->refcount)) in sched_put_rd()
499 call_rcu(&rd->rcu, free_rootdomain); in sched_put_rd()
504 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL)) in init_rootdomain()
506 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL)) in init_rootdomain()
508 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) in init_rootdomain()
510 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) in init_rootdomain()
514 rd->rto_cpu = -1; in init_rootdomain()
515 raw_spin_lock_init(&rd->rto_lock); in init_rootdomain()
516 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func); in init_rootdomain()
519 init_dl_bw(&rd->dl_bw); in init_rootdomain()
520 if (cpudl_init(&rd->cpudl) != 0) in init_rootdomain()
523 if (cpupri_init(&rd->cpupri) != 0) in init_rootdomain()
527 rd->max_cap_orig_cpu = -1; in init_rootdomain()
532 cpudl_cleanup(&rd->cpudl); in init_rootdomain()
534 free_cpumask_var(rd->rto_mask); in init_rootdomain()
536 free_cpumask_var(rd->dlo_mask); in init_rootdomain()
538 free_cpumask_var(rd->online); in init_rootdomain()
540 free_cpumask_var(rd->span); in init_rootdomain()
542 return -ENOMEM; in init_rootdomain()
546 * By default the system creates a single root-domain with all CPUs as
583 tmp = sg->next; in free_sched_groups()
585 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) in free_sched_groups()
586 kfree(sg->sgc); in free_sched_groups()
588 if (atomic_dec_and_test(&sg->ref)) in free_sched_groups()
601 free_sched_groups(sd->groups, 1); in destroy_sched_domain()
603 if (sd->shared && atomic_dec_and_test(&sd->shared->ref)) in destroy_sched_domain()
604 kfree(sd->shared); in destroy_sched_domain()
613 struct sched_domain *parent = sd->parent; in destroy_sched_domains_rcu()
622 call_rcu(&sd->rcu, destroy_sched_domains_rcu); in destroy_sched_domains()
630 * Also keep a unique ID per domain (we use the first CPU number in
643 static void update_top_cache_domain(int cpu) in update_top_cache_domain() argument
647 int id = cpu; in update_top_cache_domain()
650 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); in update_top_cache_domain()
654 sds = sd->shared; in update_top_cache_domain()
657 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); in update_top_cache_domain()
658 per_cpu(sd_llc_size, cpu) = size; in update_top_cache_domain()
659 per_cpu(sd_llc_id, cpu) = id; in update_top_cache_domain()
660 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds); in update_top_cache_domain()
662 sd = lowest_flag_domain(cpu, SD_NUMA); in update_top_cache_domain()
663 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); in update_top_cache_domain()
665 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); in update_top_cache_domain()
666 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd); in update_top_cache_domain()
668 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY); in update_top_cache_domain()
669 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd); in update_top_cache_domain()
673 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
677 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) in cpu_attach_domain() argument
679 struct rq *rq = cpu_rq(cpu); in cpu_attach_domain()
685 struct sched_domain *parent = tmp->parent; in cpu_attach_domain()
690 tmp->parent = parent->parent; in cpu_attach_domain()
691 if (parent->parent) in cpu_attach_domain()
692 parent->parent->child = tmp; in cpu_attach_domain()
698 if (parent->flags & SD_PREFER_SIBLING) in cpu_attach_domain()
699 tmp->flags |= SD_PREFER_SIBLING; in cpu_attach_domain()
702 tmp = tmp->parent; in cpu_attach_domain()
707 sd = sd->parent; in cpu_attach_domain()
710 sd->child = NULL; in cpu_attach_domain()
713 for (tmp = sd; tmp; tmp = tmp->parent) in cpu_attach_domain()
714 numa_distance += !!(tmp->flags & SD_NUMA); in cpu_attach_domain()
716 sched_domain_debug(sd, cpu); in cpu_attach_domain()
719 tmp = rq->sd; in cpu_attach_domain()
720 rcu_assign_pointer(rq->sd, sd); in cpu_attach_domain()
721 dirty_sched_domain_sysctl(cpu); in cpu_attach_domain()
724 update_top_cache_domain(cpu); in cpu_attach_domain()
740 * Return the canonical balance CPU for this group, this is the first CPU
757 * Given a node-distance table, for example:
759 * node 0 1 2 3
763 * 3: 20 30 20 10
767 * 0 ----- 1
771 * 3 ----- 2
777 * For the above NUMA topology that gives 3 levels:
779 * NUMA-2 0-3 0-3 0-3 0-3
780 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
782 * NUMA-1 0-1,3 0-2 1-3 0,2-3
783 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
785 * NUMA-0 0 1 2 3
790 * represented multiple times -- hence the "overlap" naming for this part of
794 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
798 * - the first group of each domain is its child domain; this
799 * gets us the first 0-1,3
800 * - the only uncovered node is 2, who's child domain is 1-3.
802 * However, because of the overlap, computing a unique CPU for each group is
803 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
804 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
805 * end up at those groups (they would end up in group: 0-1,3).
820 * node 0 1 2 3
824 * 3: 30 20 20 10
828 * 0 ----- 1
832 * 2 ----- 3
834 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
838 * not of the same number for each CPU. Consider:
840 * NUMA-2 0-3 0-3
841 * groups: {0-2},{1-3} {1-3},{0-2}
843 * NUMA-1 0-2 0-3 0-3 1-3
845 * NUMA-0 0 1 2 3
863 struct sd_data *sdd = sd->private; in build_balance_mask()
870 sibling = *per_cpu_ptr(sdd->sd, i); in build_balance_mask()
877 if (!sibling->child) in build_balance_mask()
881 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child))) in build_balance_mask()
892 * XXX: This creates per-node group entries; since the load-balancer will
893 * immediately access remote memory to construct this group's load-balance
897 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu) in build_group_from_child_sched_domain() argument
903 GFP_KERNEL, cpu_to_node(cpu)); in build_group_from_child_sched_domain()
909 if (sd->child) in build_group_from_child_sched_domain()
910 cpumask_copy(sg_span, sched_domain_span(sd->child)); in build_group_from_child_sched_domain()
914 atomic_inc(&sg->ref); in build_group_from_child_sched_domain()
922 struct sd_data *sdd = sd->private; in init_overlap_sched_group()
924 int cpu; in init_overlap_sched_group() local
927 cpu = cpumask_first_and(sched_group_span(sg), mask); in init_overlap_sched_group()
929 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu); in init_overlap_sched_group()
930 if (atomic_inc_return(&sg->sgc->ref) == 1) in init_overlap_sched_group()
936 * Initialize sgc->capacity such that even if we mess up the in init_overlap_sched_group()
941 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); in init_overlap_sched_group()
942 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE; in init_overlap_sched_group()
943 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE; in init_overlap_sched_group()
953 while (sibling->child && in find_descended_sibling()
954 !cpumask_subset(sched_domain_span(sibling->child), in find_descended_sibling()
956 sibling = sibling->child; in find_descended_sibling()
963 while (sibling->child && in find_descended_sibling()
964 cpumask_equal(sched_domain_span(sibling->child), in find_descended_sibling()
966 sibling = sibling->child; in find_descended_sibling()
972 build_overlap_sched_groups(struct sched_domain *sd, int cpu) in build_overlap_sched_groups() argument
977 struct sd_data *sdd = sd->private; in build_overlap_sched_groups()
983 for_each_cpu_wrap(i, span, cpu) { in build_overlap_sched_groups()
989 sibling = *per_cpu_ptr(sdd->sd, i); in build_overlap_sched_groups()
998 * Domains should always include the CPU they're built on, so in build_overlap_sched_groups()
1006 * But for machines whose NUMA diameter are 3 or above, we move in build_overlap_sched_groups()
1011 * Smallest diameter=3 topology is: in build_overlap_sched_groups()
1013 * node 0 1 2 3 in build_overlap_sched_groups()
1017 * 3: 40 30 20 10 in build_overlap_sched_groups()
1019 * 0 --- 1 --- 2 --- 3 in build_overlap_sched_groups()
1021 * NUMA-3 0-3 N/A N/A 0-3 in build_overlap_sched_groups()
1022 * groups: {0-2},{1-3} {1-3},{0-2} in build_overlap_sched_groups()
1024 * NUMA-2 0-2 0-3 0-3 1-3 in build_overlap_sched_groups()
1025 * groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2} in build_overlap_sched_groups()
1027 * NUMA-1 0-1 0-2 1-3 2-3 in build_overlap_sched_groups()
1028 * groups: {0},{1} {1},{2},{0} {2},{3},{1} {3},{2} in build_overlap_sched_groups()
1030 * NUMA-0 0 1 2 3 in build_overlap_sched_groups()
1032 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the in build_overlap_sched_groups()
1035 if (sibling->child && in build_overlap_sched_groups()
1036 !cpumask_subset(sched_domain_span(sibling->child), span)) in build_overlap_sched_groups()
1039 sg = build_group_from_child_sched_domain(sibling, cpu); in build_overlap_sched_groups()
1051 last->next = sg; in build_overlap_sched_groups()
1053 last->next = first; in build_overlap_sched_groups()
1055 sd->groups = first; in build_overlap_sched_groups()
1062 return -ENOMEM; in build_overlap_sched_groups()
1067 * Package topology (also see the load-balance blurb in fair.c)
1072 * - Simultaneous multithreading (SMT)
1073 * - Multi-Core Cache (MC)
1074 * - Package (DIE)
1078 * The tree consists of 3 primary data structures:
1080 * sched_domain -> sched_group -> sched_group_capacity
1082 * `-' `-'
1084 * The sched_domains are per-CPU and have a two way link (parent & child) and
1090 * CPU of that sched_domain [*].
1094 * CPU 0 1 2 3 4 5 6 7
1100 * - or -
1102 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1103 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1104 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1106 * CPU 0 1 2 3 4 5 6 7
1114 * There are two related construction problems, both require a CPU that
1117 * - The first is the balance_cpu (see should_we_balance() and the
1118 * load-balance blub in fair.c); for each group we only want 1 CPU to
1121 * - The second is the sched_group_capacity; we want all identical groups
1127 * for each CPU in the hierarchy.
1129 * Therefore computing a unique CPU for each group is trivial (the iteration
1131 * group), we can simply pick the first CPU in each group.
1137 static struct sched_group *get_group(int cpu, struct sd_data *sdd) in get_group() argument
1139 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); in get_group()
1140 struct sched_domain *child = sd->child; in get_group()
1145 cpu = cpumask_first(sched_domain_span(child)); in get_group()
1147 sg = *per_cpu_ptr(sdd->sg, cpu); in get_group()
1148 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu); in get_group()
1151 already_visited = atomic_inc_return(&sg->ref) > 1; in get_group()
1153 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1)); in get_group()
1163 cpumask_set_cpu(cpu, sched_group_span(sg)); in get_group()
1164 cpumask_set_cpu(cpu, group_balance_mask(sg)); in get_group()
1167 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg)); in get_group()
1168 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE; in get_group()
1169 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE; in get_group()
1176 * covered by the given span, will set each group's ->cpumask correctly,
1177 * and will initialize their ->sgc.
1182 build_sched_groups(struct sched_domain *sd, int cpu) in build_sched_groups() argument
1185 struct sd_data *sdd = sd->private; in build_sched_groups()
1195 for_each_cpu_wrap(i, span, cpu) { in build_sched_groups()
1208 last->next = sg; in build_sched_groups()
1211 last->next = first; in build_sched_groups()
1212 sd->groups = first; in build_sched_groups()
1227 void init_sched_groups_capacity(int cpu, struct sched_domain *sd) in init_sched_groups_capacity() argument
1229 struct sched_group *sg = sd->groups; in init_sched_groups_capacity()
1237 int cpu, max_cpu = -1; in init_sched_groups_capacity() local
1242 sg->group_weight = cpumask_weight(&avail_mask); in init_sched_groups_capacity()
1244 sg->group_weight = cpumask_weight(sched_group_span(sg)); in init_sched_groups_capacity()
1247 if (!(sd->flags & SD_ASYM_PACKING)) in init_sched_groups_capacity()
1250 for_each_cpu(cpu, sched_group_span(sg)) { in init_sched_groups_capacity()
1252 max_cpu = cpu; in init_sched_groups_capacity()
1253 else if (sched_asym_prefer(cpu, max_cpu)) in init_sched_groups_capacity()
1254 max_cpu = cpu; in init_sched_groups_capacity()
1256 sg->asym_prefer_cpu = max_cpu; in init_sched_groups_capacity()
1259 sg = sg->next; in init_sched_groups_capacity()
1260 } while (sg != sd->groups); in init_sched_groups_capacity()
1262 if (cpu != group_balance_cpu(sg)) in init_sched_groups_capacity()
1265 update_group_capacity(sd, cpu); in init_sched_groups_capacity()
1270 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1273 static int default_relax_domain_level = -1;
1290 if (!attr || attr->relax_domain_level < 0) { in set_domain_attribute()
1295 request = attr->relax_domain_level; in set_domain_attribute()
1297 if (sd->level > request) { in set_domain_attribute()
1299 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); in set_domain_attribute()
1311 if (!atomic_read(&d->rd->refcount)) in __free_domain_allocs()
1312 free_rootdomain(&d->rd->rcu); in __free_domain_allocs()
1315 free_percpu(d->sd); in __free_domain_allocs()
1332 d->sd = alloc_percpu(struct sched_domain *); in __visit_domain_allocation_hell()
1333 if (!d->sd) in __visit_domain_allocation_hell()
1335 d->rd = alloc_rootdomain(); in __visit_domain_allocation_hell()
1336 if (!d->rd) in __visit_domain_allocation_hell()
1347 static void claim_allocations(int cpu, struct sched_domain *sd) in claim_allocations() argument
1349 struct sd_data *sdd = sd->private; in claim_allocations()
1351 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); in claim_allocations()
1352 *per_cpu_ptr(sdd->sd, cpu) = NULL; in claim_allocations()
1354 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref)) in claim_allocations()
1355 *per_cpu_ptr(sdd->sds, cpu) = NULL; in claim_allocations()
1357 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) in claim_allocations()
1358 *per_cpu_ptr(sdd->sg, cpu) = NULL; in claim_allocations()
1360 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) in claim_allocations()
1361 *per_cpu_ptr(sdd->sgc, cpu) = NULL; in claim_allocations()
1383 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1384 * SD_SHARE_PKG_RESOURCES - describes shared caches
1385 * SD_NUMA - describes NUMA topologies
1390 * SD_ASYM_PACKING - describes SMT quirks
1401 struct sched_domain *child, int dflags, int cpu) in sd_init() argument
1403 struct sd_data *sdd = &tl->data; in sd_init()
1404 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); in sd_init()
1411 sched_domains_curr_level = tl->numa_level; in sd_init()
1414 sd_weight = cpumask_weight(tl->mask(cpu)); in sd_init()
1416 if (tl->sd_flags) in sd_init()
1417 sd_flags = (*tl->sd_flags)(); in sd_init()
1452 .name = tl->name, in sd_init()
1456 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); in sd_init()
1464 if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child) in sd_init()
1465 sd->child->flags &= ~SD_PREFER_SIBLING; in sd_init()
1467 if (sd->flags & SD_SHARE_CPUCAPACITY) { in sd_init()
1468 sd->imbalance_pct = 110; in sd_init()
1470 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { in sd_init()
1471 sd->imbalance_pct = 117; in sd_init()
1472 sd->cache_nice_tries = 1; in sd_init()
1475 } else if (sd->flags & SD_NUMA) { in sd_init()
1476 sd->cache_nice_tries = 2; in sd_init()
1478 sd->flags &= ~SD_PREFER_SIBLING; in sd_init()
1479 sd->flags |= SD_SERIALIZE; in sd_init()
1480 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) { in sd_init()
1481 sd->flags &= ~(SD_BALANCE_EXEC | in sd_init()
1488 sd->cache_nice_tries = 1; in sd_init()
1495 if (sd->flags & SD_SHARE_PKG_RESOURCES) { in sd_init()
1496 sd->shared = *per_cpu_ptr(sdd->sds, sd_id); in sd_init()
1497 atomic_inc(&sd->shared->ref); in sd_init()
1498 atomic_set(&sd->shared->nr_busy_cpus, sd_weight); in sd_init()
1501 sd->private = sdd; in sd_init()
1507 * Topology list, bottom-up.
1524 for (tl = sched_domain_topology; tl->mask; tl++)
1536 static const struct cpumask *sd_numa_mask(int cpu) in sd_numa_mask() argument
1538 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; in sd_numa_mask()
1590 * - If the maximum distance between any nodes is 1 hop, the system
1592 * - If for two nodes A and B, located N > 1 hops away from each other,
1639 * O(nr_nodes^2) deduplicating selection sort -- in order to find the in sched_init_numa()
1721 sched_numa_warn("Node-distance not symmetric"); in sched_init_numa()
1770 sched_max_numa_distance = sched_domains_numa_distance[nr_levels - 1]; in sched_init_numa()
1775 void sched_domains_numa_masks_set(unsigned int cpu) in sched_domains_numa_masks_set() argument
1777 int node = cpu_to_node(cpu); in sched_domains_numa_masks_set()
1783 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); in sched_domains_numa_masks_set()
1788 void sched_domains_numa_masks_clear(unsigned int cpu) in sched_domains_numa_masks_clear() argument
1794 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); in sched_domains_numa_masks_clear()
1799 * sched_numa_find_closest() - given the NUMA topology, find the cpu
1800 * closest to @cpu from @cpumask.
1801 * cpumask: cpumask to find a cpu from
1802 * cpu: cpu to be close to
1804 * returns: cpu, or nr_cpu_ids when nothing found.
1806 int sched_numa_find_closest(const struct cpumask *cpus, int cpu) in sched_numa_find_closest() argument
1808 int i, j = cpu_to_node(cpu); in sched_numa_find_closest()
1811 cpu = cpumask_any_and(cpus, sched_domains_numa_masks[i][j]); in sched_numa_find_closest()
1812 if (cpu < nr_cpu_ids) in sched_numa_find_closest()
1813 return cpu; in sched_numa_find_closest()
1826 struct sd_data *sdd = &tl->data; in __sdt_alloc()
1828 sdd->sd = alloc_percpu(struct sched_domain *); in __sdt_alloc()
1829 if (!sdd->sd) in __sdt_alloc()
1830 return -ENOMEM; in __sdt_alloc()
1832 sdd->sds = alloc_percpu(struct sched_domain_shared *); in __sdt_alloc()
1833 if (!sdd->sds) in __sdt_alloc()
1834 return -ENOMEM; in __sdt_alloc()
1836 sdd->sg = alloc_percpu(struct sched_group *); in __sdt_alloc()
1837 if (!sdd->sg) in __sdt_alloc()
1838 return -ENOMEM; in __sdt_alloc()
1840 sdd->sgc = alloc_percpu(struct sched_group_capacity *); in __sdt_alloc()
1841 if (!sdd->sgc) in __sdt_alloc()
1842 return -ENOMEM; in __sdt_alloc()
1853 return -ENOMEM; in __sdt_alloc()
1855 *per_cpu_ptr(sdd->sd, j) = sd; in __sdt_alloc()
1860 return -ENOMEM; in __sdt_alloc()
1862 *per_cpu_ptr(sdd->sds, j) = sds; in __sdt_alloc()
1867 return -ENOMEM; in __sdt_alloc()
1869 sg->next = sg; in __sdt_alloc()
1871 *per_cpu_ptr(sdd->sg, j) = sg; in __sdt_alloc()
1876 return -ENOMEM; in __sdt_alloc()
1879 sgc->id = j; in __sdt_alloc()
1882 *per_cpu_ptr(sdd->sgc, j) = sgc; in __sdt_alloc()
1895 struct sd_data *sdd = &tl->data; in __sdt_free()
1900 if (sdd->sd) { in __sdt_free()
1901 sd = *per_cpu_ptr(sdd->sd, j); in __sdt_free()
1902 if (sd && (sd->flags & SD_OVERLAP)) in __sdt_free()
1903 free_sched_groups(sd->groups, 0); in __sdt_free()
1904 kfree(*per_cpu_ptr(sdd->sd, j)); in __sdt_free()
1907 if (sdd->sds) in __sdt_free()
1908 kfree(*per_cpu_ptr(sdd->sds, j)); in __sdt_free()
1909 if (sdd->sg) in __sdt_free()
1910 kfree(*per_cpu_ptr(sdd->sg, j)); in __sdt_free()
1911 if (sdd->sgc) in __sdt_free()
1912 kfree(*per_cpu_ptr(sdd->sgc, j)); in __sdt_free()
1914 free_percpu(sdd->sd); in __sdt_free()
1915 sdd->sd = NULL; in __sdt_free()
1916 free_percpu(sdd->sds); in __sdt_free()
1917 sdd->sds = NULL; in __sdt_free()
1918 free_percpu(sdd->sg); in __sdt_free()
1919 sdd->sg = NULL; in __sdt_free()
1920 free_percpu(sdd->sgc); in __sdt_free()
1921 sdd->sgc = NULL; in __sdt_free()
1927 struct sched_domain *child, int dflags, int cpu) in build_sched_domain() argument
1929 struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu); in build_sched_domain()
1932 sd->level = child->level + 1; in build_sched_domain()
1933 sched_domain_level_max = max(sched_domain_level_max, sd->level); in build_sched_domain()
1934 child->parent = sd; in build_sched_domain()
1941 child->name, sd->name); in build_sched_domain()
1957 * any two given CPUs at this (non-NUMA) topology level.
1960 const struct cpumask *cpu_map, int cpu) in topology_span_sane() argument
1965 if (tl->flags & SDTL_OVERLAP) in topology_span_sane()
1969 * Non-NUMA levels cannot partially overlap - they must be either in topology_span_sane()
1971 * breaking the sched_group lists - i.e. a later get_group() pass in topology_span_sane()
1975 if (i == cpu) in topology_span_sane()
1983 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) && in topology_span_sane()
1984 cpumask_intersects(tl->mask(cpu), tl->mask(i))) in topology_span_sane()
1992 * Find the sched_domain_topology_level where all CPU capacities are visible
2017 * Examine topology from all CPU's point of views to detect the lowest in asym_cpu_capacity_level()
2018 * sched_domain_topology_level where a highest capacity CPU is visible in asym_cpu_capacity_level()
2029 for_each_cpu_and(j, tl->mask(i), cpu_map) { in asym_cpu_capacity_level()
2061 int i, ret = -ENOMEM; in build_sched_domains()
2093 if (tl->flags & SDTL_OVERLAP) in build_sched_domains()
2094 sd->flags |= SD_OVERLAP; in build_sched_domains()
2102 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { in build_sched_domains()
2103 sd->span_weight = cpumask_weight(sched_domain_span(sd)); in build_sched_domains()
2104 if (sd->flags & SD_OVERLAP) { in build_sched_domains()
2114 /* Calculate CPU capacity for physical packages and nodes */ in build_sched_domains()
2115 for (i = nr_cpumask_bits-1; i >= 0; i--) { in build_sched_domains()
2119 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { in build_sched_domains()
2129 int max_cpu = READ_ONCE(d.rd->max_cap_orig_cpu); in build_sched_domains()
2138 WRITE_ONCE(d.rd->max_cap_orig_cpu, i); in build_sched_domains()
2142 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity)) in build_sched_domains()
2143 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig); in build_sched_domains()
2154 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity); in build_sched_domains()
2182 * CPU core maps. It is supposed to return 1 if the topology changed
2245 unsigned int cpu = cpumask_any(cpu_map); in detach_destroy_domains() local
2248 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu))) in detach_destroy_domains()
2312 /* Let the architecture update CPU core mappings: */ in partition_sched_domains_locked()
2337 * its dl_bw->total_bw needs to be cleared. It in partition_sched_domains_locked()
2341 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd; in partition_sched_domains_locked()
2346 /* No match - a current sched domain not in new doms_new[] */ in partition_sched_domains_locked()
2367 /* No match - add a new doms_new */ in partition_sched_domains_locked()
2378 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) { in partition_sched_domains_locked()
2383 /* No match - add perf. domains for a new rd */ in partition_sched_domains_locked()