Lines Matching +full:child +full:- +full:node
1 // SPDX-License-Identifier: GPL-2.0
46 struct sched_group *group = sd->groups; in sched_domain_debug_one()
47 unsigned long flags = sd->flags; in sched_domain_debug_one()
52 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level); in sched_domain_debug_one()
54 cpumask_pr_args(sched_domain_span(sd)), sd->name); in sched_domain_debug_one()
57 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); in sched_domain_debug_one()
60 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); in sched_domain_debug_one()
67 if ((meta_flags & SDF_SHARED_CHILD) && sd->child && in sched_domain_debug_one()
68 !(sd->child->flags & flag)) in sched_domain_debug_one()
69 printk(KERN_ERR "ERROR: flag %s set here but not in child\n", in sched_domain_debug_one()
72 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent && in sched_domain_debug_one()
73 !(sd->parent->flags & flag)) in sched_domain_debug_one()
92 if (!(sd->flags & SD_NUMA) && in sched_domain_debug_one()
102 group->sgc->id, in sched_domain_debug_one()
105 if ((sd->flags & SD_NUMA) && in sched_domain_debug_one()
111 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) in sched_domain_debug_one()
112 printk(KERN_CONT " cap=%lu", group->sgc->capacity); in sched_domain_debug_one()
114 if (group == sd->groups && sd->child && in sched_domain_debug_one()
115 !cpumask_equal(sched_domain_span(sd->child), in sched_domain_debug_one()
117 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n"); in sched_domain_debug_one()
122 group = group->next; in sched_domain_debug_one()
124 if (group != sd->groups) in sched_domain_debug_one()
127 } while (group != sd->groups); in sched_domain_debug_one()
131 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); in sched_domain_debug_one()
133 if (sd->parent && in sched_domain_debug_one()
134 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) in sched_domain_debug_one()
135 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); in sched_domain_debug_one()
147 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); in sched_domain_debug()
151 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu); in sched_domain_debug()
157 sd = sd->parent; in sched_domain_debug()
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()
245 pr_info("rd %*pbl: Checking EAS: frequency-invariant load tracking not yet supported", in sched_is_eas_possible()
278 return -EPERM; in sched_energy_aware_handler()
282 return -EOPNOTSUPP; in sched_energy_aware_handler()
325 tmp = pd->next; in free_pd()
336 pd = pd->next; in find_pd()
356 pd->em_pd = obj; in pd_init()
373 em_pd_nr_perf_states(pd->em_pd)); in perf_domain_debug()
374 pd = pd->next; in perf_domain_debug()
414 struct root_domain *rd = cpu_rq(cpu)->rd; in build_perf_domains()
431 tmp->next = pd; in build_perf_domains()
438 tmp = rd->pd; in build_perf_domains()
439 rcu_assign_pointer(rd->pd, pd); in build_perf_domains()
441 call_rcu(&tmp->rcu, destroy_perf_domain_rcu); in build_perf_domains()
447 tmp = rd->pd; in build_perf_domains()
448 rcu_assign_pointer(rd->pd, NULL); in build_perf_domains()
450 call_rcu(&tmp->rcu, destroy_perf_domain_rcu); in build_perf_domains()
462 cpupri_cleanup(&rd->cpupri); in free_rootdomain()
463 cpudl_cleanup(&rd->cpudl); in free_rootdomain()
464 free_cpumask_var(rd->dlo_mask); in free_rootdomain()
465 free_cpumask_var(rd->rto_mask); in free_rootdomain()
466 free_cpumask_var(rd->online); in free_rootdomain()
467 free_cpumask_var(rd->span); in free_rootdomain()
468 free_pd(rd->pd); in free_rootdomain()
479 if (rq->rd) { in rq_attach_root()
480 old_rd = rq->rd; in rq_attach_root()
482 if (cpumask_test_cpu(rq->cpu, old_rd->online)) in rq_attach_root()
485 cpumask_clear_cpu(rq->cpu, old_rd->span); in rq_attach_root()
492 if (!atomic_dec_and_test(&old_rd->refcount)) in rq_attach_root()
496 atomic_inc(&rd->refcount); in rq_attach_root()
497 rq->rd = rd; in rq_attach_root()
499 cpumask_set_cpu(rq->cpu, rd->span); in rq_attach_root()
500 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) in rq_attach_root()
508 if (rq->fair_server.dl_server) in rq_attach_root()
509 __dl_server_attach_root(&rq->fair_server, rq); in rq_attach_root()
514 call_rcu(&old_rd->rcu, free_rootdomain); in rq_attach_root()
519 atomic_inc(&rd->refcount); in sched_get_rd()
524 if (!atomic_dec_and_test(&rd->refcount)) in sched_put_rd()
527 call_rcu(&rd->rcu, free_rootdomain); in sched_put_rd()
532 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL)) in init_rootdomain()
534 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL)) in init_rootdomain()
536 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) in init_rootdomain()
538 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) in init_rootdomain()
542 rd->rto_cpu = -1; in init_rootdomain()
543 raw_spin_lock_init(&rd->rto_lock); in init_rootdomain()
544 rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func); in init_rootdomain()
547 rd->visit_cookie = 0; in init_rootdomain()
548 init_dl_bw(&rd->dl_bw); in init_rootdomain()
549 if (cpudl_init(&rd->cpudl) != 0) in init_rootdomain()
552 if (cpupri_init(&rd->cpupri) != 0) in init_rootdomain()
557 cpudl_cleanup(&rd->cpudl); in init_rootdomain()
559 free_cpumask_var(rd->rto_mask); in init_rootdomain()
561 free_cpumask_var(rd->dlo_mask); in init_rootdomain()
563 free_cpumask_var(rd->online); in init_rootdomain()
565 free_cpumask_var(rd->span); in init_rootdomain()
567 return -ENOMEM; in init_rootdomain()
571 * By default the system creates a single root-domain with all CPUs as
608 tmp = sg->next; in free_sched_groups()
610 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) in free_sched_groups()
611 kfree(sg->sgc); in free_sched_groups()
613 if (atomic_dec_and_test(&sg->ref)) in free_sched_groups()
626 free_sched_groups(sd->groups, 1); in destroy_sched_domain()
628 if (sd->shared && atomic_dec_and_test(&sd->shared->ref)) in destroy_sched_domain()
629 kfree(sd->shared); in destroy_sched_domain()
638 struct sched_domain *parent = sd->parent; in destroy_sched_domains_rcu()
647 call_rcu(&sd->rcu, destroy_sched_domains_rcu); in destroy_sched_domains()
682 sds = sd->shared; in update_top_cache_domain()
697 * but equals to LLC id on non-Cluster machines. in update_top_cache_domain()
723 struct sched_domain *parent = tmp->parent; in cpu_attach_domain()
728 tmp->parent = parent->parent; in cpu_attach_domain()
730 if (parent->parent) { in cpu_attach_domain()
731 parent->parent->child = tmp; in cpu_attach_domain()
732 parent->parent->groups->flags = tmp->flags; in cpu_attach_domain()
740 if (parent->flags & SD_PREFER_SIBLING) in cpu_attach_domain()
741 tmp->flags |= SD_PREFER_SIBLING; in cpu_attach_domain()
744 tmp = tmp->parent; in cpu_attach_domain()
749 sd = sd->parent; in cpu_attach_domain()
752 struct sched_group *sg = sd->groups; in cpu_attach_domain()
755 * sched groups hold the flags of the child sched in cpu_attach_domain()
757 * the child is being destroyed. in cpu_attach_domain()
760 sg->flags = 0; in cpu_attach_domain()
761 } while (sg != sd->groups); in cpu_attach_domain()
763 sd->child = NULL; in cpu_attach_domain()
770 tmp = rq->sd; in cpu_attach_domain()
771 rcu_assign_pointer(rq->sd, sd); in cpu_attach_domain()
808 * Given a node-distance table, for example:
810 * node 0 1 2 3
816 * which represents a 4 node ring topology like:
818 * 0 ----- 1
822 * 3 ----- 2
830 * NUMA-2 0-3 0-3 0-3 0-3
831 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
833 * NUMA-1 0-1,3 0-2 1-3 0,2-3
836 * NUMA-0 0 1 2 3
840 * When we iterate a domain in child domain chunks some nodes can be
841 * represented multiple times -- hence the "overlap" naming for this part of
845 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
849 * - the first group of each domain is its child domain; this
850 * gets us the first 0-1,3
851 * - the only uncovered node is 2, who's child domain is 1-3.
854 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
855 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
856 * end up at those groups (they would end up in group: 0-1,3).
860 * (child) domain tree.
871 * node 0 1 2 3
879 * 0 ----- 1
883 * 2 ----- 3
891 * NUMA-2 0-3 0-3
892 * groups: {0-2},{1-3} {1-3},{0-2}
894 * NUMA-1 0-2 0-3 0-3 1-3
896 * NUMA-0 0 1 2 3
906 * isn't complete yet, however since each group represents a (child) domain we
914 struct sd_data *sdd = sd->private; in build_balance_mask()
921 sibling = *per_cpu_ptr(sdd->sd, i); in build_balance_mask()
928 if (!sibling->child) in build_balance_mask()
932 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child))) in build_balance_mask()
943 * XXX: This creates per-node group entries; since the load-balancer will
944 * immediately access remote memory to construct this group's load-balance
945 * statistics having the groups node local is of dubious benefit.
960 if (sd->child) { in build_group_from_child_sched_domain()
961 cpumask_copy(sg_span, sched_domain_span(sd->child)); in build_group_from_child_sched_domain()
962 sg->flags = sd->child->flags; in build_group_from_child_sched_domain()
967 atomic_inc(&sg->ref); in build_group_from_child_sched_domain()
975 struct sd_data *sdd = sd->private; in init_overlap_sched_group()
982 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu); in init_overlap_sched_group()
983 if (atomic_inc_return(&sg->sgc->ref) == 1) in init_overlap_sched_group()
989 * Initialize sgc->capacity such that even if we mess up the in init_overlap_sched_group()
994 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); in init_overlap_sched_group()
995 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE; in init_overlap_sched_group()
996 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE; in init_overlap_sched_group()
1003 * The proper descendant would be the one whose child won't span out in find_descended_sibling()
1006 while (sibling->child && in find_descended_sibling()
1007 !cpumask_subset(sched_domain_span(sibling->child), in find_descended_sibling()
1009 sibling = sibling->child; in find_descended_sibling()
1016 while (sibling->child && in find_descended_sibling()
1017 cpumask_equal(sched_domain_span(sibling->child), in find_descended_sibling()
1019 sibling = sibling->child; in find_descended_sibling()
1030 struct sd_data *sdd = sd->private; in build_overlap_sched_groups()
1042 sibling = *per_cpu_ptr(sdd->sd, i); in build_overlap_sched_groups()
1045 * Asymmetric node setups can result in situations where the in build_overlap_sched_groups()
1058 * Usually we build sched_group by sibling's child sched_domain in build_overlap_sched_groups()
1060 * to build sched_group by sibling's proper descendant's child in build_overlap_sched_groups()
1061 * domain because sibling's child sched_domain will span out of in build_overlap_sched_groups()
1066 * node 0 1 2 3 in build_overlap_sched_groups()
1072 * 0 --- 1 --- 2 --- 3 in build_overlap_sched_groups()
1074 * NUMA-3 0-3 N/A N/A 0-3 in build_overlap_sched_groups()
1075 * groups: {0-2},{1-3} {1-3},{0-2} in build_overlap_sched_groups()
1077 * NUMA-2 0-2 0-3 0-3 1-3 in build_overlap_sched_groups()
1078 * groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2} in build_overlap_sched_groups()
1080 * NUMA-1 0-1 0-2 1-3 2-3 in build_overlap_sched_groups()
1083 * NUMA-0 0 1 2 3 in build_overlap_sched_groups()
1085 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the in build_overlap_sched_groups()
1088 if (sibling->child && in build_overlap_sched_groups()
1089 !cpumask_subset(sched_domain_span(sibling->child), span)) in build_overlap_sched_groups()
1104 last->next = sg; in build_overlap_sched_groups()
1106 last->next = first; in build_overlap_sched_groups()
1108 sd->groups = first; in build_overlap_sched_groups()
1115 return -ENOMEM; in build_overlap_sched_groups()
1120 * Package topology (also see the load-balance blurb in fair.c)
1125 * - Simultaneous multithreading (SMT)
1126 * - Multi-Core Cache (MC)
1127 * - Package (PKG)
1129 * Where the last one more or less denotes everything up to a NUMA node.
1133 * sched_domain -> sched_group -> sched_group_capacity
1135 * `-' `-'
1137 * The sched_domains are per-CPU and have a two way link (parent & child) and
1153 * - or -
1155 * PKG 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1156 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1157 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1162 * topology levels, while sched_group moves you sideways through it, at child
1170 * - The first is the balance_cpu (see should_we_balance() and the
1171 * load-balance blurb in fair.c); for each group we only want 1 CPU to
1174 * - The second is the sched_group_capacity; we want all identical groups
1187 * [*] in other words, the first group of each domain is its child domain.
1192 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); in get_group()
1193 struct sched_domain *child = sd->child; in get_group() local
1197 if (child) in get_group()
1198 cpu = cpumask_first(sched_domain_span(child)); in get_group()
1200 sg = *per_cpu_ptr(sdd->sg, cpu); in get_group()
1201 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu); in get_group()
1204 already_visited = atomic_inc_return(&sg->ref) > 1; in get_group()
1206 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1)); in get_group()
1212 if (child) { in get_group()
1213 cpumask_copy(sched_group_span(sg), sched_domain_span(child)); in get_group()
1215 sg->flags = child->flags; in get_group()
1221 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg)); in get_group()
1222 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE; in get_group()
1223 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE; in get_group()
1230 * covered by the given span, will set each group's ->cpumask correctly,
1231 * and will initialize their ->sgc.
1239 struct sd_data *sdd = sd->private; in build_sched_groups()
1262 last->next = sg; in build_sched_groups()
1265 last->next = first; in build_sched_groups()
1266 sd->groups = first; in build_sched_groups()
1283 struct sched_group *sg = sd->groups; in init_sched_groups_capacity()
1289 int cpu, cores = 0, max_cpu = -1; in init_sched_groups_capacity()
1291 sg->group_weight = cpumask_weight(sched_group_span(sg)); in init_sched_groups_capacity()
1300 sg->cores = cores; in init_sched_groups_capacity()
1302 if (!(sd->flags & SD_ASYM_PACKING)) in init_sched_groups_capacity()
1311 sg->asym_prefer_cpu = max_cpu; in init_sched_groups_capacity()
1314 sg = sg->next; in init_sched_groups_capacity()
1315 } while (sg != sd->groups); in init_sched_groups_capacity()
1335 if (!(sd->flags & SD_ASYM_PACKING)) in sched_update_asym_prefer_cpu()
1340 * node and will require updating "asym_prefer_cpu" on in sched_update_asym_prefer_cpu()
1344 * "sg->asym_prefer_cpu" to "sg->sgc->asym_prefer_cpu" in sched_update_asym_prefer_cpu()
1347 WARN_ON_ONCE(sd->flags & SD_NUMA); in sched_update_asym_prefer_cpu()
1349 sg = sd->groups; in sched_update_asym_prefer_cpu()
1350 if (cpu != sg->asym_prefer_cpu) { in sched_update_asym_prefer_cpu()
1357 if (!sched_asym_prefer(cpu, sg->asym_prefer_cpu)) in sched_update_asym_prefer_cpu()
1360 WRITE_ONCE(sg->asym_prefer_cpu, cpu); in sched_update_asym_prefer_cpu()
1373 WRITE_ONCE(sg->asym_prefer_cpu, asym_prefer_cpu); in sched_update_asym_prefer_cpu()
1440 if (capacity == entry->capacity) in asym_cpu_capacity_update_data()
1442 else if (!insert_entry && capacity > entry->capacity) in asym_cpu_capacity_update_data()
1449 entry->capacity = capacity; in asym_cpu_capacity_update_data()
1453 list_add_tail_rcu(&entry->link, &asym_cap_list); in asym_cpu_capacity_update_data()
1455 list_add_rcu(&entry->link, &insert_entry->link); in asym_cpu_capacity_update_data()
1461 * Build-up/update list of CPUs grouped by their capacities
1478 list_del_rcu(&entry->link); in asym_cpu_capacity_scan()
1479 call_rcu(&entry->rcu, free_asym_cap_entry); in asym_cpu_capacity_scan()
1489 list_del_rcu(&entry->link); in asym_cpu_capacity_scan()
1490 call_rcu(&entry->rcu, free_asym_cap_entry); in asym_cpu_capacity_scan()
1496 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1499 static int default_relax_domain_level = -1;
1516 if (!attr || attr->relax_domain_level < 0) { in set_domain_attribute()
1521 request = attr->relax_domain_level; in set_domain_attribute()
1523 if (sd->level >= request) { in set_domain_attribute()
1525 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); in set_domain_attribute()
1537 if (!atomic_read(&d->rd->refcount)) in __free_domain_allocs()
1538 free_rootdomain(&d->rd->rcu); in __free_domain_allocs()
1541 free_percpu(d->sd); in __free_domain_allocs()
1558 d->sd = alloc_percpu(struct sched_domain *); in __visit_domain_allocation_hell()
1559 if (!d->sd) in __visit_domain_allocation_hell()
1561 d->rd = alloc_rootdomain(); in __visit_domain_allocation_hell()
1562 if (!d->rd) in __visit_domain_allocation_hell()
1575 struct sd_data *sdd = sd->private; in claim_allocations()
1577 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); in claim_allocations()
1578 *per_cpu_ptr(sdd->sd, cpu) = NULL; in claim_allocations()
1580 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref)) in claim_allocations()
1581 *per_cpu_ptr(sdd->sds, cpu) = NULL; in claim_allocations()
1583 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) in claim_allocations()
1584 *per_cpu_ptr(sdd->sg, cpu) = NULL; in claim_allocations()
1586 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) in claim_allocations()
1587 *per_cpu_ptr(sdd->sgc, cpu) = NULL; in claim_allocations()
1615 * SD_ASYM_PACKING - describes SMT quirks
1627 struct sched_domain *child, int cpu) in sd_init() argument
1629 struct sd_data *sdd = &tl->data; in sd_init()
1630 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); in sd_init()
1634 sd_weight = cpumask_weight(tl->mask(tl, cpu)); in sd_init()
1636 if (tl->sd_flags) in sd_init()
1637 sd_flags = (*tl->sd_flags)(); in sd_init()
1667 .child = child, in sd_init()
1668 .name = tl->name, in sd_init()
1672 cpumask_and(sd_span, cpu_map, tl->mask(tl, cpu)); in sd_init()
1675 sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map); in sd_init()
1677 WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) == in sd_init()
1685 if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child) in sd_init()
1686 sd->child->flags &= ~SD_PREFER_SIBLING; in sd_init()
1688 if (sd->flags & SD_SHARE_CPUCAPACITY) { in sd_init()
1689 sd->imbalance_pct = 110; in sd_init()
1691 } else if (sd->flags & SD_SHARE_LLC) { in sd_init()
1692 sd->imbalance_pct = 117; in sd_init()
1693 sd->cache_nice_tries = 1; in sd_init()
1696 } else if (sd->flags & SD_NUMA) { in sd_init()
1697 sd->cache_nice_tries = 2; in sd_init()
1699 sd->flags &= ~SD_PREFER_SIBLING; in sd_init()
1700 sd->flags |= SD_SERIALIZE; in sd_init()
1701 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) { in sd_init()
1702 sd->flags &= ~(SD_BALANCE_EXEC | in sd_init()
1709 sd->cache_nice_tries = 1; in sd_init()
1716 if (sd->flags & SD_SHARE_LLC) { in sd_init()
1717 sd->shared = *per_cpu_ptr(sdd->sds, sd_id); in sd_init()
1718 atomic_inc(&sd->shared->ref); in sd_init()
1719 atomic_set(&sd->shared->nr_busy_cpus, sd_weight); in sd_init()
1722 sd->private = sdd; in sd_init()
1769 * Topology list, bottom-up.
1792 for (tl = sched_domain_topology; tl->mask; tl++)
1811 return sched_domains_numa_masks[tl->numa_level][cpu_to_node(cpu)]; in sd_numa_mask()
1882 * - If the maximum distance between any nodes is 1 hop, the system
1884 * - If for two nodes A and B, located N > 1 hops away from each other,
1885 * there is an intermediary node C, which is < N hops away from both
1905 /* Is there an intermediary node between a and b? */ in init_numa_topology_type()
1937 * O(nr_nodes^2) de-duplicating selection sort -- in order to find the in sched_init_numa()
2001 * Now for each level, construct a mask per node which contains all in sched_init_numa()
2020 sched_numa_warn("Node-distance not symmetric"); in sched_init_numa()
2046 * Add the NUMA identity distance, aka single NODE. in sched_init_numa()
2048 tl[i++] = SDTL_INIT(sd_numa_mask, NULL, NODE); in sched_init_numa()
2062 WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]); in sched_init_numa()
2107 int node; in sched_update_numa() local
2109 node = cpu_to_node(cpu); in sched_update_numa()
2112 * node is onlined or the last CPU of a node is offlined. in sched_update_numa()
2114 if (cpumask_weight(cpumask_of_node(node)) != 1) in sched_update_numa()
2118 sched_init_numa(online ? NUMA_NO_NODE : node); in sched_update_numa()
2123 int node = cpu_to_node(cpu); in sched_domains_numa_masks_set() local
2131 /* Set ourselves in the remote node's masks */ in sched_domains_numa_masks_set()
2132 if (node_distance(j, node) <= sched_domains_numa_distance[i]) in sched_domains_numa_masks_set()
2151 * sched_numa_find_closest() - given the NUMA topology, find the cpu
2185 int node; member
2195 if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu) in hop_cmp()
2198 if (b == k->masks) { in hop_cmp()
2199 k->w = 0; in hop_cmp()
2203 prev_hop = *((struct cpumask ***)b - 1); in hop_cmp()
2204 k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]); in hop_cmp()
2205 if (k->w <= k->cpu) in hop_cmp()
2208 return -1; in hop_cmp()
2212 * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth closest CPU
2214 * from a given @node.
2217 * @node: NUMA node to order CPUs by distance
2221 int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node) in sched_numa_find_nth_cpu() argument
2227 if (node == NUMA_NO_NODE) in sched_numa_find_nth_cpu()
2232 /* CPU-less node entries are uninitialized in sched_domains_numa_masks */ in sched_numa_find_nth_cpu()
2233 node = numa_nearest_node(node, N_CPU); in sched_numa_find_nth_cpu()
2234 k.node = node; in sched_numa_find_nth_cpu()
2243 hop = hop_masks - k.masks; in sched_numa_find_nth_cpu()
2246 cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) : in sched_numa_find_nth_cpu()
2247 cpumask_nth_and(cpu, cpus, k.masks[0][node]); in sched_numa_find_nth_cpu()
2255 * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2256 * @node
2257 * @node: The node to count hops from.
2258 * @hops: Include CPUs up to that many hops away. 0 means local node.
2261 * @node, an error value otherwise.
2264 * read-side section, copy it if required beyond that.
2271 const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops) in sched_numa_hop_mask() argument
2275 if (node >= nr_node_ids || hops >= sched_domains_numa_levels) in sched_numa_hop_mask()
2276 return ERR_PTR(-EINVAL); in sched_numa_hop_mask()
2280 return ERR_PTR(-EBUSY); in sched_numa_hop_mask()
2282 return masks[hops][node]; in sched_numa_hop_mask()
2294 struct sd_data *sdd = &tl->data; in __sdt_alloc()
2296 sdd->sd = alloc_percpu(struct sched_domain *); in __sdt_alloc()
2297 if (!sdd->sd) in __sdt_alloc()
2298 return -ENOMEM; in __sdt_alloc()
2300 sdd->sds = alloc_percpu(struct sched_domain_shared *); in __sdt_alloc()
2301 if (!sdd->sds) in __sdt_alloc()
2302 return -ENOMEM; in __sdt_alloc()
2304 sdd->sg = alloc_percpu(struct sched_group *); in __sdt_alloc()
2305 if (!sdd->sg) in __sdt_alloc()
2306 return -ENOMEM; in __sdt_alloc()
2308 sdd->sgc = alloc_percpu(struct sched_group_capacity *); in __sdt_alloc()
2309 if (!sdd->sgc) in __sdt_alloc()
2310 return -ENOMEM; in __sdt_alloc()
2321 return -ENOMEM; in __sdt_alloc()
2323 *per_cpu_ptr(sdd->sd, j) = sd; in __sdt_alloc()
2328 return -ENOMEM; in __sdt_alloc()
2330 *per_cpu_ptr(sdd->sds, j) = sds; in __sdt_alloc()
2335 return -ENOMEM; in __sdt_alloc()
2337 sg->next = sg; in __sdt_alloc()
2339 *per_cpu_ptr(sdd->sg, j) = sg; in __sdt_alloc()
2344 return -ENOMEM; in __sdt_alloc()
2346 sgc->id = j; in __sdt_alloc()
2348 *per_cpu_ptr(sdd->sgc, j) = sgc; in __sdt_alloc()
2361 struct sd_data *sdd = &tl->data; in __sdt_free()
2366 if (sdd->sd) { in __sdt_free()
2367 sd = *per_cpu_ptr(sdd->sd, j); in __sdt_free()
2368 if (sd && (sd->flags & SD_NUMA)) in __sdt_free()
2369 free_sched_groups(sd->groups, 0); in __sdt_free()
2370 kfree(*per_cpu_ptr(sdd->sd, j)); in __sdt_free()
2373 if (sdd->sds) in __sdt_free()
2374 kfree(*per_cpu_ptr(sdd->sds, j)); in __sdt_free()
2375 if (sdd->sg) in __sdt_free()
2376 kfree(*per_cpu_ptr(sdd->sg, j)); in __sdt_free()
2377 if (sdd->sgc) in __sdt_free()
2378 kfree(*per_cpu_ptr(sdd->sgc, j)); in __sdt_free()
2380 free_percpu(sdd->sd); in __sdt_free()
2381 sdd->sd = NULL; in __sdt_free()
2382 free_percpu(sdd->sds); in __sdt_free()
2383 sdd->sds = NULL; in __sdt_free()
2384 free_percpu(sdd->sg); in __sdt_free()
2385 sdd->sg = NULL; in __sdt_free()
2386 free_percpu(sdd->sgc); in __sdt_free()
2387 sdd->sgc = NULL; in __sdt_free()
2393 struct sched_domain *child, int cpu) in build_sched_domain() argument
2395 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu); in build_sched_domain()
2397 if (child) { in build_sched_domain()
2398 sd->level = child->level + 1; in build_sched_domain()
2399 sched_domain_level_max = max(sched_domain_level_max, sd->level); in build_sched_domain()
2400 child->parent = sd; in build_sched_domain()
2402 if (!cpumask_subset(sched_domain_span(child), in build_sched_domain()
2406 child->name, sd->name); in build_sched_domain()
2407 /* Fixup, ensure @sd has at least @child CPUs. */ in build_sched_domain()
2410 sched_domain_span(child)); in build_sched_domain()
2421 * any two given CPUs on non-NUMA topology levels.
2436 if (tl->sd_flags) in topology_span_sane()
2437 tl_common_flags = (*tl->sd_flags)(); in topology_span_sane()
2447 * Non-NUMA levels cannot partially overlap - they must be either in topology_span_sane()
2449 * breaking the sched_group lists - i.e. a later get_group() pass in topology_span_sane()
2453 const struct cpumask *tl_cpu_mask = tl->mask(tl, cpu); in topology_span_sane()
2461 if (!cpumask_equal(tl->mask(tl, id), tl_cpu_mask)) in topology_span_sane()
2487 int i, ret = -ENOMEM; in build_sched_domains()
2507 has_asym |= sd->flags & SD_ASYM_CPUCAPACITY; in build_sched_domains()
2521 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { in build_sched_domains()
2522 sd->span_weight = cpumask_weight(sched_domain_span(sd)); in build_sched_domains()
2523 if (sd->flags & SD_NUMA) { in build_sched_domains()
2541 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { in build_sched_domains()
2542 struct sched_domain *child = sd->child; in build_sched_domains() local
2544 if (!(sd->flags & SD_SHARE_LLC) && child && in build_sched_domains()
2545 (child->flags & SD_SHARE_LLC)) { in build_sched_domains()
2550 * For a single LLC per node, allow an in build_sched_domains()
2551 * imbalance up to 12.5% of the node. This is in build_sched_domains()
2552 * arbitrary cutoff based two factors -- SMT and in build_sched_domains()
2553 * memory channels. For SMT-2, the intent is to in build_sched_domains()
2555 * SMT-4 or SMT-8 *may* benefit from a different in build_sched_domains()
2563 * on one node while LLCs on another node in build_sched_domains()
2569 nr_llcs = sd->span_weight / child->span_weight; in build_sched_domains()
2571 imb = sd->span_weight >> 3; in build_sched_domains()
2575 sd->imb_numa_nr = imb; in build_sched_domains()
2578 top_p = sd->parent; in build_sched_domains()
2579 while (top_p && !(top_p->flags & SD_NUMA)) { in build_sched_domains()
2580 top_p = top_p->parent; in build_sched_domains()
2582 imb_span = top_p ? top_p->span_weight : sd->span_weight; in build_sched_domains()
2584 int factor = max(1U, (sd->span_weight / imb_span)); in build_sched_domains()
2586 sd->imb_numa_nr = imb * factor; in build_sched_domains()
2592 for (i = nr_cpumask_bits-1; i >= 0; i--) { in build_sched_domains()
2596 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { in build_sched_domains()
2805 /* No match - a current sched domain not in new doms_new[] */ in partition_sched_domains_locked()
2826 /* No match - add a new doms_new */ in partition_sched_domains_locked()
2837 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) { in partition_sched_domains_locked()
2842 /* No match - add perf domains for a new rd */ in partition_sched_domains_locked()