Lines Matching +full:cpu +full:- +full:capacity
1 // SPDX-License-Identifier: GPL-2.0
35 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, in sched_domain_debug_one() argument
38 struct sched_group *group = sd->groups; in sched_domain_debug_one()
39 unsigned long flags = sd->flags; in sched_domain_debug_one()
44 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level); in sched_domain_debug_one()
46 cpumask_pr_args(sched_domain_span(sd)), sd->name); in sched_domain_debug_one()
48 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { in sched_domain_debug_one()
49 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); in sched_domain_debug_one()
51 if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) { in sched_domain_debug_one()
52 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); in sched_domain_debug_one()
59 if ((meta_flags & SDF_SHARED_CHILD) && sd->child && in sched_domain_debug_one()
60 !(sd->child->flags & flag)) in sched_domain_debug_one()
64 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent && in sched_domain_debug_one()
65 !(sd->parent->flags & flag)) in sched_domain_debug_one()
84 if (!(sd->flags & SD_OVERLAP) && in sched_domain_debug_one()
94 group->sgc->id, in sched_domain_debug_one()
97 if ((sd->flags & SD_OVERLAP) && in sched_domain_debug_one()
103 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) in sched_domain_debug_one()
104 printk(KERN_CONT " cap=%lu", group->sgc->capacity); in sched_domain_debug_one()
106 if (group == sd->groups && sd->child && in sched_domain_debug_one()
107 !cpumask_equal(sched_domain_span(sd->child), in sched_domain_debug_one()
109 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n"); in sched_domain_debug_one()
114 group = group->next; in sched_domain_debug_one()
116 if (group != sd->groups) in sched_domain_debug_one()
119 } while (group != sd->groups); in sched_domain_debug_one()
123 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); in sched_domain_debug_one()
125 if (sd->parent && in sched_domain_debug_one()
126 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) in sched_domain_debug_one()
127 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); in sched_domain_debug_one()
131 static void sched_domain_debug(struct sched_domain *sd, int cpu) in sched_domain_debug() argument
139 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); in sched_domain_debug()
143 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu); in sched_domain_debug()
146 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) in sched_domain_debug()
149 sd = sd->parent; in sched_domain_debug()
157 # define sched_domain_debug(sd, cpu) do { } while (0) argument
177 if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) && in sd_degenerate()
178 (sd->groups != sd->groups->next)) in sd_degenerate()
182 if (sd->flags & (SD_WAKE_AFFINE)) in sd_degenerate()
191 unsigned long cflags = sd->flags, pflags = parent->flags; in sd_parent_degenerate()
200 if (parent->groups == parent->groups->next) in sd_parent_degenerate()
231 return -EPERM; in sched_energy_aware_handler()
270 tmp = pd->next; in free_pd()
276 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu) in find_pd() argument
279 if (cpumask_test_cpu(cpu, perf_domain_span(pd))) in find_pd()
281 pd = pd->next; in find_pd()
287 static struct perf_domain *pd_init(int cpu) in pd_init() argument
289 struct em_perf_domain *obj = em_cpu_get(cpu); in pd_init()
294 pr_info("%s: no EM found for CPU%d\n", __func__, cpu); in pd_init()
301 pd->em_pd = obj; in pd_init()
318 em_pd_nr_perf_states(pd->em_pd)); in perf_domain_debug()
319 pd = pd->next; in perf_domain_debug()
360 * - nr_pd: the number of performance domains
361 * - nr_cpus: the number of CPUs
362 * - nr_ps: the sum of the number of performance states of all performance
366 * It is generally not a good idea to use such a model in the wake-up path on
369 * with per-CPU DVFS and less than 8 performance states each, for example.
378 int cpu = cpumask_first(cpu_map); in build_perf_domains() local
379 struct root_domain *rd = cpu_rq(cpu)->rd; in build_perf_domains()
386 /* EAS is enabled for asymmetric CPU capacity topologies. */ in build_perf_domains()
387 if (!per_cpu(sd_asym_cpucapacity, cpu)) { in build_perf_domains()
404 pr_warn("rd %*pbl: Disabling EAS: frequency-invariant load tracking not yet supported", in build_perf_domains()
419 gov = policy->governor; in build_perf_domains()
422 if (rd->pd) in build_perf_domains()
432 tmp->next = pd; in build_perf_domains()
440 nr_ps += em_pd_nr_perf_states(pd->em_pd); in build_perf_domains()
453 tmp = rd->pd; in build_perf_domains()
454 rcu_assign_pointer(rd->pd, pd); in build_perf_domains()
456 call_rcu(&tmp->rcu, destroy_perf_domain_rcu); in build_perf_domains()
462 tmp = rd->pd; in build_perf_domains()
463 rcu_assign_pointer(rd->pd, NULL); in build_perf_domains()
465 call_rcu(&tmp->rcu, destroy_perf_domain_rcu); in build_perf_domains()
477 cpupri_cleanup(&rd->cpupri); in free_rootdomain()
478 cpudl_cleanup(&rd->cpudl); in free_rootdomain()
479 free_cpumask_var(rd->dlo_mask); in free_rootdomain()
480 free_cpumask_var(rd->rto_mask); in free_rootdomain()
481 free_cpumask_var(rd->online); in free_rootdomain()
482 free_cpumask_var(rd->span); in free_rootdomain()
483 free_pd(rd->pd); in free_rootdomain()
494 if (rq->rd) { in rq_attach_root()
495 old_rd = rq->rd; in rq_attach_root()
497 if (cpumask_test_cpu(rq->cpu, old_rd->online)) in rq_attach_root()
500 cpumask_clear_cpu(rq->cpu, old_rd->span); in rq_attach_root()
507 if (!atomic_dec_and_test(&old_rd->refcount)) in rq_attach_root()
511 atomic_inc(&rd->refcount); in rq_attach_root()
512 rq->rd = rd; in rq_attach_root()
514 cpumask_set_cpu(rq->cpu, rd->span); in rq_attach_root()
515 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) in rq_attach_root()
521 call_rcu(&old_rd->rcu, free_rootdomain); in rq_attach_root()
526 atomic_inc(&rd->refcount); in sched_get_rd()
531 if (!atomic_dec_and_test(&rd->refcount)) in sched_put_rd()
534 call_rcu(&rd->rcu, free_rootdomain); in sched_put_rd()
539 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL)) in init_rootdomain()
541 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL)) in init_rootdomain()
543 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) in init_rootdomain()
545 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) in init_rootdomain()
549 rd->rto_cpu = -1; in init_rootdomain()
550 raw_spin_lock_init(&rd->rto_lock); in init_rootdomain()
551 rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func); in init_rootdomain()
554 rd->visit_gen = 0; in init_rootdomain()
555 init_dl_bw(&rd->dl_bw); in init_rootdomain()
556 if (cpudl_init(&rd->cpudl) != 0) in init_rootdomain()
559 if (cpupri_init(&rd->cpupri) != 0) in init_rootdomain()
564 cpudl_cleanup(&rd->cpudl); in init_rootdomain()
566 free_cpumask_var(rd->rto_mask); in init_rootdomain()
568 free_cpumask_var(rd->dlo_mask); in init_rootdomain()
570 free_cpumask_var(rd->online); in init_rootdomain()
572 free_cpumask_var(rd->span); in init_rootdomain()
574 return -ENOMEM; in init_rootdomain()
578 * By default the system creates a single root-domain with all CPUs as
615 tmp = sg->next; in free_sched_groups()
617 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) in free_sched_groups()
618 kfree(sg->sgc); in free_sched_groups()
620 if (atomic_dec_and_test(&sg->ref)) in free_sched_groups()
631 * dropping group/capacity references, freeing where none remain. in destroy_sched_domain()
633 free_sched_groups(sd->groups, 1); in destroy_sched_domain()
635 if (sd->shared && atomic_dec_and_test(&sd->shared->ref)) in destroy_sched_domain()
636 kfree(sd->shared); in destroy_sched_domain()
645 struct sched_domain *parent = sd->parent; in destroy_sched_domains_rcu()
654 call_rcu(&sd->rcu, destroy_sched_domains_rcu); in destroy_sched_domains()
662 * Also keep a unique ID per domain (we use the first CPU number in
675 static void update_top_cache_domain(int cpu) in update_top_cache_domain() argument
679 int id = cpu; in update_top_cache_domain()
682 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); in update_top_cache_domain()
686 sds = sd->shared; in update_top_cache_domain()
689 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); in update_top_cache_domain()
690 per_cpu(sd_llc_size, cpu) = size; in update_top_cache_domain()
691 per_cpu(sd_llc_id, cpu) = id; in update_top_cache_domain()
692 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds); in update_top_cache_domain()
694 sd = lowest_flag_domain(cpu, SD_NUMA); in update_top_cache_domain()
695 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); in update_top_cache_domain()
697 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); in update_top_cache_domain()
698 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd); in update_top_cache_domain()
700 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL); in update_top_cache_domain()
701 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd); in update_top_cache_domain()
705 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
709 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) in cpu_attach_domain() argument
711 struct rq *rq = cpu_rq(cpu); in cpu_attach_domain()
716 struct sched_domain *parent = tmp->parent; in cpu_attach_domain()
721 tmp->parent = parent->parent; in cpu_attach_domain()
723 if (parent->parent) { in cpu_attach_domain()
724 parent->parent->child = tmp; in cpu_attach_domain()
725 parent->parent->groups->flags = tmp->flags; in cpu_attach_domain()
733 if (parent->flags & SD_PREFER_SIBLING) in cpu_attach_domain()
734 tmp->flags |= SD_PREFER_SIBLING; in cpu_attach_domain()
737 tmp = tmp->parent; in cpu_attach_domain()
742 sd = sd->parent; in cpu_attach_domain()
745 struct sched_group *sg = sd->groups; in cpu_attach_domain()
753 sg->flags = 0; in cpu_attach_domain()
754 } while (sg != sd->groups); in cpu_attach_domain()
756 sd->child = NULL; in cpu_attach_domain()
760 sched_domain_debug(sd, cpu); in cpu_attach_domain()
763 tmp = rq->sd; in cpu_attach_domain()
764 rcu_assign_pointer(rq->sd, sd); in cpu_attach_domain()
765 dirty_sched_domain_sysctl(cpu); in cpu_attach_domain()
768 update_top_cache_domain(cpu); in cpu_attach_domain()
784 * Return the canonical balance CPU for this group, this is the first CPU
801 * Given a node-distance table, for example:
811 * 0 ----- 1
815 * 3 ----- 2
823 * NUMA-2 0-3 0-3 0-3 0-3
824 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
826 * NUMA-1 0-1,3 0-2 1-3 0,2-3
829 * NUMA-0 0 1 2 3
834 * represented multiple times -- hence the "overlap" naming for this part of
838 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
842 * - the first group of each domain is its child domain; this
843 * gets us the first 0-1,3
844 * - the only uncovered node is 2, who's child domain is 1-3.
846 * However, because of the overlap, computing a unique CPU for each group is
847 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
848 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
849 * end up at those groups (they would end up in group: 0-1,3).
872 * 0 ----- 1
876 * 2 ----- 3
882 * not of the same number for each CPU. Consider:
884 * NUMA-2 0-3 0-3
885 * groups: {0-2},{1-3} {1-3},{0-2}
887 * NUMA-1 0-2 0-3 0-3 1-3
889 * NUMA-0 0 1 2 3
907 struct sd_data *sdd = sd->private; in build_balance_mask()
914 sibling = *per_cpu_ptr(sdd->sd, i); in build_balance_mask()
921 if (!sibling->child) in build_balance_mask()
925 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child))) in build_balance_mask()
936 * XXX: This creates per-node group entries; since the load-balancer will
937 * immediately access remote memory to construct this group's load-balance
941 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu) in build_group_from_child_sched_domain() argument
947 GFP_KERNEL, cpu_to_node(cpu)); in build_group_from_child_sched_domain()
953 if (sd->child) { in build_group_from_child_sched_domain()
954 cpumask_copy(sg_span, sched_domain_span(sd->child)); in build_group_from_child_sched_domain()
955 sg->flags = sd->child->flags; in build_group_from_child_sched_domain()
960 atomic_inc(&sg->ref); in build_group_from_child_sched_domain()
968 struct sd_data *sdd = sd->private; in init_overlap_sched_group()
970 int cpu; in init_overlap_sched_group() local
973 cpu = cpumask_first(mask); in init_overlap_sched_group()
975 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu); in init_overlap_sched_group()
976 if (atomic_inc_return(&sg->sgc->ref) == 1) in init_overlap_sched_group()
982 * Initialize sgc->capacity such that even if we mess up the in init_overlap_sched_group()
987 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); in init_overlap_sched_group()
988 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE; in init_overlap_sched_group()
989 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE; in init_overlap_sched_group()
999 while (sibling->child && in find_descended_sibling()
1000 !cpumask_subset(sched_domain_span(sibling->child), in find_descended_sibling()
1002 sibling = sibling->child; in find_descended_sibling()
1009 while (sibling->child && in find_descended_sibling()
1010 cpumask_equal(sched_domain_span(sibling->child), in find_descended_sibling()
1012 sibling = sibling->child; in find_descended_sibling()
1018 build_overlap_sched_groups(struct sched_domain *sd, int cpu) in build_overlap_sched_groups() argument
1023 struct sd_data *sdd = sd->private; in build_overlap_sched_groups()
1029 for_each_cpu_wrap(i, span, cpu) { in build_overlap_sched_groups()
1035 sibling = *per_cpu_ptr(sdd->sd, i); in build_overlap_sched_groups()
1044 * Domains should always include the CPU they're built on, so in build_overlap_sched_groups()
1065 * 0 --- 1 --- 2 --- 3 in build_overlap_sched_groups()
1067 * NUMA-3 0-3 N/A N/A 0-3 in build_overlap_sched_groups()
1068 * groups: {0-2},{1-3} {1-3},{0-2} in build_overlap_sched_groups()
1070 * NUMA-2 0-2 0-3 0-3 1-3 in build_overlap_sched_groups()
1071 * groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2} in build_overlap_sched_groups()
1073 * NUMA-1 0-1 0-2 1-3 2-3 in build_overlap_sched_groups()
1076 * NUMA-0 0 1 2 3 in build_overlap_sched_groups()
1078 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the in build_overlap_sched_groups()
1081 if (sibling->child && in build_overlap_sched_groups()
1082 !cpumask_subset(sched_domain_span(sibling->child), span)) in build_overlap_sched_groups()
1085 sg = build_group_from_child_sched_domain(sibling, cpu); in build_overlap_sched_groups()
1097 last->next = sg; in build_overlap_sched_groups()
1099 last->next = first; in build_overlap_sched_groups()
1101 sd->groups = first; in build_overlap_sched_groups()
1108 return -ENOMEM; in build_overlap_sched_groups()
1113 * Package topology (also see the load-balance blurb in fair.c)
1118 * - Simultaneous multithreading (SMT)
1119 * - Multi-Core Cache (MC)
1120 * - Package (DIE)
1126 * sched_domain -> sched_group -> sched_group_capacity
1128 * `-' `-'
1130 * The sched_domains are per-CPU and have a two way link (parent & child) and
1136 * CPU of that sched_domain [*].
1140 * CPU 0 1 2 3 4 5 6 7
1146 * - or -
1148 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1149 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1150 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1152 * CPU 0 1 2 3 4 5 6 7
1160 * There are two related construction problems, both require a CPU that
1163 * - The first is the balance_cpu (see should_we_balance() and the
1164 * load-balance blub in fair.c); for each group we only want 1 CPU to
1167 * - The second is the sched_group_capacity; we want all identical groups
1173 * for each CPU in the hierarchy.
1175 * Therefore computing a unique CPU for each group is trivial (the iteration
1177 * group), we can simply pick the first CPU in each group.
1183 static struct sched_group *get_group(int cpu, struct sd_data *sdd) in get_group() argument
1185 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); in get_group()
1186 struct sched_domain *child = sd->child; in get_group()
1191 cpu = cpumask_first(sched_domain_span(child)); in get_group()
1193 sg = *per_cpu_ptr(sdd->sg, cpu); in get_group()
1194 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu); in get_group()
1197 already_visited = atomic_inc_return(&sg->ref) > 1; in get_group()
1199 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1)); in get_group()
1208 sg->flags = child->flags; in get_group()
1210 cpumask_set_cpu(cpu, sched_group_span(sg)); in get_group()
1211 cpumask_set_cpu(cpu, group_balance_mask(sg)); in get_group()
1214 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg)); in get_group()
1215 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE; in get_group()
1216 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE; in get_group()
1223 * covered by the given span, will set each group's ->cpumask correctly,
1224 * and will initialize their ->sgc.
1229 build_sched_groups(struct sched_domain *sd, int cpu) in build_sched_groups() argument
1232 struct sd_data *sdd = sd->private; in build_sched_groups()
1242 for_each_cpu_wrap(i, span, cpu) { in build_sched_groups()
1255 last->next = sg; in build_sched_groups()
1258 last->next = first; in build_sched_groups()
1259 sd->groups = first; in build_sched_groups()
1267 * cpu_capacity indicates the capacity of sched group, which is used while
1274 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) in init_sched_groups_capacity() argument
1276 struct sched_group *sg = sd->groups; in init_sched_groups_capacity()
1282 int cpu, cores = 0, max_cpu = -1; in init_sched_groups_capacity() local
1284 sg->group_weight = cpumask_weight(sched_group_span(sg)); in init_sched_groups_capacity()
1287 for_each_cpu(cpu, mask) { in init_sched_groups_capacity()
1290 cpumask_andnot(mask, mask, cpu_smt_mask(cpu)); in init_sched_groups_capacity()
1293 sg->cores = cores; in init_sched_groups_capacity()
1295 if (!(sd->flags & SD_ASYM_PACKING)) in init_sched_groups_capacity()
1298 for_each_cpu(cpu, sched_group_span(sg)) { in init_sched_groups_capacity()
1300 max_cpu = cpu; in init_sched_groups_capacity()
1301 else if (sched_asym_prefer(cpu, max_cpu)) in init_sched_groups_capacity()
1302 max_cpu = cpu; in init_sched_groups_capacity()
1304 sg->asym_prefer_cpu = max_cpu; in init_sched_groups_capacity()
1307 sg = sg->next; in init_sched_groups_capacity()
1308 } while (sg != sd->groups); in init_sched_groups_capacity()
1310 if (cpu != group_balance_cpu(sg)) in init_sched_groups_capacity()
1313 update_group_capacity(sd, cpu); in init_sched_groups_capacity()
1317 * Asymmetric CPU capacity bits
1321 unsigned long capacity; member
1327 * Each list entry contains a CPU mask reflecting CPUs that share the same
1328 * capacity.
1333 #define cpu_capacity_span(asym_data) to_cpumask((asym_data)->cpus)
1336 * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1347 * Count how many unique CPU capacities this domain spans across in asym_cpu_capacity_classify()
1364 /* Some of the available CPU capacity values have not been detected */ in asym_cpu_capacity_classify()
1373 static inline void asym_cpu_capacity_update_data(int cpu) in asym_cpu_capacity_update_data() argument
1375 unsigned long capacity = arch_scale_cpu_capacity(cpu); in asym_cpu_capacity_update_data() local
1379 if (capacity == entry->capacity) in asym_cpu_capacity_update_data()
1386 entry->capacity = capacity; in asym_cpu_capacity_update_data()
1387 list_add(&entry->link, &asym_cap_list); in asym_cpu_capacity_update_data()
1389 __cpumask_set_cpu(cpu, cpu_capacity_span(entry)); in asym_cpu_capacity_update_data()
1393 * Build-up/update list of CPUs grouped by their capacities
1395 * with state indicating CPU topology changes.
1400 int cpu; in asym_cpu_capacity_scan() local
1405 for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN)) in asym_cpu_capacity_scan()
1406 asym_cpu_capacity_update_data(cpu); in asym_cpu_capacity_scan()
1410 list_del(&entry->link); in asym_cpu_capacity_scan()
1416 * Only one capacity value has been detected i.e. this system is symmetric. in asym_cpu_capacity_scan()
1421 list_del(&entry->link); in asym_cpu_capacity_scan()
1428 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1431 static int default_relax_domain_level = -1;
1448 if (!attr || attr->relax_domain_level < 0) { in set_domain_attribute()
1453 request = attr->relax_domain_level; in set_domain_attribute()
1455 if (sd->level >= request) { in set_domain_attribute()
1457 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); in set_domain_attribute()
1469 if (!atomic_read(&d->rd->refcount)) in __free_domain_allocs()
1470 free_rootdomain(&d->rd->rcu); in __free_domain_allocs()
1473 free_percpu(d->sd); in __free_domain_allocs()
1490 d->sd = alloc_percpu(struct sched_domain *); in __visit_domain_allocation_hell()
1491 if (!d->sd) in __visit_domain_allocation_hell()
1493 d->rd = alloc_rootdomain(); in __visit_domain_allocation_hell()
1494 if (!d->rd) in __visit_domain_allocation_hell()
1505 static void claim_allocations(int cpu, struct sched_domain *sd) in claim_allocations() argument
1507 struct sd_data *sdd = sd->private; in claim_allocations()
1509 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); in claim_allocations()
1510 *per_cpu_ptr(sdd->sd, cpu) = NULL; in claim_allocations()
1512 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref)) in claim_allocations()
1513 *per_cpu_ptr(sdd->sds, cpu) = NULL; in claim_allocations()
1515 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) in claim_allocations()
1516 *per_cpu_ptr(sdd->sg, cpu) = NULL; in claim_allocations()
1518 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) in claim_allocations()
1519 *per_cpu_ptr(sdd->sgc, cpu) = NULL; in claim_allocations()
1540 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1541 * SD_SHARE_PKG_RESOURCES - describes shared caches
1542 * SD_NUMA - describes NUMA topologies
1547 * SD_ASYM_PACKING - describes SMT quirks
1558 struct sched_domain *child, int cpu) in sd_init() argument
1560 struct sd_data *sdd = &tl->data; in sd_init()
1561 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); in sd_init()
1569 sched_domains_curr_level = tl->numa_level; in sd_init()
1572 sd_weight = cpumask_weight(tl->mask(cpu)); in sd_init()
1574 if (tl->sd_flags) in sd_init()
1575 sd_flags = (*tl->sd_flags)(); in sd_init()
1607 .name = tl->name, in sd_init()
1612 cpumask_and(sd_span, cpu_map, tl->mask(cpu)); in sd_init()
1615 sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map); in sd_init()
1617 WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) == in sd_init()
1619 "CPU capacity asymmetry not supported on SMT\n"); in sd_init()
1625 if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child) in sd_init()
1626 sd->child->flags &= ~SD_PREFER_SIBLING; in sd_init()
1628 if (sd->flags & SD_SHARE_CPUCAPACITY) { in sd_init()
1629 sd->imbalance_pct = 110; in sd_init()
1631 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { in sd_init()
1632 sd->imbalance_pct = 117; in sd_init()
1633 sd->cache_nice_tries = 1; in sd_init()
1636 } else if (sd->flags & SD_NUMA) { in sd_init()
1637 sd->cache_nice_tries = 2; in sd_init()
1639 sd->flags &= ~SD_PREFER_SIBLING; in sd_init()
1640 sd->flags |= SD_SERIALIZE; in sd_init()
1641 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) { in sd_init()
1642 sd->flags &= ~(SD_BALANCE_EXEC | in sd_init()
1649 sd->cache_nice_tries = 1; in sd_init()
1656 if (sd->flags & SD_SHARE_PKG_RESOURCES) { in sd_init()
1657 sd->shared = *per_cpu_ptr(sdd->sds, sd_id); in sd_init()
1658 atomic_inc(&sd->shared->ref); in sd_init()
1659 atomic_set(&sd->shared->nr_busy_cpus, sd_weight); in sd_init()
1662 sd->private = sdd; in sd_init()
1668 * Topology list, bottom-up.
1691 for (tl = sched_domain_topology; tl->mask; tl++)
1704 static const struct cpumask *sd_numa_mask(int cpu) in sd_numa_mask() argument
1706 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; in sd_numa_mask()
1777 * - If the maximum distance between any nodes is 1 hop, the system
1779 * - If for two nodes A and B, located N > 1 hops away from each other,
1832 * O(nr_nodes^2) deduplicating selection sort -- in order to find the in sched_init_numa()
1915 sched_numa_warn("Node-distance not symmetric"); in sched_init_numa()
1966 WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]); in sched_init_numa()
2009 void sched_update_numa(int cpu, bool online) in sched_update_numa() argument
2013 node = cpu_to_node(cpu); in sched_update_numa()
2015 * Scheduler NUMA topology is updated when the first CPU of a in sched_update_numa()
2016 * node is onlined or the last CPU of a node is offlined. in sched_update_numa()
2025 void sched_domains_numa_masks_set(unsigned int cpu) in sched_domains_numa_masks_set() argument
2027 int node = cpu_to_node(cpu); in sched_domains_numa_masks_set()
2037 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); in sched_domains_numa_masks_set()
2042 void sched_domains_numa_masks_clear(unsigned int cpu) in sched_domains_numa_masks_clear() argument
2049 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); in sched_domains_numa_masks_clear()
2055 * sched_numa_find_closest() - given the NUMA topology, find the cpu
2056 * closest to @cpu from @cpumask.
2057 * cpumask: cpumask to find a cpu from
2058 * cpu: cpu to be close to
2060 * returns: cpu, or nr_cpu_ids when nothing found.
2062 int sched_numa_find_closest(const struct cpumask *cpus, int cpu) in sched_numa_find_closest() argument
2064 int i, j = cpu_to_node(cpu), found = nr_cpu_ids; in sched_numa_find_closest()
2074 cpu = cpumask_any_and(cpus, masks[i][j]); in sched_numa_find_closest()
2075 if (cpu < nr_cpu_ids) { in sched_numa_find_closest()
2076 found = cpu; in sched_numa_find_closest()
2090 int cpu; member
2099 if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu) in hop_cmp()
2102 if (b == k->masks) { in hop_cmp()
2103 k->w = 0; in hop_cmp()
2107 prev_hop = *((struct cpumask ***)b - 1); in hop_cmp()
2108 k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]); in hop_cmp()
2109 if (k->w <= k->cpu) in hop_cmp()
2112 return -1; in hop_cmp()
2116 * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth next cpu
2117 * closest to @cpu from @cpumask.
2118 * cpumask: cpumask to find a cpu from
2119 * cpu: Nth cpu to find
2121 * returns: cpu, or nr_cpu_ids when nothing found.
2123 int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node) in sched_numa_find_nth_cpu() argument
2125 struct __cmp_key k = { .cpus = cpus, .cpu = cpu }; in sched_numa_find_nth_cpu()
2130 return cpumask_nth_and(cpu, cpus, cpu_online_mask); in sched_numa_find_nth_cpu()
2134 /* CPU-less node entries are uninitialized in sched_domains_numa_masks */ in sched_numa_find_nth_cpu()
2143 hop = hop_masks - k.masks; in sched_numa_find_nth_cpu()
2146 cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) : in sched_numa_find_nth_cpu()
2147 cpumask_nth_and(cpu, cpus, k.masks[0][node]); in sched_numa_find_nth_cpu()
2155 * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2164 * read-side section, copy it if required beyond that.
2176 return ERR_PTR(-EINVAL); in sched_numa_hop_mask()
2180 return ERR_PTR(-EBUSY); in sched_numa_hop_mask()
2194 struct sd_data *sdd = &tl->data; in __sdt_alloc()
2196 sdd->sd = alloc_percpu(struct sched_domain *); in __sdt_alloc()
2197 if (!sdd->sd) in __sdt_alloc()
2198 return -ENOMEM; in __sdt_alloc()
2200 sdd->sds = alloc_percpu(struct sched_domain_shared *); in __sdt_alloc()
2201 if (!sdd->sds) in __sdt_alloc()
2202 return -ENOMEM; in __sdt_alloc()
2204 sdd->sg = alloc_percpu(struct sched_group *); in __sdt_alloc()
2205 if (!sdd->sg) in __sdt_alloc()
2206 return -ENOMEM; in __sdt_alloc()
2208 sdd->sgc = alloc_percpu(struct sched_group_capacity *); in __sdt_alloc()
2209 if (!sdd->sgc) in __sdt_alloc()
2210 return -ENOMEM; in __sdt_alloc()
2221 return -ENOMEM; in __sdt_alloc()
2223 *per_cpu_ptr(sdd->sd, j) = sd; in __sdt_alloc()
2228 return -ENOMEM; in __sdt_alloc()
2230 *per_cpu_ptr(sdd->sds, j) = sds; in __sdt_alloc()
2235 return -ENOMEM; in __sdt_alloc()
2237 sg->next = sg; in __sdt_alloc()
2239 *per_cpu_ptr(sdd->sg, j) = sg; in __sdt_alloc()
2244 return -ENOMEM; in __sdt_alloc()
2247 sgc->id = j; in __sdt_alloc()
2250 *per_cpu_ptr(sdd->sgc, j) = sgc; in __sdt_alloc()
2263 struct sd_data *sdd = &tl->data; in __sdt_free()
2268 if (sdd->sd) { in __sdt_free()
2269 sd = *per_cpu_ptr(sdd->sd, j); in __sdt_free()
2270 if (sd && (sd->flags & SD_OVERLAP)) in __sdt_free()
2271 free_sched_groups(sd->groups, 0); in __sdt_free()
2272 kfree(*per_cpu_ptr(sdd->sd, j)); in __sdt_free()
2275 if (sdd->sds) in __sdt_free()
2276 kfree(*per_cpu_ptr(sdd->sds, j)); in __sdt_free()
2277 if (sdd->sg) in __sdt_free()
2278 kfree(*per_cpu_ptr(sdd->sg, j)); in __sdt_free()
2279 if (sdd->sgc) in __sdt_free()
2280 kfree(*per_cpu_ptr(sdd->sgc, j)); in __sdt_free()
2282 free_percpu(sdd->sd); in __sdt_free()
2283 sdd->sd = NULL; in __sdt_free()
2284 free_percpu(sdd->sds); in __sdt_free()
2285 sdd->sds = NULL; in __sdt_free()
2286 free_percpu(sdd->sg); in __sdt_free()
2287 sdd->sg = NULL; in __sdt_free()
2288 free_percpu(sdd->sgc); in __sdt_free()
2289 sdd->sgc = NULL; in __sdt_free()
2295 struct sched_domain *child, int cpu) in build_sched_domain() argument
2297 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu); in build_sched_domain()
2300 sd->level = child->level + 1; in build_sched_domain()
2301 sched_domain_level_max = max(sched_domain_level_max, sd->level); in build_sched_domain()
2302 child->parent = sd; in build_sched_domain()
2309 child->name, sd->name); in build_sched_domain()
2325 * any two given CPUs at this (non-NUMA) topology level.
2328 const struct cpumask *cpu_map, int cpu) in topology_span_sane() argument
2333 if (tl->flags & SDTL_OVERLAP) in topology_span_sane()
2337 * Non-NUMA levels cannot partially overlap - they must be either in topology_span_sane()
2339 * breaking the sched_group lists - i.e. a later get_group() pass in topology_span_sane()
2343 if (i == cpu) in topology_span_sane()
2351 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) && in topology_span_sane()
2352 cpumask_intersects(tl->mask(cpu), tl->mask(i))) in topology_span_sane()
2370 int i, ret = -ENOMEM; in build_sched_domains()
2392 has_asym |= sd->flags & SD_ASYM_CPUCAPACITY; in build_sched_domains()
2396 if (tl->flags & SDTL_OVERLAP) in build_sched_domains()
2397 sd->flags |= SD_OVERLAP; in build_sched_domains()
2405 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { in build_sched_domains()
2406 sd->span_weight = cpumask_weight(sched_domain_span(sd)); in build_sched_domains()
2407 if (sd->flags & SD_OVERLAP) { in build_sched_domains()
2425 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { in build_sched_domains()
2426 struct sched_domain *child = sd->child; in build_sched_domains()
2428 if (!(sd->flags & SD_SHARE_PKG_RESOURCES) && child && in build_sched_domains()
2429 (child->flags & SD_SHARE_PKG_RESOURCES)) { in build_sched_domains()
2436 * arbitrary cutoff based two factors -- SMT and in build_sched_domains()
2437 * memory channels. For SMT-2, the intent is to in build_sched_domains()
2439 * SMT-4 or SMT-8 *may* benefit from a different in build_sched_domains()
2453 nr_llcs = sd->span_weight / child->span_weight; in build_sched_domains()
2455 imb = sd->span_weight >> 3; in build_sched_domains()
2459 sd->imb_numa_nr = imb; in build_sched_domains()
2462 top_p = sd->parent; in build_sched_domains()
2463 while (top_p && !(top_p->flags & SD_NUMA)) { in build_sched_domains()
2464 top_p = top_p->parent; in build_sched_domains()
2466 imb_span = top_p ? top_p->span_weight : sd->span_weight; in build_sched_domains()
2468 int factor = max(1U, (sd->span_weight / imb_span)); in build_sched_domains()
2470 sd->imb_numa_nr = imb * factor; in build_sched_domains()
2475 /* Calculate CPU capacity for physical packages and nodes */ in build_sched_domains()
2476 for (i = nr_cpumask_bits-1; i >= 0; i--) { in build_sched_domains()
2480 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { in build_sched_domains()
2493 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity)) in build_sched_domains()
2494 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig); in build_sched_domains()
2505 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity); in build_sched_domains()
2533 * CPU core maps. It is supposed to return 1 if the topology changed
2596 unsigned int cpu = cpumask_any(cpu_map); in detach_destroy_domains() local
2599 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu))) in detach_destroy_domains()
2660 /* Let the architecture update CPU core mappings: */ in partition_sched_domains_locked()
2662 /* Trigger rebuilding CPU capacity asymmetry data */ in partition_sched_domains_locked()
2688 * its dl_bw->total_bw needs to be cleared. It in partition_sched_domains_locked()
2692 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd; in partition_sched_domains_locked()
2697 /* No match - a current sched domain not in new doms_new[] */ in partition_sched_domains_locked()
2718 /* No match - add a new doms_new */ in partition_sched_domains_locked()
2729 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) { in partition_sched_domains_locked()
2734 /* No match - add perf. domains for a new rd */ in partition_sched_domains_locked()