| 1 | #include <console/printf.h> |
| 2 | #include <global.h> |
| 3 | #include <log.h> |
| 4 | #include <mem/alloc.h> |
| 5 | #include <mem/alloc_or_die.h> |
| 6 | #include <mem/numa.h> |
| 7 | #include <sch/sched.h> |
| 8 | #include <smp/domain.h> |
| 9 | #include <stdbool.h> |
| 10 | #include <sync/spinlock.h> |
| 11 | |
| 12 | LOG_SITE_DECLARE_DEFAULT(domain); |
| 13 | LOG_HANDLE_DECLARE_DEFAULT(domain); |
| 14 | |
| 15 | #define domain_log(lvl, fmt, ...) \ |
| 16 | log(LOG_SITE(domain), LOG_HANDLE(domain), lvl, fmt, ##__VA_ARGS__) |
| 17 | |
| 18 | #define domain_err(fmt, ...) domain_log(LOG_ERROR, fmt, ##__VA_ARGS__) |
| 19 | #define domain_warn(fmt, ...) domain_log(LOG_WARN, fmt, ##__VA_ARGS__) |
| 20 | #define domain_info(fmt, ...) domain_log(LOG_INFO, fmt, ##__VA_ARGS__) |
| 21 | #define domain_debug(fmt, ...) domain_log(LOG_DEBUG, fmt, ##__VA_ARGS__) |
| 22 | #define domain_trace(fmt, ...) domain_log(LOG_TRACE, fmt, ##__VA_ARGS__) |
| 23 | |
| 24 | static void init_global_domain(uint64_t domain_count) { |
| 25 | global.domain_count = domain_count; |
| 26 | global.domains = alloc_or_die( |
| 27 | kmalloc(sizeof(struct domain *) * domain_count, ALLOC_FLAGS_ZERO)); |
| 28 | |
| 29 | for (size_t i = 0; i < domain_count; i++) { |
| 30 | |
| 31 | /* We align this up to the page so that they can all be |
| 32 | * migrated later on to pages on each domain... */ |
| 33 | global.domains[i] = alloc_or_die( |
| 34 | kmalloc(PAGE_ALIGN_UP(sizeof(struct domain)), ALLOC_FLAGS_ZERO)); |
| 35 | |
| 36 | global.domains[i]->id = i; |
| 37 | } |
| 38 | } |
| 39 | |
| 40 | /* Map 1:1 with NUMA nodes */ |
| 41 | static void construct_domains_from_numa_nodes(void) { |
| 42 | init_global_domain(domain_count: global.numa_node_count); |
| 43 | for (size_t i = 0; i < global.numa_node_count; i++) { |
| 44 | struct numa_node *nn = &global.numa_nodes[i]; |
| 45 | struct domain *cd = global.domains[i]; |
| 46 | cd->num_cores = cpu_mask_popcount(m: &nn->cpus); |
| 47 | cd->associated_node = nn; |
| 48 | cd->cores = |
| 49 | kmalloc(sizeof(struct core *) * cd->num_cores, ALLOC_FLAGS_ZERO); |
| 50 | } |
| 51 | } |
| 52 | |
| 53 | static void construct_domains_after_smp() { |
| 54 | if (global.numa_node_count > 1) { |
| 55 | for (size_t i = 0; i < global.domain_count; i++) { |
| 56 | struct domain *cd = global.domains[i]; |
| 57 | struct numa_node *nn = &global.numa_nodes[i]; |
| 58 | struct cpu_mask nm = nn->cpus; |
| 59 | size_t j; |
| 60 | size_t k = 0; |
| 61 | cpu_mask_for_each(j, nm) { |
| 62 | cd->cores[k++] = global.cores[j]; |
| 63 | } |
| 64 | } |
| 65 | } else { |
| 66 | size_t core_index = 0; |
| 67 | for (size_t i = 0; i < global.domain_count; i++) { |
| 68 | struct domain *cd = global.domains[i]; |
| 69 | |
| 70 | for (size_t j = 0; j < cd->num_cores; j++) { |
| 71 | global.cores[core_index]->domain = cd; |
| 72 | cd->cores[j] = global.cores[core_index++]; |
| 73 | } |
| 74 | } |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | static void construct_domains_from_cores(void) { |
| 79 | size_t n_domains = global.core_count / CORES_PER_DOMAIN; |
| 80 | size_t remainder = global.core_count % CORES_PER_DOMAIN; |
| 81 | |
| 82 | if (remainder > 0) |
| 83 | n_domains++; /* one extra for leftover cores */ |
| 84 | |
| 85 | init_global_domain(domain_count: n_domains); |
| 86 | |
| 87 | for (size_t i = 0; i < n_domains; i++) { |
| 88 | struct domain *cd = global.domains[i]; |
| 89 | |
| 90 | cd->associated_node = NULL; |
| 91 | |
| 92 | /* Decide how many cores this domain should get */ |
| 93 | size_t cores_this_domain = CORES_PER_DOMAIN; |
| 94 | if (i == n_domains - 1 && remainder > 0) |
| 95 | cores_this_domain = remainder; /* last one gets leftovers */ |
| 96 | |
| 97 | cd->num_cores = cores_this_domain; |
| 98 | cd->cores = kmalloc_or_die(sizeof(struct core *) * cores_this_domain, |
| 99 | ALLOC_FLAGS_ZERO); |
| 100 | } |
| 101 | } |
| 102 | |
| 103 | void domain_dump(void) { |
| 104 | domain_info("Domains (%zu total)" , global.domain_count); |
| 105 | |
| 106 | for (size_t i = 0; i < global.domain_count; i++) { |
| 107 | struct domain *cd = global.domains[i]; |
| 108 | if (cd->associated_node) { |
| 109 | domain_info(" Domain %zu: Cores = %zu, NUMA node = %zu" , i, |
| 110 | cd->num_cores, cd->associated_node->topo->id); |
| 111 | } else { |
| 112 | domain_info(" Domain %zu: Cores = %zu, NUMA node = <none>" , i, |
| 113 | cd->num_cores); |
| 114 | } |
| 115 | |
| 116 | for (size_t j = 0; j < cd->num_cores; j++) { |
| 117 | if (cd->cores && cd->cores[j]) { |
| 118 | struct core *c = cd->cores[j]; |
| 119 | domain_info(" Core %zu" , c->id); |
| 120 | } else { |
| 121 | domain_info(" Core <NULL>" ); |
| 122 | } |
| 123 | } |
| 124 | } |
| 125 | } |
| 126 | |
| 127 | /* If NUMA is present, domains map 1:1 with |
| 128 | * NUMA nodes. If not, we just group cores into |
| 129 | * groups of CORES_PER_DOMAIN |
| 130 | * and construct domains from them. */ |
| 131 | void domain_init(void) { |
| 132 | if (global.numa_node_count > 1) { |
| 133 | construct_domains_from_numa_nodes(); |
| 134 | } else { |
| 135 | construct_domains_from_cores(); |
| 136 | } |
| 137 | |
| 138 | /* NOTE: This is THE exception that we make */ |
| 139 | global.cores[0]->domain = global.domains[0]; |
| 140 | } |
| 141 | |
| 142 | void domain_init_after_smp() { |
| 143 | construct_domains_after_smp(); |
| 144 | for (size_t i = 0; i < global.domain_count; i++) { |
| 145 | struct domain *domain = global.domains[i]; |
| 146 | alloc_or_die(cpu_mask_init(&domain->cpu_mask, global.core_count)); |
| 147 | |
| 148 | for (size_t j = 0; j < domain->num_cores; j++) { |
| 149 | domain->cores[j]->domain_cpu_id = j; |
| 150 | domain->cores[j]->domain = domain; |
| 151 | cpu_mask_set(m: &domain->cpu_mask, cpu: domain->cores[j]->id); |
| 152 | } |
| 153 | } |
| 154 | } |
| 155 | |
| 156 | void domain_set_cpu_mask(struct cpu_mask *mask, struct domain *domain) { |
| 157 | for (size_t i = 0; i < domain->num_cores; i++) |
| 158 | cpu_mask_set(m: mask, cpu: domain->cores[i]->id); |
| 159 | } |
| 160 | |
| 161 | /* Create a CPU mask that has all the bits for the domain set */ |
| 162 | struct cpu_mask *domain_create_cpu_mask(struct domain *domain) { |
| 163 | struct cpu_mask *ret = cpu_mask_create(); |
| 164 | if (!ret) |
| 165 | goto err; |
| 166 | |
| 167 | if (!cpu_mask_init(m: ret, nbits: global.core_count)) |
| 168 | goto err; |
| 169 | |
| 170 | domain_set_cpu_mask(mask: ret, domain); |
| 171 | |
| 172 | return ret; |
| 173 | err: |
| 174 | if (ret) |
| 175 | kfree(ret); |
| 176 | |
| 177 | return NULL; |
| 178 | } |
| 179 | |
| 180 | static void domains_move(void *a, void *b) { |
| 181 | (void) a, (void) b; |
| 182 | for (size_t i = 0; i < global.domain_count; i++) { |
| 183 | struct domain *domain = global.domains[i]; |
| 184 | movealloc(domain->id, domain); |
| 185 | } |
| 186 | } |
| 187 | |
| 188 | bool domain_idle(struct domain *domain) { |
| 189 | for (size_t i = 0; i < domain->num_cores; i++) |
| 190 | if (!scheduler_core_idle(c: domain->cores[i])) |
| 191 | return false; |
| 192 | |
| 193 | return true; |
| 194 | } |
| 195 | |
| 196 | size_t domain_for_core(size_t cpu) { |
| 197 | for (size_t i = 0; i < global.numa_node_count; i++) { |
| 198 | struct numa_node *nn = &global.numa_nodes[i]; |
| 199 | if (cpu_mask_test(m: &nn->cpus, cpu)) |
| 200 | return i; |
| 201 | } |
| 202 | |
| 203 | panic("unreachable!" ); |
| 204 | } |
| 205 | |
| 206 | MOVEALLOC_REGISTER_CALL(domain_move, domains_move, /* a = */ NULL, |
| 207 | /* b = */ NULL); |
| 208 | |