1 /* Common code for 32 and 64-bit NUMA */
2 #include <linux/topology.h>
3 #include <linux/module.h>
4 #include <linux/bootmem.h>
8 int __initdata numa_off;
9 nodemask_t numa_nodes_parsed __initdata;
11 static __init int numa_setup(char *opt)
15 if (!strncmp(opt, "off", 3))
17 #ifdef CONFIG_NUMA_EMU
18 if (!strncmp(opt, "fake=", 5))
19 numa_emu_cmdline(opt + 5);
21 #ifdef CONFIG_ACPI_NUMA
22 if (!strncmp(opt, "noacpi", 6))
27 early_param("numa", numa_setup);
30 * apicid, cpu, node mappings
32 s16 __apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
33 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
36 int __cpuinit numa_cpu_node(int cpu)
38 int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
40 if (apicid != BAD_APICID)
41 return __apicid_to_node[apicid];
45 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
46 EXPORT_SYMBOL(node_to_cpumask_map);
49 * Map cpu index to node index
51 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
52 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
54 void __cpuinit numa_set_node(int cpu, int node)
56 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
58 /* early setting, no percpu area yet */
59 if (cpu_to_node_map) {
60 cpu_to_node_map[cpu] = node;
64 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
65 if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
66 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
71 per_cpu(x86_cpu_to_node_map, cpu) = node;
73 if (node != NUMA_NO_NODE)
74 set_cpu_numa_node(cpu, node);
77 void __cpuinit numa_clear_node(int cpu)
79 numa_set_node(cpu, NUMA_NO_NODE);
83 * Allocate node_to_cpumask_map based on number of available nodes
84 * Requires node_possible_map to be valid.
86 * Note: node_to_cpumask() is not valid until after this is done.
87 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
89 void __init setup_node_to_cpumask_map(void)
91 unsigned int node, num = 0;
93 /* setup nr_node_ids if not done yet */
94 if (nr_node_ids == MAX_NUMNODES) {
95 for_each_node_mask(node, node_possible_map)
97 nr_node_ids = num + 1;
100 /* allocate the map */
101 for (node = 0; node < nr_node_ids; node++)
102 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
104 /* cpumask_of_node() will now work */
105 pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
109 * There are unfortunately some poorly designed mainboards around that
110 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
111 * mapping. To avoid this fill in the mapping for all possible CPUs,
112 * as the number of CPUs is not known yet. We round robin the existing
115 void __init numa_init_array(void)
119 rr = first_node(node_online_map);
120 for (i = 0; i < nr_cpu_ids; i++) {
121 if (early_cpu_to_node(i) != NUMA_NO_NODE)
123 numa_set_node(i, rr);
124 rr = next_node(rr, node_online_map);
125 if (rr == MAX_NUMNODES)
126 rr = first_node(node_online_map);
130 static __init int find_near_online_node(int node)
133 int min_val = INT_MAX;
136 for_each_online_node(n) {
137 val = node_distance(node, n);
149 * Setup early cpu_to_node.
151 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
152 * and apicid_to_node[] tables have valid entries for a CPU.
153 * This means we skip cpu_to_node[] initialisation for NUMA
154 * emulation and faking node case (when running a kernel compiled
155 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
156 * is already initialized in a round robin manner at numa_init_array,
157 * prior to this call, and this initialization is good enough
158 * for the fake NUMA cases.
160 * Called before the per_cpu areas are setup.
162 void __init init_cpu_to_node(void)
165 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
167 BUG_ON(cpu_to_apicid == NULL);
169 for_each_possible_cpu(cpu) {
170 int node = numa_cpu_node(cpu);
172 if (node == NUMA_NO_NODE)
174 if (!node_online(node))
175 node = find_near_online_node(node);
176 numa_set_node(cpu, node);
180 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
182 # ifndef CONFIG_NUMA_EMU
183 void __cpuinit numa_add_cpu(int cpu)
185 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
188 void __cpuinit numa_remove_cpu(int cpu)
190 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
192 # endif /* !CONFIG_NUMA_EMU */
194 #else /* !CONFIG_DEBUG_PER_CPU_MAPS */
196 int __cpu_to_node(int cpu)
198 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
200 "cpu_to_node(%d): usage too early!\n", cpu);
202 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
204 return per_cpu(x86_cpu_to_node_map, cpu);
206 EXPORT_SYMBOL(__cpu_to_node);
209 * Same function as cpu_to_node() but used if called before the
210 * per_cpu areas are setup.
212 int early_cpu_to_node(int cpu)
214 if (early_per_cpu_ptr(x86_cpu_to_node_map))
215 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
217 if (!cpu_possible(cpu)) {
219 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
223 return per_cpu(x86_cpu_to_node_map, cpu);
226 void debug_cpumask_set_cpu(int cpu, int node, bool enable)
228 struct cpumask *mask;
231 if (node == NUMA_NO_NODE) {
232 /* early_cpu_to_node() already emits a warning and trace */
235 mask = node_to_cpumask_map[node];
237 pr_err("node_to_cpumask_map[%i] NULL\n", node);
243 cpumask_set_cpu(cpu, mask);
245 cpumask_clear_cpu(cpu, mask);
247 cpulist_scnprintf(buf, sizeof(buf), mask);
248 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
249 enable ? "numa_add_cpu" : "numa_remove_cpu",
254 # ifndef CONFIG_NUMA_EMU
255 static void __cpuinit numa_set_cpumask(int cpu, bool enable)
257 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
260 void __cpuinit numa_add_cpu(int cpu)
262 numa_set_cpumask(cpu, true);
265 void __cpuinit numa_remove_cpu(int cpu)
267 numa_set_cpumask(cpu, false);
269 # endif /* !CONFIG_NUMA_EMU */
272 * Returns a pointer to the bitmask of CPUs on Node 'node'.
274 const struct cpumask *cpumask_of_node(int node)
276 if (node >= nr_node_ids) {
278 "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
281 return cpu_none_mask;
283 if (node_to_cpumask_map[node] == NULL) {
285 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
288 return cpu_online_mask;
290 return node_to_cpumask_map[node];
292 EXPORT_SYMBOL(cpumask_of_node);
294 #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */