1 #include <linux/kernel.h>
2 #include <linux/module.h>
3 #include <linux/init.h>
4 #include <linux/bootmem.h>
5 #include <linux/percpu.h>
6 #include <linux/kexec.h>
7 #include <linux/crash_dump.h>
9 #include <asm/percpu.h>
10 #include <asm/sections.h>
11 #include <asm/processor.h>
12 #include <asm/setup.h>
13 #include <asm/topology.h>
14 #include <asm/mpspec.h>
15 #include <asm/apicdef.h>
16 #include <asm/highmem.h>
18 #ifdef CONFIG_X86_LOCAL_APIC
19 unsigned int num_processors;
20 unsigned disabled_cpus __cpuinitdata;
21 /* Processor that is doing the boot up */
22 unsigned int boot_cpu_physical_apicid = -1U;
23 unsigned int max_physical_apicid;
24 EXPORT_SYMBOL(boot_cpu_physical_apicid);
26 /* Bitmask of physically existing CPUs */
27 physid_mask_t phys_cpu_present_map;
30 /* map cpu index to physical APIC ID */
31 DEFINE_EARLY_PER_CPU(u16, x86_cpu_to_apicid, BAD_APICID);
32 DEFINE_EARLY_PER_CPU(u16, x86_bios_cpu_apicid, BAD_APICID);
33 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid);
34 EXPORT_EARLY_PER_CPU_SYMBOL(x86_bios_cpu_apicid);
36 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
39 /* map cpu index to node index */
40 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
41 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
43 /* which logical CPUs are on which nodes */
44 cpumask_t *node_to_cpumask_map;
45 EXPORT_SYMBOL(node_to_cpumask_map);
47 /* setup node_to_cpumask_map */
48 static void __init setup_node_to_cpumask_map(void);
51 static inline void setup_node_to_cpumask_map(void) { }
54 #if defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) && defined(CONFIG_X86_SMP)
56 * Copy data used in early init routines from the initial arrays to the
57 * per cpu data areas. These arrays then become expendable and the
58 * *_early_ptr's are zeroed indicating that the static arrays are gone.
60 static void __init setup_per_cpu_maps(void)
64 for_each_possible_cpu(cpu) {
65 per_cpu(x86_cpu_to_apicid, cpu) =
66 early_per_cpu_map(x86_cpu_to_apicid, cpu);
67 per_cpu(x86_bios_cpu_apicid, cpu) =
68 early_per_cpu_map(x86_bios_cpu_apicid, cpu);
70 per_cpu(x86_cpu_to_node_map, cpu) =
71 early_per_cpu_map(x86_cpu_to_node_map, cpu);
75 /* indicate the early static arrays will soon be gone */
76 early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
77 early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
79 early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
85 * Great future not-so-futuristic plan: make i386 and x86_64 do it
88 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
89 EXPORT_SYMBOL(__per_cpu_offset);
90 static inline void setup_cpu_pda_map(void) { }
92 #elif !defined(CONFIG_SMP)
93 static inline void setup_cpu_pda_map(void) { }
95 #else /* CONFIG_SMP && CONFIG_X86_64 */
98 * Allocate cpu_pda pointer table and array via alloc_bootmem.
100 static void __init setup_cpu_pda_map(void)
103 struct x8664_pda **new_cpu_pda;
107 size = roundup(sizeof(struct x8664_pda), cache_line_size());
109 /* allocate cpu_pda array and pointer table */
111 unsigned long tsize = nr_cpu_ids * sizeof(void *);
112 unsigned long asize = size * (nr_cpu_ids - 1);
114 tsize = roundup(tsize, cache_line_size());
115 new_cpu_pda = alloc_bootmem(tsize + asize);
116 pda = (char *)new_cpu_pda + tsize;
119 /* initialize pointer table to static pda's */
120 for_each_possible_cpu(cpu) {
122 /* leave boot cpu pda in place */
123 new_cpu_pda[0] = cpu_pda(0);
126 new_cpu_pda[cpu] = (struct x8664_pda *)pda;
127 new_cpu_pda[cpu]->in_bootmem = 1;
131 /* point to new pointer table */
132 _cpu_pda = new_cpu_pda;
138 * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
139 * Always point %gs to its beginning
141 void __init setup_per_cpu_areas(void)
143 ssize_t size, old_size;
147 /* Setup cpu_pda map */
150 /* Copy section for each CPU (we discard the original) */
151 old_size = PERCPU_ENOUGH_ROOM;
152 size = old_size + per_cpu_dyn_array_size();
153 printk(KERN_INFO "PERCPU: Allocating %zd bytes of per cpu data\n",
156 for_each_possible_cpu(cpu) {
157 #ifndef CONFIG_NEED_MULTIPLE_NODES
158 ptr = alloc_bootmem_pages(size);
160 int node = early_cpu_to_node(cpu);
161 if (!node_online(node) || !NODE_DATA(node)) {
162 ptr = alloc_bootmem_pages(size);
164 "cpu %d has no node %d or node-local memory\n",
167 printk(KERN_DEBUG "per cpu data for cpu%d at %016lx\n",
171 ptr = alloc_bootmem_pages_node(NODE_DATA(node), size);
173 printk(KERN_DEBUG "per cpu data for cpu%d on node%d at %016lx\n",
174 cpu, node, __pa(ptr));
177 per_cpu_offset(cpu) = ptr - __per_cpu_start;
178 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
180 per_cpu_alloc_dyn_array(cpu, ptr + old_size);
184 printk(KERN_DEBUG "NR_CPUS: %d, nr_cpu_ids: %d, nr_node_ids %d\n",
185 NR_CPUS, nr_cpu_ids, nr_node_ids);
187 /* Setup percpu data maps */
188 setup_per_cpu_maps();
190 /* Setup node to cpumask map */
191 setup_node_to_cpumask_map();
199 * Allocate node_to_cpumask_map based on number of available nodes
200 * Requires node_possible_map to be valid.
202 * Note: node_to_cpumask() is not valid until after this is done.
204 static void __init setup_node_to_cpumask_map(void)
206 unsigned int node, num = 0;
209 /* setup nr_node_ids if not done yet */
210 if (nr_node_ids == MAX_NUMNODES) {
211 for_each_node_mask(node, node_possible_map)
213 nr_node_ids = num + 1;
216 /* allocate the map */
217 map = alloc_bootmem_low(nr_node_ids * sizeof(cpumask_t));
219 pr_debug(KERN_DEBUG "Node to cpumask map at %p for %d nodes\n",
222 /* node_to_cpumask() will now work */
223 node_to_cpumask_map = map;
226 void __cpuinit numa_set_node(int cpu, int node)
228 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
230 if (cpu_pda(cpu) && node != NUMA_NO_NODE)
231 cpu_pda(cpu)->nodenumber = node;
234 cpu_to_node_map[cpu] = node;
236 else if (per_cpu_offset(cpu))
237 per_cpu(x86_cpu_to_node_map, cpu) = node;
240 pr_debug("Setting node for non-present cpu %d\n", cpu);
243 void __cpuinit numa_clear_node(int cpu)
245 numa_set_node(cpu, NUMA_NO_NODE);
248 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
250 void __cpuinit numa_add_cpu(int cpu)
252 cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
255 void __cpuinit numa_remove_cpu(int cpu)
257 cpu_clear(cpu, node_to_cpumask_map[cpu_to_node(cpu)]);
260 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
263 * --------- debug versions of the numa functions ---------
265 static void __cpuinit numa_set_cpumask(int cpu, int enable)
267 int node = cpu_to_node(cpu);
271 if (node_to_cpumask_map == NULL) {
272 printk(KERN_ERR "node_to_cpumask_map NULL\n");
277 mask = &node_to_cpumask_map[node];
281 cpu_clear(cpu, *mask);
283 cpulist_scnprintf(buf, sizeof(buf), *mask);
284 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
285 enable? "numa_add_cpu":"numa_remove_cpu", cpu, node, buf);
288 void __cpuinit numa_add_cpu(int cpu)
290 numa_set_cpumask(cpu, 1);
293 void __cpuinit numa_remove_cpu(int cpu)
295 numa_set_cpumask(cpu, 0);
298 int cpu_to_node(int cpu)
300 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
302 "cpu_to_node(%d): usage too early!\n", cpu);
304 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
306 return per_cpu(x86_cpu_to_node_map, cpu);
308 EXPORT_SYMBOL(cpu_to_node);
311 * Same function as cpu_to_node() but used if called before the
312 * per_cpu areas are setup.
314 int early_cpu_to_node(int cpu)
316 if (early_per_cpu_ptr(x86_cpu_to_node_map))
317 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
319 if (!per_cpu_offset(cpu)) {
321 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
325 return per_cpu(x86_cpu_to_node_map, cpu);
330 static const cpumask_t cpu_mask_none;
333 * Returns a pointer to the bitmask of CPUs on Node 'node'.
335 const cpumask_t *_node_to_cpumask_ptr(int node)
337 if (node_to_cpumask_map == NULL) {
339 "_node_to_cpumask_ptr(%d): no node_to_cpumask_map!\n",
342 return (const cpumask_t *)&cpu_online_map;
344 if (node >= nr_node_ids) {
346 "_node_to_cpumask_ptr(%d): node > nr_node_ids(%d)\n",
349 return &cpu_mask_none;
351 return &node_to_cpumask_map[node];
353 EXPORT_SYMBOL(_node_to_cpumask_ptr);
356 * Returns a bitmask of CPUs on Node 'node'.
358 * Side note: this function creates the returned cpumask on the stack
359 * so with a high NR_CPUS count, excessive stack space is used. The
360 * node_to_cpumask_ptr function should be used whenever possible.
362 cpumask_t node_to_cpumask(int node)
364 if (node_to_cpumask_map == NULL) {
366 "node_to_cpumask(%d): no node_to_cpumask_map!\n", node);
368 return cpu_online_map;
370 if (node >= nr_node_ids) {
372 "node_to_cpumask(%d): node > nr_node_ids(%d)\n",
375 return cpu_mask_none;
377 return node_to_cpumask_map[node];
379 EXPORT_SYMBOL(node_to_cpumask);
382 * --------- end of debug versions of the numa functions ---------
385 #endif /* CONFIG_DEBUG_PER_CPU_MAPS */
387 #endif /* X86_64_NUMA */