arch/x86/kernel/setup_percpu.c

   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>
   8 #include <linux/smp.h>
   9 #include <linux/topology.h>
  10 #include <asm/sections.h>
  11 #include <asm/processor.h>
  12 #include <asm/setup.h>
  13 #include <asm/mpspec.h>
  14 #include <asm/apicdef.h>
  15 #include <asm/highmem.h>
  16 #include <asm/proto.h>
  17 #include <asm/cpumask.h>
  18 #include <asm/cpu.h>
  19 #include <asm/stackprotector.h>
  20
  21 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
  22 # define DBG(x...) printk(KERN_DEBUG x)
  23 #else
  24 # define DBG(x...)
  25 #endif
  26
  27 DEFINE_PER_CPU(int, cpu_number);
  28 EXPORT_PER_CPU_SYMBOL(cpu_number);
  29
  30 #ifdef CONFIG_X86_64
  31 #define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
  32 #else
  33 #define BOOT_PERCPU_OFFSET 0
  34 #endif
  35
  36 DEFINE_PER_CPU(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
  37 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
  38
  39 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly = {
  40         [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
  41 };
  42 EXPORT_SYMBOL(__per_cpu_offset);
  43
  44 static inline void setup_percpu_segment(int cpu)
  45 {
  46 #ifdef CONFIG_X86_32
  47         struct desc_struct gdt;
  48
  49         pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
  50                         0x2 | DESCTYPE_S, 0x8);
  51         gdt.s = 1;
  52         write_gdt_entry(get_cpu_gdt_table(cpu),
  53                         GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
  54 #endif
  55 }
  56
  57 /*
  58  * Great future plan:
  59  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
  60  * Always point %gs to its beginning
  61  */
  62 void __init setup_per_cpu_areas(void)
  63 {
  64         ssize_t size = __per_cpu_end - __per_cpu_start;
  65         unsigned int nr_cpu_pages = DIV_ROUND_UP(size, PAGE_SIZE);
  66         static struct page **pages;
  67         size_t pages_size;
  68         unsigned int cpu, i, j;
  69         unsigned long delta;
  70         size_t pcpu_unit_size;
  71
  72         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
  73                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
  74         pr_info("PERCPU: Allocating %zd bytes for static per cpu data\n", size);
  75
  76         pages_size = nr_cpu_pages * num_possible_cpus() * sizeof(pages[0]);
  77         pages = alloc_bootmem(pages_size);
  78
  79         j = 0;
  80         for_each_possible_cpu(cpu) {
  81                 void *ptr;
  82
  83                 for (i = 0; i < nr_cpu_pages; i++) {
  84 #ifndef CONFIG_NEED_MULTIPLE_NODES
  85                         ptr = alloc_bootmem_pages(PAGE_SIZE);
  86 #else
  87                         int node = early_cpu_to_node(cpu);
  88
  89                         if (!node_online(node) || !NODE_DATA(node)) {
  90                                 ptr = alloc_bootmem_pages(PAGE_SIZE);
  91                                 pr_info("cpu %d has no node %d or node-local "
  92                                         "memory\n", cpu, node);
  93                                 pr_debug("per cpu data for cpu%d at %016lx\n",
  94                                          cpu, __pa(ptr));
  95                         } else {
  96                                 ptr = alloc_bootmem_pages_node(NODE_DATA(node),
  97                                                                PAGE_SIZE);
  98                                 pr_debug("per cpu data for cpu%d on node%d "
  99                                          "at %016lx\n", cpu, node, __pa(ptr));
 100                         }
 101 #endif
 102                         memcpy(ptr, __per_cpu_load + i * PAGE_SIZE, PAGE_SIZE);
 103                         pages[j++] = virt_to_page(ptr);
 104                 }
 105         }
 106
 107         pcpu_unit_size = pcpu_setup_static(populate_extra_pte, pages, size);
 108
 109         free_bootmem(__pa(pages), pages_size);
 110
 111         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 112         for_each_possible_cpu(cpu) {
 113                 per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
 114                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 115                 per_cpu(cpu_number, cpu) = cpu;
 116                 setup_percpu_segment(cpu);
 117                 setup_stack_canary_segment(cpu);
 118                 /*
 119                  * Copy data used in early init routines from the
 120                  * initial arrays to the per cpu data areas.  These
 121                  * arrays then become expendable and the *_early_ptr's
 122                  * are zeroed indicating that the static arrays are
 123                  * gone.
 124                  */
 125 #ifdef CONFIG_X86_LOCAL_APIC
 126                 per_cpu(x86_cpu_to_apicid, cpu) =
 127                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 128                 per_cpu(x86_bios_cpu_apicid, cpu) =
 129                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 130 #endif
 131 #ifdef CONFIG_X86_64
 132                 per_cpu(irq_stack_ptr, cpu) =
 133                         per_cpu(irq_stack_union.irq_stack, cpu) +
 134                         IRQ_STACK_SIZE - 64;
 135 #ifdef CONFIG_NUMA
 136                 per_cpu(x86_cpu_to_node_map, cpu) =
 137                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
 138 #endif
 139 #endif
 140                 /*
 141                  * Up to this point, the boot CPU has been using .data.init
 142                  * area.  Reload any changed state for the boot CPU.
 143                  */
 144                 if (cpu == boot_cpu_id)
 145                         switch_to_new_gdt(cpu);
 146
 147                 DBG("PERCPU: cpu %4d %p\n", cpu, ptr);
 148         }
 149
 150         /* indicate the early static arrays will soon be gone */
 151 #ifdef CONFIG_X86_LOCAL_APIC
 152         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 153         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 154 #endif
 155 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 156         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 157 #endif
 158
 159         /* Setup node to cpumask map */
 160         setup_node_to_cpumask_map();
 161
 162         /* Setup cpu initialized, callin, callout masks */
 163         setup_cpu_local_masks();
 164 }