arch/x86/kernel/head64.c

   1 /*
   2  *  prepare to run common code
   3  *
   4  *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
   5  */
   6
   7 #include <linux/init.h>
   8 #include <linux/linkage.h>
   9 #include <linux/types.h>
  10 #include <linux/kernel.h>
  11 #include <linux/string.h>
  12 #include <linux/percpu.h>
  13 #include <linux/start_kernel.h>
  14 #include <linux/io.h>
  15 #include <linux/memblock.h>
  16
  17 #include <asm/processor.h>
  18 #include <asm/proto.h>
  19 #include <asm/smp.h>
  20 #include <asm/setup.h>
  21 #include <asm/desc.h>
  22 #include <asm/pgtable.h>
  23 #include <asm/tlbflush.h>
  24 #include <asm/sections.h>
  25 #include <asm/kdebug.h>
  26 #include <asm/e820.h>
  27 #include <asm/bios_ebda.h>
  28 #include <asm/bootparam_utils.h>
  29 #include <asm/microcode.h>
  30 #include <asm/kasan.h>
  31
  32 /*
  33  * Manage page tables very early on.
  34  */
  35 extern pgd_t early_level4_pgt[PTRS_PER_PGD];
  36 extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD];
  37 static unsigned int __initdata next_early_pgt = 2;
  38 pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX);
  39
  40 /* Wipe all early page tables except for the kernel symbol map */
  41 static void __init reset_early_page_tables(void)
  42 {
  43         unsigned long i;
  44
  45         for (i = 0; i < PTRS_PER_PGD-1; i++)
  46                 early_level4_pgt[i].pgd = 0;
  47
  48         next_early_pgt = 0;
  49
  50         write_cr3(__pa_nodebug(early_level4_pgt));
  51 }
  52
  53 /* Create a new PMD entry */
  54 int __init early_make_pgtable(unsigned long address)
  55 {
  56         unsigned long physaddr = address - __PAGE_OFFSET;
  57         unsigned long i;
  58         pgdval_t pgd, *pgd_p;
  59         pudval_t pud, *pud_p;
  60         pmdval_t pmd, *pmd_p;
  61
  62         /* Invalid address or early pgt is done ?  */
  63         if (physaddr >= MAXMEM || read_cr3() != __pa_nodebug(early_level4_pgt))
  64                 return -1;
  65
  66 again:
  67         pgd_p = &early_level4_pgt[pgd_index(address)].pgd;
  68         pgd = *pgd_p;
  69
  70         /*
  71          * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
  72          * critical -- __PAGE_OFFSET would point us back into the dynamic
  73          * range and we might end up looping forever...
  74          */
  75         if (pgd)
  76                 pud_p = (pudval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
  77         else {
  78                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
  79                         reset_early_page_tables();
  80                         goto again;
  81                 }
  82
  83                 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
  84                 for (i = 0; i < PTRS_PER_PUD; i++)
  85                         pud_p[i] = 0;
  86                 *pgd_p = (pgdval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
  87         }
  88         pud_p += pud_index(address);
  89         pud = *pud_p;
  90
  91         if (pud)
  92                 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
  93         else {
  94                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
  95                         reset_early_page_tables();
  96                         goto again;
  97                 }
  98
  99                 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
 100                 for (i = 0; i < PTRS_PER_PMD; i++)
 101                         pmd_p[i] = 0;
 102                 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 103         }
 104         pmd = (physaddr & PMD_MASK) + early_pmd_flags;
 105         pmd_p[pmd_index(address)] = pmd;
 106
 107         return 0;
 108 }
 109
 110 /* Don't add a printk in there. printk relies on the PDA which is not initialized
 111    yet. */
 112 static void __init clear_bss(void)
 113 {
 114         memset(__bss_start, 0,
 115                (unsigned long) __bss_stop - (unsigned long) __bss_start);
 116 }
 117
 118 static unsigned long get_cmd_line_ptr(void)
 119 {
 120         unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
 121
 122         cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
 123
 124         return cmd_line_ptr;
 125 }
 126
 127 static void __init copy_bootdata(char *real_mode_data)
 128 {
 129         char * command_line;
 130         unsigned long cmd_line_ptr;
 131
 132         memcpy(&boot_params, real_mode_data, sizeof boot_params);
 133         sanitize_boot_params(&boot_params);
 134         cmd_line_ptr = get_cmd_line_ptr();
 135         if (cmd_line_ptr) {
 136                 command_line = __va(cmd_line_ptr);
 137                 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 138         }
 139 }
 140
 141 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
 142 {
 143         int i;
 144
 145         /*
 146          * Build-time sanity checks on the kernel image and module
 147          * area mappings. (these are purely build-time and produce no code)
 148          */
 149         BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
 150         BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
 151         BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
 152         BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
 153         BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
 154         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 155         BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 156                                 (__START_KERNEL & PGDIR_MASK)));
 157         BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
 158
 159         cr4_init_shadow();
 160
 161         /* Kill off the identity-map trampoline */
 162         reset_early_page_tables();
 163
 164         clear_bss();
 165
 166         clear_page(init_level4_pgt);
 167
 168         kasan_early_init();
 169
 170         for (i = 0; i < NUM_EXCEPTION_VECTORS; i++)
 171                 set_intr_gate(i, early_idt_handler_array[i]);
 172         load_idt((const struct desc_ptr *)&idt_descr);
 173
 174         copy_bootdata(__va(real_mode_data));
 175
 176         /*
 177          * Load microcode early on BSP.
 178          */
 179         load_ucode_bsp();
 180
 181         /* set init_level4_pgt kernel high mapping*/
 182         init_level4_pgt[511] = early_level4_pgt[511];
 183
 184         x86_64_start_reservations(real_mode_data);
 185 }
 186
 187 void __init x86_64_start_reservations(char *real_mode_data)
 188 {
 189         /* version is always not zero if it is copied */
 190         if (!boot_params.hdr.version)
 191                 copy_bootdata(__va(real_mode_data));
 192
 193         reserve_ebda_region();
 194
 195         start_kernel();
 196 }