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 #define DISABLE_BRANCH_PROFILING
   8 #include <linux/init.h>
   9 #include <linux/linkage.h>
  10 #include <linux/types.h>
  11 #include <linux/kernel.h>
  12 #include <linux/string.h>
  13 #include <linux/percpu.h>
  14 #include <linux/start_kernel.h>
  15 #include <linux/io.h>
  16 #include <linux/memblock.h>
  17
  18 #include <asm/processor.h>
  19 #include <asm/proto.h>
  20 #include <asm/smp.h>
  21 #include <asm/setup.h>
  22 #include <asm/desc.h>
  23 #include <asm/pgtable.h>
  24 #include <asm/tlbflush.h>
  25 #include <asm/sections.h>
  26 #include <asm/kdebug.h>
  27 #include <asm/e820.h>
  28 #include <asm/bios_ebda.h>
  29 #include <asm/bootparam_utils.h>
  30 #include <asm/microcode.h>
  31 #include <asm/kasan.h>
  32
  33 /*
  34  * Manage page tables very early on.
  35  */
  36 extern pgd_t early_level4_pgt[PTRS_PER_PGD];
  37 extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD];
  38 static unsigned int __initdata next_early_pgt = 2;
  39 pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX);
  40
  41 /* Wipe all early page tables except for the kernel symbol map */
  42 static void __init reset_early_page_tables(void)
  43 {
  44         memset(early_level4_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
  45         next_early_pgt = 0;
  46         write_cr3(__pa_nodebug(early_level4_pgt));
  47 }
  48
  49 /* Create a new PMD entry */
  50 int __init early_make_pgtable(unsigned long address)
  51 {
  52         unsigned long physaddr = address - __PAGE_OFFSET;
  53         pgdval_t pgd, *pgd_p;
  54         pudval_t pud, *pud_p;
  55         pmdval_t pmd, *pmd_p;
  56
  57         /* Invalid address or early pgt is done ?  */
  58         if (physaddr >= MAXMEM || read_cr3() != __pa_nodebug(early_level4_pgt))
  59                 return -1;
  60
  61 again:
  62         pgd_p = &early_level4_pgt[pgd_index(address)].pgd;
  63         pgd = *pgd_p;
  64
  65         /*
  66          * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
  67          * critical -- __PAGE_OFFSET would point us back into the dynamic
  68          * range and we might end up looping forever...
  69          */
  70         if (pgd)
  71                 pud_p = (pudval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
  72         else {
  73                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
  74                         reset_early_page_tables();
  75                         goto again;
  76                 }
  77
  78                 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
  79                 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
  80                 *pgd_p = (pgdval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
  81         }
  82         pud_p += pud_index(address);
  83         pud = *pud_p;
  84
  85         if (pud)
  86                 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
  87         else {
  88                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
  89                         reset_early_page_tables();
  90                         goto again;
  91                 }
  92
  93                 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
  94                 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
  95                 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
  96         }
  97         pmd = (physaddr & PMD_MASK) + early_pmd_flags;
  98         pmd_p[pmd_index(address)] = pmd;
  99
 100         return 0;
 101 }
 102
 103 /* Don't add a printk in there. printk relies on the PDA which is not initialized
 104    yet. */
 105 static void __init clear_bss(void)
 106 {
 107         memset(__bss_start, 0,
 108                (unsigned long) __bss_stop - (unsigned long) __bss_start);
 109         memset(__brk_base, 0,
 110                (unsigned long) __brk_limit - (unsigned long) __brk_base);
 111 }
 112
 113 static unsigned long get_cmd_line_ptr(void)
 114 {
 115         unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
 116
 117         cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
 118
 119         return cmd_line_ptr;
 120 }
 121
 122 static void __init copy_bootdata(char *real_mode_data)
 123 {
 124         char * command_line;
 125         unsigned long cmd_line_ptr;
 126
 127         memcpy(&boot_params, real_mode_data, sizeof boot_params);
 128         sanitize_boot_params(&boot_params);
 129         cmd_line_ptr = get_cmd_line_ptr();
 130         if (cmd_line_ptr) {
 131                 command_line = __va(cmd_line_ptr);
 132                 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 133         }
 134 }
 135
 136 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
 137 {
 138         int i;
 139
 140         /*
 141          * Build-time sanity checks on the kernel image and module
 142          * area mappings. (these are purely build-time and produce no code)
 143          */
 144         BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
 145         BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
 146         BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
 147         BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
 148         BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
 149         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 150         BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 151                                 (__START_KERNEL & PGDIR_MASK)));
 152         BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
 153
 154         cr4_init_shadow();
 155
 156         /* Kill off the identity-map trampoline */
 157         reset_early_page_tables();
 158
 159         clear_bss();
 160
 161         clear_page(init_level4_pgt);
 162
 163         kasan_early_init();
 164
 165         for (i = 0; i < NUM_EXCEPTION_VECTORS; i++)
 166                 set_intr_gate(i, early_idt_handler_array[i]);
 167         load_idt((const struct desc_ptr *)&idt_descr);
 168
 169         copy_bootdata(__va(real_mode_data));
 170
 171         /*
 172          * Load microcode early on BSP.
 173          */
 174         load_ucode_bsp();
 175
 176         /* set init_level4_pgt kernel high mapping*/
 177         init_level4_pgt[511] = early_level4_pgt[511];
 178
 179         x86_64_start_reservations(real_mode_data);
 180 }
 181
 182 void __init x86_64_start_reservations(char *real_mode_data)
 183 {
 184         /* version is always not zero if it is copied */
 185         if (!boot_params.hdr.version)
 186                 copy_bootdata(__va(real_mode_data));
 187
 188         x86_early_init_platform_quirks();
 189
 190         switch (boot_params.hdr.hardware_subarch) {
 191         case X86_SUBARCH_INTEL_MID:
 192                 x86_intel_mid_early_setup();
 193                 break;
 194         default:
 195                 break;
 196         }
 197
 198         start_kernel();
 199 }