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