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/api.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;
  39 pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX);
  40
  41 static void __init *fixup_pointer(void *ptr, unsigned long physaddr)
  42 {
  43         return ptr - (void *)_text + (void *)physaddr;
  44 }
  45
  46 void __init __startup_64(unsigned long physaddr)
  47 {
  48         unsigned long load_delta, *p;
  49         pgdval_t *pgd;
  50         pudval_t *pud;
  51         pmdval_t *pmd, pmd_entry;
  52         int i;
  53
  54         /* Is the address too large? */
  55         if (physaddr >> MAX_PHYSMEM_BITS)
  56                 for (;;);
  57
  58         /*
  59          * Compute the delta between the address I am compiled to run at
  60          * and the address I am actually running at.
  61          */
  62         load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
  63
  64         /* Is the address not 2M aligned? */
  65         if (load_delta & ~PMD_PAGE_MASK)
  66                 for (;;);
  67
  68         /* Fixup the physical addresses in the page table */
  69
  70         pgd = fixup_pointer(&early_level4_pgt, physaddr);
  71         pgd[pgd_index(__START_KERNEL_map)] += load_delta;
  72
  73         pud = fixup_pointer(&level3_kernel_pgt, physaddr);
  74         pud[510] += load_delta;
  75         pud[511] += load_delta;
  76
  77         pmd = fixup_pointer(level2_fixmap_pgt, physaddr);
  78         pmd[506] += load_delta;
  79
  80         /*
  81          * Set up the identity mapping for the switchover.  These
  82          * entries should *NOT* have the global bit set!  This also
  83          * creates a bunch of nonsense entries but that is fine --
  84          * it avoids problems around wraparound.
  85          */
  86
  87         pud = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr);
  88         pmd = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr);
  89
  90         i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
  91         pgd[i + 0] = (pgdval_t)pud + _KERNPG_TABLE;
  92         pgd[i + 1] = (pgdval_t)pud + _KERNPG_TABLE;
  93
  94         i = (physaddr >> PUD_SHIFT) % PTRS_PER_PUD;
  95         pud[i + 0] = (pudval_t)pmd + _KERNPG_TABLE;
  96         pud[i + 1] = (pudval_t)pmd + _KERNPG_TABLE;
  97
  98         pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
  99         pmd_entry +=  physaddr;
 100
 101         for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
 102                 int idx = i + (physaddr >> PMD_SHIFT) % PTRS_PER_PMD;
 103                 pmd[idx] = pmd_entry + i * PMD_SIZE;
 104         }
 105
 106         /*
 107          * Fixup the kernel text+data virtual addresses. Note that
 108          * we might write invalid pmds, when the kernel is relocated
 109          * cleanup_highmap() fixes this up along with the mappings
 110          * beyond _end.
 111          */
 112
 113         pmd = fixup_pointer(level2_kernel_pgt, physaddr);
 114         for (i = 0; i < PTRS_PER_PMD; i++) {
 115                 if (pmd[i] & _PAGE_PRESENT)
 116                         pmd[i] += load_delta;
 117         }
 118
 119         /* Fixup phys_base */
 120         p = fixup_pointer(&phys_base, physaddr);
 121         *p += load_delta;
 122 }
 123
 124 /* Wipe all early page tables except for the kernel symbol map */
 125 static void __init reset_early_page_tables(void)
 126 {
 127         memset(early_level4_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
 128         next_early_pgt = 0;
 129         write_cr3(__pa_nodebug(early_level4_pgt));
 130 }
 131
 132 /* Create a new PMD entry */
 133 int __init early_make_pgtable(unsigned long address)
 134 {
 135         unsigned long physaddr = address - __PAGE_OFFSET;
 136         pgdval_t pgd, *pgd_p;
 137         pudval_t pud, *pud_p;
 138         pmdval_t pmd, *pmd_p;
 139
 140         /* Invalid address or early pgt is done ?  */
 141         if (physaddr >= MAXMEM ||
 142             read_cr3_pa() != __pa_nodebug(early_level4_pgt))
 143                 return -1;
 144
 145 again:
 146         pgd_p = &early_level4_pgt[pgd_index(address)].pgd;
 147         pgd = *pgd_p;
 148
 149         /*
 150          * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
 151          * critical -- __PAGE_OFFSET would point us back into the dynamic
 152          * range and we might end up looping forever...
 153          */
 154         if (pgd)
 155                 pud_p = (pudval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 156         else {
 157                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 158                         reset_early_page_tables();
 159                         goto again;
 160                 }
 161
 162                 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
 163                 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
 164                 *pgd_p = (pgdval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 165         }
 166         pud_p += pud_index(address);
 167         pud = *pud_p;
 168
 169         if (pud)
 170                 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 171         else {
 172                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 173                         reset_early_page_tables();
 174                         goto again;
 175                 }
 176
 177                 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
 178                 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
 179                 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 180         }
 181         pmd = (physaddr & PMD_MASK) + early_pmd_flags;
 182         pmd_p[pmd_index(address)] = pmd;
 183
 184         return 0;
 185 }
 186
 187 /* Don't add a printk in there. printk relies on the PDA which is not initialized
 188    yet. */
 189 static void __init clear_bss(void)
 190 {
 191         memset(__bss_start, 0,
 192                (unsigned long) __bss_stop - (unsigned long) __bss_start);
 193 }
 194
 195 static unsigned long get_cmd_line_ptr(void)
 196 {
 197         unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
 198
 199         cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
 200
 201         return cmd_line_ptr;
 202 }
 203
 204 static void __init copy_bootdata(char *real_mode_data)
 205 {
 206         char * command_line;
 207         unsigned long cmd_line_ptr;
 208
 209         memcpy(&boot_params, real_mode_data, sizeof boot_params);
 210         sanitize_boot_params(&boot_params);
 211         cmd_line_ptr = get_cmd_line_ptr();
 212         if (cmd_line_ptr) {
 213                 command_line = __va(cmd_line_ptr);
 214                 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 215         }
 216 }
 217
 218 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
 219 {
 220         int i;
 221
 222         /*
 223          * Build-time sanity checks on the kernel image and module
 224          * area mappings. (these are purely build-time and produce no code)
 225          */
 226         BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
 227         BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
 228         BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
 229         BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
 230         BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
 231         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 232         BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 233                                 (__START_KERNEL & PGDIR_MASK)));
 234         BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
 235
 236         cr4_init_shadow();
 237
 238         /* Kill off the identity-map trampoline */
 239         reset_early_page_tables();
 240
 241         clear_bss();
 242
 243         clear_page(init_level4_pgt);
 244
 245         kasan_early_init();
 246
 247         for (i = 0; i < NUM_EXCEPTION_VECTORS; i++)
 248                 set_intr_gate(i, early_idt_handler_array[i]);
 249         load_idt((const struct desc_ptr *)&idt_descr);
 250
 251         copy_bootdata(__va(real_mode_data));
 252
 253         /*
 254          * Load microcode early on BSP.
 255          */
 256         load_ucode_bsp();
 257
 258         /* set init_level4_pgt kernel high mapping*/
 259         init_level4_pgt[511] = early_level4_pgt[511];
 260
 261         x86_64_start_reservations(real_mode_data);
 262 }
 263
 264 void __init x86_64_start_reservations(char *real_mode_data)
 265 {
 266         /* version is always not zero if it is copied */
 267         if (!boot_params.hdr.version)
 268                 copy_bootdata(__va(real_mode_data));
 269
 270         x86_early_init_platform_quirks();
 271
 272         switch (boot_params.hdr.hardware_subarch) {
 273         case X86_SUBARCH_INTEL_MID:
 274                 x86_intel_mid_early_setup();
 275                 break;
 276         default:
 277                 break;
 278         }
 279
 280         start_kernel();
 281 }