개념 설명 전체 · v6.18.37 / arch/x86/kernel/head64.c
1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * prepare to run common code 4 * 5 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE 6 */ 7 8 /* cpu_feature_enabled() cannot be used this early */ 9 #define USE_EARLY_PGTABLE_L5 10 11 #include <linux/init.h> 12 #include <linux/linkage.h> 13 #include <linux/types.h> 14 #include <linux/kernel.h> 15 #include <linux/string.h> 16 #include <linux/percpu.h> 17 #include <linux/start_kernel.h> 18 #include <linux/io.h> 19 #include <linux/memblock.h> 20 #include <linux/cc_platform.h> 21 #include <linux/pgtable.h> 22 23 #include <asm/asm.h> 24 #include <asm/page_64.h> 25 #include <asm/processor.h> 26 #include <asm/proto.h> 27 #include <asm/smp.h> 28 #include <asm/setup.h> 29 #include <asm/desc.h> 30 #include <asm/tlbflush.h> 31 #include <asm/sections.h> 32 #include <asm/kdebug.h> 33 #include <asm/e820/api.h> 34 #include <asm/bios_ebda.h> 35 #include <asm/bootparam_utils.h> 36 #include <asm/microcode.h> 37 #include <asm/kasan.h> 38 #include <asm/fixmap.h> 39 #include <asm/realmode.h> 40 #include <asm/extable.h> 41 #include <asm/trapnr.h> 42 #include <asm/sev.h> 43 #include <asm/tdx.h> 44 #include <asm/init.h> 45 46 /* 47 * Manage page tables very early on. 48 */ 49 extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD]; 50 unsigned int __initdata next_early_pgt; 51 SYM_PIC_ALIAS(next_early_pgt); 52 pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX); 53 54 unsigned int __pgtable_l5_enabled __ro_after_init; 55 SYM_PIC_ALIAS(__pgtable_l5_enabled); 56 unsigned int pgdir_shift __ro_after_init = 39; 57 EXPORT_SYMBOL(pgdir_shift); 58 SYM_PIC_ALIAS(pgdir_shift); 59 unsigned int ptrs_per_p4d __ro_after_init = 1; 60 EXPORT_SYMBOL(ptrs_per_p4d); 61 SYM_PIC_ALIAS(ptrs_per_p4d); 62 63 unsigned long page_offset_base __ro_after_init = __PAGE_OFFSET_BASE_L4; 64 EXPORT_SYMBOL(page_offset_base); 65 unsigned long vmalloc_base __ro_after_init = __VMALLOC_BASE_L4; 66 EXPORT_SYMBOL(vmalloc_base); 67 unsigned long vmemmap_base __ro_after_init = __VMEMMAP_BASE_L4; 68 EXPORT_SYMBOL(vmemmap_base); 69 70 /* Wipe all early page tables except for the kernel symbol map */ 71 static void __init reset_early_page_tables(void) 72 { 73 memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1)); 74 next_early_pgt = 0; 75 write_cr3(__sme_pa_nodebug(early_top_pgt)); 76 } 77 78 /* Create a new PMD entry */ 79 bool __init __early_make_pgtable(unsigned long address, pmdval_t pmd) 80 { 81 unsigned long physaddr = address - __PAGE_OFFSET; 82 pgdval_t pgd, *pgd_p; 83 p4dval_t p4d, *p4d_p; 84 pudval_t pud, *pud_p; 85 pmdval_t *pmd_p; 86 87 /* Invalid address or early pgt is done ? */ 88 if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt)) 89 return false; 90 91 again: 92 pgd_p = &early_top_pgt[pgd_index(address)].pgd; 93 pgd = *pgd_p; 94 95 /* 96 * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is 97 * critical -- __PAGE_OFFSET would point us back into the dynamic 98 * range and we might end up looping forever... 99 */ 100 if (!pgtable_l5_enabled()) 101 p4d_p = pgd_p; 102 else if (pgd) 103 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); 104 else { 105 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { 106 reset_early_page_tables(); 107 goto again; 108 } 109 110 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++]; 111 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D); 112 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; 113 } 114 p4d_p += p4d_index(address); 115 p4d = *p4d_p; 116 117 if (p4d) 118 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); 119 else { 120 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { 121 reset_early_page_tables(); 122 goto again; 123 } 124 125 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++]; 126 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD); 127 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; 128 } 129 pud_p += pud_index(address); 130 pud = *pud_p; 131 132 if (pud) 133 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); 134 else { 135 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { 136 reset_early_page_tables(); 137 goto again; 138 } 139 140 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++]; 141 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD); 142 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; 143 } 144 pmd_p[pmd_index(address)] = pmd; 145 146 return true; 147 } 148 149 static bool __init early_make_pgtable(unsigned long address) 150 { 151 unsigned long physaddr = address - __PAGE_OFFSET; 152 pmdval_t pmd; 153 154 pmd = (physaddr & PMD_MASK) + early_pmd_flags; 155 156 return __early_make_pgtable(address, pmd); 157 } 158 159 void __init do_early_exception(struct pt_regs *regs, int trapnr) 160 { 161 if (trapnr == X86_TRAP_PF && 162 early_make_pgtable(native_read_cr2())) 163 return; 164 165 if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT) && 166 trapnr == X86_TRAP_VC && handle_vc_boot_ghcb(regs)) 167 return; 168 169 if (trapnr == X86_TRAP_VE && tdx_early_handle_ve(regs)) 170 return; 171 172 early_fixup_exception(regs, trapnr); 173 } 174 175 /* Don't add a printk in there. printk relies on the PDA which is not initialized 176 yet. */ 177 void __init clear_bss(void) 178 { 179 memset(__bss_start, 0, 180 (unsigned long) __bss_stop - (unsigned long) __bss_start); 181 memset(__brk_base, 0, 182 (unsigned long) __brk_limit - (unsigned long) __brk_base); 183 } 184 185 static unsigned long get_cmd_line_ptr(void) 186 { 187 unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr; 188 189 cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32; 190 191 return cmd_line_ptr; 192 } 193 194 static void __init copy_bootdata(char *real_mode_data) 195 { 196 char * command_line; 197 unsigned long cmd_line_ptr; 198 199 /* 200 * If SME is active, this will create decrypted mappings of the 201 * boot data in advance of the copy operations. 202 */ 203 sme_map_bootdata(real_mode_data); 204 205 memcpy(&boot_params, real_mode_data, sizeof(boot_params)); 206 sanitize_boot_params(&boot_params); 207 cmd_line_ptr = get_cmd_line_ptr(); 208 if (cmd_line_ptr) { 209 command_line = __va(cmd_line_ptr); 210 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE); 211 } 212 213 /* 214 * The old boot data is no longer needed and won't be reserved, 215 * freeing up that memory for use by the system. If SME is active, 216 * we need to remove the mappings that were created so that the 217 * memory doesn't remain mapped as decrypted. 218 */ 219 sme_unmap_bootdata(real_mode_data); 220 } 221 222 asmlinkage __visible void __init __noreturn x86_64_start_kernel(char * real_mode_data) 223 { 224 /* 225 * Build-time sanity checks on the kernel image and module 226 * area mappings. (these are purely build-time and produce no code) 227 */ 228 BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map); 229 BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE); 230 BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE); 231 BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0); 232 BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0); 233 BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL)); 234 MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) == 235 (__START_KERNEL & PGDIR_MASK))); 236 BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END); 237 238 cr4_init_shadow(); 239 240 /* Kill off the identity-map trampoline */ 241 reset_early_page_tables(); 242 243 if (pgtable_l5_enabled()) { 244 page_offset_base = __PAGE_OFFSET_BASE_L5; 245 vmalloc_base = __VMALLOC_BASE_L5; 246 vmemmap_base = __VMEMMAP_BASE_L5; 247 } 248 249 clear_bss(); 250 251 /* 252 * This needs to happen *before* kasan_early_init() because latter maps stuff 253 * into that page. 254 */ 255 clear_page(init_top_pgt); 256 257 /* 258 * SME support may update early_pmd_flags to include the memory 259 * encryption mask, so it needs to be called before anything 260 * that may generate a page fault. 261 */ 262 sme_early_init(); 263 264 kasan_early_init(); 265 266 /* 267 * Flush global TLB entries which could be left over from the trampoline page 268 * table. 269 * 270 * This needs to happen *after* kasan_early_init() as KASAN-enabled .configs 271 * instrument native_write_cr4() so KASAN must be initialized for that 272 * instrumentation to work. 273 */ 274 __native_tlb_flush_global(this_cpu_read(cpu_tlbstate.cr4)); 275 276 idt_setup_early_handler(); 277 278 /* Needed before cc_platform_has() can be used for TDX */ 279 tdx_early_init(); 280 281 copy_bootdata(__va(real_mode_data)); 282 283 /* 284 * Load microcode early on BSP. 285 */ 286 load_ucode_bsp(); 287 288 /* set init_top_pgt kernel high mapping*/ 289 init_top_pgt[511] = early_top_pgt[511]; 290 291 x86_64_start_reservations(real_mode_data); 292 } 293 294 void __init __noreturn x86_64_start_reservations(char *real_mode_data) 295 { 296 /* version is always not zero if it is copied */ 297 if (!boot_params.hdr.version) 298 copy_bootdata(__va(real_mode_data)); 299 300 x86_early_init_platform_quirks(); 301 302 switch (boot_params.hdr.hardware_subarch) { 303 case X86_SUBARCH_INTEL_MID: 304 x86_intel_mid_early_setup(); 305 break; 306 default: 307 break; 308 } 309 310 start_kernel(); 311 } 312 313 void early_setup_idt(void) 314 { 315 void *handler = NULL; 316 317 if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT)) { 318 setup_ghcb(); 319 handler = vc_boot_ghcb; 320 } 321 322 __pi_startup_64_load_idt(handler); 323 }