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 #define DISABLE_BRANCH_PROFILING 9 10 /* cpu_feature_enabled() cannot be used this early */ 11 #define USE_EARLY_PGTABLE_L5 12 13 #include <linux/init.h> 14 #include <linux/linkage.h> 15 #include <linux/types.h> 16 #include <linux/kernel.h> 17 #include <linux/string.h> 18 #include <linux/percpu.h> 19 #include <linux/start_kernel.h> 20 #include <linux/io.h> 21 #include <linux/memblock.h> 22 #include <linux/mem_encrypt.h> 23 24 #include <asm/processor.h> 25 #include <asm/proto.h> 26 #include <asm/smp.h> 27 #include <asm/setup.h> 28 #include <asm/desc.h> 29 #include <asm/pgtable.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 39 /* 40 * Manage page tables very early on. 41 */ 42 extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD]; 43 static unsigned int __initdata next_early_pgt; 44 pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX); 45 46 #ifdef CONFIG_X86_5LEVEL 47 unsigned int __pgtable_l5_enabled __initdata; 48 unsigned int pgdir_shift __ro_after_init = 39; 49 EXPORT_SYMBOL(pgdir_shift); 50 unsigned int ptrs_per_p4d __ro_after_init = 1; 51 EXPORT_SYMBOL(ptrs_per_p4d); 52 #endif 53 54 #ifdef CONFIG_DYNAMIC_MEMORY_LAYOUT 55 unsigned long page_offset_base __ro_after_init = __PAGE_OFFSET_BASE_L4; 56 EXPORT_SYMBOL(page_offset_base); 57 unsigned long vmalloc_base __ro_after_init = __VMALLOC_BASE_L4; 58 EXPORT_SYMBOL(vmalloc_base); 59 unsigned long vmemmap_base __ro_after_init = __VMEMMAP_BASE_L4; 60 EXPORT_SYMBOL(vmemmap_base); 61 #endif 62 63 #define __head __section(.head.text) 64 65 static void __head *fixup_pointer(void *ptr, unsigned long physaddr) 66 { 67 return ptr - (void *)_text + (void *)physaddr; 68 } 69 70 static unsigned long __head *fixup_long(void *ptr, unsigned long physaddr) 71 { 72 return fixup_pointer(ptr, physaddr); 73 } 74 75 #ifdef CONFIG_X86_5LEVEL 76 static unsigned int __head *fixup_int(void *ptr, unsigned long physaddr) 77 { 78 return fixup_pointer(ptr, physaddr); 79 } 80 81 static bool __head check_la57_support(unsigned long physaddr) 82 { 83 /* 84 * 5-level paging is detected and enabled at kernel decomression 85 * stage. Only check if it has been enabled there. 86 */ 87 if (!(native_read_cr4() & X86_CR4_LA57)) 88 return false; 89 90 *fixup_int(&__pgtable_l5_enabled, physaddr) = 1; 91 *fixup_int(&pgdir_shift, physaddr) = 48; 92 *fixup_int(&ptrs_per_p4d, physaddr) = 512; 93 *fixup_long(&page_offset_base, physaddr) = __PAGE_OFFSET_BASE_L5; 94 *fixup_long(&vmalloc_base, physaddr) = __VMALLOC_BASE_L5; 95 *fixup_long(&vmemmap_base, physaddr) = __VMEMMAP_BASE_L5; 96 97 return true; 98 } 99 #else 100 static bool __head check_la57_support(unsigned long physaddr) 101 { 102 return false; 103 } 104 #endif 105 106 /* Code in __startup_64() can be relocated during execution, but the compiler 107 * doesn't have to generate PC-relative relocations when accessing globals from 108 * that function. Clang actually does not generate them, which leads to 109 * boot-time crashes. To work around this problem, every global pointer must 110 * be adjusted using fixup_pointer(). 111 */ 112 unsigned long __head __startup_64(unsigned long physaddr, 113 struct boot_params *bp) 114 { 115 unsigned long load_delta, *p; 116 unsigned long pgtable_flags; 117 pgdval_t *pgd; 118 p4dval_t *p4d; 119 pudval_t *pud; 120 pmdval_t *pmd, pmd_entry; 121 pteval_t *mask_ptr; 122 bool la57; 123 int i; 124 unsigned int *next_pgt_ptr; 125 126 la57 = check_la57_support(physaddr); 127 128 /* Is the address too large? */ 129 if (physaddr >> MAX_PHYSMEM_BITS) 130 for (;;); 131 132 /* 133 * Compute the delta between the address I am compiled to run at 134 * and the address I am actually running at. 135 */ 136 load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map); 137 138 /* Is the address not 2M aligned? */ 139 if (load_delta & ~PMD_PAGE_MASK) 140 for (;;); 141 142 /* Activate Secure Memory Encryption (SME) if supported and enabled */ 143 sme_enable(bp); 144 145 /* Include the SME encryption mask in the fixup value */ 146 load_delta += sme_get_me_mask(); 147 148 /* Fixup the physical addresses in the page table */ 149 150 pgd = fixup_pointer(&early_top_pgt, physaddr); 151 p = pgd + pgd_index(__START_KERNEL_map); 152 if (la57) 153 *p = (unsigned long)level4_kernel_pgt; 154 else 155 *p = (unsigned long)level3_kernel_pgt; 156 *p += _PAGE_TABLE_NOENC - __START_KERNEL_map + load_delta; 157 158 if (la57) { 159 p4d = fixup_pointer(&level4_kernel_pgt, physaddr); 160 p4d[511] += load_delta; 161 } 162 163 pud = fixup_pointer(&level3_kernel_pgt, physaddr); 164 pud[510] += load_delta; 165 pud[511] += load_delta; 166 167 pmd = fixup_pointer(level2_fixmap_pgt, physaddr); 168 pmd[506] += load_delta; 169 170 /* 171 * Set up the identity mapping for the switchover. These 172 * entries should *NOT* have the global bit set! This also 173 * creates a bunch of nonsense entries but that is fine -- 174 * it avoids problems around wraparound. 175 */ 176 177 next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr); 178 pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr); 179 pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr); 180 181 pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask(); 182 183 if (la57) { 184 p4d = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr); 185 186 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD; 187 pgd[i + 0] = (pgdval_t)p4d + pgtable_flags; 188 pgd[i + 1] = (pgdval_t)p4d + pgtable_flags; 189 190 i = (physaddr >> P4D_SHIFT) % PTRS_PER_P4D; 191 p4d[i + 0] = (pgdval_t)pud + pgtable_flags; 192 p4d[i + 1] = (pgdval_t)pud + pgtable_flags; 193 } else { 194 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD; 195 pgd[i + 0] = (pgdval_t)pud + pgtable_flags; 196 pgd[i + 1] = (pgdval_t)pud + pgtable_flags; 197 } 198 199 i = (physaddr >> PUD_SHIFT) % PTRS_PER_PUD; 200 pud[i + 0] = (pudval_t)pmd + pgtable_flags; 201 pud[i + 1] = (pudval_t)pmd + pgtable_flags; 202 203 pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL; 204 /* Filter out unsupported __PAGE_KERNEL_* bits: */ 205 mask_ptr = fixup_pointer(&__supported_pte_mask, physaddr); 206 pmd_entry &= *mask_ptr; 207 pmd_entry += sme_get_me_mask(); 208 pmd_entry += physaddr; 209 210 for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) { 211 int idx = i + (physaddr >> PMD_SHIFT) % PTRS_PER_PMD; 212 pmd[idx] = pmd_entry + i * PMD_SIZE; 213 } 214 215 /* 216 * Fixup the kernel text+data virtual addresses. Note that 217 * we might write invalid pmds, when the kernel is relocated 218 * cleanup_highmap() fixes this up along with the mappings 219 * beyond _end. 220 */ 221 222 pmd = fixup_pointer(level2_kernel_pgt, physaddr); 223 for (i = 0; i < PTRS_PER_PMD; i++) { 224 if (pmd[i] & _PAGE_PRESENT) 225 pmd[i] += load_delta; 226 } 227 228 /* 229 * Fixup phys_base - remove the memory encryption mask to obtain 230 * the true physical address. 231 */ 232 *fixup_long(&phys_base, physaddr) += load_delta - sme_get_me_mask(); 233 234 /* Encrypt the kernel and related (if SME is active) */ 235 sme_encrypt_kernel(bp); 236 237 /* 238 * Return the SME encryption mask (if SME is active) to be used as a 239 * modifier for the initial pgdir entry programmed into CR3. 240 */ 241 return sme_get_me_mask(); 242 } 243 244 unsigned long __startup_secondary_64(void) 245 { 246 /* 247 * Return the SME encryption mask (if SME is active) to be used as a 248 * modifier for the initial pgdir entry programmed into CR3. 249 */ 250 return sme_get_me_mask(); 251 } 252 253 /* Wipe all early page tables except for the kernel symbol map */ 254 static void __init reset_early_page_tables(void) 255 { 256 memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1)); 257 next_early_pgt = 0; 258 write_cr3(__sme_pa_nodebug(early_top_pgt)); 259 } 260 261 /* Create a new PMD entry */ 262 int __init __early_make_pgtable(unsigned long address, pmdval_t pmd) 263 { 264 unsigned long physaddr = address - __PAGE_OFFSET; 265 pgdval_t pgd, *pgd_p; 266 p4dval_t p4d, *p4d_p; 267 pudval_t pud, *pud_p; 268 pmdval_t *pmd_p; 269 270 /* Invalid address or early pgt is done ? */ 271 if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt)) 272 return -1; 273 274 again: 275 pgd_p = &early_top_pgt[pgd_index(address)].pgd; 276 pgd = *pgd_p; 277 278 /* 279 * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is 280 * critical -- __PAGE_OFFSET would point us back into the dynamic 281 * range and we might end up looping forever... 282 */ 283 if (!pgtable_l5_enabled()) 284 p4d_p = pgd_p; 285 else if (pgd) 286 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); 287 else { 288 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { 289 reset_early_page_tables(); 290 goto again; 291 } 292 293 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++]; 294 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D); 295 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; 296 } 297 p4d_p += p4d_index(address); 298 p4d = *p4d_p; 299 300 if (p4d) 301 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); 302 else { 303 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { 304 reset_early_page_tables(); 305 goto again; 306 } 307 308 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++]; 309 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD); 310 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; 311 } 312 pud_p += pud_index(address); 313 pud = *pud_p; 314 315 if (pud) 316 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); 317 else { 318 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { 319 reset_early_page_tables(); 320 goto again; 321 } 322 323 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++]; 324 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD); 325 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; 326 } 327 pmd_p[pmd_index(address)] = pmd; 328 329 return 0; 330 } 331 332 int __init early_make_pgtable(unsigned long address) 333 { 334 unsigned long physaddr = address - __PAGE_OFFSET; 335 pmdval_t pmd; 336 337 pmd = (physaddr & PMD_MASK) + early_pmd_flags; 338 339 return __early_make_pgtable(address, pmd); 340 } 341 342 /* Don't add a printk in there. printk relies on the PDA which is not initialized 343 yet. */ 344 static void __init clear_bss(void) 345 { 346 memset(__bss_start, 0, 347 (unsigned long) __bss_stop - (unsigned long) __bss_start); 348 } 349 350 static unsigned long get_cmd_line_ptr(void) 351 { 352 unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr; 353 354 cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32; 355 356 return cmd_line_ptr; 357 } 358 359 static void __init copy_bootdata(char *real_mode_data) 360 { 361 char * command_line; 362 unsigned long cmd_line_ptr; 363 364 /* 365 * If SME is active, this will create decrypted mappings of the 366 * boot data in advance of the copy operations. 367 */ 368 sme_map_bootdata(real_mode_data); 369 370 memcpy(&boot_params, real_mode_data, sizeof boot_params); 371 sanitize_boot_params(&boot_params); 372 cmd_line_ptr = get_cmd_line_ptr(); 373 if (cmd_line_ptr) { 374 command_line = __va(cmd_line_ptr); 375 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE); 376 } 377 378 /* 379 * The old boot data is no longer needed and won't be reserved, 380 * freeing up that memory for use by the system. If SME is active, 381 * we need to remove the mappings that were created so that the 382 * memory doesn't remain mapped as decrypted. 383 */ 384 sme_unmap_bootdata(real_mode_data); 385 } 386 387 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data) 388 { 389 /* 390 * Build-time sanity checks on the kernel image and module 391 * area mappings. (these are purely build-time and produce no code) 392 */ 393 BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map); 394 BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE); 395 BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE); 396 BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0); 397 BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0); 398 BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL)); 399 MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) == 400 (__START_KERNEL & PGDIR_MASK))); 401 BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END); 402 403 cr4_init_shadow(); 404 405 /* Kill off the identity-map trampoline */ 406 reset_early_page_tables(); 407 408 clear_bss(); 409 410 clear_page(init_top_pgt); 411 412 /* 413 * SME support may update early_pmd_flags to include the memory 414 * encryption mask, so it needs to be called before anything 415 * that may generate a page fault. 416 */ 417 sme_early_init(); 418 419 kasan_early_init(); 420 421 idt_setup_early_handler(); 422 423 copy_bootdata(__va(real_mode_data)); 424 425 /* 426 * Load microcode early on BSP. 427 */ 428 load_ucode_bsp(); 429 430 /* set init_top_pgt kernel high mapping*/ 431 init_top_pgt[511] = early_top_pgt[511]; 432 433 x86_64_start_reservations(real_mode_data); 434 } 435 436 void __init x86_64_start_reservations(char *real_mode_data) 437 { 438 /* version is always not zero if it is copied */ 439 if (!boot_params.hdr.version) 440 copy_bootdata(__va(real_mode_data)); 441 442 x86_early_init_platform_quirks(); 443 444 switch (boot_params.hdr.hardware_subarch) { 445 case X86_SUBARCH_INTEL_MID: 446 x86_intel_mid_early_setup(); 447 break; 448 default: 449 break; 450 } 451 452 start_kernel(); 453 } 454