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 #include <asm/fixmap.h> 39 40 /* 41 * Manage page tables very early on. 42 */ 43 extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD]; 44 static unsigned int __initdata next_early_pgt; 45 pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX); 46 47 #ifdef CONFIG_X86_5LEVEL 48 unsigned int __pgtable_l5_enabled __ro_after_init; 49 unsigned int pgdir_shift __ro_after_init = 39; 50 EXPORT_SYMBOL(pgdir_shift); 51 unsigned int ptrs_per_p4d __ro_after_init = 1; 52 EXPORT_SYMBOL(ptrs_per_p4d); 53 #endif 54 55 #ifdef CONFIG_DYNAMIC_MEMORY_LAYOUT 56 unsigned long page_offset_base __ro_after_init = __PAGE_OFFSET_BASE_L4; 57 EXPORT_SYMBOL(page_offset_base); 58 unsigned long vmalloc_base __ro_after_init = __VMALLOC_BASE_L4; 59 EXPORT_SYMBOL(vmalloc_base); 60 unsigned long vmemmap_base __ro_after_init = __VMEMMAP_BASE_L4; 61 EXPORT_SYMBOL(vmemmap_base); 62 #endif 63 64 #define __head __section(.head.text) 65 66 static void __head *fixup_pointer(void *ptr, unsigned long physaddr) 67 { 68 return ptr - (void *)_text + (void *)physaddr; 69 } 70 71 static unsigned long __head *fixup_long(void *ptr, unsigned long physaddr) 72 { 73 return fixup_pointer(ptr, physaddr); 74 } 75 76 #ifdef CONFIG_X86_5LEVEL 77 static unsigned int __head *fixup_int(void *ptr, unsigned long physaddr) 78 { 79 return fixup_pointer(ptr, physaddr); 80 } 81 82 static bool __head check_la57_support(unsigned long physaddr) 83 { 84 /* 85 * 5-level paging is detected and enabled at kernel decomression 86 * stage. Only check if it has been enabled there. 87 */ 88 if (!(native_read_cr4() & X86_CR4_LA57)) 89 return false; 90 91 *fixup_int(&__pgtable_l5_enabled, physaddr) = 1; 92 *fixup_int(&pgdir_shift, physaddr) = 48; 93 *fixup_int(&ptrs_per_p4d, physaddr) = 512; 94 *fixup_long(&page_offset_base, physaddr) = __PAGE_OFFSET_BASE_L5; 95 *fixup_long(&vmalloc_base, physaddr) = __VMALLOC_BASE_L5; 96 *fixup_long(&vmemmap_base, physaddr) = __VMEMMAP_BASE_L5; 97 98 return true; 99 } 100 #else 101 static bool __head check_la57_support(unsigned long physaddr) 102 { 103 return false; 104 } 105 #endif 106 107 /* Code in __startup_64() can be relocated during execution, but the compiler 108 * doesn't have to generate PC-relative relocations when accessing globals from 109 * that function. Clang actually does not generate them, which leads to 110 * boot-time crashes. To work around this problem, every global pointer must 111 * be adjusted using fixup_pointer(). 112 */ 113 unsigned long __head __startup_64(unsigned long physaddr, 114 struct boot_params *bp) 115 { 116 unsigned long vaddr, vaddr_end; 117 unsigned long load_delta, *p; 118 unsigned long pgtable_flags; 119 pgdval_t *pgd; 120 p4dval_t *p4d; 121 pudval_t *pud; 122 pmdval_t *pmd, pmd_entry; 123 pteval_t *mask_ptr; 124 bool la57; 125 int i; 126 unsigned int *next_pgt_ptr; 127 128 la57 = check_la57_support(physaddr); 129 130 /* Is the address too large? */ 131 if (physaddr >> MAX_PHYSMEM_BITS) 132 for (;;); 133 134 /* 135 * Compute the delta between the address I am compiled to run at 136 * and the address I am actually running at. 137 */ 138 load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map); 139 140 /* Is the address not 2M aligned? */ 141 if (load_delta & ~PMD_PAGE_MASK) 142 for (;;); 143 144 /* Activate Secure Memory Encryption (SME) if supported and enabled */ 145 sme_enable(bp); 146 147 /* Include the SME encryption mask in the fixup value */ 148 load_delta += sme_get_me_mask(); 149 150 /* Fixup the physical addresses in the page table */ 151 152 pgd = fixup_pointer(&early_top_pgt, physaddr); 153 p = pgd + pgd_index(__START_KERNEL_map); 154 if (la57) 155 *p = (unsigned long)level4_kernel_pgt; 156 else 157 *p = (unsigned long)level3_kernel_pgt; 158 *p += _PAGE_TABLE_NOENC - __START_KERNEL_map + load_delta; 159 160 if (la57) { 161 p4d = fixup_pointer(&level4_kernel_pgt, physaddr); 162 p4d[511] += load_delta; 163 } 164 165 pud = fixup_pointer(&level3_kernel_pgt, physaddr); 166 pud[510] += load_delta; 167 pud[511] += load_delta; 168 169 pmd = fixup_pointer(level2_fixmap_pgt, physaddr); 170 for (i = FIXMAP_PMD_TOP; i > FIXMAP_PMD_TOP - FIXMAP_PMD_NUM; i--) 171 pmd[i] += load_delta; 172 173 /* 174 * Set up the identity mapping for the switchover. These 175 * entries should *NOT* have the global bit set! This also 176 * creates a bunch of nonsense entries but that is fine -- 177 * it avoids problems around wraparound. 178 */ 179 180 next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr); 181 pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr); 182 pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr); 183 184 pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask(); 185 186 if (la57) { 187 p4d = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], 188 physaddr); 189 190 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD; 191 pgd[i + 0] = (pgdval_t)p4d + pgtable_flags; 192 pgd[i + 1] = (pgdval_t)p4d + pgtable_flags; 193 194 i = physaddr >> P4D_SHIFT; 195 p4d[(i + 0) % PTRS_PER_P4D] = (pgdval_t)pud + pgtable_flags; 196 p4d[(i + 1) % PTRS_PER_P4D] = (pgdval_t)pud + pgtable_flags; 197 } else { 198 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD; 199 pgd[i + 0] = (pgdval_t)pud + pgtable_flags; 200 pgd[i + 1] = (pgdval_t)pud + pgtable_flags; 201 } 202 203 i = physaddr >> PUD_SHIFT; 204 pud[(i + 0) % PTRS_PER_PUD] = (pudval_t)pmd + pgtable_flags; 205 pud[(i + 1) % PTRS_PER_PUD] = (pudval_t)pmd + pgtable_flags; 206 207 pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL; 208 /* Filter out unsupported __PAGE_KERNEL_* bits: */ 209 mask_ptr = fixup_pointer(&__supported_pte_mask, physaddr); 210 pmd_entry &= *mask_ptr; 211 pmd_entry += sme_get_me_mask(); 212 pmd_entry += physaddr; 213 214 for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) { 215 int idx = i + (physaddr >> PMD_SHIFT); 216 217 pmd[idx % PTRS_PER_PMD] = pmd_entry + i * PMD_SIZE; 218 } 219 220 /* 221 * Fixup the kernel text+data virtual addresses. Note that 222 * we might write invalid pmds, when the kernel is relocated 223 * cleanup_highmap() fixes this up along with the mappings 224 * beyond _end. 225 */ 226 227 pmd = fixup_pointer(level2_kernel_pgt, physaddr); 228 for (i = 0; i < PTRS_PER_PMD; i++) { 229 if (pmd[i] & _PAGE_PRESENT) 230 pmd[i] += load_delta; 231 } 232 233 /* 234 * Fixup phys_base - remove the memory encryption mask to obtain 235 * the true physical address. 236 */ 237 *fixup_long(&phys_base, physaddr) += load_delta - sme_get_me_mask(); 238 239 /* Encrypt the kernel and related (if SME is active) */ 240 sme_encrypt_kernel(bp); 241 242 /* 243 * Clear the memory encryption mask from the .bss..decrypted section. 244 * The bss section will be memset to zero later in the initialization so 245 * there is no need to zero it after changing the memory encryption 246 * attribute. 247 */ 248 if (mem_encrypt_active()) { 249 vaddr = (unsigned long)__start_bss_decrypted; 250 vaddr_end = (unsigned long)__end_bss_decrypted; 251 for (; vaddr < vaddr_end; vaddr += PMD_SIZE) { 252 i = pmd_index(vaddr); 253 pmd[i] -= sme_get_me_mask(); 254 } 255 } 256 257 /* 258 * Return the SME encryption mask (if SME is active) to be used as a 259 * modifier for the initial pgdir entry programmed into CR3. 260 */ 261 return sme_get_me_mask(); 262 } 263 264 unsigned long __startup_secondary_64(void) 265 { 266 /* 267 * Return the SME encryption mask (if SME is active) to be used as a 268 * modifier for the initial pgdir entry programmed into CR3. 269 */ 270 return sme_get_me_mask(); 271 } 272 273 /* Wipe all early page tables except for the kernel symbol map */ 274 static void __init reset_early_page_tables(void) 275 { 276 memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1)); 277 next_early_pgt = 0; 278 write_cr3(__sme_pa_nodebug(early_top_pgt)); 279 } 280 281 /* Create a new PMD entry */ 282 int __init __early_make_pgtable(unsigned long address, pmdval_t pmd) 283 { 284 unsigned long physaddr = address - __PAGE_OFFSET; 285 pgdval_t pgd, *pgd_p; 286 p4dval_t p4d, *p4d_p; 287 pudval_t pud, *pud_p; 288 pmdval_t *pmd_p; 289 290 /* Invalid address or early pgt is done ? */ 291 if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt)) 292 return -1; 293 294 again: 295 pgd_p = &early_top_pgt[pgd_index(address)].pgd; 296 pgd = *pgd_p; 297 298 /* 299 * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is 300 * critical -- __PAGE_OFFSET would point us back into the dynamic 301 * range and we might end up looping forever... 302 */ 303 if (!pgtable_l5_enabled()) 304 p4d_p = pgd_p; 305 else if (pgd) 306 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); 307 else { 308 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { 309 reset_early_page_tables(); 310 goto again; 311 } 312 313 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++]; 314 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D); 315 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; 316 } 317 p4d_p += p4d_index(address); 318 p4d = *p4d_p; 319 320 if (p4d) 321 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); 322 else { 323 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { 324 reset_early_page_tables(); 325 goto again; 326 } 327 328 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++]; 329 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD); 330 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; 331 } 332 pud_p += pud_index(address); 333 pud = *pud_p; 334 335 if (pud) 336 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); 337 else { 338 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { 339 reset_early_page_tables(); 340 goto again; 341 } 342 343 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++]; 344 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD); 345 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; 346 } 347 pmd_p[pmd_index(address)] = pmd; 348 349 return 0; 350 } 351 352 int __init early_make_pgtable(unsigned long address) 353 { 354 unsigned long physaddr = address - __PAGE_OFFSET; 355 pmdval_t pmd; 356 357 pmd = (physaddr & PMD_MASK) + early_pmd_flags; 358 359 return __early_make_pgtable(address, pmd); 360 } 361 362 /* Don't add a printk in there. printk relies on the PDA which is not initialized 363 yet. */ 364 static void __init clear_bss(void) 365 { 366 memset(__bss_start, 0, 367 (unsigned long) __bss_stop - (unsigned long) __bss_start); 368 } 369 370 static unsigned long get_cmd_line_ptr(void) 371 { 372 unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr; 373 374 cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32; 375 376 return cmd_line_ptr; 377 } 378 379 static void __init copy_bootdata(char *real_mode_data) 380 { 381 char * command_line; 382 unsigned long cmd_line_ptr; 383 384 /* 385 * If SME is active, this will create decrypted mappings of the 386 * boot data in advance of the copy operations. 387 */ 388 sme_map_bootdata(real_mode_data); 389 390 memcpy(&boot_params, real_mode_data, sizeof(boot_params)); 391 sanitize_boot_params(&boot_params); 392 cmd_line_ptr = get_cmd_line_ptr(); 393 if (cmd_line_ptr) { 394 command_line = __va(cmd_line_ptr); 395 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE); 396 } 397 398 /* 399 * The old boot data is no longer needed and won't be reserved, 400 * freeing up that memory for use by the system. If SME is active, 401 * we need to remove the mappings that were created so that the 402 * memory doesn't remain mapped as decrypted. 403 */ 404 sme_unmap_bootdata(real_mode_data); 405 } 406 407 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data) 408 { 409 /* 410 * Build-time sanity checks on the kernel image and module 411 * area mappings. (these are purely build-time and produce no code) 412 */ 413 BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map); 414 BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE); 415 BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE); 416 BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0); 417 BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0); 418 BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL)); 419 MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) == 420 (__START_KERNEL & PGDIR_MASK))); 421 BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END); 422 423 cr4_init_shadow(); 424 425 /* Kill off the identity-map trampoline */ 426 reset_early_page_tables(); 427 428 clear_bss(); 429 430 clear_page(init_top_pgt); 431 432 /* 433 * SME support may update early_pmd_flags to include the memory 434 * encryption mask, so it needs to be called before anything 435 * that may generate a page fault. 436 */ 437 sme_early_init(); 438 439 kasan_early_init(); 440 441 idt_setup_early_handler(); 442 443 copy_bootdata(__va(real_mode_data)); 444 445 /* 446 * Load microcode early on BSP. 447 */ 448 load_ucode_bsp(); 449 450 /* set init_top_pgt kernel high mapping*/ 451 init_top_pgt[511] = early_top_pgt[511]; 452 453 x86_64_start_reservations(real_mode_data); 454 } 455 456 void __init x86_64_start_reservations(char *real_mode_data) 457 { 458 /* version is always not zero if it is copied */ 459 if (!boot_params.hdr.version) 460 copy_bootdata(__va(real_mode_data)); 461 462 x86_early_init_platform_quirks(); 463 464 switch (boot_params.hdr.hardware_subarch) { 465 case X86_SUBARCH_INTEL_MID: 466 x86_intel_mid_early_setup(); 467 break; 468 default: 469 break; 470 } 471 472 start_kernel(); 473 } 474