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 #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          * Only the region occupied by the kernel image has so far
 227          * been checked against the table of usable memory regions
 228          * provided by the firmware, so invalidate pages outside that
 229          * region. A page table entry that maps to a reserved area of
 230          * memory would allow processor speculation into that area,
 231          * and on some hardware (particularly the UV platform) even
 232          * speculative access to some reserved areas is caught as an
 233          * error, causing the BIOS to halt the system.
 234          */
 235
 236         pmd = fixup_pointer(level2_kernel_pgt, physaddr);
 237
 238         /* invalidate pages before the kernel image */
 239         for (i = 0; i < pmd_index((unsigned long)_text); i++)
 240                 pmd[i] &= ~_PAGE_PRESENT;
 241
 242         /* fixup pages that are part of the kernel image */
 243         for (; i <= pmd_index((unsigned long)_end); i++)
 244                 if (pmd[i] & _PAGE_PRESENT)
 245                         pmd[i] += load_delta;
 246
 247         /* invalidate pages after the kernel image */
 248         for (; i < PTRS_PER_PMD; i++)
 249                 pmd[i] &= ~_PAGE_PRESENT;
 250
 251         /*
 252          * Fixup phys_base - remove the memory encryption mask to obtain
 253          * the true physical address.
 254          */
 255         *fixup_long(&phys_base, physaddr) += load_delta - sme_get_me_mask();
 256
 257         /* Encrypt the kernel and related (if SME is active) */
 258         sme_encrypt_kernel(bp);
 259
 260         /*
 261          * Clear the memory encryption mask from the .bss..decrypted section.
 262          * The bss section will be memset to zero later in the initialization so
 263          * there is no need to zero it after changing the memory encryption
 264          * attribute.
 265          */
 266         if (mem_encrypt_active()) {
 267                 vaddr = (unsigned long)__start_bss_decrypted;
 268                 vaddr_end = (unsigned long)__end_bss_decrypted;
 269                 for (; vaddr < vaddr_end; vaddr += PMD_SIZE) {
 270                         i = pmd_index(vaddr);
 271                         pmd[i] -= sme_get_me_mask();
 272                 }
 273         }
 274
 275         /*
 276          * Return the SME encryption mask (if SME is active) to be used as a
 277          * modifier for the initial pgdir entry programmed into CR3.
 278          */
 279         return sme_get_me_mask();
 280 }
 281
 282 unsigned long __startup_secondary_64(void)
 283 {
 284         /*
 285          * Return the SME encryption mask (if SME is active) to be used as a
 286          * modifier for the initial pgdir entry programmed into CR3.
 287          */
 288         return sme_get_me_mask();
 289 }
 290
 291 /* Wipe all early page tables except for the kernel symbol map */
 292 static void __init reset_early_page_tables(void)
 293 {
 294         memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
 295         next_early_pgt = 0;
 296         write_cr3(__sme_pa_nodebug(early_top_pgt));
 297 }
 298
 299 /* Create a new PMD entry */
 300 int __init __early_make_pgtable(unsigned long address, pmdval_t pmd)
 301 {
 302         unsigned long physaddr = address - __PAGE_OFFSET;
 303         pgdval_t pgd, *pgd_p;
 304         p4dval_t p4d, *p4d_p;
 305         pudval_t pud, *pud_p;
 306         pmdval_t *pmd_p;
 307
 308         /* Invalid address or early pgt is done ?  */
 309         if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
 310                 return -1;
 311
 312 again:
 313         pgd_p = &early_top_pgt[pgd_index(address)].pgd;
 314         pgd = *pgd_p;
 315
 316         /*
 317          * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
 318          * critical -- __PAGE_OFFSET would point us back into the dynamic
 319          * range and we might end up looping forever...
 320          */
 321         if (!pgtable_l5_enabled())
 322                 p4d_p = pgd_p;
 323         else if (pgd)
 324                 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 325         else {
 326                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 327                         reset_early_page_tables();
 328                         goto again;
 329                 }
 330
 331                 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
 332                 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
 333                 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 334         }
 335         p4d_p += p4d_index(address);
 336         p4d = *p4d_p;
 337
 338         if (p4d)
 339                 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 340         else {
 341                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 342                         reset_early_page_tables();
 343                         goto again;
 344                 }
 345
 346                 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
 347                 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
 348                 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 349         }
 350         pud_p += pud_index(address);
 351         pud = *pud_p;
 352
 353         if (pud)
 354                 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 355         else {
 356                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 357                         reset_early_page_tables();
 358                         goto again;
 359                 }
 360
 361                 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
 362                 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
 363                 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 364         }
 365         pmd_p[pmd_index(address)] = pmd;
 366
 367         return 0;
 368 }
 369
 370 int __init early_make_pgtable(unsigned long address)
 371 {
 372         unsigned long physaddr = address - __PAGE_OFFSET;
 373         pmdval_t pmd;
 374
 375         pmd = (physaddr & PMD_MASK) + early_pmd_flags;
 376
 377         return __early_make_pgtable(address, pmd);
 378 }
 379
 380 /* Don't add a printk in there. printk relies on the PDA which is not initialized
 381    yet. */
 382 static void __init clear_bss(void)
 383 {
 384         memset(__bss_start, 0,
 385                (unsigned long) __bss_stop - (unsigned long) __bss_start);
 386         memset(__brk_base, 0,
 387                (unsigned long) __brk_limit - (unsigned long) __brk_base);
 388 }
 389
 390 static unsigned long get_cmd_line_ptr(void)
 391 {
 392         unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
 393
 394         cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
 395
 396         return cmd_line_ptr;
 397 }
 398
 399 static void __init copy_bootdata(char *real_mode_data)
 400 {
 401         char * command_line;
 402         unsigned long cmd_line_ptr;
 403
 404         /*
 405          * If SME is active, this will create decrypted mappings of the
 406          * boot data in advance of the copy operations.
 407          */
 408         sme_map_bootdata(real_mode_data);
 409
 410         memcpy(&boot_params, real_mode_data, sizeof boot_params);
 411         sanitize_boot_params(&boot_params);
 412         cmd_line_ptr = get_cmd_line_ptr();
 413         if (cmd_line_ptr) {
 414                 command_line = __va(cmd_line_ptr);
 415                 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 416         }
 417
 418         /*
 419          * The old boot data is no longer needed and won't be reserved,
 420          * freeing up that memory for use by the system. If SME is active,
 421          * we need to remove the mappings that were created so that the
 422          * memory doesn't remain mapped as decrypted.
 423          */
 424         sme_unmap_bootdata(real_mode_data);
 425 }
 426
 427 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
 428 {
 429         /*
 430          * Build-time sanity checks on the kernel image and module
 431          * area mappings. (these are purely build-time and produce no code)
 432          */
 433         BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
 434         BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
 435         BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
 436         BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
 437         BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
 438         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 439         MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 440                                 (__START_KERNEL & PGDIR_MASK)));
 441         BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
 442
 443         cr4_init_shadow();
 444
 445         /* Kill off the identity-map trampoline */
 446         reset_early_page_tables();
 447
 448         clear_bss();
 449
 450         clear_page(init_top_pgt);
 451
 452         /*
 453          * SME support may update early_pmd_flags to include the memory
 454          * encryption mask, so it needs to be called before anything
 455          * that may generate a page fault.
 456          */
 457         sme_early_init();
 458
 459         kasan_early_init();
 460
 461         idt_setup_early_handler();
 462
 463         copy_bootdata(__va(real_mode_data));
 464
 465         /*
 466          * Load microcode early on BSP.
 467          */
 468         load_ucode_bsp();
 469
 470         /* set init_top_pgt kernel high mapping*/
 471         init_top_pgt[511] = early_top_pgt[511];
 472
 473         x86_64_start_reservations(real_mode_data);
 474 }
 475
 476 void __init x86_64_start_reservations(char *real_mode_data)
 477 {
 478         /* version is always not zero if it is copied */
 479         if (!boot_params.hdr.version)
 480                 copy_bootdata(__va(real_mode_data));
 481
 482         x86_early_init_platform_quirks();
 483
 484         switch (boot_params.hdr.hardware_subarch) {
 485         case X86_SUBARCH_INTEL_MID:
 486                 x86_intel_mid_early_setup();
 487                 break;
 488         default:
 489                 break;
 490         }
 491
 492         start_kernel();
 493 }