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
  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 __ro_after_init;
  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         if (native_cpuid_eax(0) < 7)
  84                 return false;
  85
  86         if (!(native_cpuid_ecx(7) & (1 << (X86_FEATURE_LA57 & 31))))
  87                 return false;
  88
  89         *fixup_int(&__pgtable_l5_enabled, physaddr) = 1;
  90         *fixup_int(&pgdir_shift, physaddr) = 48;
  91         *fixup_int(&ptrs_per_p4d, physaddr) = 512;
  92         *fixup_long(&page_offset_base, physaddr) = __PAGE_OFFSET_BASE_L5;
  93         *fixup_long(&vmalloc_base, physaddr) = __VMALLOC_BASE_L5;
  94         *fixup_long(&vmemmap_base, physaddr) = __VMEMMAP_BASE_L5;
  95
  96         return true;
  97 }
  98 #else
  99 static bool __head check_la57_support(unsigned long physaddr)
 100 {
 101         return false;
 102 }
 103 #endif
 104
 105 /* Code in __startup_64() can be relocated during execution, but the compiler
 106  * doesn't have to generate PC-relative relocations when accessing globals from
 107  * that function. Clang actually does not generate them, which leads to
 108  * boot-time crashes. To work around this problem, every global pointer must
 109  * be adjusted using fixup_pointer().
 110  */
 111 unsigned long __head __startup_64(unsigned long physaddr,
 112                                   struct boot_params *bp)
 113 {
 114         unsigned long load_delta, *p;
 115         unsigned long pgtable_flags;
 116         pgdval_t *pgd;
 117         p4dval_t *p4d;
 118         pudval_t *pud;
 119         pmdval_t *pmd, pmd_entry;
 120         pteval_t *mask_ptr;
 121         bool la57;
 122         int i;
 123         unsigned int *next_pgt_ptr;
 124
 125         la57 = check_la57_support(physaddr);
 126
 127         /* Is the address too large? */
 128         if (physaddr >> MAX_PHYSMEM_BITS)
 129                 for (;;);
 130
 131         /*
 132          * Compute the delta between the address I am compiled to run at
 133          * and the address I am actually running at.
 134          */
 135         load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
 136
 137         /* Is the address not 2M aligned? */
 138         if (load_delta & ~PMD_PAGE_MASK)
 139                 for (;;);
 140
 141         /* Activate Secure Memory Encryption (SME) if supported and enabled */
 142         sme_enable(bp);
 143
 144         /* Include the SME encryption mask in the fixup value */
 145         load_delta += sme_get_me_mask();
 146
 147         /* Fixup the physical addresses in the page table */
 148
 149         pgd = fixup_pointer(&early_top_pgt, physaddr);
 150         p = pgd + pgd_index(__START_KERNEL_map);
 151         if (la57)
 152                 *p = (unsigned long)level4_kernel_pgt;
 153         else
 154                 *p = (unsigned long)level3_kernel_pgt;
 155         *p += _PAGE_TABLE_NOENC - __START_KERNEL_map + load_delta;
 156
 157         if (la57) {
 158                 p4d = fixup_pointer(&level4_kernel_pgt, physaddr);
 159                 p4d[511] += load_delta;
 160         }
 161
 162         pud = fixup_pointer(&level3_kernel_pgt, physaddr);
 163         pud[510] += load_delta;
 164         pud[511] += load_delta;
 165
 166         pmd = fixup_pointer(level2_fixmap_pgt, physaddr);
 167         pmd[506] += load_delta;
 168
 169         /*
 170          * Set up the identity mapping for the switchover.  These
 171          * entries should *NOT* have the global bit set!  This also
 172          * creates a bunch of nonsense entries but that is fine --
 173          * it avoids problems around wraparound.
 174          */
 175
 176         next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr);
 177         pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
 178         pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
 179
 180         pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask();
 181
 182         if (la57) {
 183                 p4d = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr);
 184
 185                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 186                 pgd[i + 0] = (pgdval_t)p4d + pgtable_flags;
 187                 pgd[i + 1] = (pgdval_t)p4d + pgtable_flags;
 188
 189                 i = (physaddr >> P4D_SHIFT) % PTRS_PER_P4D;
 190                 p4d[i + 0] = (pgdval_t)pud + pgtable_flags;
 191                 p4d[i + 1] = (pgdval_t)pud + pgtable_flags;
 192         } else {
 193                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 194                 pgd[i + 0] = (pgdval_t)pud + pgtable_flags;
 195                 pgd[i + 1] = (pgdval_t)pud + pgtable_flags;
 196         }
 197
 198         i = (physaddr >> PUD_SHIFT) % PTRS_PER_PUD;
 199         pud[i + 0] = (pudval_t)pmd + pgtable_flags;
 200         pud[i + 1] = (pudval_t)pmd + pgtable_flags;
 201
 202         pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
 203         /* Filter out unsupported __PAGE_KERNEL_* bits: */
 204         mask_ptr = fixup_pointer(&__supported_pte_mask, physaddr);
 205         pmd_entry &= *mask_ptr;
 206         pmd_entry += sme_get_me_mask();
 207         pmd_entry +=  physaddr;
 208
 209         for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
 210                 int idx = i + (physaddr >> PMD_SHIFT) % PTRS_PER_PMD;
 211                 pmd[idx] = pmd_entry + i * PMD_SIZE;
 212         }
 213
 214         /*
 215          * Fixup the kernel text+data virtual addresses. Note that
 216          * we might write invalid pmds, when the kernel is relocated
 217          * cleanup_highmap() fixes this up along with the mappings
 218          * beyond _end.
 219          */
 220
 221         pmd = fixup_pointer(level2_kernel_pgt, physaddr);
 222         for (i = 0; i < PTRS_PER_PMD; i++) {
 223                 if (pmd[i] & _PAGE_PRESENT)
 224                         pmd[i] += load_delta;
 225         }
 226
 227         /*
 228          * Fixup phys_base - remove the memory encryption mask to obtain
 229          * the true physical address.
 230          */
 231         *fixup_long(&phys_base, physaddr) += load_delta - sme_get_me_mask();
 232
 233         /* Encrypt the kernel and related (if SME is active) */
 234         sme_encrypt_kernel(bp);
 235
 236         /*
 237          * Return the SME encryption mask (if SME is active) to be used as a
 238          * modifier for the initial pgdir entry programmed into CR3.
 239          */
 240         return sme_get_me_mask();
 241 }
 242
 243 unsigned long __startup_secondary_64(void)
 244 {
 245         /*
 246          * Return the SME encryption mask (if SME is active) to be used as a
 247          * modifier for the initial pgdir entry programmed into CR3.
 248          */
 249         return sme_get_me_mask();
 250 }
 251
 252 /* Wipe all early page tables except for the kernel symbol map */
 253 static void __init reset_early_page_tables(void)
 254 {
 255         memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
 256         next_early_pgt = 0;
 257         write_cr3(__sme_pa_nodebug(early_top_pgt));
 258 }
 259
 260 /* Create a new PMD entry */
 261 int __init __early_make_pgtable(unsigned long address, pmdval_t pmd)
 262 {
 263         unsigned long physaddr = address - __PAGE_OFFSET;
 264         pgdval_t pgd, *pgd_p;
 265         p4dval_t p4d, *p4d_p;
 266         pudval_t pud, *pud_p;
 267         pmdval_t *pmd_p;
 268
 269         /* Invalid address or early pgt is done ?  */
 270         if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
 271                 return -1;
 272
 273 again:
 274         pgd_p = &early_top_pgt[pgd_index(address)].pgd;
 275         pgd = *pgd_p;
 276
 277         /*
 278          * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
 279          * critical -- __PAGE_OFFSET would point us back into the dynamic
 280          * range and we might end up looping forever...
 281          */
 282         if (!pgtable_l5_enabled)
 283                 p4d_p = pgd_p;
 284         else if (pgd)
 285                 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 286         else {
 287                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 288                         reset_early_page_tables();
 289                         goto again;
 290                 }
 291
 292                 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
 293                 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
 294                 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 295         }
 296         p4d_p += p4d_index(address);
 297         p4d = *p4d_p;
 298
 299         if (p4d)
 300                 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 301         else {
 302                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 303                         reset_early_page_tables();
 304                         goto again;
 305                 }
 306
 307                 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
 308                 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
 309                 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 310         }
 311         pud_p += pud_index(address);
 312         pud = *pud_p;
 313
 314         if (pud)
 315                 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 316         else {
 317                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 318                         reset_early_page_tables();
 319                         goto again;
 320                 }
 321
 322                 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
 323                 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
 324                 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 325         }
 326         pmd_p[pmd_index(address)] = pmd;
 327
 328         return 0;
 329 }
 330
 331 int __init early_make_pgtable(unsigned long address)
 332 {
 333         unsigned long physaddr = address - __PAGE_OFFSET;
 334         pmdval_t pmd;
 335
 336         pmd = (physaddr & PMD_MASK) + early_pmd_flags;
 337
 338         return __early_make_pgtable(address, pmd);
 339 }
 340
 341 /* Don't add a printk in there. printk relies on the PDA which is not initialized
 342    yet. */
 343 static void __init clear_bss(void)
 344 {
 345         memset(__bss_start, 0,
 346                (unsigned long) __bss_stop - (unsigned long) __bss_start);
 347 }
 348
 349 static unsigned long get_cmd_line_ptr(void)
 350 {
 351         unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
 352
 353         cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
 354
 355         return cmd_line_ptr;
 356 }
 357
 358 static void __init copy_bootdata(char *real_mode_data)
 359 {
 360         char * command_line;
 361         unsigned long cmd_line_ptr;
 362
 363         /*
 364          * If SME is active, this will create decrypted mappings of the
 365          * boot data in advance of the copy operations.
 366          */
 367         sme_map_bootdata(real_mode_data);
 368
 369         memcpy(&boot_params, real_mode_data, sizeof boot_params);
 370         sanitize_boot_params(&boot_params);
 371         cmd_line_ptr = get_cmd_line_ptr();
 372         if (cmd_line_ptr) {
 373                 command_line = __va(cmd_line_ptr);
 374                 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 375         }
 376
 377         /*
 378          * The old boot data is no longer needed and won't be reserved,
 379          * freeing up that memory for use by the system. If SME is active,
 380          * we need to remove the mappings that were created so that the
 381          * memory doesn't remain mapped as decrypted.
 382          */
 383         sme_unmap_bootdata(real_mode_data);
 384 }
 385
 386 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
 387 {
 388         /*
 389          * Build-time sanity checks on the kernel image and module
 390          * area mappings. (these are purely build-time and produce no code)
 391          */
 392         BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
 393         BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
 394         BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
 395         BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
 396         BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
 397         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 398         MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 399                                 (__START_KERNEL & PGDIR_MASK)));
 400         BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
 401
 402         cr4_init_shadow();
 403
 404         /* Kill off the identity-map trampoline */
 405         reset_early_page_tables();
 406
 407         clear_bss();
 408
 409         clear_page(init_top_pgt);
 410
 411         /*
 412          * SME support may update early_pmd_flags to include the memory
 413          * encryption mask, so it needs to be called before anything
 414          * that may generate a page fault.
 415          */
 416         sme_early_init();
 417
 418         kasan_early_init();
 419
 420         idt_setup_early_handler();
 421
 422         copy_bootdata(__va(real_mode_data));
 423
 424         /*
 425          * Load microcode early on BSP.
 426          */
 427         load_ucode_bsp();
 428
 429         /* set init_top_pgt kernel high mapping*/
 430         init_top_pgt[511] = early_top_pgt[511];
 431
 432         x86_64_start_reservations(real_mode_data);
 433 }
 434
 435 void __init x86_64_start_reservations(char *real_mode_data)
 436 {
 437         /* version is always not zero if it is copied */
 438         if (!boot_params.hdr.version)
 439                 copy_bootdata(__va(real_mode_data));
 440
 441         x86_early_init_platform_quirks();
 442
 443         switch (boot_params.hdr.hardware_subarch) {
 444         case X86_SUBARCH_INTEL_MID:
 445                 x86_intel_mid_early_setup();
 446                 break;
 447         default:
 448                 break;
 449         }
 450
 451         start_kernel();
 452 }