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 #include <linux/init.h>
  10 #include <linux/linkage.h>
  11 #include <linux/types.h>
  12 #include <linux/kernel.h>
  13 #include <linux/string.h>
  14 #include <linux/percpu.h>
  15 #include <linux/start_kernel.h>
  16 #include <linux/io.h>
  17 #include <linux/memblock.h>
  18 #include <linux/mem_encrypt.h>
  19
  20 #include <asm/processor.h>
  21 #include <asm/proto.h>
  22 #include <asm/smp.h>
  23 #include <asm/setup.h>
  24 #include <asm/desc.h>
  25 #include <asm/pgtable.h>
  26 #include <asm/tlbflush.h>
  27 #include <asm/sections.h>
  28 #include <asm/kdebug.h>
  29 #include <asm/e820/api.h>
  30 #include <asm/bios_ebda.h>
  31 #include <asm/bootparam_utils.h>
  32 #include <asm/microcode.h>
  33 #include <asm/kasan.h>
  34
  35 #ifdef CONFIG_X86_5LEVEL
  36 #undef pgtable_l5_enabled
  37 #define pgtable_l5_enabled __pgtable_l5_enabled
  38 #endif
  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 EXPORT_SYMBOL(__pgtable_l5_enabled);
  50 unsigned int pgdir_shift __ro_after_init = 39;
  51 EXPORT_SYMBOL(pgdir_shift);
  52 unsigned int ptrs_per_p4d __ro_after_init = 1;
  53 EXPORT_SYMBOL(ptrs_per_p4d);
  54 #endif
  55
  56 #ifdef CONFIG_DYNAMIC_MEMORY_LAYOUT
  57 unsigned long page_offset_base __ro_after_init = __PAGE_OFFSET_BASE_L4;
  58 EXPORT_SYMBOL(page_offset_base);
  59 unsigned long vmalloc_base __ro_after_init = __VMALLOC_BASE_L4;
  60 EXPORT_SYMBOL(vmalloc_base);
  61 unsigned long vmemmap_base __ro_after_init = __VMEMMAP_BASE_L4;
  62 EXPORT_SYMBOL(vmemmap_base);
  63 #endif
  64
  65 #define __head  __section(.head.text)
  66
  67 static void __head *fixup_pointer(void *ptr, unsigned long physaddr)
  68 {
  69         return ptr - (void *)_text + (void *)physaddr;
  70 }
  71
  72 static unsigned long __head *fixup_long(void *ptr, unsigned long physaddr)
  73 {
  74         return fixup_pointer(ptr, physaddr);
  75 }
  76
  77 #ifdef CONFIG_X86_5LEVEL
  78 static unsigned int __head *fixup_int(void *ptr, unsigned long physaddr)
  79 {
  80         return fixup_pointer(ptr, physaddr);
  81 }
  82
  83 static bool __head check_la57_support(unsigned long physaddr)
  84 {
  85         if (native_cpuid_eax(0) < 7)
  86                 return false;
  87
  88         if (!(native_cpuid_ecx(7) & (1 << (X86_FEATURE_LA57 & 31))))
  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 unsigned long __head __startup_64(unsigned long physaddr,
 108                                   struct boot_params *bp)
 109 {
 110         unsigned long load_delta, *p;
 111         unsigned long pgtable_flags;
 112         pgdval_t *pgd;
 113         p4dval_t *p4d;
 114         pudval_t *pud;
 115         pmdval_t *pmd, pmd_entry;
 116         bool la57;
 117         int i;
 118         unsigned int *next_pgt_ptr;
 119
 120         la57 = check_la57_support(physaddr);
 121
 122         /* Is the address too large? */
 123         if (physaddr >> MAX_PHYSMEM_BITS)
 124                 for (;;);
 125
 126         /*
 127          * Compute the delta between the address I am compiled to run at
 128          * and the address I am actually running at.
 129          */
 130         load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
 131
 132         /* Is the address not 2M aligned? */
 133         if (load_delta & ~PMD_PAGE_MASK)
 134                 for (;;);
 135
 136         /* Activate Secure Memory Encryption (SME) if supported and enabled */
 137         sme_enable(bp);
 138
 139         /* Include the SME encryption mask in the fixup value */
 140         load_delta += sme_get_me_mask();
 141
 142         /* Fixup the physical addresses in the page table */
 143
 144         pgd = fixup_pointer(&early_top_pgt, physaddr);
 145         p = pgd + pgd_index(__START_KERNEL_map);
 146         if (la57)
 147                 *p = (unsigned long)level4_kernel_pgt;
 148         else
 149                 *p = (unsigned long)level3_kernel_pgt;
 150         *p += _PAGE_TABLE_NOENC - __START_KERNEL_map + load_delta;
 151
 152         if (la57) {
 153                 p4d = fixup_pointer(&level4_kernel_pgt, physaddr);
 154                 p4d[511] += load_delta;
 155         }
 156
 157         pud = fixup_pointer(&level3_kernel_pgt, physaddr);
 158         pud[510] += load_delta;
 159         pud[511] += load_delta;
 160
 161         pmd = fixup_pointer(level2_fixmap_pgt, physaddr);
 162         pmd[506] += load_delta;
 163
 164         /*
 165          * Set up the identity mapping for the switchover.  These
 166          * entries should *NOT* have the global bit set!  This also
 167          * creates a bunch of nonsense entries but that is fine --
 168          * it avoids problems around wraparound.
 169          */
 170
 171         next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr);
 172         pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
 173         pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
 174
 175         pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask();
 176
 177         if (la57) {
 178                 p4d = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr);
 179
 180                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 181                 pgd[i + 0] = (pgdval_t)p4d + pgtable_flags;
 182                 pgd[i + 1] = (pgdval_t)p4d + pgtable_flags;
 183
 184                 i = (physaddr >> P4D_SHIFT) % PTRS_PER_P4D;
 185                 p4d[i + 0] = (pgdval_t)pud + pgtable_flags;
 186                 p4d[i + 1] = (pgdval_t)pud + pgtable_flags;
 187         } else {
 188                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 189                 pgd[i + 0] = (pgdval_t)pud + pgtable_flags;
 190                 pgd[i + 1] = (pgdval_t)pud + pgtable_flags;
 191         }
 192
 193         i = (physaddr >> PUD_SHIFT) % PTRS_PER_PUD;
 194         pud[i + 0] = (pudval_t)pmd + pgtable_flags;
 195         pud[i + 1] = (pudval_t)pmd + pgtable_flags;
 196
 197         pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
 198         /* Filter out unsupported __PAGE_KERNEL_* bits: */
 199         pmd_entry &= __supported_pte_mask;
 200         pmd_entry += sme_get_me_mask();
 201         pmd_entry +=  physaddr;
 202
 203         for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
 204                 int idx = i + (physaddr >> PMD_SHIFT) % PTRS_PER_PMD;
 205                 pmd[idx] = pmd_entry + i * PMD_SIZE;
 206         }
 207
 208         /*
 209          * Fixup the kernel text+data virtual addresses. Note that
 210          * we might write invalid pmds, when the kernel is relocated
 211          * cleanup_highmap() fixes this up along with the mappings
 212          * beyond _end.
 213          */
 214
 215         pmd = fixup_pointer(level2_kernel_pgt, physaddr);
 216         for (i = 0; i < PTRS_PER_PMD; i++) {
 217                 if (pmd[i] & _PAGE_PRESENT)
 218                         pmd[i] += load_delta;
 219         }
 220
 221         /*
 222          * Fixup phys_base - remove the memory encryption mask to obtain
 223          * the true physical address.
 224          */
 225         *fixup_long(&phys_base, physaddr) += load_delta - sme_get_me_mask();
 226
 227         /* Encrypt the kernel and related (if SME is active) */
 228         sme_encrypt_kernel(bp);
 229
 230         /*
 231          * Return the SME encryption mask (if SME is active) to be used as a
 232          * modifier for the initial pgdir entry programmed into CR3.
 233          */
 234         return sme_get_me_mask();
 235 }
 236
 237 unsigned long __startup_secondary_64(void)
 238 {
 239         /*
 240          * Return the SME encryption mask (if SME is active) to be used as a
 241          * modifier for the initial pgdir entry programmed into CR3.
 242          */
 243         return sme_get_me_mask();
 244 }
 245
 246 /* Wipe all early page tables except for the kernel symbol map */
 247 static void __init reset_early_page_tables(void)
 248 {
 249         memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
 250         next_early_pgt = 0;
 251         write_cr3(__sme_pa_nodebug(early_top_pgt));
 252 }
 253
 254 /* Create a new PMD entry */
 255 int __init __early_make_pgtable(unsigned long address, pmdval_t pmd)
 256 {
 257         unsigned long physaddr = address - __PAGE_OFFSET;
 258         pgdval_t pgd, *pgd_p;
 259         p4dval_t p4d, *p4d_p;
 260         pudval_t pud, *pud_p;
 261         pmdval_t *pmd_p;
 262
 263         /* Invalid address or early pgt is done ?  */
 264         if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
 265                 return -1;
 266
 267 again:
 268         pgd_p = &early_top_pgt[pgd_index(address)].pgd;
 269         pgd = *pgd_p;
 270
 271         /*
 272          * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
 273          * critical -- __PAGE_OFFSET would point us back into the dynamic
 274          * range and we might end up looping forever...
 275          */
 276         if (!pgtable_l5_enabled)
 277                 p4d_p = pgd_p;
 278         else if (pgd)
 279                 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 280         else {
 281                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 282                         reset_early_page_tables();
 283                         goto again;
 284                 }
 285
 286                 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
 287                 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
 288                 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 289         }
 290         p4d_p += p4d_index(address);
 291         p4d = *p4d_p;
 292
 293         if (p4d)
 294                 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 295         else {
 296                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 297                         reset_early_page_tables();
 298                         goto again;
 299                 }
 300
 301                 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
 302                 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
 303                 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 304         }
 305         pud_p += pud_index(address);
 306         pud = *pud_p;
 307
 308         if (pud)
 309                 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 310         else {
 311                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 312                         reset_early_page_tables();
 313                         goto again;
 314                 }
 315
 316                 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
 317                 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
 318                 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 319         }
 320         pmd_p[pmd_index(address)] = pmd;
 321
 322         return 0;
 323 }
 324
 325 int __init early_make_pgtable(unsigned long address)
 326 {
 327         unsigned long physaddr = address - __PAGE_OFFSET;
 328         pmdval_t pmd;
 329
 330         pmd = (physaddr & PMD_MASK) + early_pmd_flags;
 331
 332         return __early_make_pgtable(address, pmd);
 333 }
 334
 335 /* Don't add a printk in there. printk relies on the PDA which is not initialized
 336    yet. */
 337 static void __init clear_bss(void)
 338 {
 339         memset(__bss_start, 0,
 340                (unsigned long) __bss_stop - (unsigned long) __bss_start);
 341 }
 342
 343 static unsigned long get_cmd_line_ptr(void)
 344 {
 345         unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
 346
 347         cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
 348
 349         return cmd_line_ptr;
 350 }
 351
 352 static void __init copy_bootdata(char *real_mode_data)
 353 {
 354         char * command_line;
 355         unsigned long cmd_line_ptr;
 356
 357         /*
 358          * If SME is active, this will create decrypted mappings of the
 359          * boot data in advance of the copy operations.
 360          */
 361         sme_map_bootdata(real_mode_data);
 362
 363         memcpy(&boot_params, real_mode_data, sizeof boot_params);
 364         sanitize_boot_params(&boot_params);
 365         cmd_line_ptr = get_cmd_line_ptr();
 366         if (cmd_line_ptr) {
 367                 command_line = __va(cmd_line_ptr);
 368                 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 369         }
 370
 371         /*
 372          * The old boot data is no longer needed and won't be reserved,
 373          * freeing up that memory for use by the system. If SME is active,
 374          * we need to remove the mappings that were created so that the
 375          * memory doesn't remain mapped as decrypted.
 376          */
 377         sme_unmap_bootdata(real_mode_data);
 378 }
 379
 380 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
 381 {
 382         /*
 383          * Build-time sanity checks on the kernel image and module
 384          * area mappings. (these are purely build-time and produce no code)
 385          */
 386         BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
 387         BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
 388         BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
 389         BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
 390         BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
 391         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 392         MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 393                                 (__START_KERNEL & PGDIR_MASK)));
 394         BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
 395
 396         cr4_init_shadow();
 397
 398         /* Kill off the identity-map trampoline */
 399         reset_early_page_tables();
 400
 401         clear_bss();
 402
 403         clear_page(init_top_pgt);
 404
 405         /*
 406          * SME support may update early_pmd_flags to include the memory
 407          * encryption mask, so it needs to be called before anything
 408          * that may generate a page fault.
 409          */
 410         sme_early_init();
 411
 412         kasan_early_init();
 413
 414         idt_setup_early_handler();
 415
 416         copy_bootdata(__va(real_mode_data));
 417
 418         /*
 419          * Load microcode early on BSP.
 420          */
 421         load_ucode_bsp();
 422
 423         /* set init_top_pgt kernel high mapping*/
 424         init_top_pgt[511] = early_top_pgt[511];
 425
 426         x86_64_start_reservations(real_mode_data);
 427 }
 428
 429 void __init x86_64_start_reservations(char *real_mode_data)
 430 {
 431         /* version is always not zero if it is copied */
 432         if (!boot_params.hdr.version)
 433                 copy_bootdata(__va(real_mode_data));
 434
 435         x86_early_init_platform_quirks();
 436
 437         switch (boot_params.hdr.hardware_subarch) {
 438         case X86_SUBARCH_INTEL_MID:
 439                 x86_intel_mid_early_setup();
 440                 break;
 441         default:
 442                 break;
 443         }
 444
 445         start_kernel();
 446 }