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 #include <linux/pgtable.h>
  24
  25 #include <asm/processor.h>
  26 #include <asm/proto.h>
  27 #include <asm/smp.h>
  28 #include <asm/setup.h>
  29 #include <asm/desc.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 /*
  65  * GDT used on the boot CPU before switching to virtual addresses.
  66  */
  67 static struct desc_struct startup_gdt[GDT_ENTRIES] = {
  68         [GDT_ENTRY_KERNEL32_CS]         = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
  69         [GDT_ENTRY_KERNEL_CS]           = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
  70         [GDT_ENTRY_KERNEL_DS]           = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
  71 };
  72
  73 /*
  74  * Address needs to be set at runtime because it references the startup_gdt
  75  * while the kernel still uses a direct mapping.
  76  */
  77 static struct desc_ptr startup_gdt_descr = {
  78         .size = sizeof(startup_gdt),
  79         .address = 0,
  80 };
  81
  82 #define __head  __section(.head.text)
  83
  84 static void __head *fixup_pointer(void *ptr, unsigned long physaddr)
  85 {
  86         return ptr - (void *)_text + (void *)physaddr;
  87 }
  88
  89 static unsigned long __head *fixup_long(void *ptr, unsigned long physaddr)
  90 {
  91         return fixup_pointer(ptr, physaddr);
  92 }
  93
  94 #ifdef CONFIG_X86_5LEVEL
  95 static unsigned int __head *fixup_int(void *ptr, unsigned long physaddr)
  96 {
  97         return fixup_pointer(ptr, physaddr);
  98 }
  99
 100 static bool __head check_la57_support(unsigned long physaddr)
 101 {
 102         /*
 103          * 5-level paging is detected and enabled at kernel decomression
 104          * stage. Only check if it has been enabled there.
 105          */
 106         if (!(native_read_cr4() & X86_CR4_LA57))
 107                 return false;
 108
 109         *fixup_int(&__pgtable_l5_enabled, physaddr) = 1;
 110         *fixup_int(&pgdir_shift, physaddr) = 48;
 111         *fixup_int(&ptrs_per_p4d, physaddr) = 512;
 112         *fixup_long(&page_offset_base, physaddr) = __PAGE_OFFSET_BASE_L5;
 113         *fixup_long(&vmalloc_base, physaddr) = __VMALLOC_BASE_L5;
 114         *fixup_long(&vmemmap_base, physaddr) = __VMEMMAP_BASE_L5;
 115
 116         return true;
 117 }
 118 #else
 119 static bool __head check_la57_support(unsigned long physaddr)
 120 {
 121         return false;
 122 }
 123 #endif
 124
 125 /* Code in __startup_64() can be relocated during execution, but the compiler
 126  * doesn't have to generate PC-relative relocations when accessing globals from
 127  * that function. Clang actually does not generate them, which leads to
 128  * boot-time crashes. To work around this problem, every global pointer must
 129  * be adjusted using fixup_pointer().
 130  */
 131 unsigned long __head __startup_64(unsigned long physaddr,
 132                                   struct boot_params *bp)
 133 {
 134         unsigned long vaddr, vaddr_end;
 135         unsigned long load_delta, *p;
 136         unsigned long pgtable_flags;
 137         pgdval_t *pgd;
 138         p4dval_t *p4d;
 139         pudval_t *pud;
 140         pmdval_t *pmd, pmd_entry;
 141         pteval_t *mask_ptr;
 142         bool la57;
 143         int i;
 144         unsigned int *next_pgt_ptr;
 145
 146         la57 = check_la57_support(physaddr);
 147
 148         /* Is the address too large? */
 149         if (physaddr >> MAX_PHYSMEM_BITS)
 150                 for (;;);
 151
 152         /*
 153          * Compute the delta between the address I am compiled to run at
 154          * and the address I am actually running at.
 155          */
 156         load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
 157
 158         /* Is the address not 2M aligned? */
 159         if (load_delta & ~PMD_PAGE_MASK)
 160                 for (;;);
 161
 162         /* Activate Secure Memory Encryption (SME) if supported and enabled */
 163         sme_enable(bp);
 164
 165         /* Include the SME encryption mask in the fixup value */
 166         load_delta += sme_get_me_mask();
 167
 168         /* Fixup the physical addresses in the page table */
 169
 170         pgd = fixup_pointer(&early_top_pgt, physaddr);
 171         p = pgd + pgd_index(__START_KERNEL_map);
 172         if (la57)
 173                 *p = (unsigned long)level4_kernel_pgt;
 174         else
 175                 *p = (unsigned long)level3_kernel_pgt;
 176         *p += _PAGE_TABLE_NOENC - __START_KERNEL_map + load_delta;
 177
 178         if (la57) {
 179                 p4d = fixup_pointer(&level4_kernel_pgt, physaddr);
 180                 p4d[511] += load_delta;
 181         }
 182
 183         pud = fixup_pointer(&level3_kernel_pgt, physaddr);
 184         pud[510] += load_delta;
 185         pud[511] += load_delta;
 186
 187         pmd = fixup_pointer(level2_fixmap_pgt, physaddr);
 188         for (i = FIXMAP_PMD_TOP; i > FIXMAP_PMD_TOP - FIXMAP_PMD_NUM; i--)
 189                 pmd[i] += load_delta;
 190
 191         /*
 192          * Set up the identity mapping for the switchover.  These
 193          * entries should *NOT* have the global bit set!  This also
 194          * creates a bunch of nonsense entries but that is fine --
 195          * it avoids problems around wraparound.
 196          */
 197
 198         next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr);
 199         pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
 200         pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
 201
 202         pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask();
 203
 204         if (la57) {
 205                 p4d = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++],
 206                                     physaddr);
 207
 208                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 209                 pgd[i + 0] = (pgdval_t)p4d + pgtable_flags;
 210                 pgd[i + 1] = (pgdval_t)p4d + pgtable_flags;
 211
 212                 i = physaddr >> P4D_SHIFT;
 213                 p4d[(i + 0) % PTRS_PER_P4D] = (pgdval_t)pud + pgtable_flags;
 214                 p4d[(i + 1) % PTRS_PER_P4D] = (pgdval_t)pud + pgtable_flags;
 215         } else {
 216                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 217                 pgd[i + 0] = (pgdval_t)pud + pgtable_flags;
 218                 pgd[i + 1] = (pgdval_t)pud + pgtable_flags;
 219         }
 220
 221         i = physaddr >> PUD_SHIFT;
 222         pud[(i + 0) % PTRS_PER_PUD] = (pudval_t)pmd + pgtable_flags;
 223         pud[(i + 1) % PTRS_PER_PUD] = (pudval_t)pmd + pgtable_flags;
 224
 225         pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
 226         /* Filter out unsupported __PAGE_KERNEL_* bits: */
 227         mask_ptr = fixup_pointer(&__supported_pte_mask, physaddr);
 228         pmd_entry &= *mask_ptr;
 229         pmd_entry += sme_get_me_mask();
 230         pmd_entry +=  physaddr;
 231
 232         for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
 233                 int idx = i + (physaddr >> PMD_SHIFT);
 234
 235                 pmd[idx % PTRS_PER_PMD] = pmd_entry + i * PMD_SIZE;
 236         }
 237
 238         /*
 239          * Fixup the kernel text+data virtual addresses. Note that
 240          * we might write invalid pmds, when the kernel is relocated
 241          * cleanup_highmap() fixes this up along with the mappings
 242          * beyond _end.
 243          *
 244          * Only the region occupied by the kernel image has so far
 245          * been checked against the table of usable memory regions
 246          * provided by the firmware, so invalidate pages outside that
 247          * region. A page table entry that maps to a reserved area of
 248          * memory would allow processor speculation into that area,
 249          * and on some hardware (particularly the UV platform) even
 250          * speculative access to some reserved areas is caught as an
 251          * error, causing the BIOS to halt the system.
 252          */
 253
 254         pmd = fixup_pointer(level2_kernel_pgt, physaddr);
 255
 256         /* invalidate pages before the kernel image */
 257         for (i = 0; i < pmd_index((unsigned long)_text); i++)
 258                 pmd[i] &= ~_PAGE_PRESENT;
 259
 260         /* fixup pages that are part of the kernel image */
 261         for (; i <= pmd_index((unsigned long)_end); i++)
 262                 if (pmd[i] & _PAGE_PRESENT)
 263                         pmd[i] += load_delta;
 264
 265         /* invalidate pages after the kernel image */
 266         for (; i < PTRS_PER_PMD; i++)
 267                 pmd[i] &= ~_PAGE_PRESENT;
 268
 269         /*
 270          * Fixup phys_base - remove the memory encryption mask to obtain
 271          * the true physical address.
 272          */
 273         *fixup_long(&phys_base, physaddr) += load_delta - sme_get_me_mask();
 274
 275         /* Encrypt the kernel and related (if SME is active) */
 276         sme_encrypt_kernel(bp);
 277
 278         /*
 279          * Clear the memory encryption mask from the .bss..decrypted section.
 280          * The bss section will be memset to zero later in the initialization so
 281          * there is no need to zero it after changing the memory encryption
 282          * attribute.
 283          */
 284         if (mem_encrypt_active()) {
 285                 vaddr = (unsigned long)__start_bss_decrypted;
 286                 vaddr_end = (unsigned long)__end_bss_decrypted;
 287                 for (; vaddr < vaddr_end; vaddr += PMD_SIZE) {
 288                         i = pmd_index(vaddr);
 289                         pmd[i] -= sme_get_me_mask();
 290                 }
 291         }
 292
 293         /*
 294          * Return the SME encryption mask (if SME is active) to be used as a
 295          * modifier for the initial pgdir entry programmed into CR3.
 296          */
 297         return sme_get_me_mask();
 298 }
 299
 300 unsigned long __startup_secondary_64(void)
 301 {
 302         /*
 303          * Return the SME encryption mask (if SME is active) to be used as a
 304          * modifier for the initial pgdir entry programmed into CR3.
 305          */
 306         return sme_get_me_mask();
 307 }
 308
 309 /* Wipe all early page tables except for the kernel symbol map */
 310 static void __init reset_early_page_tables(void)
 311 {
 312         memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
 313         next_early_pgt = 0;
 314         write_cr3(__sme_pa_nodebug(early_top_pgt));
 315 }
 316
 317 /* Create a new PMD entry */
 318 int __init __early_make_pgtable(unsigned long address, pmdval_t pmd)
 319 {
 320         unsigned long physaddr = address - __PAGE_OFFSET;
 321         pgdval_t pgd, *pgd_p;
 322         p4dval_t p4d, *p4d_p;
 323         pudval_t pud, *pud_p;
 324         pmdval_t *pmd_p;
 325
 326         /* Invalid address or early pgt is done ?  */
 327         if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
 328                 return -1;
 329
 330 again:
 331         pgd_p = &early_top_pgt[pgd_index(address)].pgd;
 332         pgd = *pgd_p;
 333
 334         /*
 335          * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
 336          * critical -- __PAGE_OFFSET would point us back into the dynamic
 337          * range and we might end up looping forever...
 338          */
 339         if (!pgtable_l5_enabled())
 340                 p4d_p = pgd_p;
 341         else if (pgd)
 342                 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 343         else {
 344                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 345                         reset_early_page_tables();
 346                         goto again;
 347                 }
 348
 349                 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
 350                 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
 351                 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 352         }
 353         p4d_p += p4d_index(address);
 354         p4d = *p4d_p;
 355
 356         if (p4d)
 357                 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 358         else {
 359                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 360                         reset_early_page_tables();
 361                         goto again;
 362                 }
 363
 364                 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
 365                 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
 366                 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 367         }
 368         pud_p += pud_index(address);
 369         pud = *pud_p;
 370
 371         if (pud)
 372                 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 373         else {
 374                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 375                         reset_early_page_tables();
 376                         goto again;
 377                 }
 378
 379                 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
 380                 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
 381                 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 382         }
 383         pmd_p[pmd_index(address)] = pmd;
 384
 385         return 0;
 386 }
 387
 388 int __init early_make_pgtable(unsigned long address)
 389 {
 390         unsigned long physaddr = address - __PAGE_OFFSET;
 391         pmdval_t pmd;
 392
 393         pmd = (physaddr & PMD_MASK) + early_pmd_flags;
 394
 395         return __early_make_pgtable(address, pmd);
 396 }
 397
 398 /* Don't add a printk in there. printk relies on the PDA which is not initialized
 399    yet. */
 400 static void __init clear_bss(void)
 401 {
 402         memset(__bss_start, 0,
 403                (unsigned long) __bss_stop - (unsigned long) __bss_start);
 404 }
 405
 406 static unsigned long get_cmd_line_ptr(void)
 407 {
 408         unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
 409
 410         cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
 411
 412         return cmd_line_ptr;
 413 }
 414
 415 static void __init copy_bootdata(char *real_mode_data)
 416 {
 417         char * command_line;
 418         unsigned long cmd_line_ptr;
 419
 420         /*
 421          * If SME is active, this will create decrypted mappings of the
 422          * boot data in advance of the copy operations.
 423          */
 424         sme_map_bootdata(real_mode_data);
 425
 426         memcpy(&boot_params, real_mode_data, sizeof(boot_params));
 427         sanitize_boot_params(&boot_params);
 428         cmd_line_ptr = get_cmd_line_ptr();
 429         if (cmd_line_ptr) {
 430                 command_line = __va(cmd_line_ptr);
 431                 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 432         }
 433
 434         /*
 435          * The old boot data is no longer needed and won't be reserved,
 436          * freeing up that memory for use by the system. If SME is active,
 437          * we need to remove the mappings that were created so that the
 438          * memory doesn't remain mapped as decrypted.
 439          */
 440         sme_unmap_bootdata(real_mode_data);
 441 }
 442
 443 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
 444 {
 445         /*
 446          * Build-time sanity checks on the kernel image and module
 447          * area mappings. (these are purely build-time and produce no code)
 448          */
 449         BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
 450         BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
 451         BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
 452         BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
 453         BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
 454         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 455         MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 456                                 (__START_KERNEL & PGDIR_MASK)));
 457         BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
 458
 459         cr4_init_shadow();
 460
 461         /* Kill off the identity-map trampoline */
 462         reset_early_page_tables();
 463
 464         clear_bss();
 465
 466         clear_page(init_top_pgt);
 467
 468         /*
 469          * SME support may update early_pmd_flags to include the memory
 470          * encryption mask, so it needs to be called before anything
 471          * that may generate a page fault.
 472          */
 473         sme_early_init();
 474
 475         kasan_early_init();
 476
 477         idt_setup_early_handler();
 478
 479         copy_bootdata(__va(real_mode_data));
 480
 481         /*
 482          * Load microcode early on BSP.
 483          */
 484         load_ucode_bsp();
 485
 486         /* set init_top_pgt kernel high mapping*/
 487         init_top_pgt[511] = early_top_pgt[511];
 488
 489         x86_64_start_reservations(real_mode_data);
 490 }
 491
 492 void __init x86_64_start_reservations(char *real_mode_data)
 493 {
 494         /* version is always not zero if it is copied */
 495         if (!boot_params.hdr.version)
 496                 copy_bootdata(__va(real_mode_data));
 497
 498         x86_early_init_platform_quirks();
 499
 500         switch (boot_params.hdr.hardware_subarch) {
 501         case X86_SUBARCH_INTEL_MID:
 502                 x86_intel_mid_early_setup();
 503                 break;
 504         default:
 505                 break;
 506         }
 507
 508         start_kernel();
 509 }
 510
 511 /*
 512  * Setup boot CPU state needed before kernel switches to virtual addresses.
 513  */
 514 void __head startup_64_setup_env(unsigned long physbase)
 515 {
 516         /* Load GDT */
 517         startup_gdt_descr.address = (unsigned long)fixup_pointer(startup_gdt, physbase);
 518         native_load_gdt(&startup_gdt_descr);
 519
 520         /* New GDT is live - reload data segment registers */
 521         asm volatile("movl %%eax, %%ds\n"
 522                      "movl %%eax, %%ss\n"
 523                      "movl %%eax, %%es\n" : : "a"(__KERNEL_DS) : "memory");
 524 }