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 /*
  36  * Manage page tables very early on.
  37  */
  38 extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD];
  39 static unsigned int __initdata next_early_pgt;
  40 pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX);
  41
  42 #ifdef CONFIG_X86_5LEVEL
  43 unsigned int pgtable_l5_enabled __ro_after_init;
  44 EXPORT_SYMBOL(pgtable_l5_enabled);
  45 unsigned int pgdir_shift __ro_after_init = 48;
  46 EXPORT_SYMBOL(pgdir_shift);
  47 unsigned int ptrs_per_p4d __ro_after_init = 512;
  48 EXPORT_SYMBOL(ptrs_per_p4d);
  49 #endif
  50
  51 #ifdef CONFIG_DYNAMIC_MEMORY_LAYOUT
  52 unsigned long page_offset_base __ro_after_init = __PAGE_OFFSET_BASE;
  53 EXPORT_SYMBOL(page_offset_base);
  54 unsigned long vmalloc_base __ro_after_init = __VMALLOC_BASE;
  55 EXPORT_SYMBOL(vmalloc_base);
  56 unsigned long vmemmap_base __ro_after_init = __VMEMMAP_BASE;
  57 EXPORT_SYMBOL(vmemmap_base);
  58 #endif
  59
  60 #define __head  __section(.head.text)
  61
  62 static void __head *fixup_pointer(void *ptr, unsigned long physaddr)
  63 {
  64         return ptr - (void *)_text + (void *)physaddr;
  65 }
  66
  67 #ifdef CONFIG_X86_5LEVEL
  68 static unsigned int __head *fixup_int(void *ptr, unsigned long physaddr)
  69 {
  70         return fixup_pointer(ptr, physaddr);
  71 }
  72
  73 static void __head check_la57_support(unsigned long physaddr)
  74 {
  75         if (native_cpuid_eax(0) < 7)
  76                 return;
  77
  78         if (!(native_cpuid_ecx(7) & (1 << (X86_FEATURE_LA57 & 31))))
  79                 return;
  80
  81         *fixup_int(&pgtable_l5_enabled, physaddr) = 1;
  82 }
  83 #else
  84 static void __head check_la57_support(unsigned long physaddr) {}
  85 #endif
  86
  87 unsigned long __head __startup_64(unsigned long physaddr,
  88                                   struct boot_params *bp)
  89 {
  90         unsigned long load_delta, *p;
  91         unsigned long pgtable_flags;
  92         pgdval_t *pgd;
  93         p4dval_t *p4d;
  94         pudval_t *pud;
  95         pmdval_t *pmd, pmd_entry;
  96         int i;
  97         unsigned int *next_pgt_ptr;
  98
  99         check_la57_support(physaddr);
 100
 101         /* Is the address too large? */
 102         if (physaddr >> MAX_PHYSMEM_BITS)
 103                 for (;;);
 104
 105         /*
 106          * Compute the delta between the address I am compiled to run at
 107          * and the address I am actually running at.
 108          */
 109         load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
 110
 111         /* Is the address not 2M aligned? */
 112         if (load_delta & ~PMD_PAGE_MASK)
 113                 for (;;);
 114
 115         /* Activate Secure Memory Encryption (SME) if supported and enabled */
 116         sme_enable(bp);
 117
 118         /* Include the SME encryption mask in the fixup value */
 119         load_delta += sme_get_me_mask();
 120
 121         /* Fixup the physical addresses in the page table */
 122
 123         pgd = fixup_pointer(&early_top_pgt, physaddr);
 124         pgd[pgd_index(__START_KERNEL_map)] += load_delta;
 125
 126         if (IS_ENABLED(CONFIG_X86_5LEVEL)) {
 127                 p4d = fixup_pointer(&level4_kernel_pgt, physaddr);
 128                 p4d[511] += load_delta;
 129         }
 130
 131         pud = fixup_pointer(&level3_kernel_pgt, physaddr);
 132         pud[510] += load_delta;
 133         pud[511] += load_delta;
 134
 135         pmd = fixup_pointer(level2_fixmap_pgt, physaddr);
 136         pmd[506] += load_delta;
 137
 138         /*
 139          * Set up the identity mapping for the switchover.  These
 140          * entries should *NOT* have the global bit set!  This also
 141          * creates a bunch of nonsense entries but that is fine --
 142          * it avoids problems around wraparound.
 143          */
 144
 145         next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr);
 146         pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
 147         pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
 148
 149         pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask();
 150
 151         if (IS_ENABLED(CONFIG_X86_5LEVEL)) {
 152                 p4d = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr);
 153
 154                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 155                 pgd[i + 0] = (pgdval_t)p4d + pgtable_flags;
 156                 pgd[i + 1] = (pgdval_t)p4d + pgtable_flags;
 157
 158                 i = (physaddr >> P4D_SHIFT) % PTRS_PER_P4D;
 159                 p4d[i + 0] = (pgdval_t)pud + pgtable_flags;
 160                 p4d[i + 1] = (pgdval_t)pud + pgtable_flags;
 161         } else {
 162                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 163                 pgd[i + 0] = (pgdval_t)pud + pgtable_flags;
 164                 pgd[i + 1] = (pgdval_t)pud + pgtable_flags;
 165         }
 166
 167         i = (physaddr >> PUD_SHIFT) % PTRS_PER_PUD;
 168         pud[i + 0] = (pudval_t)pmd + pgtable_flags;
 169         pud[i + 1] = (pudval_t)pmd + pgtable_flags;
 170
 171         pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
 172         pmd_entry += sme_get_me_mask();
 173         pmd_entry +=  physaddr;
 174
 175         for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
 176                 int idx = i + (physaddr >> PMD_SHIFT) % PTRS_PER_PMD;
 177                 pmd[idx] = pmd_entry + i * PMD_SIZE;
 178         }
 179
 180         /*
 181          * Fixup the kernel text+data virtual addresses. Note that
 182          * we might write invalid pmds, when the kernel is relocated
 183          * cleanup_highmap() fixes this up along with the mappings
 184          * beyond _end.
 185          */
 186
 187         pmd = fixup_pointer(level2_kernel_pgt, physaddr);
 188         for (i = 0; i < PTRS_PER_PMD; i++) {
 189                 if (pmd[i] & _PAGE_PRESENT)
 190                         pmd[i] += load_delta;
 191         }
 192
 193         /*
 194          * Fixup phys_base - remove the memory encryption mask to obtain
 195          * the true physical address.
 196          */
 197         p = fixup_pointer(&phys_base, physaddr);
 198         *p += load_delta - sme_get_me_mask();
 199
 200         /* Encrypt the kernel and related (if SME is active) */
 201         sme_encrypt_kernel(bp);
 202
 203         /*
 204          * Return the SME encryption mask (if SME is active) to be used as a
 205          * modifier for the initial pgdir entry programmed into CR3.
 206          */
 207         return sme_get_me_mask();
 208 }
 209
 210 unsigned long __startup_secondary_64(void)
 211 {
 212         /*
 213          * Return the SME encryption mask (if SME is active) to be used as a
 214          * modifier for the initial pgdir entry programmed into CR3.
 215          */
 216         return sme_get_me_mask();
 217 }
 218
 219 /* Wipe all early page tables except for the kernel symbol map */
 220 static void __init reset_early_page_tables(void)
 221 {
 222         memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
 223         next_early_pgt = 0;
 224         write_cr3(__sme_pa_nodebug(early_top_pgt));
 225 }
 226
 227 /* Create a new PMD entry */
 228 int __init __early_make_pgtable(unsigned long address, pmdval_t pmd)
 229 {
 230         unsigned long physaddr = address - __PAGE_OFFSET;
 231         pgdval_t pgd, *pgd_p;
 232         p4dval_t p4d, *p4d_p;
 233         pudval_t pud, *pud_p;
 234         pmdval_t *pmd_p;
 235
 236         /* Invalid address or early pgt is done ?  */
 237         if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
 238                 return -1;
 239
 240 again:
 241         pgd_p = &early_top_pgt[pgd_index(address)].pgd;
 242         pgd = *pgd_p;
 243
 244         /*
 245          * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
 246          * critical -- __PAGE_OFFSET would point us back into the dynamic
 247          * range and we might end up looping forever...
 248          */
 249         if (!IS_ENABLED(CONFIG_X86_5LEVEL))
 250                 p4d_p = pgd_p;
 251         else if (pgd)
 252                 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 253         else {
 254                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 255                         reset_early_page_tables();
 256                         goto again;
 257                 }
 258
 259                 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
 260                 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
 261                 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 262         }
 263         p4d_p += p4d_index(address);
 264         p4d = *p4d_p;
 265
 266         if (p4d)
 267                 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 268         else {
 269                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 270                         reset_early_page_tables();
 271                         goto again;
 272                 }
 273
 274                 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
 275                 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
 276                 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 277         }
 278         pud_p += pud_index(address);
 279         pud = *pud_p;
 280
 281         if (pud)
 282                 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 283         else {
 284                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 285                         reset_early_page_tables();
 286                         goto again;
 287                 }
 288
 289                 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
 290                 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
 291                 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 292         }
 293         pmd_p[pmd_index(address)] = pmd;
 294
 295         return 0;
 296 }
 297
 298 int __init early_make_pgtable(unsigned long address)
 299 {
 300         unsigned long physaddr = address - __PAGE_OFFSET;
 301         pmdval_t pmd;
 302
 303         pmd = (physaddr & PMD_MASK) + early_pmd_flags;
 304
 305         return __early_make_pgtable(address, pmd);
 306 }
 307
 308 /* Don't add a printk in there. printk relies on the PDA which is not initialized
 309    yet. */
 310 static void __init clear_bss(void)
 311 {
 312         memset(__bss_start, 0,
 313                (unsigned long) __bss_stop - (unsigned long) __bss_start);
 314 }
 315
 316 static unsigned long get_cmd_line_ptr(void)
 317 {
 318         unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
 319
 320         cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
 321
 322         return cmd_line_ptr;
 323 }
 324
 325 static void __init copy_bootdata(char *real_mode_data)
 326 {
 327         char * command_line;
 328         unsigned long cmd_line_ptr;
 329
 330         /*
 331          * If SME is active, this will create decrypted mappings of the
 332          * boot data in advance of the copy operations.
 333          */
 334         sme_map_bootdata(real_mode_data);
 335
 336         memcpy(&boot_params, real_mode_data, sizeof boot_params);
 337         sanitize_boot_params(&boot_params);
 338         cmd_line_ptr = get_cmd_line_ptr();
 339         if (cmd_line_ptr) {
 340                 command_line = __va(cmd_line_ptr);
 341                 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 342         }
 343
 344         /*
 345          * The old boot data is no longer needed and won't be reserved,
 346          * freeing up that memory for use by the system. If SME is active,
 347          * we need to remove the mappings that were created so that the
 348          * memory doesn't remain mapped as decrypted.
 349          */
 350         sme_unmap_bootdata(real_mode_data);
 351 }
 352
 353 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
 354 {
 355         /*
 356          * Build-time sanity checks on the kernel image and module
 357          * area mappings. (these are purely build-time and produce no code)
 358          */
 359         BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
 360         BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
 361         BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
 362         BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
 363         BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
 364         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 365         MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 366                                 (__START_KERNEL & PGDIR_MASK)));
 367         BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
 368
 369         cr4_init_shadow();
 370
 371         /* Kill off the identity-map trampoline */
 372         reset_early_page_tables();
 373
 374         clear_bss();
 375
 376         clear_page(init_top_pgt);
 377
 378         /*
 379          * SME support may update early_pmd_flags to include the memory
 380          * encryption mask, so it needs to be called before anything
 381          * that may generate a page fault.
 382          */
 383         sme_early_init();
 384
 385         kasan_early_init();
 386
 387         idt_setup_early_handler();
 388
 389         copy_bootdata(__va(real_mode_data));
 390
 391         /*
 392          * Load microcode early on BSP.
 393          */
 394         load_ucode_bsp();
 395
 396         /* set init_top_pgt kernel high mapping*/
 397         init_top_pgt[511] = early_top_pgt[511];
 398
 399         x86_64_start_reservations(real_mode_data);
 400 }
 401
 402 void __init x86_64_start_reservations(char *real_mode_data)
 403 {
 404         /* version is always not zero if it is copied */
 405         if (!boot_params.hdr.version)
 406                 copy_bootdata(__va(real_mode_data));
 407
 408         x86_early_init_platform_quirks();
 409
 410         switch (boot_params.hdr.hardware_subarch) {
 411         case X86_SUBARCH_INTEL_MID:
 412                 x86_intel_mid_early_setup();
 413                 break;
 414         default:
 415                 break;
 416         }
 417
 418         start_kernel();
 419 }