arch/x86/kernel/head64.c

   1 /*
   2  *  prepare to run common code
   3  *
   4  *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
   5  */
   6
   7 #define DISABLE_BRANCH_PROFILING
   8 #include <linux/init.h>
   9 #include <linux/linkage.h>
  10 #include <linux/types.h>
  11 #include <linux/kernel.h>
  12 #include <linux/string.h>
  13 #include <linux/percpu.h>
  14 #include <linux/start_kernel.h>
  15 #include <linux/io.h>
  16 #include <linux/memblock.h>
  17 #include <linux/mem_encrypt.h>
  18
  19 #include <asm/processor.h>
  20 #include <asm/proto.h>
  21 #include <asm/smp.h>
  22 #include <asm/setup.h>
  23 #include <asm/desc.h>
  24 #include <asm/pgtable.h>
  25 #include <asm/tlbflush.h>
  26 #include <asm/sections.h>
  27 #include <asm/kdebug.h>
  28 #include <asm/e820/api.h>
  29 #include <asm/bios_ebda.h>
  30 #include <asm/bootparam_utils.h>
  31 #include <asm/microcode.h>
  32 #include <asm/kasan.h>
  33
  34 /*
  35  * Manage page tables very early on.
  36  */
  37 extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD];
  38 static unsigned int __initdata next_early_pgt;
  39 pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX);
  40
  41 #define __head  __section(.head.text)
  42
  43 static void __head *fixup_pointer(void *ptr, unsigned long physaddr)
  44 {
  45         return ptr - (void *)_text + (void *)physaddr;
  46 }
  47
  48 unsigned long __head __startup_64(unsigned long physaddr,
  49                                   struct boot_params *bp)
  50 {
  51         unsigned long load_delta, *p;
  52         unsigned long pgtable_flags;
  53         pgdval_t *pgd;
  54         p4dval_t *p4d;
  55         pudval_t *pud;
  56         pmdval_t *pmd, pmd_entry;
  57         int i;
  58         unsigned int *next_pgt_ptr;
  59
  60         /* Is the address too large? */
  61         if (physaddr >> MAX_PHYSMEM_BITS)
  62                 for (;;);
  63
  64         /*
  65          * Compute the delta between the address I am compiled to run at
  66          * and the address I am actually running at.
  67          */
  68         load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
  69
  70         /* Is the address not 2M aligned? */
  71         if (load_delta & ~PMD_PAGE_MASK)
  72                 for (;;);
  73
  74         /* Activate Secure Memory Encryption (SME) if supported and enabled */
  75         sme_enable(bp);
  76
  77         /* Include the SME encryption mask in the fixup value */
  78         load_delta += sme_get_me_mask();
  79
  80         /* Fixup the physical addresses in the page table */
  81
  82         pgd = fixup_pointer(&early_top_pgt, physaddr);
  83         pgd[pgd_index(__START_KERNEL_map)] += load_delta;
  84
  85         if (IS_ENABLED(CONFIG_X86_5LEVEL)) {
  86                 p4d = fixup_pointer(&level4_kernel_pgt, physaddr);
  87                 p4d[511] += load_delta;
  88         }
  89
  90         pud = fixup_pointer(&level3_kernel_pgt, physaddr);
  91         pud[510] += load_delta;
  92         pud[511] += load_delta;
  93
  94         pmd = fixup_pointer(level2_fixmap_pgt, physaddr);
  95         pmd[506] += load_delta;
  96
  97         /*
  98          * Set up the identity mapping for the switchover.  These
  99          * entries should *NOT* have the global bit set!  This also
 100          * creates a bunch of nonsense entries but that is fine --
 101          * it avoids problems around wraparound.
 102          */
 103
 104         next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr);
 105         pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
 106         pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
 107
 108         pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask();
 109
 110         if (IS_ENABLED(CONFIG_X86_5LEVEL)) {
 111                 p4d = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr);
 112
 113                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 114                 pgd[i + 0] = (pgdval_t)p4d + pgtable_flags;
 115                 pgd[i + 1] = (pgdval_t)p4d + pgtable_flags;
 116
 117                 i = (physaddr >> P4D_SHIFT) % PTRS_PER_P4D;
 118                 p4d[i + 0] = (pgdval_t)pud + pgtable_flags;
 119                 p4d[i + 1] = (pgdval_t)pud + pgtable_flags;
 120         } else {
 121                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 122                 pgd[i + 0] = (pgdval_t)pud + pgtable_flags;
 123                 pgd[i + 1] = (pgdval_t)pud + pgtable_flags;
 124         }
 125
 126         i = (physaddr >> PUD_SHIFT) % PTRS_PER_PUD;
 127         pud[i + 0] = (pudval_t)pmd + pgtable_flags;
 128         pud[i + 1] = (pudval_t)pmd + pgtable_flags;
 129
 130         pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
 131         pmd_entry += sme_get_me_mask();
 132         pmd_entry +=  physaddr;
 133
 134         for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
 135                 int idx = i + (physaddr >> PMD_SHIFT) % PTRS_PER_PMD;
 136                 pmd[idx] = pmd_entry + i * PMD_SIZE;
 137         }
 138
 139         /*
 140          * Fixup the kernel text+data virtual addresses. Note that
 141          * we might write invalid pmds, when the kernel is relocated
 142          * cleanup_highmap() fixes this up along with the mappings
 143          * beyond _end.
 144          */
 145
 146         pmd = fixup_pointer(level2_kernel_pgt, physaddr);
 147         for (i = 0; i < PTRS_PER_PMD; i++) {
 148                 if (pmd[i] & _PAGE_PRESENT)
 149                         pmd[i] += load_delta;
 150         }
 151
 152         /*
 153          * Fixup phys_base - remove the memory encryption mask to obtain
 154          * the true physical address.
 155          */
 156         p = fixup_pointer(&phys_base, physaddr);
 157         *p += load_delta - sme_get_me_mask();
 158
 159         /* Encrypt the kernel (if SME is active) */
 160         sme_encrypt_kernel();
 161
 162         /*
 163          * Return the SME encryption mask (if SME is active) to be used as a
 164          * modifier for the initial pgdir entry programmed into CR3.
 165          */
 166         return sme_get_me_mask();
 167 }
 168
 169 unsigned long __startup_secondary_64(void)
 170 {
 171         /*
 172          * Return the SME encryption mask (if SME is active) to be used as a
 173          * modifier for the initial pgdir entry programmed into CR3.
 174          */
 175         return sme_get_me_mask();
 176 }
 177
 178 /* Wipe all early page tables except for the kernel symbol map */
 179 static void __init reset_early_page_tables(void)
 180 {
 181         memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
 182         next_early_pgt = 0;
 183         write_cr3(__sme_pa_nodebug(early_top_pgt));
 184 }
 185
 186 /* Create a new PMD entry */
 187 int __init __early_make_pgtable(unsigned long address, pmdval_t pmd)
 188 {
 189         unsigned long physaddr = address - __PAGE_OFFSET;
 190         pgdval_t pgd, *pgd_p;
 191         p4dval_t p4d, *p4d_p;
 192         pudval_t pud, *pud_p;
 193         pmdval_t *pmd_p;
 194
 195         /* Invalid address or early pgt is done ?  */
 196         if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
 197                 return -1;
 198
 199 again:
 200         pgd_p = &early_top_pgt[pgd_index(address)].pgd;
 201         pgd = *pgd_p;
 202
 203         /*
 204          * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
 205          * critical -- __PAGE_OFFSET would point us back into the dynamic
 206          * range and we might end up looping forever...
 207          */
 208         if (!IS_ENABLED(CONFIG_X86_5LEVEL))
 209                 p4d_p = pgd_p;
 210         else if (pgd)
 211                 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 212         else {
 213                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 214                         reset_early_page_tables();
 215                         goto again;
 216                 }
 217
 218                 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
 219                 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
 220                 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 221         }
 222         p4d_p += p4d_index(address);
 223         p4d = *p4d_p;
 224
 225         if (p4d)
 226                 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 227         else {
 228                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 229                         reset_early_page_tables();
 230                         goto again;
 231                 }
 232
 233                 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
 234                 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
 235                 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 236         }
 237         pud_p += pud_index(address);
 238         pud = *pud_p;
 239
 240         if (pud)
 241                 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 242         else {
 243                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 244                         reset_early_page_tables();
 245                         goto again;
 246                 }
 247
 248                 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
 249                 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
 250                 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 251         }
 252         pmd_p[pmd_index(address)] = pmd;
 253
 254         return 0;
 255 }
 256
 257 int __init early_make_pgtable(unsigned long address)
 258 {
 259         unsigned long physaddr = address - __PAGE_OFFSET;
 260         pmdval_t pmd;
 261
 262         pmd = (physaddr & PMD_MASK) + early_pmd_flags;
 263
 264         return __early_make_pgtable(address, pmd);
 265 }
 266
 267 /* Don't add a printk in there. printk relies on the PDA which is not initialized
 268    yet. */
 269 static void __init clear_bss(void)
 270 {
 271         memset(__bss_start, 0,
 272                (unsigned long) __bss_stop - (unsigned long) __bss_start);
 273 }
 274
 275 static unsigned long get_cmd_line_ptr(void)
 276 {
 277         unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
 278
 279         cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
 280
 281         return cmd_line_ptr;
 282 }
 283
 284 static void __init copy_bootdata(char *real_mode_data)
 285 {
 286         char * command_line;
 287         unsigned long cmd_line_ptr;
 288
 289         /*
 290          * If SME is active, this will create decrypted mappings of the
 291          * boot data in advance of the copy operations.
 292          */
 293         sme_map_bootdata(real_mode_data);
 294
 295         memcpy(&boot_params, real_mode_data, sizeof boot_params);
 296         sanitize_boot_params(&boot_params);
 297         cmd_line_ptr = get_cmd_line_ptr();
 298         if (cmd_line_ptr) {
 299                 command_line = __va(cmd_line_ptr);
 300                 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 301         }
 302
 303         /*
 304          * The old boot data is no longer needed and won't be reserved,
 305          * freeing up that memory for use by the system. If SME is active,
 306          * we need to remove the mappings that were created so that the
 307          * memory doesn't remain mapped as decrypted.
 308          */
 309         sme_unmap_bootdata(real_mode_data);
 310 }
 311
 312 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
 313 {
 314         int i;
 315
 316         /*
 317          * Build-time sanity checks on the kernel image and module
 318          * area mappings. (these are purely build-time and produce no code)
 319          */
 320         BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
 321         BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
 322         BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
 323         BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
 324         BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
 325         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 326         BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 327                                 (__START_KERNEL & PGDIR_MASK)));
 328         BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
 329
 330         cr4_init_shadow();
 331
 332         /* Kill off the identity-map trampoline */
 333         reset_early_page_tables();
 334
 335         clear_bss();
 336
 337         clear_page(init_top_pgt);
 338
 339         /*
 340          * SME support may update early_pmd_flags to include the memory
 341          * encryption mask, so it needs to be called before anything
 342          * that may generate a page fault.
 343          */
 344         sme_early_init();
 345
 346         kasan_early_init();
 347
 348         for (i = 0; i < NUM_EXCEPTION_VECTORS; i++)
 349                 set_intr_gate(i, early_idt_handler_array[i]);
 350         load_idt((const struct desc_ptr *)&idt_descr);
 351
 352         copy_bootdata(__va(real_mode_data));
 353
 354         /*
 355          * Load microcode early on BSP.
 356          */
 357         load_ucode_bsp();
 358
 359         /* set init_top_pgt kernel high mapping*/
 360         init_top_pgt[511] = early_top_pgt[511];
 361
 362         x86_64_start_reservations(real_mode_data);
 363 }
 364
 365 void __init x86_64_start_reservations(char *real_mode_data)
 366 {
 367         /* version is always not zero if it is copied */
 368         if (!boot_params.hdr.version)
 369                 copy_bootdata(__va(real_mode_data));
 370
 371         x86_early_init_platform_quirks();
 372
 373         switch (boot_params.hdr.hardware_subarch) {
 374         case X86_SUBARCH_INTEL_MID:
 375                 x86_intel_mid_early_setup();
 376                 break;
 377         default:
 378                 break;
 379         }
 380
 381         start_kernel();
 382 }