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 pgd_t early_top_pgt[PTRS_PER_PGD];
  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 #define __head  __section(.head.text)
  43
  44 static void __head *fixup_pointer(void *ptr, unsigned long physaddr)
  45 {
  46         return ptr - (void *)_text + (void *)physaddr;
  47 }
  48
  49 unsigned long __head __startup_64(unsigned long physaddr)
  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
  59         /* Is the address too large? */
  60         if (physaddr >> MAX_PHYSMEM_BITS)
  61                 for (;;);
  62
  63         /*
  64          * Compute the delta between the address I am compiled to run at
  65          * and the address I am actually running at.
  66          */
  67         load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
  68
  69         /* Is the address not 2M aligned? */
  70         if (load_delta & ~PMD_PAGE_MASK)
  71                 for (;;);
  72
  73         /* Activate Secure Memory Encryption (SME) if supported and enabled */
  74         sme_enable();
  75
  76         /* Include the SME encryption mask in the fixup value */
  77         load_delta += sme_get_me_mask();
  78
  79         /* Fixup the physical addresses in the page table */
  80
  81         pgd = fixup_pointer(&early_top_pgt, physaddr);
  82         pgd[pgd_index(__START_KERNEL_map)] += load_delta;
  83
  84         if (IS_ENABLED(CONFIG_X86_5LEVEL)) {
  85                 p4d = fixup_pointer(&level4_kernel_pgt, physaddr);
  86                 p4d[511] += load_delta;
  87         }
  88
  89         pud = fixup_pointer(&level3_kernel_pgt, physaddr);
  90         pud[510] += load_delta;
  91         pud[511] += load_delta;
  92
  93         pmd = fixup_pointer(level2_fixmap_pgt, physaddr);
  94         pmd[506] += load_delta;
  95
  96         /*
  97          * Set up the identity mapping for the switchover.  These
  98          * entries should *NOT* have the global bit set!  This also
  99          * creates a bunch of nonsense entries but that is fine --
 100          * it avoids problems around wraparound.
 101          */
 102
 103         pud = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr);
 104         pmd = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr);
 105         pgtable_flags = _KERNPG_TABLE + sme_get_me_mask();
 106
 107         if (IS_ENABLED(CONFIG_X86_5LEVEL)) {
 108                 p4d = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr);
 109
 110                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 111                 pgd[i + 0] = (pgdval_t)p4d + pgtable_flags;
 112                 pgd[i + 1] = (pgdval_t)p4d + pgtable_flags;
 113
 114                 i = (physaddr >> P4D_SHIFT) % PTRS_PER_P4D;
 115                 p4d[i + 0] = (pgdval_t)pud + pgtable_flags;
 116                 p4d[i + 1] = (pgdval_t)pud + pgtable_flags;
 117         } else {
 118                 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
 119                 pgd[i + 0] = (pgdval_t)pud + pgtable_flags;
 120                 pgd[i + 1] = (pgdval_t)pud + pgtable_flags;
 121         }
 122
 123         i = (physaddr >> PUD_SHIFT) % PTRS_PER_PUD;
 124         pud[i + 0] = (pudval_t)pmd + pgtable_flags;
 125         pud[i + 1] = (pudval_t)pmd + pgtable_flags;
 126
 127         pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
 128         pmd_entry += sme_get_me_mask();
 129         pmd_entry +=  physaddr;
 130
 131         for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
 132                 int idx = i + (physaddr >> PMD_SHIFT) % PTRS_PER_PMD;
 133                 pmd[idx] = pmd_entry + i * PMD_SIZE;
 134         }
 135
 136         /*
 137          * Fixup the kernel text+data virtual addresses. Note that
 138          * we might write invalid pmds, when the kernel is relocated
 139          * cleanup_highmap() fixes this up along with the mappings
 140          * beyond _end.
 141          */
 142
 143         pmd = fixup_pointer(level2_kernel_pgt, physaddr);
 144         for (i = 0; i < PTRS_PER_PMD; i++) {
 145                 if (pmd[i] & _PAGE_PRESENT)
 146                         pmd[i] += load_delta;
 147         }
 148
 149         /*
 150          * Fixup phys_base - remove the memory encryption mask to obtain
 151          * the true physical address.
 152          */
 153         p = fixup_pointer(&phys_base, physaddr);
 154         *p += load_delta - sme_get_me_mask();
 155
 156         /* Encrypt the kernel (if SME is active) */
 157         sme_encrypt_kernel();
 158
 159         /*
 160          * Return the SME encryption mask (if SME is active) to be used as a
 161          * modifier for the initial pgdir entry programmed into CR3.
 162          */
 163         return sme_get_me_mask();
 164 }
 165
 166 unsigned long __startup_secondary_64(void)
 167 {
 168         /*
 169          * Return the SME encryption mask (if SME is active) to be used as a
 170          * modifier for the initial pgdir entry programmed into CR3.
 171          */
 172         return sme_get_me_mask();
 173 }
 174
 175 /* Wipe all early page tables except for the kernel symbol map */
 176 static void __init reset_early_page_tables(void)
 177 {
 178         memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
 179         next_early_pgt = 0;
 180         write_cr3(__pa_nodebug(early_top_pgt));
 181 }
 182
 183 /* Create a new PMD entry */
 184 int __init early_make_pgtable(unsigned long address)
 185 {
 186         unsigned long physaddr = address - __PAGE_OFFSET;
 187         pgdval_t pgd, *pgd_p;
 188         p4dval_t p4d, *p4d_p;
 189         pudval_t pud, *pud_p;
 190         pmdval_t pmd, *pmd_p;
 191
 192         /* Invalid address or early pgt is done ?  */
 193         if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
 194                 return -1;
 195
 196 again:
 197         pgd_p = &early_top_pgt[pgd_index(address)].pgd;
 198         pgd = *pgd_p;
 199
 200         /*
 201          * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
 202          * critical -- __PAGE_OFFSET would point us back into the dynamic
 203          * range and we might end up looping forever...
 204          */
 205         if (!IS_ENABLED(CONFIG_X86_5LEVEL))
 206                 p4d_p = pgd_p;
 207         else if (pgd)
 208                 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 209         else {
 210                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 211                         reset_early_page_tables();
 212                         goto again;
 213                 }
 214
 215                 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
 216                 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
 217                 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 218         }
 219         p4d_p += p4d_index(address);
 220         p4d = *p4d_p;
 221
 222         if (p4d)
 223                 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 224         else {
 225                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 226                         reset_early_page_tables();
 227                         goto again;
 228                 }
 229
 230                 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
 231                 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
 232                 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 233         }
 234         pud_p += pud_index(address);
 235         pud = *pud_p;
 236
 237         if (pud)
 238                 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
 239         else {
 240                 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
 241                         reset_early_page_tables();
 242                         goto again;
 243                 }
 244
 245                 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
 246                 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
 247                 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
 248         }
 249         pmd = (physaddr & PMD_MASK) + early_pmd_flags;
 250         pmd_p[pmd_index(address)] = pmd;
 251
 252         return 0;
 253 }
 254
 255 /* Don't add a printk in there. printk relies on the PDA which is not initialized
 256    yet. */
 257 static void __init clear_bss(void)
 258 {
 259         memset(__bss_start, 0,
 260                (unsigned long) __bss_stop - (unsigned long) __bss_start);
 261 }
 262
 263 static unsigned long get_cmd_line_ptr(void)
 264 {
 265         unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
 266
 267         cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
 268
 269         return cmd_line_ptr;
 270 }
 271
 272 static void __init copy_bootdata(char *real_mode_data)
 273 {
 274         char * command_line;
 275         unsigned long cmd_line_ptr;
 276
 277         memcpy(&boot_params, real_mode_data, sizeof boot_params);
 278         sanitize_boot_params(&boot_params);
 279         cmd_line_ptr = get_cmd_line_ptr();
 280         if (cmd_line_ptr) {
 281                 command_line = __va(cmd_line_ptr);
 282                 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 283         }
 284 }
 285
 286 asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
 287 {
 288         int i;
 289
 290         /*
 291          * Build-time sanity checks on the kernel image and module
 292          * area mappings. (these are purely build-time and produce no code)
 293          */
 294         BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
 295         BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
 296         BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
 297         BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
 298         BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
 299         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 300         BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 301                                 (__START_KERNEL & PGDIR_MASK)));
 302         BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
 303
 304         cr4_init_shadow();
 305
 306         /* Kill off the identity-map trampoline */
 307         reset_early_page_tables();
 308
 309         clear_bss();
 310
 311         clear_page(init_top_pgt);
 312
 313         kasan_early_init();
 314
 315         for (i = 0; i < NUM_EXCEPTION_VECTORS; i++)
 316                 set_intr_gate(i, early_idt_handler_array[i]);
 317         load_idt((const struct desc_ptr *)&idt_descr);
 318
 319         copy_bootdata(__va(real_mode_data));
 320
 321         /*
 322          * Load microcode early on BSP.
 323          */
 324         load_ucode_bsp();
 325
 326         /* set init_top_pgt kernel high mapping*/
 327         init_top_pgt[511] = early_top_pgt[511];
 328
 329         x86_64_start_reservations(real_mode_data);
 330 }
 331
 332 void __init x86_64_start_reservations(char *real_mode_data)
 333 {
 334         /* version is always not zero if it is copied */
 335         if (!boot_params.hdr.version)
 336                 copy_bootdata(__va(real_mode_data));
 337
 338         x86_early_init_platform_quirks();
 339
 340         switch (boot_params.hdr.hardware_subarch) {
 341         case X86_SUBARCH_INTEL_MID:
 342                 x86_intel_mid_early_setup();
 343                 break;
 344         default:
 345                 break;
 346         }
 347
 348         start_kernel();
 349 }