arch/x86/kernel/machine_kexec_64.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * handle transition of Linux booting another kernel
   4  * Copyright (C) 2002-2005 Eric Biederman  <ebiederm@xmission.com>
   5  */
   6
   7 #define pr_fmt(fmt)     "kexec: " fmt
   8
   9 #include <linux/mm.h>
  10 #include <linux/kexec.h>
  11 #include <linux/string.h>
  12 #include <linux/gfp.h>
  13 #include <linux/reboot.h>
  14 #include <linux/numa.h>
  15 #include <linux/ftrace.h>
  16 #include <linux/io.h>
  17 #include <linux/suspend.h>
  18 #include <linux/vmalloc.h>
  19 #include <linux/efi.h>
  20 #include <linux/cc_platform.h>
  21
  22 #include <asm/init.h>
  23 #include <asm/tlbflush.h>
  24 #include <asm/mmu_context.h>
  25 #include <asm/io_apic.h>
  26 #include <asm/debugreg.h>
  27 #include <asm/kexec-bzimage64.h>
  28 #include <asm/setup.h>
  29 #include <asm/set_memory.h>
  30 #include <asm/cpu.h>
  31
  32 #ifdef CONFIG_ACPI
  33 /*
  34  * Used while adding mapping for ACPI tables.
  35  * Can be reused when other iomem regions need be mapped
  36  */
  37 struct init_pgtable_data {
  38         struct x86_mapping_info *info;
  39         pgd_t *level4p;
  40 };
  41
  42 static int mem_region_callback(struct resource *res, void *arg)
  43 {
  44         struct init_pgtable_data *data = arg;
  45
  46         return kernel_ident_mapping_init(data->info, data->level4p,
  47                                          res->start, res->end + 1);
  48 }
  49
  50 static int
  51 map_acpi_tables(struct x86_mapping_info *info, pgd_t *level4p)
  52 {
  53         struct init_pgtable_data data;
  54         unsigned long flags;
  55         int ret;
  56
  57         data.info = info;
  58         data.level4p = level4p;
  59         flags = IORESOURCE_MEM | IORESOURCE_BUSY;
  60
  61         ret = walk_iomem_res_desc(IORES_DESC_ACPI_TABLES, flags, 0, -1,
  62                                   &data, mem_region_callback);
  63         if (ret && ret != -EINVAL)
  64                 return ret;
  65
  66         /* ACPI tables could be located in ACPI Non-volatile Storage region */
  67         ret = walk_iomem_res_desc(IORES_DESC_ACPI_NV_STORAGE, flags, 0, -1,
  68                                   &data, mem_region_callback);
  69         if (ret && ret != -EINVAL)
  70                 return ret;
  71
  72         return 0;
  73 }
  74 #else
  75 static int map_acpi_tables(struct x86_mapping_info *info, pgd_t *level4p) { return 0; }
  76 #endif
  77
  78 #ifdef CONFIG_KEXEC_FILE
  79 const struct kexec_file_ops * const kexec_file_loaders[] = {
  80                 &kexec_bzImage64_ops,
  81                 NULL
  82 };
  83 #endif
  84
  85 static int
  86 map_efi_systab(struct x86_mapping_info *info, pgd_t *level4p)
  87 {
  88 #ifdef CONFIG_EFI
  89         unsigned long mstart, mend;
  90
  91         if (!efi_enabled(EFI_BOOT))
  92                 return 0;
  93
  94         mstart = (boot_params.efi_info.efi_systab |
  95                         ((u64)boot_params.efi_info.efi_systab_hi<<32));
  96
  97         if (efi_enabled(EFI_64BIT))
  98                 mend = mstart + sizeof(efi_system_table_64_t);
  99         else
 100                 mend = mstart + sizeof(efi_system_table_32_t);
 101
 102         if (!mstart)
 103                 return 0;
 104
 105         return kernel_ident_mapping_init(info, level4p, mstart, mend);
 106 #endif
 107         return 0;
 108 }
 109
 110 static void free_transition_pgtable(struct kimage *image)
 111 {
 112         free_page((unsigned long)image->arch.p4d);
 113         image->arch.p4d = NULL;
 114         free_page((unsigned long)image->arch.pud);
 115         image->arch.pud = NULL;
 116         free_page((unsigned long)image->arch.pmd);
 117         image->arch.pmd = NULL;
 118         free_page((unsigned long)image->arch.pte);
 119         image->arch.pte = NULL;
 120 }
 121
 122 static int init_transition_pgtable(struct kimage *image, pgd_t *pgd)
 123 {
 124         pgprot_t prot = PAGE_KERNEL_EXEC_NOENC;
 125         unsigned long vaddr, paddr;
 126         int result = -ENOMEM;
 127         p4d_t *p4d;
 128         pud_t *pud;
 129         pmd_t *pmd;
 130         pte_t *pte;
 131
 132         vaddr = (unsigned long)relocate_kernel;
 133         paddr = __pa(page_address(image->control_code_page)+PAGE_SIZE);
 134         pgd += pgd_index(vaddr);
 135         if (!pgd_present(*pgd)) {
 136                 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
 137                 if (!p4d)
 138                         goto err;
 139                 image->arch.p4d = p4d;
 140                 set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE));
 141         }
 142         p4d = p4d_offset(pgd, vaddr);
 143         if (!p4d_present(*p4d)) {
 144                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
 145                 if (!pud)
 146                         goto err;
 147                 image->arch.pud = pud;
 148                 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
 149         }
 150         pud = pud_offset(p4d, vaddr);
 151         if (!pud_present(*pud)) {
 152                 pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
 153                 if (!pmd)
 154                         goto err;
 155                 image->arch.pmd = pmd;
 156                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 157         }
 158         pmd = pmd_offset(pud, vaddr);
 159         if (!pmd_present(*pmd)) {
 160                 pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
 161                 if (!pte)
 162                         goto err;
 163                 image->arch.pte = pte;
 164                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 165         }
 166         pte = pte_offset_kernel(pmd, vaddr);
 167
 168         if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
 169                 prot = PAGE_KERNEL_EXEC;
 170
 171         set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
 172         return 0;
 173 err:
 174         return result;
 175 }
 176
 177 static void *alloc_pgt_page(void *data)
 178 {
 179         struct kimage *image = (struct kimage *)data;
 180         struct page *page;
 181         void *p = NULL;
 182
 183         page = kimage_alloc_control_pages(image, 0);
 184         if (page) {
 185                 p = page_address(page);
 186                 clear_page(p);
 187         }
 188
 189         return p;
 190 }
 191
 192 static int init_pgtable(struct kimage *image, unsigned long start_pgtable)
 193 {
 194         struct x86_mapping_info info = {
 195                 .alloc_pgt_page = alloc_pgt_page,
 196                 .context        = image,
 197                 .page_flag      = __PAGE_KERNEL_LARGE_EXEC,
 198                 .kernpg_flag    = _KERNPG_TABLE_NOENC,
 199         };
 200         unsigned long mstart, mend;
 201         pgd_t *level4p;
 202         int result;
 203         int i;
 204
 205         level4p = (pgd_t *)__va(start_pgtable);
 206         clear_page(level4p);
 207
 208         if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) {
 209                 info.page_flag   |= _PAGE_ENC;
 210                 info.kernpg_flag |= _PAGE_ENC;
 211         }
 212
 213         if (direct_gbpages)
 214                 info.direct_gbpages = true;
 215
 216         for (i = 0; i < nr_pfn_mapped; i++) {
 217                 mstart = pfn_mapped[i].start << PAGE_SHIFT;
 218                 mend   = pfn_mapped[i].end << PAGE_SHIFT;
 219
 220                 result = kernel_ident_mapping_init(&info,
 221                                                  level4p, mstart, mend);
 222                 if (result)
 223                         return result;
 224         }
 225
 226         /*
 227          * segments's mem ranges could be outside 0 ~ max_pfn,
 228          * for example when jump back to original kernel from kexeced kernel.
 229          * or first kernel is booted with user mem map, and second kernel
 230          * could be loaded out of that range.
 231          */
 232         for (i = 0; i < image->nr_segments; i++) {
 233                 mstart = image->segment[i].mem;
 234                 mend   = mstart + image->segment[i].memsz;
 235
 236                 result = kernel_ident_mapping_init(&info,
 237                                                  level4p, mstart, mend);
 238
 239                 if (result)
 240                         return result;
 241         }
 242
 243         /*
 244          * Prepare EFI systab and ACPI tables for kexec kernel since they are
 245          * not covered by pfn_mapped.
 246          */
 247         result = map_efi_systab(&info, level4p);
 248         if (result)
 249                 return result;
 250
 251         result = map_acpi_tables(&info, level4p);
 252         if (result)
 253                 return result;
 254
 255         return init_transition_pgtable(image, level4p);
 256 }
 257
 258 static void load_segments(void)
 259 {
 260         __asm__ __volatile__ (
 261                 "\tmovl %0,%%ds\n"
 262                 "\tmovl %0,%%es\n"
 263                 "\tmovl %0,%%ss\n"
 264                 "\tmovl %0,%%fs\n"
 265                 "\tmovl %0,%%gs\n"
 266                 : : "a" (__KERNEL_DS) : "memory"
 267                 );
 268 }
 269
 270 int machine_kexec_prepare(struct kimage *image)
 271 {
 272         unsigned long start_pgtable;
 273         int result;
 274
 275         /* Calculate the offsets */
 276         start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 277
 278         /* Setup the identity mapped 64bit page table */
 279         result = init_pgtable(image, start_pgtable);
 280         if (result)
 281                 return result;
 282
 283         return 0;
 284 }
 285
 286 void machine_kexec_cleanup(struct kimage *image)
 287 {
 288         free_transition_pgtable(image);
 289 }
 290
 291 /*
 292  * Do not allocate memory (or fail in any way) in machine_kexec().
 293  * We are past the point of no return, committed to rebooting now.
 294  */
 295 void machine_kexec(struct kimage *image)
 296 {
 297         unsigned long page_list[PAGES_NR];
 298         unsigned int host_mem_enc_active;
 299         int save_ftrace_enabled;
 300         void *control_page;
 301
 302         /*
 303          * This must be done before load_segments() since if call depth tracking
 304          * is used then GS must be valid to make any function calls.
 305          */
 306         host_mem_enc_active = cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT);
 307
 308 #ifdef CONFIG_KEXEC_JUMP
 309         if (image->preserve_context)
 310                 save_processor_state();
 311 #endif
 312
 313         save_ftrace_enabled = __ftrace_enabled_save();
 314
 315         /* Interrupts aren't acceptable while we reboot */
 316         local_irq_disable();
 317         hw_breakpoint_disable();
 318         cet_disable();
 319
 320         if (image->preserve_context) {
 321 #ifdef CONFIG_X86_IO_APIC
 322                 /*
 323                  * We need to put APICs in legacy mode so that we can
 324                  * get timer interrupts in second kernel. kexec/kdump
 325                  * paths already have calls to restore_boot_irq_mode()
 326                  * in one form or other. kexec jump path also need one.
 327                  */
 328                 clear_IO_APIC();
 329                 restore_boot_irq_mode();
 330 #endif
 331         }
 332
 333         control_page = page_address(image->control_code_page) + PAGE_SIZE;
 334         __memcpy(control_page, relocate_kernel, KEXEC_CONTROL_CODE_MAX_SIZE);
 335
 336         page_list[PA_CONTROL_PAGE] = virt_to_phys(control_page);
 337         page_list[VA_CONTROL_PAGE] = (unsigned long)control_page;
 338         page_list[PA_TABLE_PAGE] =
 339           (unsigned long)__pa(page_address(image->control_code_page));
 340
 341         if (image->type == KEXEC_TYPE_DEFAULT)
 342                 page_list[PA_SWAP_PAGE] = (page_to_pfn(image->swap_page)
 343                                                 << PAGE_SHIFT);
 344
 345         /*
 346          * The segment registers are funny things, they have both a
 347          * visible and an invisible part.  Whenever the visible part is
 348          * set to a specific selector, the invisible part is loaded
 349          * with from a table in memory.  At no other time is the
 350          * descriptor table in memory accessed.
 351          *
 352          * I take advantage of this here by force loading the
 353          * segments, before I zap the gdt with an invalid value.
 354          */
 355         load_segments();
 356         /*
 357          * The gdt & idt are now invalid.
 358          * If you want to load them you must set up your own idt & gdt.
 359          */
 360         native_idt_invalidate();
 361         native_gdt_invalidate();
 362
 363         /* now call it */
 364         image->start = relocate_kernel((unsigned long)image->head,
 365                                        (unsigned long)page_list,
 366                                        image->start,
 367                                        image->preserve_context,
 368                                        host_mem_enc_active);
 369
 370 #ifdef CONFIG_KEXEC_JUMP
 371         if (image->preserve_context)
 372                 restore_processor_state();
 373 #endif
 374
 375         __ftrace_enabled_restore(save_ftrace_enabled);
 376 }
 377
 378 /* arch-dependent functionality related to kexec file-based syscall */
 379
 380 #ifdef CONFIG_KEXEC_FILE
 381 /*
 382  * Apply purgatory relocations.
 383  *
 384  * @pi:         Purgatory to be relocated.
 385  * @section:    Section relocations applying to.
 386  * @relsec:     Section containing RELAs.
 387  * @symtabsec:  Corresponding symtab.
 388  *
 389  * TODO: Some of the code belongs to generic code. Move that in kexec.c.
 390  */
 391 int arch_kexec_apply_relocations_add(struct purgatory_info *pi,
 392                                      Elf_Shdr *section, const Elf_Shdr *relsec,
 393                                      const Elf_Shdr *symtabsec)
 394 {
 395         unsigned int i;
 396         Elf64_Rela *rel;
 397         Elf64_Sym *sym;
 398         void *location;
 399         unsigned long address, sec_base, value;
 400         const char *strtab, *name, *shstrtab;
 401         const Elf_Shdr *sechdrs;
 402
 403         /* String & section header string table */
 404         sechdrs = (void *)pi->ehdr + pi->ehdr->e_shoff;
 405         strtab = (char *)pi->ehdr + sechdrs[symtabsec->sh_link].sh_offset;
 406         shstrtab = (char *)pi->ehdr + sechdrs[pi->ehdr->e_shstrndx].sh_offset;
 407
 408         rel = (void *)pi->ehdr + relsec->sh_offset;
 409
 410         pr_debug("Applying relocate section %s to %u\n",
 411                  shstrtab + relsec->sh_name, relsec->sh_info);
 412
 413         for (i = 0; i < relsec->sh_size / sizeof(*rel); i++) {
 414
 415                 /*
 416                  * rel[i].r_offset contains byte offset from beginning
 417                  * of section to the storage unit affected.
 418                  *
 419                  * This is location to update. This is temporary buffer
 420                  * where section is currently loaded. This will finally be
 421                  * loaded to a different address later, pointed to by
 422                  * ->sh_addr. kexec takes care of moving it
 423                  *  (kexec_load_segment()).
 424                  */
 425                 location = pi->purgatory_buf;
 426                 location += section->sh_offset;
 427                 location += rel[i].r_offset;
 428
 429                 /* Final address of the location */
 430                 address = section->sh_addr + rel[i].r_offset;
 431
 432                 /*
 433                  * rel[i].r_info contains information about symbol table index
 434                  * w.r.t which relocation must be made and type of relocation
 435                  * to apply. ELF64_R_SYM() and ELF64_R_TYPE() macros get
 436                  * these respectively.
 437                  */
 438                 sym = (void *)pi->ehdr + symtabsec->sh_offset;
 439                 sym += ELF64_R_SYM(rel[i].r_info);
 440
 441                 if (sym->st_name)
 442                         name = strtab + sym->st_name;
 443                 else
 444                         name = shstrtab + sechdrs[sym->st_shndx].sh_name;
 445
 446                 pr_debug("Symbol: %s info: %02x shndx: %02x value=%llx size: %llx\n",
 447                          name, sym->st_info, sym->st_shndx, sym->st_value,
 448                          sym->st_size);
 449
 450                 if (sym->st_shndx == SHN_UNDEF) {
 451                         pr_err("Undefined symbol: %s\n", name);
 452                         return -ENOEXEC;
 453                 }
 454
 455                 if (sym->st_shndx == SHN_COMMON) {
 456                         pr_err("symbol '%s' in common section\n", name);
 457                         return -ENOEXEC;
 458                 }
 459
 460                 if (sym->st_shndx == SHN_ABS)
 461                         sec_base = 0;
 462                 else if (sym->st_shndx >= pi->ehdr->e_shnum) {
 463                         pr_err("Invalid section %d for symbol %s\n",
 464                                sym->st_shndx, name);
 465                         return -ENOEXEC;
 466                 } else
 467                         sec_base = pi->sechdrs[sym->st_shndx].sh_addr;
 468
 469                 value = sym->st_value;
 470                 value += sec_base;
 471                 value += rel[i].r_addend;
 472
 473                 switch (ELF64_R_TYPE(rel[i].r_info)) {
 474                 case R_X86_64_NONE:
 475                         break;
 476                 case R_X86_64_64:
 477                         *(u64 *)location = value;
 478                         break;
 479                 case R_X86_64_32:
 480                         *(u32 *)location = value;
 481                         if (value != *(u32 *)location)
 482                                 goto overflow;
 483                         break;
 484                 case R_X86_64_32S:
 485                         *(s32 *)location = value;
 486                         if ((s64)value != *(s32 *)location)
 487                                 goto overflow;
 488                         break;
 489                 case R_X86_64_PC32:
 490                 case R_X86_64_PLT32:
 491                         value -= (u64)address;
 492                         *(u32 *)location = value;
 493                         break;
 494                 default:
 495                         pr_err("Unknown rela relocation: %llu\n",
 496                                ELF64_R_TYPE(rel[i].r_info));
 497                         return -ENOEXEC;
 498                 }
 499         }
 500         return 0;
 501
 502 overflow:
 503         pr_err("Overflow in relocation type %d value 0x%lx\n",
 504                (int)ELF64_R_TYPE(rel[i].r_info), value);
 505         return -ENOEXEC;
 506 }
 507
 508 int arch_kimage_file_post_load_cleanup(struct kimage *image)
 509 {
 510         vfree(image->elf_headers);
 511         image->elf_headers = NULL;
 512         image->elf_headers_sz = 0;
 513
 514         return kexec_image_post_load_cleanup_default(image);
 515 }
 516 #endif /* CONFIG_KEXEC_FILE */
 517
 518 #ifdef CONFIG_CRASH_DUMP
 519
 520 static int
 521 kexec_mark_range(unsigned long start, unsigned long end, bool protect)
 522 {
 523         struct page *page;
 524         unsigned int nr_pages;
 525
 526         /*
 527          * For physical range: [start, end]. We must skip the unassigned
 528          * crashk resource with zero-valued "end" member.
 529          */
 530         if (!end || start > end)
 531                 return 0;
 532
 533         page = pfn_to_page(start >> PAGE_SHIFT);
 534         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 535         if (protect)
 536                 return set_pages_ro(page, nr_pages);
 537         else
 538                 return set_pages_rw(page, nr_pages);
 539 }
 540
 541 static void kexec_mark_crashkres(bool protect)
 542 {
 543         unsigned long control;
 544
 545         kexec_mark_range(crashk_low_res.start, crashk_low_res.end, protect);
 546
 547         /* Don't touch the control code page used in crash_kexec().*/
 548         control = PFN_PHYS(page_to_pfn(kexec_crash_image->control_code_page));
 549         /* Control code page is located in the 2nd page. */
 550         kexec_mark_range(crashk_res.start, control + PAGE_SIZE - 1, protect);
 551         control += KEXEC_CONTROL_PAGE_SIZE;
 552         kexec_mark_range(control, crashk_res.end, protect);
 553 }
 554
 555 void arch_kexec_protect_crashkres(void)
 556 {
 557         kexec_mark_crashkres(true);
 558 }
 559
 560 void arch_kexec_unprotect_crashkres(void)
 561 {
 562         kexec_mark_crashkres(false);
 563 }
 564 #endif
 565
 566 /*
 567  * During a traditional boot under SME, SME will encrypt the kernel,
 568  * so the SME kexec kernel also needs to be un-encrypted in order to
 569  * replicate a normal SME boot.
 570  *
 571  * During a traditional boot under SEV, the kernel has already been
 572  * loaded encrypted, so the SEV kexec kernel needs to be encrypted in
 573  * order to replicate a normal SEV boot.
 574  */
 575 int arch_kexec_post_alloc_pages(void *vaddr, unsigned int pages, gfp_t gfp)
 576 {
 577         if (!cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT))
 578                 return 0;
 579
 580         /*
 581          * If host memory encryption is active we need to be sure that kexec
 582          * pages are not encrypted because when we boot to the new kernel the
 583          * pages won't be accessed encrypted (initially).
 584          */
 585         return set_memory_decrypted((unsigned long)vaddr, pages);
 586 }
 587
 588 void arch_kexec_pre_free_pages(void *vaddr, unsigned int pages)
 589 {
 590         if (!cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT))
 591                 return;
 592
 593         /*
 594          * If host memory encryption is active we need to reset the pages back
 595          * to being an encrypted mapping before freeing them.
 596          */
 597         set_memory_encrypted((unsigned long)vaddr, pages);
 598 }