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  <[email protected]>
   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 #include <asm/efi.h>
  32
  33 #ifdef CONFIG_ACPI
  34 /*
  35  * Used while adding mapping for ACPI tables.
  36  * Can be reused when other iomem regions need be mapped
  37  */
  38 struct init_pgtable_data {
  39         struct x86_mapping_info *info;
  40         pgd_t *level4p;
  41 };
  42
  43 static int mem_region_callback(struct resource *res, void *arg)
  44 {
  45         struct init_pgtable_data *data = arg;
  46
  47         return kernel_ident_mapping_init(data->info, data->level4p,
  48                                          res->start, res->end + 1);
  49 }
  50
  51 static int
  52 map_acpi_tables(struct x86_mapping_info *info, pgd_t *level4p)
  53 {
  54         struct init_pgtable_data data;
  55         unsigned long flags;
  56         int ret;
  57
  58         data.info = info;
  59         data.level4p = level4p;
  60         flags = IORESOURCE_MEM | IORESOURCE_BUSY;
  61
  62         ret = walk_iomem_res_desc(IORES_DESC_ACPI_TABLES, flags, 0, -1,
  63                                   &data, mem_region_callback);
  64         if (ret && ret != -EINVAL)
  65                 return ret;
  66
  67         /* ACPI tables could be located in ACPI Non-volatile Storage region */
  68         ret = walk_iomem_res_desc(IORES_DESC_ACPI_NV_STORAGE, flags, 0, -1,
  69                                   &data, mem_region_callback);
  70         if (ret && ret != -EINVAL)
  71                 return ret;
  72
  73         return 0;
  74 }
  75 #else
  76 static int map_acpi_tables(struct x86_mapping_info *info, pgd_t *level4p) { return 0; }
  77 #endif
  78
  79 #ifdef CONFIG_KEXEC_FILE
  80 const struct kexec_file_ops * const kexec_file_loaders[] = {
  81                 &kexec_bzImage64_ops,
  82                 NULL
  83 };
  84 #endif
  85
  86 static int
  87 map_efi_systab(struct x86_mapping_info *info, pgd_t *level4p)
  88 {
  89 #ifdef CONFIG_EFI
  90         unsigned long mstart, mend;
  91         void *kaddr;
  92         int ret;
  93
  94         if (!efi_enabled(EFI_BOOT))
  95                 return 0;
  96
  97         mstart = (boot_params.efi_info.efi_systab |
  98                         ((u64)boot_params.efi_info.efi_systab_hi<<32));
  99
 100         if (efi_enabled(EFI_64BIT))
 101                 mend = mstart + sizeof(efi_system_table_64_t);
 102         else
 103                 mend = mstart + sizeof(efi_system_table_32_t);
 104
 105         if (!mstart)
 106                 return 0;
 107
 108         ret = kernel_ident_mapping_init(info, level4p, mstart, mend);
 109         if (ret)
 110                 return ret;
 111
 112         kaddr = memremap(mstart, mend - mstart, MEMREMAP_WB);
 113         if (!kaddr) {
 114                 pr_err("Could not map UEFI system table\n");
 115                 return -ENOMEM;
 116         }
 117
 118         mstart = efi_config_table;
 119
 120         if (efi_enabled(EFI_64BIT)) {
 121                 efi_system_table_64_t *stbl = (efi_system_table_64_t *)kaddr;
 122
 123                 mend = mstart + sizeof(efi_config_table_64_t) * stbl->nr_tables;
 124         } else {
 125                 efi_system_table_32_t *stbl = (efi_system_table_32_t *)kaddr;
 126
 127                 mend = mstart + sizeof(efi_config_table_32_t) * stbl->nr_tables;
 128         }
 129
 130         memunmap(kaddr);
 131
 132         return kernel_ident_mapping_init(info, level4p, mstart, mend);
 133 #endif
 134         return 0;
 135 }
 136
 137 static void free_transition_pgtable(struct kimage *image)
 138 {
 139         free_page((unsigned long)image->arch.p4d);
 140         image->arch.p4d = NULL;
 141         free_page((unsigned long)image->arch.pud);
 142         image->arch.pud = NULL;
 143         free_page((unsigned long)image->arch.pmd);
 144         image->arch.pmd = NULL;
 145         free_page((unsigned long)image->arch.pte);
 146         image->arch.pte = NULL;
 147 }
 148
 149 static int init_transition_pgtable(struct kimage *image, pgd_t *pgd)
 150 {
 151         pgprot_t prot = PAGE_KERNEL_EXEC_NOENC;
 152         unsigned long vaddr, paddr;
 153         int result = -ENOMEM;
 154         p4d_t *p4d;
 155         pud_t *pud;
 156         pmd_t *pmd;
 157         pte_t *pte;
 158
 159         vaddr = (unsigned long)relocate_kernel;
 160         paddr = __pa(page_address(image->control_code_page)+PAGE_SIZE);
 161         pgd += pgd_index(vaddr);
 162         if (!pgd_present(*pgd)) {
 163                 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
 164                 if (!p4d)
 165                         goto err;
 166                 image->arch.p4d = p4d;
 167                 set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE));
 168         }
 169         p4d = p4d_offset(pgd, vaddr);
 170         if (!p4d_present(*p4d)) {
 171                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
 172                 if (!pud)
 173                         goto err;
 174                 image->arch.pud = pud;
 175                 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
 176         }
 177         pud = pud_offset(p4d, vaddr);
 178         if (!pud_present(*pud)) {
 179                 pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
 180                 if (!pmd)
 181                         goto err;
 182                 image->arch.pmd = pmd;
 183                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 184         }
 185         pmd = pmd_offset(pud, vaddr);
 186         if (!pmd_present(*pmd)) {
 187                 pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
 188                 if (!pte)
 189                         goto err;
 190                 image->arch.pte = pte;
 191                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 192         }
 193         pte = pte_offset_kernel(pmd, vaddr);
 194
 195         if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
 196                 prot = PAGE_KERNEL_EXEC;
 197
 198         set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
 199         return 0;
 200 err:
 201         return result;
 202 }
 203
 204 static void *alloc_pgt_page(void *data)
 205 {
 206         struct kimage *image = (struct kimage *)data;
 207         struct page *page;
 208         void *p = NULL;
 209
 210         page = kimage_alloc_control_pages(image, 0);
 211         if (page) {
 212                 p = page_address(page);
 213                 clear_page(p);
 214         }
 215
 216         return p;
 217 }
 218
 219 static int init_pgtable(struct kimage *image, unsigned long start_pgtable)
 220 {
 221         struct x86_mapping_info info = {
 222                 .alloc_pgt_page = alloc_pgt_page,
 223                 .context        = image,
 224                 .page_flag      = __PAGE_KERNEL_LARGE_EXEC,
 225                 .kernpg_flag    = _KERNPG_TABLE_NOENC,
 226         };
 227         unsigned long mstart, mend;
 228         pgd_t *level4p;
 229         int result;
 230         int i;
 231
 232         level4p = (pgd_t *)__va(start_pgtable);
 233         clear_page(level4p);
 234
 235         if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) {
 236                 info.page_flag   |= _PAGE_ENC;
 237                 info.kernpg_flag |= _PAGE_ENC;
 238         }
 239
 240         if (direct_gbpages)
 241                 info.direct_gbpages = true;
 242
 243         for (i = 0; i < nr_pfn_mapped; i++) {
 244                 mstart = pfn_mapped[i].start << PAGE_SHIFT;
 245                 mend   = pfn_mapped[i].end << PAGE_SHIFT;
 246
 247                 result = kernel_ident_mapping_init(&info,
 248                                                  level4p, mstart, mend);
 249                 if (result)
 250                         return result;
 251         }
 252
 253         /*
 254          * segments's mem ranges could be outside 0 ~ max_pfn,
 255          * for example when jump back to original kernel from kexeced kernel.
 256          * or first kernel is booted with user mem map, and second kernel
 257          * could be loaded out of that range.
 258          */
 259         for (i = 0; i < image->nr_segments; i++) {
 260                 mstart = image->segment[i].mem;
 261                 mend   = mstart + image->segment[i].memsz;
 262
 263                 result = kernel_ident_mapping_init(&info,
 264                                                  level4p, mstart, mend);
 265
 266                 if (result)
 267                         return result;
 268         }
 269
 270         /*
 271          * Prepare EFI systab and ACPI tables for kexec kernel since they are
 272          * not covered by pfn_mapped.
 273          */
 274         result = map_efi_systab(&info, level4p);
 275         if (result)
 276                 return result;
 277
 278         result = map_acpi_tables(&info, level4p);
 279         if (result)
 280                 return result;
 281
 282         return init_transition_pgtable(image, level4p);
 283 }
 284
 285 static void load_segments(void)
 286 {
 287         __asm__ __volatile__ (
 288                 "\tmovl %0,%%ds\n"
 289                 "\tmovl %0,%%es\n"
 290                 "\tmovl %0,%%ss\n"
 291                 "\tmovl %0,%%fs\n"
 292                 "\tmovl %0,%%gs\n"
 293                 : : "a" (__KERNEL_DS) : "memory"
 294                 );
 295 }
 296
 297 int machine_kexec_prepare(struct kimage *image)
 298 {
 299         unsigned long start_pgtable;
 300         int result;
 301
 302         /* Calculate the offsets */
 303         start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 304
 305         /* Setup the identity mapped 64bit page table */
 306         result = init_pgtable(image, start_pgtable);
 307         if (result)
 308                 return result;
 309
 310         return 0;
 311 }
 312
 313 void machine_kexec_cleanup(struct kimage *image)
 314 {
 315         free_transition_pgtable(image);
 316 }
 317
 318 /*
 319  * Do not allocate memory (or fail in any way) in machine_kexec().
 320  * We are past the point of no return, committed to rebooting now.
 321  */
 322 void machine_kexec(struct kimage *image)
 323 {
 324         unsigned long page_list[PAGES_NR];
 325         unsigned int host_mem_enc_active;
 326         int save_ftrace_enabled;
 327         void *control_page;
 328
 329         /*
 330          * This must be done before load_segments() since if call depth tracking
 331          * is used then GS must be valid to make any function calls.
 332          */
 333         host_mem_enc_active = cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT);
 334
 335 #ifdef CONFIG_KEXEC_JUMP
 336         if (image->preserve_context)
 337                 save_processor_state();
 338 #endif
 339
 340         save_ftrace_enabled = __ftrace_enabled_save();
 341
 342         /* Interrupts aren't acceptable while we reboot */
 343         local_irq_disable();
 344         hw_breakpoint_disable();
 345         cet_disable();
 346
 347         if (image->preserve_context) {
 348 #ifdef CONFIG_X86_IO_APIC
 349                 /*
 350                  * We need to put APICs in legacy mode so that we can
 351                  * get timer interrupts in second kernel. kexec/kdump
 352                  * paths already have calls to restore_boot_irq_mode()
 353                  * in one form or other. kexec jump path also need one.
 354                  */
 355                 clear_IO_APIC();
 356                 restore_boot_irq_mode();
 357 #endif
 358         }
 359
 360         control_page = page_address(image->control_code_page) + PAGE_SIZE;
 361         __memcpy(control_page, relocate_kernel, KEXEC_CONTROL_CODE_MAX_SIZE);
 362
 363         page_list[PA_CONTROL_PAGE] = virt_to_phys(control_page);
 364         page_list[VA_CONTROL_PAGE] = (unsigned long)control_page;
 365         page_list[PA_TABLE_PAGE] =
 366           (unsigned long)__pa(page_address(image->control_code_page));
 367
 368         if (image->type == KEXEC_TYPE_DEFAULT)
 369                 page_list[PA_SWAP_PAGE] = (page_to_pfn(image->swap_page)
 370                                                 << PAGE_SHIFT);
 371
 372         /*
 373          * The segment registers are funny things, they have both a
 374          * visible and an invisible part.  Whenever the visible part is
 375          * set to a specific selector, the invisible part is loaded
 376          * with from a table in memory.  At no other time is the
 377          * descriptor table in memory accessed.
 378          *
 379          * I take advantage of this here by force loading the
 380          * segments, before I zap the gdt with an invalid value.
 381          */
 382         load_segments();
 383         /*
 384          * The gdt & idt are now invalid.
 385          * If you want to load them you must set up your own idt & gdt.
 386          */
 387         native_idt_invalidate();
 388         native_gdt_invalidate();
 389
 390         /* now call it */
 391         image->start = relocate_kernel((unsigned long)image->head,
 392                                        (unsigned long)page_list,
 393                                        image->start,
 394                                        image->preserve_context,
 395                                        host_mem_enc_active);
 396
 397 #ifdef CONFIG_KEXEC_JUMP
 398         if (image->preserve_context)
 399                 restore_processor_state();
 400 #endif
 401
 402         __ftrace_enabled_restore(save_ftrace_enabled);
 403 }
 404
 405 /* arch-dependent functionality related to kexec file-based syscall */
 406
 407 #ifdef CONFIG_KEXEC_FILE
 408 /*
 409  * Apply purgatory relocations.
 410  *
 411  * @pi:         Purgatory to be relocated.
 412  * @section:    Section relocations applying to.
 413  * @relsec:     Section containing RELAs.
 414  * @symtabsec:  Corresponding symtab.
 415  *
 416  * TODO: Some of the code belongs to generic code. Move that in kexec.c.
 417  */
 418 int arch_kexec_apply_relocations_add(struct purgatory_info *pi,
 419                                      Elf_Shdr *section, const Elf_Shdr *relsec,
 420                                      const Elf_Shdr *symtabsec)
 421 {
 422         unsigned int i;
 423         Elf64_Rela *rel;
 424         Elf64_Sym *sym;
 425         void *location;
 426         unsigned long address, sec_base, value;
 427         const char *strtab, *name, *shstrtab;
 428         const Elf_Shdr *sechdrs;
 429
 430         /* String & section header string table */
 431         sechdrs = (void *)pi->ehdr + pi->ehdr->e_shoff;
 432         strtab = (char *)pi->ehdr + sechdrs[symtabsec->sh_link].sh_offset;
 433         shstrtab = (char *)pi->ehdr + sechdrs[pi->ehdr->e_shstrndx].sh_offset;
 434
 435         rel = (void *)pi->ehdr + relsec->sh_offset;
 436
 437         pr_debug("Applying relocate section %s to %u\n",
 438                  shstrtab + relsec->sh_name, relsec->sh_info);
 439
 440         for (i = 0; i < relsec->sh_size / sizeof(*rel); i++) {
 441
 442                 /*
 443                  * rel[i].r_offset contains byte offset from beginning
 444                  * of section to the storage unit affected.
 445                  *
 446                  * This is location to update. This is temporary buffer
 447                  * where section is currently loaded. This will finally be
 448                  * loaded to a different address later, pointed to by
 449                  * ->sh_addr. kexec takes care of moving it
 450                  *  (kexec_load_segment()).
 451                  */
 452                 location = pi->purgatory_buf;
 453                 location += section->sh_offset;
 454                 location += rel[i].r_offset;
 455
 456                 /* Final address of the location */
 457                 address = section->sh_addr + rel[i].r_offset;
 458
 459                 /*
 460                  * rel[i].r_info contains information about symbol table index
 461                  * w.r.t which relocation must be made and type of relocation
 462                  * to apply. ELF64_R_SYM() and ELF64_R_TYPE() macros get
 463                  * these respectively.
 464                  */
 465                 sym = (void *)pi->ehdr + symtabsec->sh_offset;
 466                 sym += ELF64_R_SYM(rel[i].r_info);
 467
 468                 if (sym->st_name)
 469                         name = strtab + sym->st_name;
 470                 else
 471                         name = shstrtab + sechdrs[sym->st_shndx].sh_name;
 472
 473                 pr_debug("Symbol: %s info: %02x shndx: %02x value=%llx size: %llx\n",
 474                          name, sym->st_info, sym->st_shndx, sym->st_value,
 475                          sym->st_size);
 476
 477                 if (sym->st_shndx == SHN_UNDEF) {
 478                         pr_err("Undefined symbol: %s\n", name);
 479                         return -ENOEXEC;
 480                 }
 481
 482                 if (sym->st_shndx == SHN_COMMON) {
 483                         pr_err("symbol '%s' in common section\n", name);
 484                         return -ENOEXEC;
 485                 }
 486
 487                 if (sym->st_shndx == SHN_ABS)
 488                         sec_base = 0;
 489                 else if (sym->st_shndx >= pi->ehdr->e_shnum) {
 490                         pr_err("Invalid section %d for symbol %s\n",
 491                                sym->st_shndx, name);
 492                         return -ENOEXEC;
 493                 } else
 494                         sec_base = pi->sechdrs[sym->st_shndx].sh_addr;
 495
 496                 value = sym->st_value;
 497                 value += sec_base;
 498                 value += rel[i].r_addend;
 499
 500                 switch (ELF64_R_TYPE(rel[i].r_info)) {
 501                 case R_X86_64_NONE:
 502                         break;
 503                 case R_X86_64_64:
 504                         *(u64 *)location = value;
 505                         break;
 506                 case R_X86_64_32:
 507                         *(u32 *)location = value;
 508                         if (value != *(u32 *)location)
 509                                 goto overflow;
 510                         break;
 511                 case R_X86_64_32S:
 512                         *(s32 *)location = value;
 513                         if ((s64)value != *(s32 *)location)
 514                                 goto overflow;
 515                         break;
 516                 case R_X86_64_PC32:
 517                 case R_X86_64_PLT32:
 518                         value -= (u64)address;
 519                         *(u32 *)location = value;
 520                         break;
 521                 default:
 522                         pr_err("Unknown rela relocation: %llu\n",
 523                                ELF64_R_TYPE(rel[i].r_info));
 524                         return -ENOEXEC;
 525                 }
 526         }
 527         return 0;
 528
 529 overflow:
 530         pr_err("Overflow in relocation type %d value 0x%lx\n",
 531                (int)ELF64_R_TYPE(rel[i].r_info), value);
 532         return -ENOEXEC;
 533 }
 534
 535 int arch_kimage_file_post_load_cleanup(struct kimage *image)
 536 {
 537         vfree(image->elf_headers);
 538         image->elf_headers = NULL;
 539         image->elf_headers_sz = 0;
 540
 541         return kexec_image_post_load_cleanup_default(image);
 542 }
 543 #endif /* CONFIG_KEXEC_FILE */
 544
 545 #ifdef CONFIG_CRASH_DUMP
 546
 547 static int
 548 kexec_mark_range(unsigned long start, unsigned long end, bool protect)
 549 {
 550         struct page *page;
 551         unsigned int nr_pages;
 552
 553         /*
 554          * For physical range: [start, end]. We must skip the unassigned
 555          * crashk resource with zero-valued "end" member.
 556          */
 557         if (!end || start > end)
 558                 return 0;
 559
 560         page = pfn_to_page(start >> PAGE_SHIFT);
 561         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 562         if (protect)
 563                 return set_pages_ro(page, nr_pages);
 564         else
 565                 return set_pages_rw(page, nr_pages);
 566 }
 567
 568 static void kexec_mark_crashkres(bool protect)
 569 {
 570         unsigned long control;
 571
 572         kexec_mark_range(crashk_low_res.start, crashk_low_res.end, protect);
 573
 574         /* Don't touch the control code page used in crash_kexec().*/
 575         control = PFN_PHYS(page_to_pfn(kexec_crash_image->control_code_page));
 576         /* Control code page is located in the 2nd page. */
 577         kexec_mark_range(crashk_res.start, control + PAGE_SIZE - 1, protect);
 578         control += KEXEC_CONTROL_PAGE_SIZE;
 579         kexec_mark_range(control, crashk_res.end, protect);
 580 }
 581
 582 void arch_kexec_protect_crashkres(void)
 583 {
 584         kexec_mark_crashkres(true);
 585 }
 586
 587 void arch_kexec_unprotect_crashkres(void)
 588 {
 589         kexec_mark_crashkres(false);
 590 }
 591 #endif
 592
 593 /*
 594  * During a traditional boot under SME, SME will encrypt the kernel,
 595  * so the SME kexec kernel also needs to be un-encrypted in order to
 596  * replicate a normal SME boot.
 597  *
 598  * During a traditional boot under SEV, the kernel has already been
 599  * loaded encrypted, so the SEV kexec kernel needs to be encrypted in
 600  * order to replicate a normal SEV boot.
 601  */
 602 int arch_kexec_post_alloc_pages(void *vaddr, unsigned int pages, gfp_t gfp)
 603 {
 604         if (!cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT))
 605                 return 0;
 606
 607         /*
 608          * If host memory encryption is active we need to be sure that kexec
 609          * pages are not encrypted because when we boot to the new kernel the
 610          * pages won't be accessed encrypted (initially).
 611          */
 612         return set_memory_decrypted((unsigned long)vaddr, pages);
 613 }
 614
 615 void arch_kexec_pre_free_pages(void *vaddr, unsigned int pages)
 616 {
 617         if (!cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT))
 618                 return;
 619
 620         /*
 621          * If host memory encryption is active we need to reset the pages back
 622          * to being an encrypted mapping before freeing them.
 623          */
 624         set_memory_encrypted((unsigned long)vaddr, pages);
 625 }