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
  21 #include <asm/init.h>
  22 #include <asm/tlbflush.h>
  23 #include <asm/mmu_context.h>
  24 #include <asm/io_apic.h>
  25 #include <asm/debugreg.h>
  26 #include <asm/kexec-bzimage64.h>
  27 #include <asm/setup.h>
  28 #include <asm/set_memory.h>
  29
  30 #ifdef CONFIG_ACPI
  31 /*
  32  * Used while adding mapping for ACPI tables.
  33  * Can be reused when other iomem regions need be mapped
  34  */
  35 struct init_pgtable_data {
  36         struct x86_mapping_info *info;
  37         pgd_t *level4p;
  38 };
  39
  40 static int mem_region_callback(struct resource *res, void *arg)
  41 {
  42         struct init_pgtable_data *data = arg;
  43         unsigned long mstart, mend;
  44
  45         mstart = res->start;
  46         mend = mstart + resource_size(res) - 1;
  47
  48         return kernel_ident_mapping_init(data->info, data->level4p, mstart, mend);
  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
  92         if (!efi_enabled(EFI_BOOT))
  93                 return 0;
  94
  95         mstart = (boot_params.efi_info.efi_systab |
  96                         ((u64)boot_params.efi_info.efi_systab_hi<<32));
  97
  98         if (efi_enabled(EFI_64BIT))
  99                 mend = mstart + sizeof(efi_system_table_64_t);
 100         else
 101                 mend = mstart + sizeof(efi_system_table_32_t);
 102
 103         if (!mstart)
 104                 return 0;
 105
 106         return kernel_ident_mapping_init(info, level4p, mstart, mend);
 107 #endif
 108         return 0;
 109 }
 110
 111 static void free_transition_pgtable(struct kimage *image)
 112 {
 113         free_page((unsigned long)image->arch.p4d);
 114         image->arch.p4d = NULL;
 115         free_page((unsigned long)image->arch.pud);
 116         image->arch.pud = NULL;
 117         free_page((unsigned long)image->arch.pmd);
 118         image->arch.pmd = NULL;
 119         free_page((unsigned long)image->arch.pte);
 120         image->arch.pte = NULL;
 121 }
 122
 123 static int init_transition_pgtable(struct kimage *image, pgd_t *pgd)
 124 {
 125         pgprot_t prot = PAGE_KERNEL_EXEC_NOENC;
 126         unsigned long vaddr, paddr;
 127         int result = -ENOMEM;
 128         p4d_t *p4d;
 129         pud_t *pud;
 130         pmd_t *pmd;
 131         pte_t *pte;
 132
 133         vaddr = (unsigned long)relocate_kernel;
 134         paddr = __pa(page_address(image->control_code_page)+PAGE_SIZE);
 135         pgd += pgd_index(vaddr);
 136         if (!pgd_present(*pgd)) {
 137                 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
 138                 if (!p4d)
 139                         goto err;
 140                 image->arch.p4d = p4d;
 141                 set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE));
 142         }
 143         p4d = p4d_offset(pgd, vaddr);
 144         if (!p4d_present(*p4d)) {
 145                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
 146                 if (!pud)
 147                         goto err;
 148                 image->arch.pud = pud;
 149                 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
 150         }
 151         pud = pud_offset(p4d, vaddr);
 152         if (!pud_present(*pud)) {
 153                 pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
 154                 if (!pmd)
 155                         goto err;
 156                 image->arch.pmd = pmd;
 157                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 158         }
 159         pmd = pmd_offset(pud, vaddr);
 160         if (!pmd_present(*pmd)) {
 161                 pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
 162                 if (!pte)
 163                         goto err;
 164                 image->arch.pte = pte;
 165                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 166         }
 167         pte = pte_offset_kernel(pmd, vaddr);
 168
 169         if (sev_active())
 170                 prot = PAGE_KERNEL_EXEC;
 171
 172         set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
 173         return 0;
 174 err:
 175         return result;
 176 }
 177
 178 static void *alloc_pgt_page(void *data)
 179 {
 180         struct kimage *image = (struct kimage *)data;
 181         struct page *page;
 182         void *p = NULL;
 183
 184         page = kimage_alloc_control_pages(image, 0);
 185         if (page) {
 186                 p = page_address(page);
 187                 clear_page(p);
 188         }
 189
 190         return p;
 191 }
 192
 193 static int init_pgtable(struct kimage *image, unsigned long start_pgtable)
 194 {
 195         struct x86_mapping_info info = {
 196                 .alloc_pgt_page = alloc_pgt_page,
 197                 .context        = image,
 198                 .page_flag      = __PAGE_KERNEL_LARGE_EXEC,
 199                 .kernpg_flag    = _KERNPG_TABLE_NOENC,
 200         };
 201         unsigned long mstart, mend;
 202         pgd_t *level4p;
 203         int result;
 204         int i;
 205
 206         level4p = (pgd_t *)__va(start_pgtable);
 207         clear_page(level4p);
 208
 209         if (sev_active()) {
 210                 info.page_flag   |= _PAGE_ENC;
 211                 info.kernpg_flag |= _PAGE_ENC;
 212         }
 213
 214         if (direct_gbpages)
 215                 info.direct_gbpages = true;
 216
 217         for (i = 0; i < nr_pfn_mapped; i++) {
 218                 mstart = pfn_mapped[i].start << PAGE_SHIFT;
 219                 mend   = pfn_mapped[i].end << PAGE_SHIFT;
 220
 221                 result = kernel_ident_mapping_init(&info,
 222                                                  level4p, mstart, mend);
 223                 if (result)
 224                         return result;
 225         }
 226
 227         /*
 228          * segments's mem ranges could be outside 0 ~ max_pfn,
 229          * for example when jump back to original kernel from kexeced kernel.
 230          * or first kernel is booted with user mem map, and second kernel
 231          * could be loaded out of that range.
 232          */
 233         for (i = 0; i < image->nr_segments; i++) {
 234                 mstart = image->segment[i].mem;
 235                 mend   = mstart + image->segment[i].memsz;
 236
 237                 result = kernel_ident_mapping_init(&info,
 238                                                  level4p, mstart, mend);
 239
 240                 if (result)
 241                         return result;
 242         }
 243
 244         /*
 245          * Prepare EFI systab and ACPI tables for kexec kernel since they are
 246          * not covered by pfn_mapped.
 247          */
 248         result = map_efi_systab(&info, level4p);
 249         if (result)
 250                 return result;
 251
 252         result = map_acpi_tables(&info, level4p);
 253         if (result)
 254                 return result;
 255
 256         return init_transition_pgtable(image, level4p);
 257 }
 258
 259 static void set_idt(void *newidt, u16 limit)
 260 {
 261         struct desc_ptr curidt;
 262
 263         /* x86-64 supports unaligned loads & stores */
 264         curidt.size    = limit;
 265         curidt.address = (unsigned long)newidt;
 266
 267         __asm__ __volatile__ (
 268                 "lidtq %0\n"
 269                 : : "m" (curidt)
 270                 );
 271 };
 272
 273
 274 static void set_gdt(void *newgdt, u16 limit)
 275 {
 276         struct desc_ptr curgdt;
 277
 278         /* x86-64 supports unaligned loads & stores */
 279         curgdt.size    = limit;
 280         curgdt.address = (unsigned long)newgdt;
 281
 282         __asm__ __volatile__ (
 283                 "lgdtq %0\n"
 284                 : : "m" (curgdt)
 285                 );
 286 };
 287
 288 static void load_segments(void)
 289 {
 290         __asm__ __volatile__ (
 291                 "\tmovl %0,%%ds\n"
 292                 "\tmovl %0,%%es\n"
 293                 "\tmovl %0,%%ss\n"
 294                 "\tmovl %0,%%fs\n"
 295                 "\tmovl %0,%%gs\n"
 296                 : : "a" (__KERNEL_DS) : "memory"
 297                 );
 298 }
 299
 300 int machine_kexec_prepare(struct kimage *image)
 301 {
 302         unsigned long start_pgtable;
 303         int result;
 304
 305         /* Calculate the offsets */
 306         start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 307
 308         /* Setup the identity mapped 64bit page table */
 309         result = init_pgtable(image, start_pgtable);
 310         if (result)
 311                 return result;
 312
 313         return 0;
 314 }
 315
 316 void machine_kexec_cleanup(struct kimage *image)
 317 {
 318         free_transition_pgtable(image);
 319 }
 320
 321 /*
 322  * Do not allocate memory (or fail in any way) in machine_kexec().
 323  * We are past the point of no return, committed to rebooting now.
 324  */
 325 void machine_kexec(struct kimage *image)
 326 {
 327         unsigned long page_list[PAGES_NR];
 328         void *control_page;
 329         int save_ftrace_enabled;
 330
 331 #ifdef CONFIG_KEXEC_JUMP
 332         if (image->preserve_context)
 333                 save_processor_state();
 334 #endif
 335
 336         save_ftrace_enabled = __ftrace_enabled_save();
 337
 338         /* Interrupts aren't acceptable while we reboot */
 339         local_irq_disable();
 340         hw_breakpoint_disable();
 341
 342         if (image->preserve_context) {
 343 #ifdef CONFIG_X86_IO_APIC
 344                 /*
 345                  * We need to put APICs in legacy mode so that we can
 346                  * get timer interrupts in second kernel. kexec/kdump
 347                  * paths already have calls to restore_boot_irq_mode()
 348                  * in one form or other. kexec jump path also need one.
 349                  */
 350                 clear_IO_APIC();
 351                 restore_boot_irq_mode();
 352 #endif
 353         }
 354
 355         control_page = page_address(image->control_code_page) + PAGE_SIZE;
 356         memcpy(control_page, relocate_kernel, KEXEC_CONTROL_CODE_MAX_SIZE);
 357
 358         page_list[PA_CONTROL_PAGE] = virt_to_phys(control_page);
 359         page_list[VA_CONTROL_PAGE] = (unsigned long)control_page;
 360         page_list[PA_TABLE_PAGE] =
 361           (unsigned long)__pa(page_address(image->control_code_page));
 362
 363         if (image->type == KEXEC_TYPE_DEFAULT)
 364                 page_list[PA_SWAP_PAGE] = (page_to_pfn(image->swap_page)
 365                                                 << PAGE_SHIFT);
 366
 367         /*
 368          * The segment registers are funny things, they have both a
 369          * visible and an invisible part.  Whenever the visible part is
 370          * set to a specific selector, the invisible part is loaded
 371          * with from a table in memory.  At no other time is the
 372          * descriptor table in memory accessed.
 373          *
 374          * I take advantage of this here by force loading the
 375          * segments, before I zap the gdt with an invalid value.
 376          */
 377         load_segments();
 378         /*
 379          * The gdt & idt are now invalid.
 380          * If you want to load them you must set up your own idt & gdt.
 381          */
 382         set_gdt(phys_to_virt(0), 0);
 383         set_idt(phys_to_virt(0), 0);
 384
 385         /* now call it */
 386         image->start = relocate_kernel((unsigned long)image->head,
 387                                        (unsigned long)page_list,
 388                                        image->start,
 389                                        image->preserve_context,
 390                                        sme_active());
 391
 392 #ifdef CONFIG_KEXEC_JUMP
 393         if (image->preserve_context)
 394                 restore_processor_state();
 395 #endif
 396
 397         __ftrace_enabled_restore(save_ftrace_enabled);
 398 }
 399
 400 /* arch-dependent functionality related to kexec file-based syscall */
 401
 402 #ifdef CONFIG_KEXEC_FILE
 403 void *arch_kexec_kernel_image_load(struct kimage *image)
 404 {
 405         vfree(image->elf_headers);
 406         image->elf_headers = NULL;
 407
 408         if (!image->fops || !image->fops->load)
 409                 return ERR_PTR(-ENOEXEC);
 410
 411         return image->fops->load(image, image->kernel_buf,
 412                                  image->kernel_buf_len, image->initrd_buf,
 413                                  image->initrd_buf_len, image->cmdline_buf,
 414                                  image->cmdline_buf_len);
 415 }
 416
 417 /*
 418  * Apply purgatory relocations.
 419  *
 420  * @pi:         Purgatory to be relocated.
 421  * @section:    Section relocations applying to.
 422  * @relsec:     Section containing RELAs.
 423  * @symtabsec:  Corresponding symtab.
 424  *
 425  * TODO: Some of the code belongs to generic code. Move that in kexec.c.
 426  */
 427 int arch_kexec_apply_relocations_add(struct purgatory_info *pi,
 428                                      Elf_Shdr *section, const Elf_Shdr *relsec,
 429                                      const Elf_Shdr *symtabsec)
 430 {
 431         unsigned int i;
 432         Elf64_Rela *rel;
 433         Elf64_Sym *sym;
 434         void *location;
 435         unsigned long address, sec_base, value;
 436         const char *strtab, *name, *shstrtab;
 437         const Elf_Shdr *sechdrs;
 438
 439         /* String & section header string table */
 440         sechdrs = (void *)pi->ehdr + pi->ehdr->e_shoff;
 441         strtab = (char *)pi->ehdr + sechdrs[symtabsec->sh_link].sh_offset;
 442         shstrtab = (char *)pi->ehdr + sechdrs[pi->ehdr->e_shstrndx].sh_offset;
 443
 444         rel = (void *)pi->ehdr + relsec->sh_offset;
 445
 446         pr_debug("Applying relocate section %s to %u\n",
 447                  shstrtab + relsec->sh_name, relsec->sh_info);
 448
 449         for (i = 0; i < relsec->sh_size / sizeof(*rel); i++) {
 450
 451                 /*
 452                  * rel[i].r_offset contains byte offset from beginning
 453                  * of section to the storage unit affected.
 454                  *
 455                  * This is location to update. This is temporary buffer
 456                  * where section is currently loaded. This will finally be
 457                  * loaded to a different address later, pointed to by
 458                  * ->sh_addr. kexec takes care of moving it
 459                  *  (kexec_load_segment()).
 460                  */
 461                 location = pi->purgatory_buf;
 462                 location += section->sh_offset;
 463                 location += rel[i].r_offset;
 464
 465                 /* Final address of the location */
 466                 address = section->sh_addr + rel[i].r_offset;
 467
 468                 /*
 469                  * rel[i].r_info contains information about symbol table index
 470                  * w.r.t which relocation must be made and type of relocation
 471                  * to apply. ELF64_R_SYM() and ELF64_R_TYPE() macros get
 472                  * these respectively.
 473                  */
 474                 sym = (void *)pi->ehdr + symtabsec->sh_offset;
 475                 sym += ELF64_R_SYM(rel[i].r_info);
 476
 477                 if (sym->st_name)
 478                         name = strtab + sym->st_name;
 479                 else
 480                         name = shstrtab + sechdrs[sym->st_shndx].sh_name;
 481
 482                 pr_debug("Symbol: %s info: %02x shndx: %02x value=%llx size: %llx\n",
 483                          name, sym->st_info, sym->st_shndx, sym->st_value,
 484                          sym->st_size);
 485
 486                 if (sym->st_shndx == SHN_UNDEF) {
 487                         pr_err("Undefined symbol: %s\n", name);
 488                         return -ENOEXEC;
 489                 }
 490
 491                 if (sym->st_shndx == SHN_COMMON) {
 492                         pr_err("symbol '%s' in common section\n", name);
 493                         return -ENOEXEC;
 494                 }
 495
 496                 if (sym->st_shndx == SHN_ABS)
 497                         sec_base = 0;
 498                 else if (sym->st_shndx >= pi->ehdr->e_shnum) {
 499                         pr_err("Invalid section %d for symbol %s\n",
 500                                sym->st_shndx, name);
 501                         return -ENOEXEC;
 502                 } else
 503                         sec_base = pi->sechdrs[sym->st_shndx].sh_addr;
 504
 505                 value = sym->st_value;
 506                 value += sec_base;
 507                 value += rel[i].r_addend;
 508
 509                 switch (ELF64_R_TYPE(rel[i].r_info)) {
 510                 case R_X86_64_NONE:
 511                         break;
 512                 case R_X86_64_64:
 513                         *(u64 *)location = value;
 514                         break;
 515                 case R_X86_64_32:
 516                         *(u32 *)location = value;
 517                         if (value != *(u32 *)location)
 518                                 goto overflow;
 519                         break;
 520                 case R_X86_64_32S:
 521                         *(s32 *)location = value;
 522                         if ((s64)value != *(s32 *)location)
 523                                 goto overflow;
 524                         break;
 525                 case R_X86_64_PC32:
 526                 case R_X86_64_PLT32:
 527                         value -= (u64)address;
 528                         *(u32 *)location = value;
 529                         break;
 530                 default:
 531                         pr_err("Unknown rela relocation: %llu\n",
 532                                ELF64_R_TYPE(rel[i].r_info));
 533                         return -ENOEXEC;
 534                 }
 535         }
 536         return 0;
 537
 538 overflow:
 539         pr_err("Overflow in relocation type %d value 0x%lx\n",
 540                (int)ELF64_R_TYPE(rel[i].r_info), value);
 541         return -ENOEXEC;
 542 }
 543 #endif /* CONFIG_KEXEC_FILE */
 544
 545 static int
 546 kexec_mark_range(unsigned long start, unsigned long end, bool protect)
 547 {
 548         struct page *page;
 549         unsigned int nr_pages;
 550
 551         /*
 552          * For physical range: [start, end]. We must skip the unassigned
 553          * crashk resource with zero-valued "end" member.
 554          */
 555         if (!end || start > end)
 556                 return 0;
 557
 558         page = pfn_to_page(start >> PAGE_SHIFT);
 559         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 560         if (protect)
 561                 return set_pages_ro(page, nr_pages);
 562         else
 563                 return set_pages_rw(page, nr_pages);
 564 }
 565
 566 static void kexec_mark_crashkres(bool protect)
 567 {
 568         unsigned long control;
 569
 570         kexec_mark_range(crashk_low_res.start, crashk_low_res.end, protect);
 571
 572         /* Don't touch the control code page used in crash_kexec().*/
 573         control = PFN_PHYS(page_to_pfn(kexec_crash_image->control_code_page));
 574         /* Control code page is located in the 2nd page. */
 575         kexec_mark_range(crashk_res.start, control + PAGE_SIZE - 1, protect);
 576         control += KEXEC_CONTROL_PAGE_SIZE;
 577         kexec_mark_range(control, crashk_res.end, protect);
 578 }
 579
 580 void arch_kexec_protect_crashkres(void)
 581 {
 582         kexec_mark_crashkres(true);
 583 }
 584
 585 void arch_kexec_unprotect_crashkres(void)
 586 {
 587         kexec_mark_crashkres(false);
 588 }
 589
 590 /*
 591  * During a traditional boot under SME, SME will encrypt the kernel,
 592  * so the SME kexec kernel also needs to be un-encrypted in order to
 593  * replicate a normal SME boot.
 594  *
 595  * During a traditional boot under SEV, the kernel has already been
 596  * loaded encrypted, so the SEV kexec kernel needs to be encrypted in
 597  * order to replicate a normal SEV boot.
 598  */
 599 int arch_kexec_post_alloc_pages(void *vaddr, unsigned int pages, gfp_t gfp)
 600 {
 601         if (sev_active())
 602                 return 0;
 603
 604         /*
 605          * If SME is active we need to be sure that kexec pages are
 606          * not encrypted because when we boot to the new kernel the
 607          * pages won't be accessed encrypted (initially).
 608          */
 609         return set_memory_decrypted((unsigned long)vaddr, pages);
 610 }
 611
 612 void arch_kexec_pre_free_pages(void *vaddr, unsigned int pages)
 613 {
 614         if (sev_active())
 615                 return;
 616
 617         /*
 618          * If SME is active we need to reset the pages back to being
 619          * an encrypted mapping before freeing them.
 620          */
 621         set_memory_encrypted((unsigned long)vaddr, pages);
 622 }