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/pgtable.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
  31 #ifdef CONFIG_ACPI
  32 /*
  33  * Used while adding mapping for ACPI tables.
  34  * Can be reused when other iomem regions need be mapped
  35  */
  36 struct init_pgtable_data {
  37         struct x86_mapping_info *info;
  38         pgd_t *level4p;
  39 };
  40
  41 static int mem_region_callback(struct resource *res, void *arg)
  42 {
  43         struct init_pgtable_data *data = arg;
  44         unsigned long mstart, mend;
  45
  46         mstart = res->start;
  47         mend = mstart + resource_size(res) - 1;
  48
  49         return kernel_ident_mapping_init(data->info, data->level4p, mstart, mend);
  50 }
  51
  52 static int
  53 map_acpi_tables(struct x86_mapping_info *info, pgd_t *level4p)
  54 {
  55         struct init_pgtable_data data;
  56         unsigned long flags;
  57         int ret;
  58
  59         data.info = info;
  60         data.level4p = level4p;
  61         flags = IORESOURCE_MEM | IORESOURCE_BUSY;
  62
  63         ret = walk_iomem_res_desc(IORES_DESC_ACPI_TABLES, flags, 0, -1,
  64                                   &data, mem_region_callback);
  65         if (ret && ret != -EINVAL)
  66                 return ret;
  67
  68         /* ACPI tables could be located in ACPI Non-volatile Storage region */
  69         ret = walk_iomem_res_desc(IORES_DESC_ACPI_NV_STORAGE, flags, 0, -1,
  70                                   &data, mem_region_callback);
  71         if (ret && ret != -EINVAL)
  72                 return ret;
  73
  74         return 0;
  75 }
  76 #else
  77 static int map_acpi_tables(struct x86_mapping_info *info, pgd_t *level4p) { return 0; }
  78 #endif
  79
  80 #ifdef CONFIG_KEXEC_FILE
  81 const struct kexec_file_ops * const kexec_file_loaders[] = {
  82                 &kexec_bzImage64_ops,
  83                 NULL
  84 };
  85 #endif
  86
  87 static int
  88 map_efi_systab(struct x86_mapping_info *info, pgd_t *level4p)
  89 {
  90 #ifdef CONFIG_EFI
  91         unsigned long mstart, mend;
  92
  93         if (!efi_enabled(EFI_BOOT))
  94                 return 0;
  95
  96         mstart = (boot_params.efi_info.efi_systab |
  97                         ((u64)boot_params.efi_info.efi_systab_hi<<32));
  98
  99         if (efi_enabled(EFI_64BIT))
 100                 mend = mstart + sizeof(efi_system_table_64_t);
 101         else
 102                 mend = mstart + sizeof(efi_system_table_32_t);
 103
 104         if (!mstart)
 105                 return 0;
 106
 107         return kernel_ident_mapping_init(info, level4p, mstart, mend);
 108 #endif
 109         return 0;
 110 }
 111
 112 static void free_transition_pgtable(struct kimage *image)
 113 {
 114         free_page((unsigned long)image->arch.p4d);
 115         image->arch.p4d = NULL;
 116         free_page((unsigned long)image->arch.pud);
 117         image->arch.pud = NULL;
 118         free_page((unsigned long)image->arch.pmd);
 119         image->arch.pmd = NULL;
 120         free_page((unsigned long)image->arch.pte);
 121         image->arch.pte = NULL;
 122 }
 123
 124 static int init_transition_pgtable(struct kimage *image, pgd_t *pgd)
 125 {
 126         pgprot_t prot = PAGE_KERNEL_EXEC_NOENC;
 127         unsigned long vaddr, paddr;
 128         int result = -ENOMEM;
 129         p4d_t *p4d;
 130         pud_t *pud;
 131         pmd_t *pmd;
 132         pte_t *pte;
 133
 134         vaddr = (unsigned long)relocate_kernel;
 135         paddr = __pa(page_address(image->control_code_page)+PAGE_SIZE);
 136         pgd += pgd_index(vaddr);
 137         if (!pgd_present(*pgd)) {
 138                 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
 139                 if (!p4d)
 140                         goto err;
 141                 image->arch.p4d = p4d;
 142                 set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE));
 143         }
 144         p4d = p4d_offset(pgd, vaddr);
 145         if (!p4d_present(*p4d)) {
 146                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
 147                 if (!pud)
 148                         goto err;
 149                 image->arch.pud = pud;
 150                 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
 151         }
 152         pud = pud_offset(p4d, vaddr);
 153         if (!pud_present(*pud)) {
 154                 pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
 155                 if (!pmd)
 156                         goto err;
 157                 image->arch.pmd = pmd;
 158                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 159         }
 160         pmd = pmd_offset(pud, vaddr);
 161         if (!pmd_present(*pmd)) {
 162                 pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
 163                 if (!pte)
 164                         goto err;
 165                 image->arch.pte = pte;
 166                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 167         }
 168         pte = pte_offset_kernel(pmd, vaddr);
 169
 170         if (sev_active())
 171                 prot = PAGE_KERNEL_EXEC;
 172
 173         set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
 174         return 0;
 175 err:
 176         return result;
 177 }
 178
 179 static void *alloc_pgt_page(void *data)
 180 {
 181         struct kimage *image = (struct kimage *)data;
 182         struct page *page;
 183         void *p = NULL;
 184
 185         page = kimage_alloc_control_pages(image, 0);
 186         if (page) {
 187                 p = page_address(page);
 188                 clear_page(p);
 189         }
 190
 191         return p;
 192 }
 193
 194 static int init_pgtable(struct kimage *image, unsigned long start_pgtable)
 195 {
 196         struct x86_mapping_info info = {
 197                 .alloc_pgt_page = alloc_pgt_page,
 198                 .context        = image,
 199                 .page_flag      = __PAGE_KERNEL_LARGE_EXEC,
 200                 .kernpg_flag    = _KERNPG_TABLE_NOENC,
 201         };
 202         unsigned long mstart, mend;
 203         pgd_t *level4p;
 204         int result;
 205         int i;
 206
 207         level4p = (pgd_t *)__va(start_pgtable);
 208         clear_page(level4p);
 209
 210         if (sev_active()) {
 211                 info.page_flag   |= _PAGE_ENC;
 212                 info.kernpg_flag |= _PAGE_ENC;
 213         }
 214
 215         if (direct_gbpages)
 216                 info.direct_gbpages = true;
 217
 218         for (i = 0; i < nr_pfn_mapped; i++) {
 219                 mstart = pfn_mapped[i].start << PAGE_SHIFT;
 220                 mend   = pfn_mapped[i].end << PAGE_SHIFT;
 221
 222                 result = kernel_ident_mapping_init(&info,
 223                                                  level4p, mstart, mend);
 224                 if (result)
 225                         return result;
 226         }
 227
 228         /*
 229          * segments's mem ranges could be outside 0 ~ max_pfn,
 230          * for example when jump back to original kernel from kexeced kernel.
 231          * or first kernel is booted with user mem map, and second kernel
 232          * could be loaded out of that range.
 233          */
 234         for (i = 0; i < image->nr_segments; i++) {
 235                 mstart = image->segment[i].mem;
 236                 mend   = mstart + image->segment[i].memsz;
 237
 238                 result = kernel_ident_mapping_init(&info,
 239                                                  level4p, mstart, mend);
 240
 241                 if (result)
 242                         return result;
 243         }
 244
 245         /*
 246          * Prepare EFI systab and ACPI tables for kexec kernel since they are
 247          * not covered by pfn_mapped.
 248          */
 249         result = map_efi_systab(&info, level4p);
 250         if (result)
 251                 return result;
 252
 253         result = map_acpi_tables(&info, level4p);
 254         if (result)
 255                 return result;
 256
 257         return init_transition_pgtable(image, level4p);
 258 }
 259
 260 static void set_idt(void *newidt, u16 limit)
 261 {
 262         struct desc_ptr curidt;
 263
 264         /* x86-64 supports unaliged loads & stores */
 265         curidt.size    = limit;
 266         curidt.address = (unsigned long)newidt;
 267
 268         __asm__ __volatile__ (
 269                 "lidtq %0\n"
 270                 : : "m" (curidt)
 271                 );
 272 };
 273
 274
 275 static void set_gdt(void *newgdt, u16 limit)
 276 {
 277         struct desc_ptr curgdt;
 278
 279         /* x86-64 supports unaligned loads & stores */
 280         curgdt.size    = limit;
 281         curgdt.address = (unsigned long)newgdt;
 282
 283         __asm__ __volatile__ (
 284                 "lgdtq %0\n"
 285                 : : "m" (curgdt)
 286                 );
 287 };
 288
 289 static void load_segments(void)
 290 {
 291         __asm__ __volatile__ (
 292                 "\tmovl %0,%%ds\n"
 293                 "\tmovl %0,%%es\n"
 294                 "\tmovl %0,%%ss\n"
 295                 "\tmovl %0,%%fs\n"
 296                 "\tmovl %0,%%gs\n"
 297                 : : "a" (__KERNEL_DS) : "memory"
 298                 );
 299 }
 300
 301 int machine_kexec_prepare(struct kimage *image)
 302 {
 303         unsigned long start_pgtable;
 304         int result;
 305
 306         /* Calculate the offsets */
 307         start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 308
 309         /* Setup the identity mapped 64bit page table */
 310         result = init_pgtable(image, start_pgtable);
 311         if (result)
 312                 return result;
 313
 314         return 0;
 315 }
 316
 317 void machine_kexec_cleanup(struct kimage *image)
 318 {
 319         free_transition_pgtable(image);
 320 }
 321
 322 /*
 323  * Do not allocate memory (or fail in any way) in machine_kexec().
 324  * We are past the point of no return, committed to rebooting now.
 325  */
 326 void machine_kexec(struct kimage *image)
 327 {
 328         unsigned long page_list[PAGES_NR];
 329         void *control_page;
 330         int save_ftrace_enabled;
 331
 332 #ifdef CONFIG_KEXEC_JUMP
 333         if (image->preserve_context)
 334                 save_processor_state();
 335 #endif
 336
 337         save_ftrace_enabled = __ftrace_enabled_save();
 338
 339         /* Interrupts aren't acceptable while we reboot */
 340         local_irq_disable();
 341         hw_breakpoint_disable();
 342
 343         if (image->preserve_context) {
 344 #ifdef CONFIG_X86_IO_APIC
 345                 /*
 346                  * We need to put APICs in legacy mode so that we can
 347                  * get timer interrupts in second kernel. kexec/kdump
 348                  * paths already have calls to restore_boot_irq_mode()
 349                  * in one form or other. kexec jump path also need one.
 350                  */
 351                 clear_IO_APIC();
 352                 restore_boot_irq_mode();
 353 #endif
 354         }
 355
 356         control_page = page_address(image->control_code_page) + PAGE_SIZE;
 357         memcpy(control_page, relocate_kernel, KEXEC_CONTROL_CODE_MAX_SIZE);
 358
 359         page_list[PA_CONTROL_PAGE] = virt_to_phys(control_page);
 360         page_list[VA_CONTROL_PAGE] = (unsigned long)control_page;
 361         page_list[PA_TABLE_PAGE] =
 362           (unsigned long)__pa(page_address(image->control_code_page));
 363
 364         if (image->type == KEXEC_TYPE_DEFAULT)
 365                 page_list[PA_SWAP_PAGE] = (page_to_pfn(image->swap_page)
 366                                                 << PAGE_SHIFT);
 367
 368         /*
 369          * The segment registers are funny things, they have both a
 370          * visible and an invisible part.  Whenever the visible part is
 371          * set to a specific selector, the invisible part is loaded
 372          * with from a table in memory.  At no other time is the
 373          * descriptor table in memory accessed.
 374          *
 375          * I take advantage of this here by force loading the
 376          * segments, before I zap the gdt with an invalid value.
 377          */
 378         load_segments();
 379         /*
 380          * The gdt & idt are now invalid.
 381          * If you want to load them you must set up your own idt & gdt.
 382          */
 383         set_gdt(phys_to_virt(0), 0);
 384         set_idt(phys_to_virt(0), 0);
 385
 386         /* now call it */
 387         image->start = relocate_kernel((unsigned long)image->head,
 388                                        (unsigned long)page_list,
 389                                        image->start,
 390                                        image->preserve_context,
 391                                        sme_active());
 392
 393 #ifdef CONFIG_KEXEC_JUMP
 394         if (image->preserve_context)
 395                 restore_processor_state();
 396 #endif
 397
 398         __ftrace_enabled_restore(save_ftrace_enabled);
 399 }
 400
 401 void arch_crash_save_vmcoreinfo(void)
 402 {
 403         u64 sme_mask = sme_me_mask;
 404
 405         VMCOREINFO_NUMBER(phys_base);
 406         VMCOREINFO_SYMBOL(init_top_pgt);
 407         vmcoreinfo_append_str("NUMBER(pgtable_l5_enabled)=%d\n",
 408                         pgtable_l5_enabled());
 409
 410 #ifdef CONFIG_NUMA
 411         VMCOREINFO_SYMBOL(node_data);
 412         VMCOREINFO_LENGTH(node_data, MAX_NUMNODES);
 413 #endif
 414         vmcoreinfo_append_str("KERNELOFFSET=%lx\n",
 415                               kaslr_offset());
 416         VMCOREINFO_NUMBER(KERNEL_IMAGE_SIZE);
 417         VMCOREINFO_NUMBER(sme_mask);
 418 }
 419
 420 /* arch-dependent functionality related to kexec file-based syscall */
 421
 422 #ifdef CONFIG_KEXEC_FILE
 423 void *arch_kexec_kernel_image_load(struct kimage *image)
 424 {
 425         vfree(image->arch.elf_headers);
 426         image->arch.elf_headers = NULL;
 427
 428         if (!image->fops || !image->fops->load)
 429                 return ERR_PTR(-ENOEXEC);
 430
 431         return image->fops->load(image, image->kernel_buf,
 432                                  image->kernel_buf_len, image->initrd_buf,
 433                                  image->initrd_buf_len, image->cmdline_buf,
 434                                  image->cmdline_buf_len);
 435 }
 436
 437 /*
 438  * Apply purgatory relocations.
 439  *
 440  * @pi:         Purgatory to be relocated.
 441  * @section:    Section relocations applying to.
 442  * @relsec:     Section containing RELAs.
 443  * @symtabsec:  Corresponding symtab.
 444  *
 445  * TODO: Some of the code belongs to generic code. Move that in kexec.c.
 446  */
 447 int arch_kexec_apply_relocations_add(struct purgatory_info *pi,
 448                                      Elf_Shdr *section, const Elf_Shdr *relsec,
 449                                      const Elf_Shdr *symtabsec)
 450 {
 451         unsigned int i;
 452         Elf64_Rela *rel;
 453         Elf64_Sym *sym;
 454         void *location;
 455         unsigned long address, sec_base, value;
 456         const char *strtab, *name, *shstrtab;
 457         const Elf_Shdr *sechdrs;
 458
 459         /* String & section header string table */
 460         sechdrs = (void *)pi->ehdr + pi->ehdr->e_shoff;
 461         strtab = (char *)pi->ehdr + sechdrs[symtabsec->sh_link].sh_offset;
 462         shstrtab = (char *)pi->ehdr + sechdrs[pi->ehdr->e_shstrndx].sh_offset;
 463
 464         rel = (void *)pi->ehdr + relsec->sh_offset;
 465
 466         pr_debug("Applying relocate section %s to %u\n",
 467                  shstrtab + relsec->sh_name, relsec->sh_info);
 468
 469         for (i = 0; i < relsec->sh_size / sizeof(*rel); i++) {
 470
 471                 /*
 472                  * rel[i].r_offset contains byte offset from beginning
 473                  * of section to the storage unit affected.
 474                  *
 475                  * This is location to update. This is temporary buffer
 476                  * where section is currently loaded. This will finally be
 477                  * loaded to a different address later, pointed to by
 478                  * ->sh_addr. kexec takes care of moving it
 479                  *  (kexec_load_segment()).
 480                  */
 481                 location = pi->purgatory_buf;
 482                 location += section->sh_offset;
 483                 location += rel[i].r_offset;
 484
 485                 /* Final address of the location */
 486                 address = section->sh_addr + rel[i].r_offset;
 487
 488                 /*
 489                  * rel[i].r_info contains information about symbol table index
 490                  * w.r.t which relocation must be made and type of relocation
 491                  * to apply. ELF64_R_SYM() and ELF64_R_TYPE() macros get
 492                  * these respectively.
 493                  */
 494                 sym = (void *)pi->ehdr + symtabsec->sh_offset;
 495                 sym += ELF64_R_SYM(rel[i].r_info);
 496
 497                 if (sym->st_name)
 498                         name = strtab + sym->st_name;
 499                 else
 500                         name = shstrtab + sechdrs[sym->st_shndx].sh_name;
 501
 502                 pr_debug("Symbol: %s info: %02x shndx: %02x value=%llx size: %llx\n",
 503                          name, sym->st_info, sym->st_shndx, sym->st_value,
 504                          sym->st_size);
 505
 506                 if (sym->st_shndx == SHN_UNDEF) {
 507                         pr_err("Undefined symbol: %s\n", name);
 508                         return -ENOEXEC;
 509                 }
 510
 511                 if (sym->st_shndx == SHN_COMMON) {
 512                         pr_err("symbol '%s' in common section\n", name);
 513                         return -ENOEXEC;
 514                 }
 515
 516                 if (sym->st_shndx == SHN_ABS)
 517                         sec_base = 0;
 518                 else if (sym->st_shndx >= pi->ehdr->e_shnum) {
 519                         pr_err("Invalid section %d for symbol %s\n",
 520                                sym->st_shndx, name);
 521                         return -ENOEXEC;
 522                 } else
 523                         sec_base = pi->sechdrs[sym->st_shndx].sh_addr;
 524
 525                 value = sym->st_value;
 526                 value += sec_base;
 527                 value += rel[i].r_addend;
 528
 529                 switch (ELF64_R_TYPE(rel[i].r_info)) {
 530                 case R_X86_64_NONE:
 531                         break;
 532                 case R_X86_64_64:
 533                         *(u64 *)location = value;
 534                         break;
 535                 case R_X86_64_32:
 536                         *(u32 *)location = value;
 537                         if (value != *(u32 *)location)
 538                                 goto overflow;
 539                         break;
 540                 case R_X86_64_32S:
 541                         *(s32 *)location = value;
 542                         if ((s64)value != *(s32 *)location)
 543                                 goto overflow;
 544                         break;
 545                 case R_X86_64_PC32:
 546                 case R_X86_64_PLT32:
 547                         value -= (u64)address;
 548                         *(u32 *)location = value;
 549                         break;
 550                 default:
 551                         pr_err("Unknown rela relocation: %llu\n",
 552                                ELF64_R_TYPE(rel[i].r_info));
 553                         return -ENOEXEC;
 554                 }
 555         }
 556         return 0;
 557
 558 overflow:
 559         pr_err("Overflow in relocation type %d value 0x%lx\n",
 560                (int)ELF64_R_TYPE(rel[i].r_info), value);
 561         return -ENOEXEC;
 562 }
 563 #endif /* CONFIG_KEXEC_FILE */
 564
 565 static int
 566 kexec_mark_range(unsigned long start, unsigned long end, bool protect)
 567 {
 568         struct page *page;
 569         unsigned int nr_pages;
 570
 571         /*
 572          * For physical range: [start, end]. We must skip the unassigned
 573          * crashk resource with zero-valued "end" member.
 574          */
 575         if (!end || start > end)
 576                 return 0;
 577
 578         page = pfn_to_page(start >> PAGE_SHIFT);
 579         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 580         if (protect)
 581                 return set_pages_ro(page, nr_pages);
 582         else
 583                 return set_pages_rw(page, nr_pages);
 584 }
 585
 586 static void kexec_mark_crashkres(bool protect)
 587 {
 588         unsigned long control;
 589
 590         kexec_mark_range(crashk_low_res.start, crashk_low_res.end, protect);
 591
 592         /* Don't touch the control code page used in crash_kexec().*/
 593         control = PFN_PHYS(page_to_pfn(kexec_crash_image->control_code_page));
 594         /* Control code page is located in the 2nd page. */
 595         kexec_mark_range(crashk_res.start, control + PAGE_SIZE - 1, protect);
 596         control += KEXEC_CONTROL_PAGE_SIZE;
 597         kexec_mark_range(control, crashk_res.end, protect);
 598 }
 599
 600 void arch_kexec_protect_crashkres(void)
 601 {
 602         kexec_mark_crashkres(true);
 603 }
 604
 605 void arch_kexec_unprotect_crashkres(void)
 606 {
 607         kexec_mark_crashkres(false);
 608 }
 609
 610 /*
 611  * During a traditional boot under SME, SME will encrypt the kernel,
 612  * so the SME kexec kernel also needs to be un-encrypted in order to
 613  * replicate a normal SME boot.
 614  *
 615  * During a traditional boot under SEV, the kernel has already been
 616  * loaded encrypted, so the SEV kexec kernel needs to be encrypted in
 617  * order to replicate a normal SEV boot.
 618  */
 619 int arch_kexec_post_alloc_pages(void *vaddr, unsigned int pages, gfp_t gfp)
 620 {
 621         if (sev_active())
 622                 return 0;
 623
 624         /*
 625          * If SME is active we need to be sure that kexec pages are
 626          * not encrypted because when we boot to the new kernel the
 627          * pages won't be accessed encrypted (initially).
 628          */
 629         return set_memory_decrypted((unsigned long)vaddr, pages);
 630 }
 631
 632 void arch_kexec_pre_free_pages(void *vaddr, unsigned int pages)
 633 {
 634         if (sev_active())
 635                 return;
 636
 637         /*
 638          * If SME is active we need to reset the pages back to being
 639          * an encrypted mapping before freeing them.
 640          */
 641         set_memory_encrypted((unsigned long)vaddr, pages);
 642 }