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