arch/x86/kernel/machine_kexec_64.c

   1 /*
   2  * handle transition of Linux booting another kernel
   3  * Copyright (C) 2002-2005 Eric Biederman  <ebiederm@xmission.com>
   4  *
   5  * This source code is licensed under the GNU General Public License,
   6  * Version 2.  See the file COPYING for more details.
   7  */
   8
   9 #define pr_fmt(fmt)     "kexec: " fmt
  10
  11 #include <linux/mm.h>
  12 #include <linux/kexec.h>
  13 #include <linux/string.h>
  14 #include <linux/gfp.h>
  15 #include <linux/reboot.h>
  16 #include <linux/numa.h>
  17 #include <linux/ftrace.h>
  18 #include <linux/io.h>
  19 #include <linux/suspend.h>
  20 #include <linux/vmalloc.h>
  21
  22 #include <asm/init.h>
  23 #include <asm/pgtable.h>
  24 #include <asm/tlbflush.h>
  25 #include <asm/mmu_context.h>
  26 #include <asm/io_apic.h>
  27 #include <asm/debugreg.h>
  28 #include <asm/kexec-bzimage64.h>
  29 #include <asm/setup.h>
  30
  31 #ifdef CONFIG_KEXEC_FILE
  32 static struct kexec_file_ops *kexec_file_loaders[] = {
  33                 &kexec_bzImage64_ops,
  34 };
  35 #endif
  36
  37 static void free_transition_pgtable(struct kimage *image)
  38 {
  39         free_page((unsigned long)image->arch.pud);
  40         image->arch.pud = NULL;
  41         free_page((unsigned long)image->arch.pmd);
  42         image->arch.pmd = NULL;
  43         free_page((unsigned long)image->arch.pte);
  44         image->arch.pte = NULL;
  45 }
  46
  47 static int init_transition_pgtable(struct kimage *image, pgd_t *pgd)
  48 {
  49         pud_t *pud;
  50         pmd_t *pmd;
  51         pte_t *pte;
  52         unsigned long vaddr, paddr;
  53         int result = -ENOMEM;
  54
  55         vaddr = (unsigned long)relocate_kernel;
  56         paddr = __pa(page_address(image->control_code_page)+PAGE_SIZE);
  57         pgd += pgd_index(vaddr);
  58         if (!pgd_present(*pgd)) {
  59                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
  60                 if (!pud)
  61                         goto err;
  62                 image->arch.pud = pud;
  63                 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE));
  64         }
  65         pud = pud_offset(pgd, vaddr);
  66         if (!pud_present(*pud)) {
  67                 pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
  68                 if (!pmd)
  69                         goto err;
  70                 image->arch.pmd = pmd;
  71                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
  72         }
  73         pmd = pmd_offset(pud, vaddr);
  74         if (!pmd_present(*pmd)) {
  75                 pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
  76                 if (!pte)
  77                         goto err;
  78                 image->arch.pte = pte;
  79                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
  80         }
  81         pte = pte_offset_kernel(pmd, vaddr);
  82         set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL_EXEC));
  83         return 0;
  84 err:
  85         return result;
  86 }
  87
  88 static void *alloc_pgt_page(void *data)
  89 {
  90         struct kimage *image = (struct kimage *)data;
  91         struct page *page;
  92         void *p = NULL;
  93
  94         page = kimage_alloc_control_pages(image, 0);
  95         if (page) {
  96                 p = page_address(page);
  97                 clear_page(p);
  98         }
  99
 100         return p;
 101 }
 102
 103 static int init_pgtable(struct kimage *image, unsigned long start_pgtable)
 104 {
 105         struct x86_mapping_info info = {
 106                 .alloc_pgt_page = alloc_pgt_page,
 107                 .context        = image,
 108                 .pmd_flag       = __PAGE_KERNEL_LARGE_EXEC,
 109         };
 110         unsigned long mstart, mend;
 111         pgd_t *level4p;
 112         int result;
 113         int i;
 114
 115         level4p = (pgd_t *)__va(start_pgtable);
 116         clear_page(level4p);
 117         for (i = 0; i < nr_pfn_mapped; i++) {
 118                 mstart = pfn_mapped[i].start << PAGE_SHIFT;
 119                 mend   = pfn_mapped[i].end << PAGE_SHIFT;
 120
 121                 result = kernel_ident_mapping_init(&info,
 122                                                  level4p, mstart, mend);
 123                 if (result)
 124                         return result;
 125         }
 126
 127         /*
 128          * segments's mem ranges could be outside 0 ~ max_pfn,
 129          * for example when jump back to original kernel from kexeced kernel.
 130          * or first kernel is booted with user mem map, and second kernel
 131          * could be loaded out of that range.
 132          */
 133         for (i = 0; i < image->nr_segments; i++) {
 134                 mstart = image->segment[i].mem;
 135                 mend   = mstart + image->segment[i].memsz;
 136
 137                 result = kernel_ident_mapping_init(&info,
 138                                                  level4p, mstart, mend);
 139
 140                 if (result)
 141                         return result;
 142         }
 143
 144         return init_transition_pgtable(image, level4p);
 145 }
 146
 147 static void set_idt(void *newidt, u16 limit)
 148 {
 149         struct desc_ptr curidt;
 150
 151         /* x86-64 supports unaliged loads & stores */
 152         curidt.size    = limit;
 153         curidt.address = (unsigned long)newidt;
 154
 155         __asm__ __volatile__ (
 156                 "lidtq %0\n"
 157                 : : "m" (curidt)
 158                 );
 159 };
 160
 161
 162 static void set_gdt(void *newgdt, u16 limit)
 163 {
 164         struct desc_ptr curgdt;
 165
 166         /* x86-64 supports unaligned loads & stores */
 167         curgdt.size    = limit;
 168         curgdt.address = (unsigned long)newgdt;
 169
 170         __asm__ __volatile__ (
 171                 "lgdtq %0\n"
 172                 : : "m" (curgdt)
 173                 );
 174 };
 175
 176 static void load_segments(void)
 177 {
 178         __asm__ __volatile__ (
 179                 "\tmovl %0,%%ds\n"
 180                 "\tmovl %0,%%es\n"
 181                 "\tmovl %0,%%ss\n"
 182                 "\tmovl %0,%%fs\n"
 183                 "\tmovl %0,%%gs\n"
 184                 : : "a" (__KERNEL_DS) : "memory"
 185                 );
 186 }
 187
 188 #ifdef CONFIG_KEXEC_FILE
 189 /* Update purgatory as needed after various image segments have been prepared */
 190 static int arch_update_purgatory(struct kimage *image)
 191 {
 192         int ret = 0;
 193
 194         if (!image->file_mode)
 195                 return 0;
 196
 197         /* Setup copying of backup region */
 198         if (image->type == KEXEC_TYPE_CRASH) {
 199                 ret = kexec_purgatory_get_set_symbol(image, "backup_dest",
 200                                 &image->arch.backup_load_addr,
 201                                 sizeof(image->arch.backup_load_addr), 0);
 202                 if (ret)
 203                         return ret;
 204
 205                 ret = kexec_purgatory_get_set_symbol(image, "backup_src",
 206                                 &image->arch.backup_src_start,
 207                                 sizeof(image->arch.backup_src_start), 0);
 208                 if (ret)
 209                         return ret;
 210
 211                 ret = kexec_purgatory_get_set_symbol(image, "backup_sz",
 212                                 &image->arch.backup_src_sz,
 213                                 sizeof(image->arch.backup_src_sz), 0);
 214                 if (ret)
 215                         return ret;
 216         }
 217
 218         return ret;
 219 }
 220 #else /* !CONFIG_KEXEC_FILE */
 221 static inline int arch_update_purgatory(struct kimage *image)
 222 {
 223         return 0;
 224 }
 225 #endif /* CONFIG_KEXEC_FILE */
 226
 227 int machine_kexec_prepare(struct kimage *image)
 228 {
 229         unsigned long start_pgtable;
 230         int result;
 231
 232         /* Calculate the offsets */
 233         start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 234
 235         /* Setup the identity mapped 64bit page table */
 236         result = init_pgtable(image, start_pgtable);
 237         if (result)
 238                 return result;
 239
 240         /* update purgatory as needed */
 241         result = arch_update_purgatory(image);
 242         if (result)
 243                 return result;
 244
 245         return 0;
 246 }
 247
 248 void machine_kexec_cleanup(struct kimage *image)
 249 {
 250         free_transition_pgtable(image);
 251 }
 252
 253 /*
 254  * Do not allocate memory (or fail in any way) in machine_kexec().
 255  * We are past the point of no return, committed to rebooting now.
 256  */
 257 void machine_kexec(struct kimage *image)
 258 {
 259         unsigned long page_list[PAGES_NR];
 260         void *control_page;
 261         int save_ftrace_enabled;
 262
 263 #ifdef CONFIG_KEXEC_JUMP
 264         if (image->preserve_context)
 265                 save_processor_state();
 266 #endif
 267
 268         save_ftrace_enabled = __ftrace_enabled_save();
 269
 270         /* Interrupts aren't acceptable while we reboot */
 271         local_irq_disable();
 272         hw_breakpoint_disable();
 273
 274         if (image->preserve_context) {
 275 #ifdef CONFIG_X86_IO_APIC
 276                 /*
 277                  * We need to put APICs in legacy mode so that we can
 278                  * get timer interrupts in second kernel. kexec/kdump
 279                  * paths already have calls to disable_IO_APIC() in
 280                  * one form or other. kexec jump path also need
 281                  * one.
 282                  */
 283                 disable_IO_APIC();
 284 #endif
 285         }
 286
 287         control_page = page_address(image->control_code_page) + PAGE_SIZE;
 288         memcpy(control_page, relocate_kernel, KEXEC_CONTROL_CODE_MAX_SIZE);
 289
 290         page_list[PA_CONTROL_PAGE] = virt_to_phys(control_page);
 291         page_list[VA_CONTROL_PAGE] = (unsigned long)control_page;
 292         page_list[PA_TABLE_PAGE] =
 293           (unsigned long)__pa(page_address(image->control_code_page));
 294
 295         if (image->type == KEXEC_TYPE_DEFAULT)
 296                 page_list[PA_SWAP_PAGE] = (page_to_pfn(image->swap_page)
 297                                                 << PAGE_SHIFT);
 298
 299         /*
 300          * The segment registers are funny things, they have both a
 301          * visible and an invisible part.  Whenever the visible part is
 302          * set to a specific selector, the invisible part is loaded
 303          * with from a table in memory.  At no other time is the
 304          * descriptor table in memory accessed.
 305          *
 306          * I take advantage of this here by force loading the
 307          * segments, before I zap the gdt with an invalid value.
 308          */
 309         load_segments();
 310         /*
 311          * The gdt & idt are now invalid.
 312          * If you want to load them you must set up your own idt & gdt.
 313          */
 314         set_gdt(phys_to_virt(0), 0);
 315         set_idt(phys_to_virt(0), 0);
 316
 317         /* now call it */
 318         image->start = relocate_kernel((unsigned long)image->head,
 319                                        (unsigned long)page_list,
 320                                        image->start,
 321                                        image->preserve_context);
 322
 323 #ifdef CONFIG_KEXEC_JUMP
 324         if (image->preserve_context)
 325                 restore_processor_state();
 326 #endif
 327
 328         __ftrace_enabled_restore(save_ftrace_enabled);
 329 }
 330
 331 void arch_crash_save_vmcoreinfo(void)
 332 {
 333         VMCOREINFO_SYMBOL(phys_base);
 334         VMCOREINFO_SYMBOL(init_level4_pgt);
 335
 336 #ifdef CONFIG_NUMA
 337         VMCOREINFO_SYMBOL(node_data);
 338         VMCOREINFO_LENGTH(node_data, MAX_NUMNODES);
 339 #endif
 340         vmcoreinfo_append_str("KERNELOFFSET=%lx\n",
 341                               kaslr_offset());
 342         VMCOREINFO_PAGE_OFFSET(PAGE_OFFSET);
 343         VMCOREINFO_VMALLOC_START(VMALLOC_START);
 344         VMCOREINFO_VMEMMAP_START(VMEMMAP_START);
 345 }
 346
 347 /* arch-dependent functionality related to kexec file-based syscall */
 348
 349 #ifdef CONFIG_KEXEC_FILE
 350 int arch_kexec_kernel_image_probe(struct kimage *image, void *buf,
 351                                   unsigned long buf_len)
 352 {
 353         int i, ret = -ENOEXEC;
 354         struct kexec_file_ops *fops;
 355
 356         for (i = 0; i < ARRAY_SIZE(kexec_file_loaders); i++) {
 357                 fops = kexec_file_loaders[i];
 358                 if (!fops || !fops->probe)
 359                         continue;
 360
 361                 ret = fops->probe(buf, buf_len);
 362                 if (!ret) {
 363                         image->fops = fops;
 364                         return ret;
 365                 }
 366         }
 367
 368         return ret;
 369 }
 370
 371 void *arch_kexec_kernel_image_load(struct kimage *image)
 372 {
 373         vfree(image->arch.elf_headers);
 374         image->arch.elf_headers = NULL;
 375
 376         if (!image->fops || !image->fops->load)
 377                 return ERR_PTR(-ENOEXEC);
 378
 379         return image->fops->load(image, image->kernel_buf,
 380                                  image->kernel_buf_len, image->initrd_buf,
 381                                  image->initrd_buf_len, image->cmdline_buf,
 382                                  image->cmdline_buf_len);
 383 }
 384
 385 int arch_kimage_file_post_load_cleanup(struct kimage *image)
 386 {
 387         if (!image->fops || !image->fops->cleanup)
 388                 return 0;
 389
 390         return image->fops->cleanup(image->image_loader_data);
 391 }
 392
 393 #ifdef CONFIG_KEXEC_VERIFY_SIG
 394 int arch_kexec_kernel_verify_sig(struct kimage *image, void *kernel,
 395                                  unsigned long kernel_len)
 396 {
 397         if (!image->fops || !image->fops->verify_sig) {
 398                 pr_debug("kernel loader does not support signature verification.");
 399                 return -EKEYREJECTED;
 400         }
 401
 402         return image->fops->verify_sig(kernel, kernel_len);
 403 }
 404 #endif
 405
 406 /*
 407  * Apply purgatory relocations.
 408  *
 409  * ehdr: Pointer to elf headers
 410  * sechdrs: Pointer to section headers.
 411  * relsec: section index of SHT_RELA section.
 412  *
 413  * TODO: Some of the code belongs to generic code. Move that in kexec.c.
 414  */
 415 int arch_kexec_apply_relocations_add(const Elf64_Ehdr *ehdr,
 416                                      Elf64_Shdr *sechdrs, unsigned int relsec)
 417 {
 418         unsigned int i;
 419         Elf64_Rela *rel;
 420         Elf64_Sym *sym;
 421         void *location;
 422         Elf64_Shdr *section, *symtabsec;
 423         unsigned long address, sec_base, value;
 424         const char *strtab, *name, *shstrtab;
 425
 426         /*
 427          * ->sh_offset has been modified to keep the pointer to section
 428          * contents in memory
 429          */
 430         rel = (void *)sechdrs[relsec].sh_offset;
 431
 432         /* Section to which relocations apply */
 433         section = &sechdrs[sechdrs[relsec].sh_info];
 434
 435         pr_debug("Applying relocate section %u to %u\n", relsec,
 436                  sechdrs[relsec].sh_info);
 437
 438         /* Associated symbol table */
 439         symtabsec = &sechdrs[sechdrs[relsec].sh_link];
 440
 441         /* String table */
 442         if (symtabsec->sh_link >= ehdr->e_shnum) {
 443                 /* Invalid strtab section number */
 444                 pr_err("Invalid string table section index %d\n",
 445                        symtabsec->sh_link);
 446                 return -ENOEXEC;
 447         }
 448
 449         strtab = (char *)sechdrs[symtabsec->sh_link].sh_offset;
 450
 451         /* section header string table */
 452         shstrtab = (char *)sechdrs[ehdr->e_shstrndx].sh_offset;
 453
 454         for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rel); i++) {
 455
 456                 /*
 457                  * rel[i].r_offset contains byte offset from beginning
 458                  * of section to the storage unit affected.
 459                  *
 460                  * This is location to update (->sh_offset). This is temporary
 461                  * buffer where section is currently loaded. This will finally
 462                  * be loaded to a different address later, pointed to by
 463                  * ->sh_addr. kexec takes care of moving it
 464                  *  (kexec_load_segment()).
 465                  */
 466                 location = (void *)(section->sh_offset + rel[i].r_offset);
 467
 468                 /* Final address of the location */
 469                 address = section->sh_addr + rel[i].r_offset;
 470
 471                 /*
 472                  * rel[i].r_info contains information about symbol table index
 473                  * w.r.t which relocation must be made and type of relocation
 474                  * to apply. ELF64_R_SYM() and ELF64_R_TYPE() macros get
 475                  * these respectively.
 476                  */
 477                 sym = (Elf64_Sym *)symtabsec->sh_offset +
 478                                 ELF64_R_SYM(rel[i].r_info);
 479
 480                 if (sym->st_name)
 481                         name = strtab + sym->st_name;
 482                 else
 483                         name = shstrtab + sechdrs[sym->st_shndx].sh_name;
 484
 485                 pr_debug("Symbol: %s info: %02x shndx: %02x value=%llx size: %llx\n",
 486                          name, sym->st_info, sym->st_shndx, sym->st_value,
 487                          sym->st_size);
 488
 489                 if (sym->st_shndx == SHN_UNDEF) {
 490                         pr_err("Undefined symbol: %s\n", name);
 491                         return -ENOEXEC;
 492                 }
 493
 494                 if (sym->st_shndx == SHN_COMMON) {
 495                         pr_err("symbol '%s' in common section\n", name);
 496                         return -ENOEXEC;
 497                 }
 498
 499                 if (sym->st_shndx == SHN_ABS)
 500                         sec_base = 0;
 501                 else if (sym->st_shndx >= ehdr->e_shnum) {
 502                         pr_err("Invalid section %d for symbol %s\n",
 503                                sym->st_shndx, name);
 504                         return -ENOEXEC;
 505                 } else
 506                         sec_base = sechdrs[sym->st_shndx].sh_addr;
 507
 508                 value = sym->st_value;
 509                 value += sec_base;
 510                 value += rel[i].r_addend;
 511
 512                 switch (ELF64_R_TYPE(rel[i].r_info)) {
 513                 case R_X86_64_NONE:
 514                         break;
 515                 case R_X86_64_64:
 516                         *(u64 *)location = value;
 517                         break;
 518                 case R_X86_64_32:
 519                         *(u32 *)location = value;
 520                         if (value != *(u32 *)location)
 521                                 goto overflow;
 522                         break;
 523                 case R_X86_64_32S:
 524                         *(s32 *)location = value;
 525                         if ((s64)value != *(s32 *)location)
 526                                 goto overflow;
 527                         break;
 528                 case R_X86_64_PC32:
 529                 case R_X86_64_PLT32:
 530                         value -= (u64)address;
 531                         *(u32 *)location = value;
 532                         break;
 533                 default:
 534                         pr_err("Unknown rela relocation: %llu\n",
 535                                ELF64_R_TYPE(rel[i].r_info));
 536                         return -ENOEXEC;
 537                 }
 538         }
 539         return 0;
 540
 541 overflow:
 542         pr_err("Overflow in relocation type %d value 0x%lx\n",
 543                (int)ELF64_R_TYPE(rel[i].r_info), value);
 544         return -ENOEXEC;
 545 }
 546 #endif /* CONFIG_KEXEC_FILE */
 547
 548 static int
 549 kexec_mark_range(unsigned long start, unsigned long end, bool protect)
 550 {
 551         struct page *page;
 552         unsigned int nr_pages;
 553
 554         /*
 555          * For physical range: [start, end]. We must skip the unassigned
 556          * crashk resource with zero-valued "end" member.
 557          */
 558         if (!end || start > end)
 559                 return 0;
 560
 561         page = pfn_to_page(start >> PAGE_SHIFT);
 562         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 563         if (protect)
 564                 return set_pages_ro(page, nr_pages);
 565         else
 566                 return set_pages_rw(page, nr_pages);
 567 }
 568
 569 static void kexec_mark_crashkres(bool protect)
 570 {
 571         unsigned long control;
 572
 573         kexec_mark_range(crashk_low_res.start, crashk_low_res.end, protect);
 574
 575         /* Don't touch the control code page used in crash_kexec().*/
 576         control = PFN_PHYS(page_to_pfn(kexec_crash_image->control_code_page));
 577         /* Control code page is located in the 2nd page. */
 578         kexec_mark_range(crashk_res.start, control + PAGE_SIZE - 1, protect);
 579         control += KEXEC_CONTROL_PAGE_SIZE;
 580         kexec_mark_range(control, crashk_res.end, protect);
 581 }
 582
 583 void arch_kexec_protect_crashkres(void)
 584 {
 585         kexec_mark_crashkres(true);
 586 }
 587
 588 void arch_kexec_unprotect_crashkres(void)
 589 {
 590         kexec_mark_crashkres(false);
 591 }