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