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