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