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