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