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 #include <asm/set_memory.h>
  31
  32 #ifdef CONFIG_KEXEC_FILE
  33 static struct kexec_file_ops *kexec_file_loaders[] = {
  34                 &kexec_bzImage64_ops,
  35 };
  36 #endif
  37
  38 static void free_transition_pgtable(struct kimage *image)
  39 {
  40         free_page((unsigned long)image->arch.p4d);
  41         free_page((unsigned long)image->arch.pud);
  42         free_page((unsigned long)image->arch.pmd);
  43         free_page((unsigned long)image->arch.pte);
  44 }
  45
  46 static int init_transition_pgtable(struct kimage *image, pgd_t *pgd)
  47 {
  48         p4d_t *p4d;
  49         pud_t *pud;
  50         pmd_t *pmd;
  51         pte_t *pte;
  52         unsigned long vaddr, paddr;
  53         int result = -ENOMEM;
  54
  55         vaddr = (unsigned long)relocate_kernel;
  56         paddr = __pa(page_address(image->control_code_page)+PAGE_SIZE);
  57         pgd += pgd_index(vaddr);
  58         if (!pgd_present(*pgd)) {
  59                 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
  60                 if (!p4d)
  61                         goto err;
  62                 image->arch.p4d = p4d;
  63                 set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE));
  64         }
  65         p4d = p4d_offset(pgd, vaddr);
  66         if (!p4d_present(*p4d)) {
  67                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
  68                 if (!pud)
  69                         goto err;
  70                 image->arch.pud = pud;
  71                 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
  72         }
  73         pud = pud_offset(p4d, vaddr);
  74         if (!pud_present(*pud)) {
  75                 pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
  76                 if (!pmd)
  77                         goto err;
  78                 image->arch.pmd = pmd;
  79                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
  80         }
  81         pmd = pmd_offset(pud, vaddr);
  82         if (!pmd_present(*pmd)) {
  83                 pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
  84                 if (!pte)
  85                         goto err;
  86                 image->arch.pte = pte;
  87                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
  88         }
  89         pte = pte_offset_kernel(pmd, vaddr);
  90         set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL_EXEC));
  91         return 0;
  92 err:
  93         free_transition_pgtable(image);
  94         return result;
  95 }
  96
  97 static void *alloc_pgt_page(void *data)
  98 {
  99         struct kimage *image = (struct kimage *)data;
 100         struct page *page;
 101         void *p = NULL;
 102
 103         page = kimage_alloc_control_pages(image, 0);
 104         if (page) {
 105                 p = page_address(page);
 106                 clear_page(p);
 107         }
 108
 109         return p;
 110 }
 111
 112 static int init_pgtable(struct kimage *image, unsigned long start_pgtable)
 113 {
 114         struct x86_mapping_info info = {
 115                 .alloc_pgt_page = alloc_pgt_page,
 116                 .context        = image,
 117                 .pmd_flag       = __PAGE_KERNEL_LARGE_EXEC,
 118         };
 119         unsigned long mstart, mend;
 120         pgd_t *level4p;
 121         int result;
 122         int i;
 123
 124         level4p = (pgd_t *)__va(start_pgtable);
 125         clear_page(level4p);
 126         for (i = 0; i < nr_pfn_mapped; i++) {
 127                 mstart = pfn_mapped[i].start << PAGE_SHIFT;
 128                 mend   = pfn_mapped[i].end << PAGE_SHIFT;
 129
 130                 result = kernel_ident_mapping_init(&info,
 131                                                  level4p, mstart, mend);
 132                 if (result)
 133                         return result;
 134         }
 135
 136         /*
 137          * segments's mem ranges could be outside 0 ~ max_pfn,
 138          * for example when jump back to original kernel from kexeced kernel.
 139          * or first kernel is booted with user mem map, and second kernel
 140          * could be loaded out of that range.
 141          */
 142         for (i = 0; i < image->nr_segments; i++) {
 143                 mstart = image->segment[i].mem;
 144                 mend   = mstart + image->segment[i].memsz;
 145
 146                 result = kernel_ident_mapping_init(&info,
 147                                                  level4p, mstart, mend);
 148
 149                 if (result)
 150                         return result;
 151         }
 152
 153         return init_transition_pgtable(image, level4p);
 154 }
 155
 156 static void set_idt(void *newidt, u16 limit)
 157 {
 158         struct desc_ptr curidt;
 159
 160         /* x86-64 supports unaliged loads & stores */
 161         curidt.size    = limit;
 162         curidt.address = (unsigned long)newidt;
 163
 164         __asm__ __volatile__ (
 165                 "lidtq %0\n"
 166                 : : "m" (curidt)
 167                 );
 168 };
 169
 170
 171 static void set_gdt(void *newgdt, u16 limit)
 172 {
 173         struct desc_ptr curgdt;
 174
 175         /* x86-64 supports unaligned loads & stores */
 176         curgdt.size    = limit;
 177         curgdt.address = (unsigned long)newgdt;
 178
 179         __asm__ __volatile__ (
 180                 "lgdtq %0\n"
 181                 : : "m" (curgdt)
 182                 );
 183 };
 184
 185 static void load_segments(void)
 186 {
 187         __asm__ __volatile__ (
 188                 "\tmovl %0,%%ds\n"
 189                 "\tmovl %0,%%es\n"
 190                 "\tmovl %0,%%ss\n"
 191                 "\tmovl %0,%%fs\n"
 192                 "\tmovl %0,%%gs\n"
 193                 : : "a" (__KERNEL_DS) : "memory"
 194                 );
 195 }
 196
 197 #ifdef CONFIG_KEXEC_FILE
 198 /* Update purgatory as needed after various image segments have been prepared */
 199 static int arch_update_purgatory(struct kimage *image)
 200 {
 201         int ret = 0;
 202
 203         if (!image->file_mode)
 204                 return 0;
 205
 206         /* Setup copying of backup region */
 207         if (image->type == KEXEC_TYPE_CRASH) {
 208                 ret = kexec_purgatory_get_set_symbol(image,
 209                                 "purgatory_backup_dest",
 210                                 &image->arch.backup_load_addr,
 211                                 sizeof(image->arch.backup_load_addr), 0);
 212                 if (ret)
 213                         return ret;
 214
 215                 ret = kexec_purgatory_get_set_symbol(image,
 216                                 "purgatory_backup_src",
 217                                 &image->arch.backup_src_start,
 218                                 sizeof(image->arch.backup_src_start), 0);
 219                 if (ret)
 220                         return ret;
 221
 222                 ret = kexec_purgatory_get_set_symbol(image,
 223                                 "purgatory_backup_sz",
 224                                 &image->arch.backup_src_sz,
 225                                 sizeof(image->arch.backup_src_sz), 0);
 226                 if (ret)
 227                         return ret;
 228         }
 229
 230         return ret;
 231 }
 232 #else /* !CONFIG_KEXEC_FILE */
 233 static inline int arch_update_purgatory(struct kimage *image)
 234 {
 235         return 0;
 236 }
 237 #endif /* CONFIG_KEXEC_FILE */
 238
 239 int machine_kexec_prepare(struct kimage *image)
 240 {
 241         unsigned long start_pgtable;
 242         int result;
 243
 244         /* Calculate the offsets */
 245         start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 246
 247         /* Setup the identity mapped 64bit page table */
 248         result = init_pgtable(image, start_pgtable);
 249         if (result)
 250                 return result;
 251
 252         /* update purgatory as needed */
 253         result = arch_update_purgatory(image);
 254         if (result)
 255                 return result;
 256
 257         return 0;
 258 }
 259
 260 void machine_kexec_cleanup(struct kimage *image)
 261 {
 262         free_transition_pgtable(image);
 263 }
 264
 265 /*
 266  * Do not allocate memory (or fail in any way) in machine_kexec().
 267  * We are past the point of no return, committed to rebooting now.
 268  */
 269 void machine_kexec(struct kimage *image)
 270 {
 271         unsigned long page_list[PAGES_NR];
 272         void *control_page;
 273         int save_ftrace_enabled;
 274
 275 #ifdef CONFIG_KEXEC_JUMP
 276         if (image->preserve_context)
 277                 save_processor_state();
 278 #endif
 279
 280         save_ftrace_enabled = __ftrace_enabled_save();
 281
 282         /* Interrupts aren't acceptable while we reboot */
 283         local_irq_disable();
 284         hw_breakpoint_disable();
 285
 286         if (image->preserve_context) {
 287 #ifdef CONFIG_X86_IO_APIC
 288                 /*
 289                  * We need to put APICs in legacy mode so that we can
 290                  * get timer interrupts in second kernel. kexec/kdump
 291                  * paths already have calls to disable_IO_APIC() in
 292                  * one form or other. kexec jump path also need
 293                  * one.
 294                  */
 295                 disable_IO_APIC();
 296 #endif
 297         }
 298
 299         control_page = page_address(image->control_code_page) + PAGE_SIZE;
 300         memcpy(control_page, relocate_kernel, KEXEC_CONTROL_CODE_MAX_SIZE);
 301
 302         page_list[PA_CONTROL_PAGE] = virt_to_phys(control_page);
 303         page_list[VA_CONTROL_PAGE] = (unsigned long)control_page;
 304         page_list[PA_TABLE_PAGE] =
 305           (unsigned long)__pa(page_address(image->control_code_page));
 306
 307         if (image->type == KEXEC_TYPE_DEFAULT)
 308                 page_list[PA_SWAP_PAGE] = (page_to_pfn(image->swap_page)
 309                                                 << PAGE_SHIFT);
 310
 311         /*
 312          * The segment registers are funny things, they have both a
 313          * visible and an invisible part.  Whenever the visible part is
 314          * set to a specific selector, the invisible part is loaded
 315          * with from a table in memory.  At no other time is the
 316          * descriptor table in memory accessed.
 317          *
 318          * I take advantage of this here by force loading the
 319          * segments, before I zap the gdt with an invalid value.
 320          */
 321         load_segments();
 322         /*
 323          * The gdt & idt are now invalid.
 324          * If you want to load them you must set up your own idt & gdt.
 325          */
 326         set_gdt(phys_to_virt(0), 0);
 327         set_idt(phys_to_virt(0), 0);
 328
 329         /* now call it */
 330         image->start = relocate_kernel((unsigned long)image->head,
 331                                        (unsigned long)page_list,
 332                                        image->start,
 333                                        image->preserve_context);
 334
 335 #ifdef CONFIG_KEXEC_JUMP
 336         if (image->preserve_context)
 337                 restore_processor_state();
 338 #endif
 339
 340         __ftrace_enabled_restore(save_ftrace_enabled);
 341 }
 342
 343 void arch_crash_save_vmcoreinfo(void)
 344 {
 345         VMCOREINFO_NUMBER(phys_base);
 346         VMCOREINFO_SYMBOL(init_level4_pgt);
 347
 348 #ifdef CONFIG_NUMA
 349         VMCOREINFO_SYMBOL(node_data);
 350         VMCOREINFO_LENGTH(node_data, MAX_NUMNODES);
 351 #endif
 352         vmcoreinfo_append_str("KERNELOFFSET=%lx\n",
 353                               kaslr_offset());
 354         VMCOREINFO_NUMBER(KERNEL_IMAGE_SIZE);
 355 }
 356
 357 /* arch-dependent functionality related to kexec file-based syscall */
 358
 359 #ifdef CONFIG_KEXEC_FILE
 360 int arch_kexec_kernel_image_probe(struct kimage *image, void *buf,
 361                                   unsigned long buf_len)
 362 {
 363         int i, ret = -ENOEXEC;
 364         struct kexec_file_ops *fops;
 365
 366         for (i = 0; i < ARRAY_SIZE(kexec_file_loaders); i++) {
 367                 fops = kexec_file_loaders[i];
 368                 if (!fops || !fops->probe)
 369                         continue;
 370
 371                 ret = fops->probe(buf, buf_len);
 372                 if (!ret) {
 373                         image->fops = fops;
 374                         return ret;
 375                 }
 376         }
 377
 378         return ret;
 379 }
 380
 381 void *arch_kexec_kernel_image_load(struct kimage *image)
 382 {
 383         vfree(image->arch.elf_headers);
 384         image->arch.elf_headers = NULL;
 385
 386         if (!image->fops || !image->fops->load)
 387                 return ERR_PTR(-ENOEXEC);
 388
 389         return image->fops->load(image, image->kernel_buf,
 390                                  image->kernel_buf_len, image->initrd_buf,
 391                                  image->initrd_buf_len, image->cmdline_buf,
 392                                  image->cmdline_buf_len);
 393 }
 394
 395 int arch_kimage_file_post_load_cleanup(struct kimage *image)
 396 {
 397         if (!image->fops || !image->fops->cleanup)
 398                 return 0;
 399
 400         return image->fops->cleanup(image->image_loader_data);
 401 }
 402
 403 #ifdef CONFIG_KEXEC_VERIFY_SIG
 404 int arch_kexec_kernel_verify_sig(struct kimage *image, void *kernel,
 405                                  unsigned long kernel_len)
 406 {
 407         if (!image->fops || !image->fops->verify_sig) {
 408                 pr_debug("kernel loader does not support signature verification.");
 409                 return -EKEYREJECTED;
 410         }
 411
 412         return image->fops->verify_sig(kernel, kernel_len);
 413 }
 414 #endif
 415
 416 /*
 417  * Apply purgatory relocations.
 418  *
 419  * ehdr: Pointer to elf headers
 420  * sechdrs: Pointer to section headers.
 421  * relsec: section index of SHT_RELA section.
 422  *
 423  * TODO: Some of the code belongs to generic code. Move that in kexec.c.
 424  */
 425 int arch_kexec_apply_relocations_add(const Elf64_Ehdr *ehdr,
 426                                      Elf64_Shdr *sechdrs, unsigned int relsec)
 427 {
 428         unsigned int i;
 429         Elf64_Rela *rel;
 430         Elf64_Sym *sym;
 431         void *location;
 432         Elf64_Shdr *section, *symtabsec;
 433         unsigned long address, sec_base, value;
 434         const char *strtab, *name, *shstrtab;
 435
 436         /*
 437          * ->sh_offset has been modified to keep the pointer to section
 438          * contents in memory
 439          */
 440         rel = (void *)sechdrs[relsec].sh_offset;
 441
 442         /* Section to which relocations apply */
 443         section = &sechdrs[sechdrs[relsec].sh_info];
 444
 445         pr_debug("Applying relocate section %u to %u\n", relsec,
 446                  sechdrs[relsec].sh_info);
 447
 448         /* Associated symbol table */
 449         symtabsec = &sechdrs[sechdrs[relsec].sh_link];
 450
 451         /* String table */
 452         if (symtabsec->sh_link >= ehdr->e_shnum) {
 453                 /* Invalid strtab section number */
 454                 pr_err("Invalid string table section index %d\n",
 455                        symtabsec->sh_link);
 456                 return -ENOEXEC;
 457         }
 458
 459         strtab = (char *)sechdrs[symtabsec->sh_link].sh_offset;
 460
 461         /* section header string table */
 462         shstrtab = (char *)sechdrs[ehdr->e_shstrndx].sh_offset;
 463
 464         for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rel); i++) {
 465
 466                 /*
 467                  * rel[i].r_offset contains byte offset from beginning
 468                  * of section to the storage unit affected.
 469                  *
 470                  * This is location to update (->sh_offset). This is temporary
 471                  * buffer where section is currently loaded. This will finally
 472                  * be loaded to a different address later, pointed to by
 473                  * ->sh_addr. kexec takes care of moving it
 474                  *  (kexec_load_segment()).
 475                  */
 476                 location = (void *)(section->sh_offset + rel[i].r_offset);
 477
 478                 /* Final address of the location */
 479                 address = section->sh_addr + rel[i].r_offset;
 480
 481                 /*
 482                  * rel[i].r_info contains information about symbol table index
 483                  * w.r.t which relocation must be made and type of relocation
 484                  * to apply. ELF64_R_SYM() and ELF64_R_TYPE() macros get
 485                  * these respectively.
 486                  */
 487                 sym = (Elf64_Sym *)symtabsec->sh_offset +
 488                                 ELF64_R_SYM(rel[i].r_info);
 489
 490                 if (sym->st_name)
 491                         name = strtab + sym->st_name;
 492                 else
 493                         name = shstrtab + sechdrs[sym->st_shndx].sh_name;
 494
 495                 pr_debug("Symbol: %s info: %02x shndx: %02x value=%llx size: %llx\n",
 496                          name, sym->st_info, sym->st_shndx, sym->st_value,
 497                          sym->st_size);
 498
 499                 if (sym->st_shndx == SHN_UNDEF) {
 500                         pr_err("Undefined symbol: %s\n", name);
 501                         return -ENOEXEC;
 502                 }
 503
 504                 if (sym->st_shndx == SHN_COMMON) {
 505                         pr_err("symbol '%s' in common section\n", name);
 506                         return -ENOEXEC;
 507                 }
 508
 509                 if (sym->st_shndx == SHN_ABS)
 510                         sec_base = 0;
 511                 else if (sym->st_shndx >= ehdr->e_shnum) {
 512                         pr_err("Invalid section %d for symbol %s\n",
 513                                sym->st_shndx, name);
 514                         return -ENOEXEC;
 515                 } else
 516                         sec_base = sechdrs[sym->st_shndx].sh_addr;
 517
 518                 value = sym->st_value;
 519                 value += sec_base;
 520                 value += rel[i].r_addend;
 521
 522                 switch (ELF64_R_TYPE(rel[i].r_info)) {
 523                 case R_X86_64_NONE:
 524                         break;
 525                 case R_X86_64_64:
 526                         *(u64 *)location = value;
 527                         break;
 528                 case R_X86_64_32:
 529                         *(u32 *)location = value;
 530                         if (value != *(u32 *)location)
 531                                 goto overflow;
 532                         break;
 533                 case R_X86_64_32S:
 534                         *(s32 *)location = value;
 535                         if ((s64)value != *(s32 *)location)
 536                                 goto overflow;
 537                         break;
 538                 case R_X86_64_PC32:
 539                         value -= (u64)address;
 540                         *(u32 *)location = value;
 541                         break;
 542                 default:
 543                         pr_err("Unknown rela relocation: %llu\n",
 544                                ELF64_R_TYPE(rel[i].r_info));
 545                         return -ENOEXEC;
 546                 }
 547         }
 548         return 0;
 549
 550 overflow:
 551         pr_err("Overflow in relocation type %d value 0x%lx\n",
 552                (int)ELF64_R_TYPE(rel[i].r_info), value);
 553         return -ENOEXEC;
 554 }
 555 #endif /* CONFIG_KEXEC_FILE */
 556
 557 static int
 558 kexec_mark_range(unsigned long start, unsigned long end, bool protect)
 559 {
 560         struct page *page;
 561         unsigned int nr_pages;
 562
 563         /*
 564          * For physical range: [start, end]. We must skip the unassigned
 565          * crashk resource with zero-valued "end" member.
 566          */
 567         if (!end || start > end)
 568                 return 0;
 569
 570         page = pfn_to_page(start >> PAGE_SHIFT);
 571         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 572         if (protect)
 573                 return set_pages_ro(page, nr_pages);
 574         else
 575                 return set_pages_rw(page, nr_pages);
 576 }
 577
 578 static void kexec_mark_crashkres(bool protect)
 579 {
 580         unsigned long control;
 581
 582         kexec_mark_range(crashk_low_res.start, crashk_low_res.end, protect);
 583
 584         /* Don't touch the control code page used in crash_kexec().*/
 585         control = PFN_PHYS(page_to_pfn(kexec_crash_image->control_code_page));
 586         /* Control code page is located in the 2nd page. */
 587         kexec_mark_range(crashk_res.start, control + PAGE_SIZE - 1, protect);
 588         control += KEXEC_CONTROL_PAGE_SIZE;
 589         kexec_mark_range(control, crashk_res.end, protect);
 590 }
 591
 592 void arch_kexec_protect_crashkres(void)
 593 {
 594         kexec_mark_crashkres(true);
 595 }
 596
 597 void arch_kexec_unprotect_crashkres(void)
 598 {
 599         kexec_mark_crashkres(false);
 600 }