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                 .page_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
 127         if (direct_gbpages)
 128                 info.direct_gbpages = true;
 129
 130         for (i = 0; i < nr_pfn_mapped; i++) {
 131                 mstart = pfn_mapped[i].start << PAGE_SHIFT;
 132                 mend   = pfn_mapped[i].end << PAGE_SHIFT;
 133
 134                 result = kernel_ident_mapping_init(&info,
 135                                                  level4p, mstart, mend);
 136                 if (result)
 137                         return result;
 138         }
 139
 140         /*
 141          * segments's mem ranges could be outside 0 ~ max_pfn,
 142          * for example when jump back to original kernel from kexeced kernel.
 143          * or first kernel is booted with user mem map, and second kernel
 144          * could be loaded out of that range.
 145          */
 146         for (i = 0; i < image->nr_segments; i++) {
 147                 mstart = image->segment[i].mem;
 148                 mend   = mstart + image->segment[i].memsz;
 149
 150                 result = kernel_ident_mapping_init(&info,
 151                                                  level4p, mstart, mend);
 152
 153                 if (result)
 154                         return result;
 155         }
 156
 157         return init_transition_pgtable(image, level4p);
 158 }
 159
 160 static void set_idt(void *newidt, u16 limit)
 161 {
 162         struct desc_ptr curidt;
 163
 164         /* x86-64 supports unaliged loads & stores */
 165         curidt.size    = limit;
 166         curidt.address = (unsigned long)newidt;
 167
 168         __asm__ __volatile__ (
 169                 "lidtq %0\n"
 170                 : : "m" (curidt)
 171                 );
 172 };
 173
 174
 175 static void set_gdt(void *newgdt, u16 limit)
 176 {
 177         struct desc_ptr curgdt;
 178
 179         /* x86-64 supports unaligned loads & stores */
 180         curgdt.size    = limit;
 181         curgdt.address = (unsigned long)newgdt;
 182
 183         __asm__ __volatile__ (
 184                 "lgdtq %0\n"
 185                 : : "m" (curgdt)
 186                 );
 187 };
 188
 189 static void load_segments(void)
 190 {
 191         __asm__ __volatile__ (
 192                 "\tmovl %0,%%ds\n"
 193                 "\tmovl %0,%%es\n"
 194                 "\tmovl %0,%%ss\n"
 195                 "\tmovl %0,%%fs\n"
 196                 "\tmovl %0,%%gs\n"
 197                 : : "a" (__KERNEL_DS) : "memory"
 198                 );
 199 }
 200
 201 #ifdef CONFIG_KEXEC_FILE
 202 /* Update purgatory as needed after various image segments have been prepared */
 203 static int arch_update_purgatory(struct kimage *image)
 204 {
 205         int ret = 0;
 206
 207         if (!image->file_mode)
 208                 return 0;
 209
 210         /* Setup copying of backup region */
 211         if (image->type == KEXEC_TYPE_CRASH) {
 212                 ret = kexec_purgatory_get_set_symbol(image,
 213                                 "purgatory_backup_dest",
 214                                 &image->arch.backup_load_addr,
 215                                 sizeof(image->arch.backup_load_addr), 0);
 216                 if (ret)
 217                         return ret;
 218
 219                 ret = kexec_purgatory_get_set_symbol(image,
 220                                 "purgatory_backup_src",
 221                                 &image->arch.backup_src_start,
 222                                 sizeof(image->arch.backup_src_start), 0);
 223                 if (ret)
 224                         return ret;
 225
 226                 ret = kexec_purgatory_get_set_symbol(image,
 227                                 "purgatory_backup_sz",
 228                                 &image->arch.backup_src_sz,
 229                                 sizeof(image->arch.backup_src_sz), 0);
 230                 if (ret)
 231                         return ret;
 232         }
 233
 234         return ret;
 235 }
 236 #else /* !CONFIG_KEXEC_FILE */
 237 static inline int arch_update_purgatory(struct kimage *image)
 238 {
 239         return 0;
 240 }
 241 #endif /* CONFIG_KEXEC_FILE */
 242
 243 int machine_kexec_prepare(struct kimage *image)
 244 {
 245         unsigned long start_pgtable;
 246         int result;
 247
 248         /* Calculate the offsets */
 249         start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 250
 251         /* Setup the identity mapped 64bit page table */
 252         result = init_pgtable(image, start_pgtable);
 253         if (result)
 254                 return result;
 255
 256         /* update purgatory as needed */
 257         result = arch_update_purgatory(image);
 258         if (result)
 259                 return result;
 260
 261         return 0;
 262 }
 263
 264 void machine_kexec_cleanup(struct kimage *image)
 265 {
 266         free_transition_pgtable(image);
 267 }
 268
 269 /*
 270  * Do not allocate memory (or fail in any way) in machine_kexec().
 271  * We are past the point of no return, committed to rebooting now.
 272  */
 273 void machine_kexec(struct kimage *image)
 274 {
 275         unsigned long page_list[PAGES_NR];
 276         void *control_page;
 277         int save_ftrace_enabled;
 278
 279 #ifdef CONFIG_KEXEC_JUMP
 280         if (image->preserve_context)
 281                 save_processor_state();
 282 #endif
 283
 284         save_ftrace_enabled = __ftrace_enabled_save();
 285
 286         /* Interrupts aren't acceptable while we reboot */
 287         local_irq_disable();
 288         hw_breakpoint_disable();
 289
 290         if (image->preserve_context) {
 291 #ifdef CONFIG_X86_IO_APIC
 292                 /*
 293                  * We need to put APICs in legacy mode so that we can
 294                  * get timer interrupts in second kernel. kexec/kdump
 295                  * paths already have calls to disable_IO_APIC() in
 296                  * one form or other. kexec jump path also need
 297                  * one.
 298                  */
 299                 disable_IO_APIC();
 300 #endif
 301         }
 302
 303         control_page = page_address(image->control_code_page) + PAGE_SIZE;
 304         memcpy(control_page, relocate_kernel, KEXEC_CONTROL_CODE_MAX_SIZE);
 305
 306         page_list[PA_CONTROL_PAGE] = virt_to_phys(control_page);
 307         page_list[VA_CONTROL_PAGE] = (unsigned long)control_page;
 308         page_list[PA_TABLE_PAGE] =
 309           (unsigned long)__pa(page_address(image->control_code_page));
 310
 311         if (image->type == KEXEC_TYPE_DEFAULT)
 312                 page_list[PA_SWAP_PAGE] = (page_to_pfn(image->swap_page)
 313                                                 << PAGE_SHIFT);
 314
 315         /*
 316          * The segment registers are funny things, they have both a
 317          * visible and an invisible part.  Whenever the visible part is
 318          * set to a specific selector, the invisible part is loaded
 319          * with from a table in memory.  At no other time is the
 320          * descriptor table in memory accessed.
 321          *
 322          * I take advantage of this here by force loading the
 323          * segments, before I zap the gdt with an invalid value.
 324          */
 325         load_segments();
 326         /*
 327          * The gdt & idt are now invalid.
 328          * If you want to load them you must set up your own idt & gdt.
 329          */
 330         set_gdt(phys_to_virt(0), 0);
 331         set_idt(phys_to_virt(0), 0);
 332
 333         /* now call it */
 334         image->start = relocate_kernel((unsigned long)image->head,
 335                                        (unsigned long)page_list,
 336                                        image->start,
 337                                        image->preserve_context);
 338
 339 #ifdef CONFIG_KEXEC_JUMP
 340         if (image->preserve_context)
 341                 restore_processor_state();
 342 #endif
 343
 344         __ftrace_enabled_restore(save_ftrace_enabled);
 345 }
 346
 347 void arch_crash_save_vmcoreinfo(void)
 348 {
 349         VMCOREINFO_NUMBER(phys_base);
 350         VMCOREINFO_SYMBOL(init_top_pgt);
 351
 352 #ifdef CONFIG_NUMA
 353         VMCOREINFO_SYMBOL(node_data);
 354         VMCOREINFO_LENGTH(node_data, MAX_NUMNODES);
 355 #endif
 356         vmcoreinfo_append_str("KERNELOFFSET=%lx\n",
 357                               kaslr_offset());
 358         VMCOREINFO_NUMBER(KERNEL_IMAGE_SIZE);
 359 }
 360
 361 /* arch-dependent functionality related to kexec file-based syscall */
 362
 363 #ifdef CONFIG_KEXEC_FILE
 364 int arch_kexec_kernel_image_probe(struct kimage *image, void *buf,
 365                                   unsigned long buf_len)
 366 {
 367         int i, ret = -ENOEXEC;
 368         struct kexec_file_ops *fops;
 369
 370         for (i = 0; i < ARRAY_SIZE(kexec_file_loaders); i++) {
 371                 fops = kexec_file_loaders[i];
 372                 if (!fops || !fops->probe)
 373                         continue;
 374
 375                 ret = fops->probe(buf, buf_len);
 376                 if (!ret) {
 377                         image->fops = fops;
 378                         return ret;
 379                 }
 380         }
 381
 382         return ret;
 383 }
 384
 385 void *arch_kexec_kernel_image_load(struct kimage *image)
 386 {
 387         vfree(image->arch.elf_headers);
 388         image->arch.elf_headers = NULL;
 389
 390         if (!image->fops || !image->fops->load)
 391                 return ERR_PTR(-ENOEXEC);
 392
 393         return image->fops->load(image, image->kernel_buf,
 394                                  image->kernel_buf_len, image->initrd_buf,
 395                                  image->initrd_buf_len, image->cmdline_buf,
 396                                  image->cmdline_buf_len);
 397 }
 398
 399 int arch_kimage_file_post_load_cleanup(struct kimage *image)
 400 {
 401         if (!image->fops || !image->fops->cleanup)
 402                 return 0;
 403
 404         return image->fops->cleanup(image->image_loader_data);
 405 }
 406
 407 #ifdef CONFIG_KEXEC_VERIFY_SIG
 408 int arch_kexec_kernel_verify_sig(struct kimage *image, void *kernel,
 409                                  unsigned long kernel_len)
 410 {
 411         if (!image->fops || !image->fops->verify_sig) {
 412                 pr_debug("kernel loader does not support signature verification.");
 413                 return -EKEYREJECTED;
 414         }
 415
 416         return image->fops->verify_sig(kernel, kernel_len);
 417 }
 418 #endif
 419
 420 /*
 421  * Apply purgatory relocations.
 422  *
 423  * ehdr: Pointer to elf headers
 424  * sechdrs: Pointer to section headers.
 425  * relsec: section index of SHT_RELA section.
 426  *
 427  * TODO: Some of the code belongs to generic code. Move that in kexec.c.
 428  */
 429 int arch_kexec_apply_relocations_add(const Elf64_Ehdr *ehdr,
 430                                      Elf64_Shdr *sechdrs, unsigned int relsec)
 431 {
 432         unsigned int i;
 433         Elf64_Rela *rel;
 434         Elf64_Sym *sym;
 435         void *location;
 436         Elf64_Shdr *section, *symtabsec;
 437         unsigned long address, sec_base, value;
 438         const char *strtab, *name, *shstrtab;
 439
 440         /*
 441          * ->sh_offset has been modified to keep the pointer to section
 442          * contents in memory
 443          */
 444         rel = (void *)sechdrs[relsec].sh_offset;
 445
 446         /* Section to which relocations apply */
 447         section = &sechdrs[sechdrs[relsec].sh_info];
 448
 449         pr_debug("Applying relocate section %u to %u\n", relsec,
 450                  sechdrs[relsec].sh_info);
 451
 452         /* Associated symbol table */
 453         symtabsec = &sechdrs[sechdrs[relsec].sh_link];
 454
 455         /* String table */
 456         if (symtabsec->sh_link >= ehdr->e_shnum) {
 457                 /* Invalid strtab section number */
 458                 pr_err("Invalid string table section index %d\n",
 459                        symtabsec->sh_link);
 460                 return -ENOEXEC;
 461         }
 462
 463         strtab = (char *)sechdrs[symtabsec->sh_link].sh_offset;
 464
 465         /* section header string table */
 466         shstrtab = (char *)sechdrs[ehdr->e_shstrndx].sh_offset;
 467
 468         for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rel); i++) {
 469
 470                 /*
 471                  * rel[i].r_offset contains byte offset from beginning
 472                  * of section to the storage unit affected.
 473                  *
 474                  * This is location to update (->sh_offset). This is temporary
 475                  * buffer where section is currently loaded. This will finally
 476                  * be loaded to a different address later, pointed to by
 477                  * ->sh_addr. kexec takes care of moving it
 478                  *  (kexec_load_segment()).
 479                  */
 480                 location = (void *)(section->sh_offset + rel[i].r_offset);
 481
 482                 /* Final address of the location */
 483                 address = section->sh_addr + rel[i].r_offset;
 484
 485                 /*
 486                  * rel[i].r_info contains information about symbol table index
 487                  * w.r.t which relocation must be made and type of relocation
 488                  * to apply. ELF64_R_SYM() and ELF64_R_TYPE() macros get
 489                  * these respectively.
 490                  */
 491                 sym = (Elf64_Sym *)symtabsec->sh_offset +
 492                                 ELF64_R_SYM(rel[i].r_info);
 493
 494                 if (sym->st_name)
 495                         name = strtab + sym->st_name;
 496                 else
 497                         name = shstrtab + sechdrs[sym->st_shndx].sh_name;
 498
 499                 pr_debug("Symbol: %s info: %02x shndx: %02x value=%llx size: %llx\n",
 500                          name, sym->st_info, sym->st_shndx, sym->st_value,
 501                          sym->st_size);
 502
 503                 if (sym->st_shndx == SHN_UNDEF) {
 504                         pr_err("Undefined symbol: %s\n", name);
 505                         return -ENOEXEC;
 506                 }
 507
 508                 if (sym->st_shndx == SHN_COMMON) {
 509                         pr_err("symbol '%s' in common section\n", name);
 510                         return -ENOEXEC;
 511                 }
 512
 513                 if (sym->st_shndx == SHN_ABS)
 514                         sec_base = 0;
 515                 else if (sym->st_shndx >= ehdr->e_shnum) {
 516                         pr_err("Invalid section %d for symbol %s\n",
 517                                sym->st_shndx, name);
 518                         return -ENOEXEC;
 519                 } else
 520                         sec_base = sechdrs[sym->st_shndx].sh_addr;
 521
 522                 value = sym->st_value;
 523                 value += sec_base;
 524                 value += rel[i].r_addend;
 525
 526                 switch (ELF64_R_TYPE(rel[i].r_info)) {
 527                 case R_X86_64_NONE:
 528                         break;
 529                 case R_X86_64_64:
 530                         *(u64 *)location = value;
 531                         break;
 532                 case R_X86_64_32:
 533                         *(u32 *)location = value;
 534                         if (value != *(u32 *)location)
 535                                 goto overflow;
 536                         break;
 537                 case R_X86_64_32S:
 538                         *(s32 *)location = value;
 539                         if ((s64)value != *(s32 *)location)
 540                                 goto overflow;
 541                         break;
 542                 case R_X86_64_PC32:
 543                         value -= (u64)address;
 544                         *(u32 *)location = value;
 545                         break;
 546                 default:
 547                         pr_err("Unknown rela relocation: %llu\n",
 548                                ELF64_R_TYPE(rel[i].r_info));
 549                         return -ENOEXEC;
 550                 }
 551         }
 552         return 0;
 553
 554 overflow:
 555         pr_err("Overflow in relocation type %d value 0x%lx\n",
 556                (int)ELF64_R_TYPE(rel[i].r_info), value);
 557         return -ENOEXEC;
 558 }
 559 #endif /* CONFIG_KEXEC_FILE */
 560
 561 static int
 562 kexec_mark_range(unsigned long start, unsigned long end, bool protect)
 563 {
 564         struct page *page;
 565         unsigned int nr_pages;
 566
 567         /*
 568          * For physical range: [start, end]. We must skip the unassigned
 569          * crashk resource with zero-valued "end" member.
 570          */
 571         if (!end || start > end)
 572                 return 0;
 573
 574         page = pfn_to_page(start >> PAGE_SHIFT);
 575         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 576         if (protect)
 577                 return set_pages_ro(page, nr_pages);
 578         else
 579                 return set_pages_rw(page, nr_pages);
 580 }
 581
 582 static void kexec_mark_crashkres(bool protect)
 583 {
 584         unsigned long control;
 585
 586         kexec_mark_range(crashk_low_res.start, crashk_low_res.end, protect);
 587
 588         /* Don't touch the control code page used in crash_kexec().*/
 589         control = PFN_PHYS(page_to_pfn(kexec_crash_image->control_code_page));
 590         /* Control code page is located in the 2nd page. */
 591         kexec_mark_range(crashk_res.start, control + PAGE_SIZE - 1, protect);
 592         control += KEXEC_CONTROL_PAGE_SIZE;
 593         kexec_mark_range(control, crashk_res.end, protect);
 594 }
 595
 596 void arch_kexec_protect_crashkres(void)
 597 {
 598         kexec_mark_crashkres(true);
 599 }
 600
 601 void arch_kexec_unprotect_crashkres(void)
 602 {
 603         kexec_mark_crashkres(false);
 604 }