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         image->arch.p4d = NULL;
  42         free_page((unsigned long)image->arch.pud);
  43         image->arch.pud = NULL;
  44         free_page((unsigned long)image->arch.pmd);
  45         image->arch.pmd = NULL;
  46         free_page((unsigned long)image->arch.pte);
  47         image->arch.pte = NULL;
  48 }
  49
  50 static int init_transition_pgtable(struct kimage *image, pgd_t *pgd)
  51 {
  52         p4d_t *p4d;
  53         pud_t *pud;
  54         pmd_t *pmd;
  55         pte_t *pte;
  56         unsigned long vaddr, paddr;
  57         int result = -ENOMEM;
  58
  59         vaddr = (unsigned long)relocate_kernel;
  60         paddr = __pa(page_address(image->control_code_page)+PAGE_SIZE);
  61         pgd += pgd_index(vaddr);
  62         if (!pgd_present(*pgd)) {
  63                 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
  64                 if (!p4d)
  65                         goto err;
  66                 image->arch.p4d = p4d;
  67                 set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE));
  68         }
  69         p4d = p4d_offset(pgd, vaddr);
  70         if (!p4d_present(*p4d)) {
  71                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
  72                 if (!pud)
  73                         goto err;
  74                 image->arch.pud = pud;
  75                 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
  76         }
  77         pud = pud_offset(p4d, vaddr);
  78         if (!pud_present(*pud)) {
  79                 pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
  80                 if (!pmd)
  81                         goto err;
  82                 image->arch.pmd = pmd;
  83                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
  84         }
  85         pmd = pmd_offset(pud, vaddr);
  86         if (!pmd_present(*pmd)) {
  87                 pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
  88                 if (!pte)
  89                         goto err;
  90                 image->arch.pte = pte;
  91                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
  92         }
  93         pte = pte_offset_kernel(pmd, vaddr);
  94         set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL_EXEC_NOENC));
  95         return 0;
  96 err:
  97         return result;
  98 }
  99
 100 static void *alloc_pgt_page(void *data)
 101 {
 102         struct kimage *image = (struct kimage *)data;
 103         struct page *page;
 104         void *p = NULL;
 105
 106         page = kimage_alloc_control_pages(image, 0);
 107         if (page) {
 108                 p = page_address(page);
 109                 clear_page(p);
 110         }
 111
 112         return p;
 113 }
 114
 115 static int init_pgtable(struct kimage *image, unsigned long start_pgtable)
 116 {
 117         struct x86_mapping_info info = {
 118                 .alloc_pgt_page = alloc_pgt_page,
 119                 .context        = image,
 120                 .page_flag      = __PAGE_KERNEL_LARGE_EXEC,
 121                 .kernpg_flag    = _KERNPG_TABLE_NOENC,
 122         };
 123         unsigned long mstart, mend;
 124         pgd_t *level4p;
 125         int result;
 126         int i;
 127
 128         level4p = (pgd_t *)__va(start_pgtable);
 129         clear_page(level4p);
 130
 131         if (direct_gbpages)
 132                 info.direct_gbpages = true;
 133
 134         for (i = 0; i < nr_pfn_mapped; i++) {
 135                 mstart = pfn_mapped[i].start << PAGE_SHIFT;
 136                 mend   = pfn_mapped[i].end << PAGE_SHIFT;
 137
 138                 result = kernel_ident_mapping_init(&info,
 139                                                  level4p, mstart, mend);
 140                 if (result)
 141                         return result;
 142         }
 143
 144         /*
 145          * segments's mem ranges could be outside 0 ~ max_pfn,
 146          * for example when jump back to original kernel from kexeced kernel.
 147          * or first kernel is booted with user mem map, and second kernel
 148          * could be loaded out of that range.
 149          */
 150         for (i = 0; i < image->nr_segments; i++) {
 151                 mstart = image->segment[i].mem;
 152                 mend   = mstart + image->segment[i].memsz;
 153
 154                 result = kernel_ident_mapping_init(&info,
 155                                                  level4p, mstart, mend);
 156
 157                 if (result)
 158                         return result;
 159         }
 160
 161         return init_transition_pgtable(image, level4p);
 162 }
 163
 164 static void set_idt(void *newidt, u16 limit)
 165 {
 166         struct desc_ptr curidt;
 167
 168         /* x86-64 supports unaliged loads & stores */
 169         curidt.size    = limit;
 170         curidt.address = (unsigned long)newidt;
 171
 172         __asm__ __volatile__ (
 173                 "lidtq %0\n"
 174                 : : "m" (curidt)
 175                 );
 176 };
 177
 178
 179 static void set_gdt(void *newgdt, u16 limit)
 180 {
 181         struct desc_ptr curgdt;
 182
 183         /* x86-64 supports unaligned loads & stores */
 184         curgdt.size    = limit;
 185         curgdt.address = (unsigned long)newgdt;
 186
 187         __asm__ __volatile__ (
 188                 "lgdtq %0\n"
 189                 : : "m" (curgdt)
 190                 );
 191 };
 192
 193 static void load_segments(void)
 194 {
 195         __asm__ __volatile__ (
 196                 "\tmovl %0,%%ds\n"
 197                 "\tmovl %0,%%es\n"
 198                 "\tmovl %0,%%ss\n"
 199                 "\tmovl %0,%%fs\n"
 200                 "\tmovl %0,%%gs\n"
 201                 : : "a" (__KERNEL_DS) : "memory"
 202                 );
 203 }
 204
 205 #ifdef CONFIG_KEXEC_FILE
 206 /* Update purgatory as needed after various image segments have been prepared */
 207 static int arch_update_purgatory(struct kimage *image)
 208 {
 209         int ret = 0;
 210
 211         if (!image->file_mode)
 212                 return 0;
 213
 214         /* Setup copying of backup region */
 215         if (image->type == KEXEC_TYPE_CRASH) {
 216                 ret = kexec_purgatory_get_set_symbol(image,
 217                                 "purgatory_backup_dest",
 218                                 &image->arch.backup_load_addr,
 219                                 sizeof(image->arch.backup_load_addr), 0);
 220                 if (ret)
 221                         return ret;
 222
 223                 ret = kexec_purgatory_get_set_symbol(image,
 224                                 "purgatory_backup_src",
 225                                 &image->arch.backup_src_start,
 226                                 sizeof(image->arch.backup_src_start), 0);
 227                 if (ret)
 228                         return ret;
 229
 230                 ret = kexec_purgatory_get_set_symbol(image,
 231                                 "purgatory_backup_sz",
 232                                 &image->arch.backup_src_sz,
 233                                 sizeof(image->arch.backup_src_sz), 0);
 234                 if (ret)
 235                         return ret;
 236         }
 237
 238         return ret;
 239 }
 240 #else /* !CONFIG_KEXEC_FILE */
 241 static inline int arch_update_purgatory(struct kimage *image)
 242 {
 243         return 0;
 244 }
 245 #endif /* CONFIG_KEXEC_FILE */
 246
 247 int machine_kexec_prepare(struct kimage *image)
 248 {
 249         unsigned long start_pgtable;
 250         int result;
 251
 252         /* Calculate the offsets */
 253         start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 254
 255         /* Setup the identity mapped 64bit page table */
 256         result = init_pgtable(image, start_pgtable);
 257         if (result)
 258                 return result;
 259
 260         /* update purgatory as needed */
 261         result = arch_update_purgatory(image);
 262         if (result)
 263                 return result;
 264
 265         return 0;
 266 }
 267
 268 void machine_kexec_cleanup(struct kimage *image)
 269 {
 270         free_transition_pgtable(image);
 271 }
 272
 273 /*
 274  * Do not allocate memory (or fail in any way) in machine_kexec().
 275  * We are past the point of no return, committed to rebooting now.
 276  */
 277 void machine_kexec(struct kimage *image)
 278 {
 279         unsigned long page_list[PAGES_NR];
 280         void *control_page;
 281         int save_ftrace_enabled;
 282
 283 #ifdef CONFIG_KEXEC_JUMP
 284         if (image->preserve_context)
 285                 save_processor_state();
 286 #endif
 287
 288         save_ftrace_enabled = __ftrace_enabled_save();
 289
 290         /* Interrupts aren't acceptable while we reboot */
 291         local_irq_disable();
 292         hw_breakpoint_disable();
 293
 294         if (image->preserve_context) {
 295 #ifdef CONFIG_X86_IO_APIC
 296                 /*
 297                  * We need to put APICs in legacy mode so that we can
 298                  * get timer interrupts in second kernel. kexec/kdump
 299                  * paths already have calls to disable_IO_APIC() in
 300                  * one form or other. kexec jump path also need
 301                  * one.
 302                  */
 303                 disable_IO_APIC();
 304 #endif
 305         }
 306
 307         control_page = page_address(image->control_code_page) + PAGE_SIZE;
 308         memcpy(control_page, relocate_kernel, KEXEC_CONTROL_CODE_MAX_SIZE);
 309
 310         page_list[PA_CONTROL_PAGE] = virt_to_phys(control_page);
 311         page_list[VA_CONTROL_PAGE] = (unsigned long)control_page;
 312         page_list[PA_TABLE_PAGE] =
 313           (unsigned long)__pa(page_address(image->control_code_page));
 314
 315         if (image->type == KEXEC_TYPE_DEFAULT)
 316                 page_list[PA_SWAP_PAGE] = (page_to_pfn(image->swap_page)
 317                                                 << PAGE_SHIFT);
 318
 319         /*
 320          * The segment registers are funny things, they have both a
 321          * visible and an invisible part.  Whenever the visible part is
 322          * set to a specific selector, the invisible part is loaded
 323          * with from a table in memory.  At no other time is the
 324          * descriptor table in memory accessed.
 325          *
 326          * I take advantage of this here by force loading the
 327          * segments, before I zap the gdt with an invalid value.
 328          */
 329         load_segments();
 330         /*
 331          * The gdt & idt are now invalid.
 332          * If you want to load them you must set up your own idt & gdt.
 333          */
 334         set_gdt(phys_to_virt(0), 0);
 335         set_idt(phys_to_virt(0), 0);
 336
 337         /* now call it */
 338         image->start = relocate_kernel((unsigned long)image->head,
 339                                        (unsigned long)page_list,
 340                                        image->start,
 341                                        image->preserve_context,
 342                                        sme_active());
 343
 344 #ifdef CONFIG_KEXEC_JUMP
 345         if (image->preserve_context)
 346                 restore_processor_state();
 347 #endif
 348
 349         __ftrace_enabled_restore(save_ftrace_enabled);
 350 }
 351
 352 void arch_crash_save_vmcoreinfo(void)
 353 {
 354         VMCOREINFO_NUMBER(phys_base);
 355         VMCOREINFO_SYMBOL(init_top_pgt);
 356
 357 #ifdef CONFIG_NUMA
 358         VMCOREINFO_SYMBOL(node_data);
 359         VMCOREINFO_LENGTH(node_data, MAX_NUMNODES);
 360 #endif
 361         vmcoreinfo_append_str("KERNELOFFSET=%lx\n",
 362                               kaslr_offset());
 363         VMCOREINFO_NUMBER(KERNEL_IMAGE_SIZE);
 364 }
 365
 366 /* arch-dependent functionality related to kexec file-based syscall */
 367
 368 #ifdef CONFIG_KEXEC_FILE
 369 int arch_kexec_kernel_image_probe(struct kimage *image, void *buf,
 370                                   unsigned long buf_len)
 371 {
 372         int i, ret = -ENOEXEC;
 373         struct kexec_file_ops *fops;
 374
 375         for (i = 0; i < ARRAY_SIZE(kexec_file_loaders); i++) {
 376                 fops = kexec_file_loaders[i];
 377                 if (!fops || !fops->probe)
 378                         continue;
 379
 380                 ret = fops->probe(buf, buf_len);
 381                 if (!ret) {
 382                         image->fops = fops;
 383                         return ret;
 384                 }
 385         }
 386
 387         return ret;
 388 }
 389
 390 void *arch_kexec_kernel_image_load(struct kimage *image)
 391 {
 392         vfree(image->arch.elf_headers);
 393         image->arch.elf_headers = NULL;
 394
 395         if (!image->fops || !image->fops->load)
 396                 return ERR_PTR(-ENOEXEC);
 397
 398         return image->fops->load(image, image->kernel_buf,
 399                                  image->kernel_buf_len, image->initrd_buf,
 400                                  image->initrd_buf_len, image->cmdline_buf,
 401                                  image->cmdline_buf_len);
 402 }
 403
 404 int arch_kimage_file_post_load_cleanup(struct kimage *image)
 405 {
 406         if (!image->fops || !image->fops->cleanup)
 407                 return 0;
 408
 409         return image->fops->cleanup(image->image_loader_data);
 410 }
 411
 412 #ifdef CONFIG_KEXEC_VERIFY_SIG
 413 int arch_kexec_kernel_verify_sig(struct kimage *image, void *kernel,
 414                                  unsigned long kernel_len)
 415 {
 416         if (!image->fops || !image->fops->verify_sig) {
 417                 pr_debug("kernel loader does not support signature verification.");
 418                 return -EKEYREJECTED;
 419         }
 420
 421         return image->fops->verify_sig(kernel, kernel_len);
 422 }
 423 #endif
 424
 425 /*
 426  * Apply purgatory relocations.
 427  *
 428  * ehdr: Pointer to elf headers
 429  * sechdrs: Pointer to section headers.
 430  * relsec: section index of SHT_RELA section.
 431  *
 432  * TODO: Some of the code belongs to generic code. Move that in kexec.c.
 433  */
 434 int arch_kexec_apply_relocations_add(const Elf64_Ehdr *ehdr,
 435                                      Elf64_Shdr *sechdrs, unsigned int relsec)
 436 {
 437         unsigned int i;
 438         Elf64_Rela *rel;
 439         Elf64_Sym *sym;
 440         void *location;
 441         Elf64_Shdr *section, *symtabsec;
 442         unsigned long address, sec_base, value;
 443         const char *strtab, *name, *shstrtab;
 444
 445         /*
 446          * ->sh_offset has been modified to keep the pointer to section
 447          * contents in memory
 448          */
 449         rel = (void *)sechdrs[relsec].sh_offset;
 450
 451         /* Section to which relocations apply */
 452         section = &sechdrs[sechdrs[relsec].sh_info];
 453
 454         pr_debug("Applying relocate section %u to %u\n", relsec,
 455                  sechdrs[relsec].sh_info);
 456
 457         /* Associated symbol table */
 458         symtabsec = &sechdrs[sechdrs[relsec].sh_link];
 459
 460         /* String table */
 461         if (symtabsec->sh_link >= ehdr->e_shnum) {
 462                 /* Invalid strtab section number */
 463                 pr_err("Invalid string table section index %d\n",
 464                        symtabsec->sh_link);
 465                 return -ENOEXEC;
 466         }
 467
 468         strtab = (char *)sechdrs[symtabsec->sh_link].sh_offset;
 469
 470         /* section header string table */
 471         shstrtab = (char *)sechdrs[ehdr->e_shstrndx].sh_offset;
 472
 473         for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rel); i++) {
 474
 475                 /*
 476                  * rel[i].r_offset contains byte offset from beginning
 477                  * of section to the storage unit affected.
 478                  *
 479                  * This is location to update (->sh_offset). This is temporary
 480                  * buffer where section is currently loaded. This will finally
 481                  * be loaded to a different address later, pointed to by
 482                  * ->sh_addr. kexec takes care of moving it
 483                  *  (kexec_load_segment()).
 484                  */
 485                 location = (void *)(section->sh_offset + rel[i].r_offset);
 486
 487                 /* Final address of the location */
 488                 address = section->sh_addr + rel[i].r_offset;
 489
 490                 /*
 491                  * rel[i].r_info contains information about symbol table index
 492                  * w.r.t which relocation must be made and type of relocation
 493                  * to apply. ELF64_R_SYM() and ELF64_R_TYPE() macros get
 494                  * these respectively.
 495                  */
 496                 sym = (Elf64_Sym *)symtabsec->sh_offset +
 497                                 ELF64_R_SYM(rel[i].r_info);
 498
 499                 if (sym->st_name)
 500                         name = strtab + sym->st_name;
 501                 else
 502                         name = shstrtab + sechdrs[sym->st_shndx].sh_name;
 503
 504                 pr_debug("Symbol: %s info: %02x shndx: %02x value=%llx size: %llx\n",
 505                          name, sym->st_info, sym->st_shndx, sym->st_value,
 506                          sym->st_size);
 507
 508                 if (sym->st_shndx == SHN_UNDEF) {
 509                         pr_err("Undefined symbol: %s\n", name);
 510                         return -ENOEXEC;
 511                 }
 512
 513                 if (sym->st_shndx == SHN_COMMON) {
 514                         pr_err("symbol '%s' in common section\n", name);
 515                         return -ENOEXEC;
 516                 }
 517
 518                 if (sym->st_shndx == SHN_ABS)
 519                         sec_base = 0;
 520                 else if (sym->st_shndx >= ehdr->e_shnum) {
 521                         pr_err("Invalid section %d for symbol %s\n",
 522                                sym->st_shndx, name);
 523                         return -ENOEXEC;
 524                 } else
 525                         sec_base = sechdrs[sym->st_shndx].sh_addr;
 526
 527                 value = sym->st_value;
 528                 value += sec_base;
 529                 value += rel[i].r_addend;
 530
 531                 switch (ELF64_R_TYPE(rel[i].r_info)) {
 532                 case R_X86_64_NONE:
 533                         break;
 534                 case R_X86_64_64:
 535                         *(u64 *)location = value;
 536                         break;
 537                 case R_X86_64_32:
 538                         *(u32 *)location = value;
 539                         if (value != *(u32 *)location)
 540                                 goto overflow;
 541                         break;
 542                 case R_X86_64_32S:
 543                         *(s32 *)location = value;
 544                         if ((s64)value != *(s32 *)location)
 545                                 goto overflow;
 546                         break;
 547                 case R_X86_64_PC32:
 548                 case R_X86_64_PLT32:
 549                         value -= (u64)address;
 550                         *(u32 *)location = value;
 551                         break;
 552                 default:
 553                         pr_err("Unknown rela relocation: %llu\n",
 554                                ELF64_R_TYPE(rel[i].r_info));
 555                         return -ENOEXEC;
 556                 }
 557         }
 558         return 0;
 559
 560 overflow:
 561         pr_err("Overflow in relocation type %d value 0x%lx\n",
 562                (int)ELF64_R_TYPE(rel[i].r_info), value);
 563         return -ENOEXEC;
 564 }
 565 #endif /* CONFIG_KEXEC_FILE */
 566
 567 static int
 568 kexec_mark_range(unsigned long start, unsigned long end, bool protect)
 569 {
 570         struct page *page;
 571         unsigned int nr_pages;
 572
 573         /*
 574          * For physical range: [start, end]. We must skip the unassigned
 575          * crashk resource with zero-valued "end" member.
 576          */
 577         if (!end || start > end)
 578                 return 0;
 579
 580         page = pfn_to_page(start >> PAGE_SHIFT);
 581         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 582         if (protect)
 583                 return set_pages_ro(page, nr_pages);
 584         else
 585                 return set_pages_rw(page, nr_pages);
 586 }
 587
 588 static void kexec_mark_crashkres(bool protect)
 589 {
 590         unsigned long control;
 591
 592         kexec_mark_range(crashk_low_res.start, crashk_low_res.end, protect);
 593
 594         /* Don't touch the control code page used in crash_kexec().*/
 595         control = PFN_PHYS(page_to_pfn(kexec_crash_image->control_code_page));
 596         /* Control code page is located in the 2nd page. */
 597         kexec_mark_range(crashk_res.start, control + PAGE_SIZE - 1, protect);
 598         control += KEXEC_CONTROL_PAGE_SIZE;
 599         kexec_mark_range(control, crashk_res.end, protect);
 600 }
 601
 602 void arch_kexec_protect_crashkres(void)
 603 {
 604         kexec_mark_crashkres(true);
 605 }
 606
 607 void arch_kexec_unprotect_crashkres(void)
 608 {
 609         kexec_mark_crashkres(false);
 610 }
 611
 612 int arch_kexec_post_alloc_pages(void *vaddr, unsigned int pages, gfp_t gfp)
 613 {
 614         /*
 615          * If SME is active we need to be sure that kexec pages are
 616          * not encrypted because when we boot to the new kernel the
 617          * pages won't be accessed encrypted (initially).
 618          */
 619         return set_memory_decrypted((unsigned long)vaddr, pages);
 620 }
 621
 622 void arch_kexec_pre_free_pages(void *vaddr, unsigned int pages)
 623 {
 624         /*
 625          * If SME is active we need to reset the pages back to being
 626          * an encrypted mapping before freeing them.
 627          */
 628         set_memory_encrypted((unsigned long)vaddr, pages);
 629 }