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 const struct kexec_file_ops * const kexec_file_loaders[] = {
  34                 &kexec_bzImage64_ops,
  35                 NULL
  36 };
  37 #endif
  38
  39 static void free_transition_pgtable(struct kimage *image)
  40 {
  41         free_page((unsigned long)image->arch.p4d);
  42         free_page((unsigned long)image->arch.pud);
  43         free_page((unsigned long)image->arch.pmd);
  44         free_page((unsigned long)image->arch.pte);
  45 }
  46
  47 static int init_transition_pgtable(struct kimage *image, pgd_t *pgd)
  48 {
  49         p4d_t *p4d;
  50         pud_t *pud;
  51         pmd_t *pmd;
  52         pte_t *pte;
  53         unsigned long vaddr, paddr;
  54         int result = -ENOMEM;
  55
  56         vaddr = (unsigned long)relocate_kernel;
  57         paddr = __pa(page_address(image->control_code_page)+PAGE_SIZE);
  58         pgd += pgd_index(vaddr);
  59         if (!pgd_present(*pgd)) {
  60                 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
  61                 if (!p4d)
  62                         goto err;
  63                 image->arch.p4d = p4d;
  64                 set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE));
  65         }
  66         p4d = p4d_offset(pgd, vaddr);
  67         if (!p4d_present(*p4d)) {
  68                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
  69                 if (!pud)
  70                         goto err;
  71                 image->arch.pud = pud;
  72                 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
  73         }
  74         pud = pud_offset(p4d, vaddr);
  75         if (!pud_present(*pud)) {
  76                 pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
  77                 if (!pmd)
  78                         goto err;
  79                 image->arch.pmd = pmd;
  80                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
  81         }
  82         pmd = pmd_offset(pud, vaddr);
  83         if (!pmd_present(*pmd)) {
  84                 pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
  85                 if (!pte)
  86                         goto err;
  87                 image->arch.pte = pte;
  88                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
  89         }
  90         pte = pte_offset_kernel(pmd, vaddr);
  91         set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL_EXEC_NOENC));
  92         return 0;
  93 err:
  94         free_transition_pgtable(image);
  95         return result;
  96 }
  97
  98 static void *alloc_pgt_page(void *data)
  99 {
 100         struct kimage *image = (struct kimage *)data;
 101         struct page *page;
 102         void *p = NULL;
 103
 104         page = kimage_alloc_control_pages(image, 0);
 105         if (page) {
 106                 p = page_address(page);
 107                 clear_page(p);
 108         }
 109
 110         return p;
 111 }
 112
 113 static int init_pgtable(struct kimage *image, unsigned long start_pgtable)
 114 {
 115         struct x86_mapping_info info = {
 116                 .alloc_pgt_page = alloc_pgt_page,
 117                 .context        = image,
 118                 .page_flag      = __PAGE_KERNEL_LARGE_EXEC,
 119                 .kernpg_flag    = _KERNPG_TABLE_NOENC,
 120         };
 121         unsigned long mstart, mend;
 122         pgd_t *level4p;
 123         int result;
 124         int i;
 125
 126         level4p = (pgd_t *)__va(start_pgtable);
 127         clear_page(level4p);
 128
 129         if (direct_gbpages)
 130                 info.direct_gbpages = true;
 131
 132         for (i = 0; i < nr_pfn_mapped; i++) {
 133                 mstart = pfn_mapped[i].start << PAGE_SHIFT;
 134                 mend   = pfn_mapped[i].end << PAGE_SHIFT;
 135
 136                 result = kernel_ident_mapping_init(&info,
 137                                                  level4p, mstart, mend);
 138                 if (result)
 139                         return result;
 140         }
 141
 142         /*
 143          * segments's mem ranges could be outside 0 ~ max_pfn,
 144          * for example when jump back to original kernel from kexeced kernel.
 145          * or first kernel is booted with user mem map, and second kernel
 146          * could be loaded out of that range.
 147          */
 148         for (i = 0; i < image->nr_segments; i++) {
 149                 mstart = image->segment[i].mem;
 150                 mend   = mstart + image->segment[i].memsz;
 151
 152                 result = kernel_ident_mapping_init(&info,
 153                                                  level4p, mstart, mend);
 154
 155                 if (result)
 156                         return result;
 157         }
 158
 159         return init_transition_pgtable(image, level4p);
 160 }
 161
 162 static void set_idt(void *newidt, u16 limit)
 163 {
 164         struct desc_ptr curidt;
 165
 166         /* x86-64 supports unaliged loads & stores */
 167         curidt.size    = limit;
 168         curidt.address = (unsigned long)newidt;
 169
 170         __asm__ __volatile__ (
 171                 "lidtq %0\n"
 172                 : : "m" (curidt)
 173                 );
 174 };
 175
 176
 177 static void set_gdt(void *newgdt, u16 limit)
 178 {
 179         struct desc_ptr curgdt;
 180
 181         /* x86-64 supports unaligned loads & stores */
 182         curgdt.size    = limit;
 183         curgdt.address = (unsigned long)newgdt;
 184
 185         __asm__ __volatile__ (
 186                 "lgdtq %0\n"
 187                 : : "m" (curgdt)
 188                 );
 189 };
 190
 191 static void load_segments(void)
 192 {
 193         __asm__ __volatile__ (
 194                 "\tmovl %0,%%ds\n"
 195                 "\tmovl %0,%%es\n"
 196                 "\tmovl %0,%%ss\n"
 197                 "\tmovl %0,%%fs\n"
 198                 "\tmovl %0,%%gs\n"
 199                 : : "a" (__KERNEL_DS) : "memory"
 200                 );
 201 }
 202
 203 #ifdef CONFIG_KEXEC_FILE
 204 /* Update purgatory as needed after various image segments have been prepared */
 205 static int arch_update_purgatory(struct kimage *image)
 206 {
 207         int ret = 0;
 208
 209         if (!image->file_mode)
 210                 return 0;
 211
 212         /* Setup copying of backup region */
 213         if (image->type == KEXEC_TYPE_CRASH) {
 214                 ret = kexec_purgatory_get_set_symbol(image,
 215                                 "purgatory_backup_dest",
 216                                 &image->arch.backup_load_addr,
 217                                 sizeof(image->arch.backup_load_addr), 0);
 218                 if (ret)
 219                         return ret;
 220
 221                 ret = kexec_purgatory_get_set_symbol(image,
 222                                 "purgatory_backup_src",
 223                                 &image->arch.backup_src_start,
 224                                 sizeof(image->arch.backup_src_start), 0);
 225                 if (ret)
 226                         return ret;
 227
 228                 ret = kexec_purgatory_get_set_symbol(image,
 229                                 "purgatory_backup_sz",
 230                                 &image->arch.backup_src_sz,
 231                                 sizeof(image->arch.backup_src_sz), 0);
 232                 if (ret)
 233                         return ret;
 234         }
 235
 236         return ret;
 237 }
 238 #else /* !CONFIG_KEXEC_FILE */
 239 static inline int arch_update_purgatory(struct kimage *image)
 240 {
 241         return 0;
 242 }
 243 #endif /* CONFIG_KEXEC_FILE */
 244
 245 int machine_kexec_prepare(struct kimage *image)
 246 {
 247         unsigned long start_pgtable;
 248         int result;
 249
 250         /* Calculate the offsets */
 251         start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 252
 253         /* Setup the identity mapped 64bit page table */
 254         result = init_pgtable(image, start_pgtable);
 255         if (result)
 256                 return result;
 257
 258         /* update purgatory as needed */
 259         result = arch_update_purgatory(image);
 260         if (result)
 261                 return result;
 262
 263         return 0;
 264 }
 265
 266 void machine_kexec_cleanup(struct kimage *image)
 267 {
 268         free_transition_pgtable(image);
 269 }
 270
 271 /*
 272  * Do not allocate memory (or fail in any way) in machine_kexec().
 273  * We are past the point of no return, committed to rebooting now.
 274  */
 275 void machine_kexec(struct kimage *image)
 276 {
 277         unsigned long page_list[PAGES_NR];
 278         void *control_page;
 279         int save_ftrace_enabled;
 280
 281 #ifdef CONFIG_KEXEC_JUMP
 282         if (image->preserve_context)
 283                 save_processor_state();
 284 #endif
 285
 286         save_ftrace_enabled = __ftrace_enabled_save();
 287
 288         /* Interrupts aren't acceptable while we reboot */
 289         local_irq_disable();
 290         hw_breakpoint_disable();
 291
 292         if (image->preserve_context) {
 293 #ifdef CONFIG_X86_IO_APIC
 294                 /*
 295                  * We need to put APICs in legacy mode so that we can
 296                  * get timer interrupts in second kernel. kexec/kdump
 297                  * paths already have calls to restore_boot_irq_mode()
 298                  * in one form or other. kexec jump path also need one.
 299                  */
 300                 clear_IO_APIC();
 301                 restore_boot_irq_mode();
 302 #endif
 303         }
 304
 305         control_page = page_address(image->control_code_page) + PAGE_SIZE;
 306         memcpy(control_page, relocate_kernel, KEXEC_CONTROL_CODE_MAX_SIZE);
 307
 308         page_list[PA_CONTROL_PAGE] = virt_to_phys(control_page);
 309         page_list[VA_CONTROL_PAGE] = (unsigned long)control_page;
 310         page_list[PA_TABLE_PAGE] =
 311           (unsigned long)__pa(page_address(image->control_code_page));
 312
 313         if (image->type == KEXEC_TYPE_DEFAULT)
 314                 page_list[PA_SWAP_PAGE] = (page_to_pfn(image->swap_page)
 315                                                 << PAGE_SHIFT);
 316
 317         /*
 318          * The segment registers are funny things, they have both a
 319          * visible and an invisible part.  Whenever the visible part is
 320          * set to a specific selector, the invisible part is loaded
 321          * with from a table in memory.  At no other time is the
 322          * descriptor table in memory accessed.
 323          *
 324          * I take advantage of this here by force loading the
 325          * segments, before I zap the gdt with an invalid value.
 326          */
 327         load_segments();
 328         /*
 329          * The gdt & idt are now invalid.
 330          * If you want to load them you must set up your own idt & gdt.
 331          */
 332         set_gdt(phys_to_virt(0), 0);
 333         set_idt(phys_to_virt(0), 0);
 334
 335         /* now call it */
 336         image->start = relocate_kernel((unsigned long)image->head,
 337                                        (unsigned long)page_list,
 338                                        image->start,
 339                                        image->preserve_context,
 340                                        sme_active());
 341
 342 #ifdef CONFIG_KEXEC_JUMP
 343         if (image->preserve_context)
 344                 restore_processor_state();
 345 #endif
 346
 347         __ftrace_enabled_restore(save_ftrace_enabled);
 348 }
 349
 350 void arch_crash_save_vmcoreinfo(void)
 351 {
 352         VMCOREINFO_NUMBER(phys_base);
 353         VMCOREINFO_SYMBOL(init_top_pgt);
 354         VMCOREINFO_NUMBER(pgtable_l5_enabled);
 355
 356 #ifdef CONFIG_NUMA
 357         VMCOREINFO_SYMBOL(node_data);
 358         VMCOREINFO_LENGTH(node_data, MAX_NUMNODES);
 359 #endif
 360         vmcoreinfo_append_str("KERNELOFFSET=%lx\n",
 361                               kaslr_offset());
 362         VMCOREINFO_NUMBER(KERNEL_IMAGE_SIZE);
 363 }
 364
 365 /* arch-dependent functionality related to kexec file-based syscall */
 366
 367 #ifdef CONFIG_KEXEC_FILE
 368 void *arch_kexec_kernel_image_load(struct kimage *image)
 369 {
 370         vfree(image->arch.elf_headers);
 371         image->arch.elf_headers = NULL;
 372
 373         if (!image->fops || !image->fops->load)
 374                 return ERR_PTR(-ENOEXEC);
 375
 376         return image->fops->load(image, image->kernel_buf,
 377                                  image->kernel_buf_len, image->initrd_buf,
 378                                  image->initrd_buf_len, image->cmdline_buf,
 379                                  image->cmdline_buf_len);
 380 }
 381
 382 /*
 383  * Apply purgatory relocations.
 384  *
 385  * ehdr: Pointer to elf headers
 386  * sechdrs: Pointer to section headers.
 387  * relsec: section index of SHT_RELA section.
 388  *
 389  * TODO: Some of the code belongs to generic code. Move that in kexec.c.
 390  */
 391 int arch_kexec_apply_relocations_add(const Elf64_Ehdr *ehdr,
 392                                      Elf64_Shdr *sechdrs, unsigned int relsec)
 393 {
 394         unsigned int i;
 395         Elf64_Rela *rel;
 396         Elf64_Sym *sym;
 397         void *location;
 398         Elf64_Shdr *section, *symtabsec;
 399         unsigned long address, sec_base, value;
 400         const char *strtab, *name, *shstrtab;
 401
 402         /*
 403          * ->sh_offset has been modified to keep the pointer to section
 404          * contents in memory
 405          */
 406         rel = (void *)sechdrs[relsec].sh_offset;
 407
 408         /* Section to which relocations apply */
 409         section = &sechdrs[sechdrs[relsec].sh_info];
 410
 411         pr_debug("Applying relocate section %u to %u\n", relsec,
 412                  sechdrs[relsec].sh_info);
 413
 414         /* Associated symbol table */
 415         symtabsec = &sechdrs[sechdrs[relsec].sh_link];
 416
 417         /* String table */
 418         if (symtabsec->sh_link >= ehdr->e_shnum) {
 419                 /* Invalid strtab section number */
 420                 pr_err("Invalid string table section index %d\n",
 421                        symtabsec->sh_link);
 422                 return -ENOEXEC;
 423         }
 424
 425         strtab = (char *)sechdrs[symtabsec->sh_link].sh_offset;
 426
 427         /* section header string table */
 428         shstrtab = (char *)sechdrs[ehdr->e_shstrndx].sh_offset;
 429
 430         for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rel); i++) {
 431
 432                 /*
 433                  * rel[i].r_offset contains byte offset from beginning
 434                  * of section to the storage unit affected.
 435                  *
 436                  * This is location to update (->sh_offset). This is temporary
 437                  * buffer where section is currently loaded. This will finally
 438                  * be loaded to a different address later, pointed to by
 439                  * ->sh_addr. kexec takes care of moving it
 440                  *  (kexec_load_segment()).
 441                  */
 442                 location = (void *)(section->sh_offset + rel[i].r_offset);
 443
 444                 /* Final address of the location */
 445                 address = section->sh_addr + rel[i].r_offset;
 446
 447                 /*
 448                  * rel[i].r_info contains information about symbol table index
 449                  * w.r.t which relocation must be made and type of relocation
 450                  * to apply. ELF64_R_SYM() and ELF64_R_TYPE() macros get
 451                  * these respectively.
 452                  */
 453                 sym = (Elf64_Sym *)symtabsec->sh_offset +
 454                                 ELF64_R_SYM(rel[i].r_info);
 455
 456                 if (sym->st_name)
 457                         name = strtab + sym->st_name;
 458                 else
 459                         name = shstrtab + sechdrs[sym->st_shndx].sh_name;
 460
 461                 pr_debug("Symbol: %s info: %02x shndx: %02x value=%llx size: %llx\n",
 462                          name, sym->st_info, sym->st_shndx, sym->st_value,
 463                          sym->st_size);
 464
 465                 if (sym->st_shndx == SHN_UNDEF) {
 466                         pr_err("Undefined symbol: %s\n", name);
 467                         return -ENOEXEC;
 468                 }
 469
 470                 if (sym->st_shndx == SHN_COMMON) {
 471                         pr_err("symbol '%s' in common section\n", name);
 472                         return -ENOEXEC;
 473                 }
 474
 475                 if (sym->st_shndx == SHN_ABS)
 476                         sec_base = 0;
 477                 else if (sym->st_shndx >= ehdr->e_shnum) {
 478                         pr_err("Invalid section %d for symbol %s\n",
 479                                sym->st_shndx, name);
 480                         return -ENOEXEC;
 481                 } else
 482                         sec_base = sechdrs[sym->st_shndx].sh_addr;
 483
 484                 value = sym->st_value;
 485                 value += sec_base;
 486                 value += rel[i].r_addend;
 487
 488                 switch (ELF64_R_TYPE(rel[i].r_info)) {
 489                 case R_X86_64_NONE:
 490                         break;
 491                 case R_X86_64_64:
 492                         *(u64 *)location = value;
 493                         break;
 494                 case R_X86_64_32:
 495                         *(u32 *)location = value;
 496                         if (value != *(u32 *)location)
 497                                 goto overflow;
 498                         break;
 499                 case R_X86_64_32S:
 500                         *(s32 *)location = value;
 501                         if ((s64)value != *(s32 *)location)
 502                                 goto overflow;
 503                         break;
 504                 case R_X86_64_PC32:
 505                 case R_X86_64_PLT32:
 506                         value -= (u64)address;
 507                         *(u32 *)location = value;
 508                         break;
 509                 default:
 510                         pr_err("Unknown rela relocation: %llu\n",
 511                                ELF64_R_TYPE(rel[i].r_info));
 512                         return -ENOEXEC;
 513                 }
 514         }
 515         return 0;
 516
 517 overflow:
 518         pr_err("Overflow in relocation type %d value 0x%lx\n",
 519                (int)ELF64_R_TYPE(rel[i].r_info), value);
 520         return -ENOEXEC;
 521 }
 522 #endif /* CONFIG_KEXEC_FILE */
 523
 524 static int
 525 kexec_mark_range(unsigned long start, unsigned long end, bool protect)
 526 {
 527         struct page *page;
 528         unsigned int nr_pages;
 529
 530         /*
 531          * For physical range: [start, end]. We must skip the unassigned
 532          * crashk resource with zero-valued "end" member.
 533          */
 534         if (!end || start > end)
 535                 return 0;
 536
 537         page = pfn_to_page(start >> PAGE_SHIFT);
 538         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 539         if (protect)
 540                 return set_pages_ro(page, nr_pages);
 541         else
 542                 return set_pages_rw(page, nr_pages);
 543 }
 544
 545 static void kexec_mark_crashkres(bool protect)
 546 {
 547         unsigned long control;
 548
 549         kexec_mark_range(crashk_low_res.start, crashk_low_res.end, protect);
 550
 551         /* Don't touch the control code page used in crash_kexec().*/
 552         control = PFN_PHYS(page_to_pfn(kexec_crash_image->control_code_page));
 553         /* Control code page is located in the 2nd page. */
 554         kexec_mark_range(crashk_res.start, control + PAGE_SIZE - 1, protect);
 555         control += KEXEC_CONTROL_PAGE_SIZE;
 556         kexec_mark_range(control, crashk_res.end, protect);
 557 }
 558
 559 void arch_kexec_protect_crashkres(void)
 560 {
 561         kexec_mark_crashkres(true);
 562 }
 563
 564 void arch_kexec_unprotect_crashkres(void)
 565 {
 566         kexec_mark_crashkres(false);
 567 }
 568
 569 int arch_kexec_post_alloc_pages(void *vaddr, unsigned int pages, gfp_t gfp)
 570 {
 571         /*
 572          * If SME is active we need to be sure that kexec pages are
 573          * not encrypted because when we boot to the new kernel the
 574          * pages won't be accessed encrypted (initially).
 575          */
 576         return set_memory_decrypted((unsigned long)vaddr, pages);
 577 }
 578
 579 void arch_kexec_pre_free_pages(void *vaddr, unsigned int pages)
 580 {
 581         /*
 582          * If SME is active we need to reset the pages back to being
 583          * an encrypted mapping before freeing them.
 584          */
 585         set_memory_encrypted((unsigned long)vaddr, pages);
 586 }