kernel/arch/arm64/mm/mmu.c

   1 /*
   2  * Based on arch/arm/mm/mmu.c
   3  *
   4  * Copyright (C) 1995-2005 Russell King
   5  * Copyright (C) 2012 ARM Ltd.
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
  18  */
  19
  20 #include <linux/export.h>
  21 #include <linux/kernel.h>
  22 #include <linux/errno.h>
  23 #include <linux/init.h>
  24 #include <linux/libfdt.h>
  25 #include <linux/mman.h>
  26 #include <linux/nodemask.h>
  27 #include <linux/memblock.h>
  28 #include <linux/fs.h>
  29 #include <linux/io.h>
  30 #include <linux/slab.h>
  31 #include <linux/stop_machine.h>
  32
  33 #include <asm/cputype.h>
  34 #include <asm/fixmap.h>
  35 #include <asm/kernel-pgtable.h>
  36 #include <asm/sections.h>
  37 #include <asm/setup.h>
  38 #include <asm/sizes.h>
  39 #include <asm/tlb.h>
  40 #include <asm/memblock.h>
  41 #include <asm/mmu_context.h>
  42
  43 #include "mm.h"
  44
  45 u64 idmap_t0sz = TCR_T0SZ(VA_BITS);
  46
  47 /*
  48  * Empty_zero_page is a special page that is used for zero-initialized data
  49  * and COW.
  50  */
  51 struct page *empty_zero_page;
  52 EXPORT_SYMBOL(empty_zero_page);
  53
  54 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
  55                               unsigned long size, pgprot_t vma_prot)
  56 {
  57         if (!pfn_valid(pfn))
  58                 return pgprot_noncached(vma_prot);
  59         else if (file->f_flags & O_SYNC)
  60                 return pgprot_writecombine(vma_prot);
  61         return vma_prot;
  62 }
  63 EXPORT_SYMBOL(phys_mem_access_prot);
  64
  65 static void __init *early_alloc(unsigned long sz)
  66 {
  67         phys_addr_t phys;
  68         void *ptr;
  69
  70         phys = memblock_alloc(sz, sz);
  71         BUG_ON(!phys);
  72         ptr = __va(phys);
  73         memset(ptr, 0, sz);
  74         return ptr;
  75 }
  76
  77 /*
  78  * remap a PMD into pages
  79  */
  80 static void split_pmd(pmd_t *pmd, pte_t *pte)
  81 {
  82         unsigned long pfn = pmd_pfn(*pmd);
  83         int i = 0;
  84
  85         do {
  86                 /*
  87                  * Need to have the least restrictive permissions available
  88                  * permissions will be fixed up later
  89                  */
  90                 set_pte(pte, pfn_pte(pfn, PAGE_KERNEL_EXEC));
  91                 pfn++;
  92         } while (pte++, i++, i < PTRS_PER_PTE);
  93 }
  94
  95 static void alloc_init_pte(pmd_t *pmd, unsigned long addr,
  96                                   unsigned long end, unsigned long pfn,
  97                                   pgprot_t prot,
  98                                   void *(*alloc)(unsigned long size))
  99 {
 100         pte_t *pte;
 101
 102         if (pmd_none(*pmd) || pmd_sect(*pmd)) {
 103                 pte = alloc(PTRS_PER_PTE * sizeof(pte_t));
 104                 if (pmd_sect(*pmd))
 105                         split_pmd(pmd, pte);
 106                 __pmd_populate(pmd, __pa(pte), PMD_TYPE_TABLE);
 107                 flush_tlb_all();
 108         }
 109         BUG_ON(pmd_bad(*pmd));
 110
 111         pte = pte_offset_kernel(pmd, addr);
 112         do {
 113                 set_pte(pte, pfn_pte(pfn, prot));
 114                 pfn++;
 115         } while (pte++, addr += PAGE_SIZE, addr != end);
 116 }
 117
 118 static void split_pud(pud_t *old_pud, pmd_t *pmd)
 119 {
 120         unsigned long addr = pud_pfn(*old_pud) << PAGE_SHIFT;
 121         pgprot_t prot = __pgprot(pud_val(*old_pud) ^ addr);
 122         int i = 0;
 123
 124         do {
 125                 set_pmd(pmd, __pmd(addr | pgprot_val(prot)));
 126                 addr += PMD_SIZE;
 127         } while (pmd++, i++, i < PTRS_PER_PMD);
 128 }
 129
 130 static void alloc_init_pmd(struct mm_struct *mm, pud_t *pud,
 131                                   unsigned long addr, unsigned long end,
 132                                   phys_addr_t phys, pgprot_t prot,
 133                                   void *(*alloc)(unsigned long size))
 134 {
 135         pmd_t *pmd;
 136         unsigned long next;
 137
 138         /*
 139          * Check for initial section mappings in the pgd/pud and remove them.
 140          */
 141         if (pud_none(*pud) || pud_sect(*pud)) {
 142                 pmd = alloc(PTRS_PER_PMD * sizeof(pmd_t));
 143                 if (pud_sect(*pud)) {
 144                         /*
 145                          * need to have the 1G of mappings continue to be
 146                          * present
 147                          */
 148                         split_pud(pud, pmd);
 149                 }
 150                 pud_populate(mm, pud, pmd);
 151                 flush_tlb_all();
 152         }
 153         BUG_ON(pud_bad(*pud));
 154
 155         pmd = pmd_offset(pud, addr);
 156         do {
 157                 next = pmd_addr_end(addr, end);
 158                 /* try section mapping first */
 159                 if (((addr | next | phys) & ~SECTION_MASK) == 0) {
 160                         pmd_t old_pmd =*pmd;
 161                         set_pmd(pmd, __pmd(phys |
 162                                            pgprot_val(mk_sect_prot(prot))));
 163                         /*
 164                          * Check for previous table entries created during
 165                          * boot (__create_page_tables) and flush them.
 166                          */
 167                         if (!pmd_none(old_pmd)) {
 168                                 flush_tlb_all();
 169                                 if (pmd_table(old_pmd)) {
 170                                         phys_addr_t table = __pa(pte_offset_map(&old_pmd, 0));
 171                                         if (!WARN_ON_ONCE(slab_is_available()))
 172                                                 memblock_free(table, PAGE_SIZE);
 173                                 }
 174                         }
 175                 } else {
 176                         alloc_init_pte(pmd, addr, next, __phys_to_pfn(phys),
 177                                        prot, alloc);
 178                 }
 179                 phys += next - addr;
 180         } while (pmd++, addr = next, addr != end);
 181 }
 182
 183 static inline bool use_1G_block(unsigned long addr, unsigned long next,
 184                         unsigned long phys)
 185 {
 186         if (PAGE_SHIFT != 12)
 187                 return false;
 188
 189         if (((addr | next | phys) & ~PUD_MASK) != 0)
 190                 return false;
 191
 192         return true;
 193 }
 194
 195 static void alloc_init_pud(struct mm_struct *mm, pgd_t *pgd,
 196                                   unsigned long addr, unsigned long end,
 197                                   phys_addr_t phys, pgprot_t prot,
 198                                   void *(*alloc)(unsigned long size))
 199 {
 200         pud_t *pud;
 201         unsigned long next;
 202
 203         if (pgd_none(*pgd)) {
 204                 pud = alloc(PTRS_PER_PUD * sizeof(pud_t));
 205                 pgd_populate(mm, pgd, pud);
 206         }
 207         BUG_ON(pgd_bad(*pgd));
 208
 209         pud = pud_offset(pgd, addr);
 210         do {
 211                 next = pud_addr_end(addr, end);
 212
 213                 /*
 214                  * For 4K granule only, attempt to put down a 1GB block
 215                  */
 216                 if (use_1G_block(addr, next, phys)) {
 217                         pud_t old_pud = *pud;
 218                         set_pud(pud, __pud(phys |
 219                                            pgprot_val(mk_sect_prot(prot))));
 220
 221                         /*
 222                          * If we have an old value for a pud, it will
 223                          * be pointing to a pmd table that we no longer
 224                          * need (from swapper_pg_dir).
 225                          *
 226                          * Look up the old pmd table and free it.
 227                          */
 228                         if (!pud_none(old_pud)) {
 229                                 flush_tlb_all();
 230                                 if (pud_table(old_pud)) {
 231                                         phys_addr_t table = __pa(pmd_offset(&old_pud, 0));
 232                                         if (!WARN_ON_ONCE(slab_is_available()))
 233                                                 memblock_free(table, PAGE_SIZE);
 234                                 }
 235                         }
 236                 } else {
 237                         alloc_init_pmd(mm, pud, addr, next, phys, prot, alloc);
 238                 }
 239                 phys += next - addr;
 240         } while (pud++, addr = next, addr != end);
 241 }
 242
 243 /*
 244  * Create the page directory entries and any necessary page tables for the
 245  * mapping specified by 'md'.
 246  */
 247 static void  __create_mapping(struct mm_struct *mm, pgd_t *pgd,
 248                                     phys_addr_t phys, unsigned long virt,
 249                                     phys_addr_t size, pgprot_t prot,
 250                                     void *(*alloc)(unsigned long size))
 251 {
 252         unsigned long addr, length, end, next;
 253
 254         addr = virt & PAGE_MASK;
 255         length = PAGE_ALIGN(size + (virt & ~PAGE_MASK));
 256
 257         end = addr + length;
 258         do {
 259                 next = pgd_addr_end(addr, end);
 260                 alloc_init_pud(mm, pgd, addr, next, phys, prot, alloc);
 261                 phys += next - addr;
 262         } while (pgd++, addr = next, addr != end);
 263 }
 264
 265 static void *late_alloc(unsigned long size)
 266 {
 267         void *ptr;
 268
 269         BUG_ON(size > PAGE_SIZE);
 270         ptr = (void *)__get_free_page(PGALLOC_GFP);
 271         BUG_ON(!ptr);
 272         return ptr;
 273 }
 274
 275 static void __init create_mapping(phys_addr_t phys, unsigned long virt,
 276                                   phys_addr_t size, pgprot_t prot)
 277 {
 278         if (virt < VMALLOC_START) {
 279                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 280                         &phys, virt);
 281                 return;
 282         }
 283         __create_mapping(&init_mm, pgd_offset_k(virt & PAGE_MASK), phys, virt,
 284                          size, prot, early_alloc);
 285 }
 286
 287 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
 288                                unsigned long virt, phys_addr_t size,
 289                                pgprot_t prot)
 290 {
 291         __create_mapping(mm, pgd_offset(mm, virt), phys, virt, size, prot,
 292                                 late_alloc);
 293 }
 294
 295 static void create_mapping_late(phys_addr_t phys, unsigned long virt,
 296                                   phys_addr_t size, pgprot_t prot)
 297 {
 298         if (virt < VMALLOC_START) {
 299                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 300                         &phys, virt);
 301                 return;
 302         }
 303
 304         return __create_mapping(&init_mm, pgd_offset_k(virt & PAGE_MASK),
 305                                 phys, virt, size, prot, late_alloc);
 306 }
 307
 308 #ifdef CONFIG_DEBUG_RODATA
 309 static void __init __map_memblock(phys_addr_t start, phys_addr_t end)
 310 {
 311         /*
 312          * Set up the executable regions using the existing section mappings
 313          * for now. This will get more fine grained later once all memory
 314          * is mapped
 315          */
 316         unsigned long kernel_x_start = round_down(__pa(_stext), SWAPPER_BLOCK_SIZE);
 317         unsigned long kernel_x_end = round_up(__pa(__init_end), SWAPPER_BLOCK_SIZE);
 318
 319         if (end < kernel_x_start) {
 320                 create_mapping(start, __phys_to_virt(start),
 321                         end - start, PAGE_KERNEL);
 322         } else if (start >= kernel_x_end) {
 323                 create_mapping(start, __phys_to_virt(start),
 324                         end - start, PAGE_KERNEL);
 325         } else {
 326                 if (start < kernel_x_start)
 327                         create_mapping(start, __phys_to_virt(start),
 328                                 kernel_x_start - start,
 329                                 PAGE_KERNEL);
 330                 create_mapping(kernel_x_start,
 331                                 __phys_to_virt(kernel_x_start),
 332                                 kernel_x_end - kernel_x_start,
 333                                 PAGE_KERNEL_EXEC);
 334                 if (kernel_x_end < end)
 335                         create_mapping(kernel_x_end,
 336                                 __phys_to_virt(kernel_x_end),
 337                                 end - kernel_x_end,
 338                                 PAGE_KERNEL);
 339         }
 340
 341 }
 342 #else
 343 static void __init __map_memblock(phys_addr_t start, phys_addr_t end)
 344 {
 345         create_mapping(start, __phys_to_virt(start), end - start,
 346                         PAGE_KERNEL_EXEC);
 347 }
 348 #endif
 349
 350 static void __init map_mem(void)
 351 {
 352         struct memblock_region *reg;
 353         phys_addr_t limit;
 354
 355         /*
 356          * Temporarily limit the memblock range. We need to do this as
 357          * create_mapping requires puds, pmds and ptes to be allocated from
 358          * memory addressable from the initial direct kernel mapping.
 359          *
 360          * The initial direct kernel mapping, located at swapper_pg_dir, gives
 361          * us PUD_SIZE (with SECTION maps) or PMD_SIZE (without SECTION maps,
 362          * memory starting from PHYS_OFFSET (which must be aligned to 2MB as
 363          * per Documentation/arm64/booting.txt).
 364          */
 365         limit = PHYS_OFFSET + SWAPPER_INIT_MAP_SIZE;
 366         memblock_set_current_limit(limit);
 367
 368         /* map all the memory banks */
 369         for_each_memblock(memory, reg) {
 370                 phys_addr_t start = reg->base;
 371                 phys_addr_t end = start + reg->size;
 372
 373                 if (start >= end)
 374                         break;
 375
 376                 if (ARM64_SWAPPER_USES_SECTION_MAPS) {
 377                         /*
 378                          * For the first memory bank align the start address and
 379                          * current memblock limit to prevent create_mapping() from
 380                          * allocating pte page tables from unmapped memory. With
 381                          * the section maps, if the first block doesn't end on section
 382                          * size boundary, create_mapping() will try to allocate a pte
 383                          * page, which may be returned from an unmapped area.
 384                          * When section maps are not used, the pte page table for the
 385                          * current limit is already present in swapper_pg_dir.
 386                          */
 387                         if (start < limit)
 388                                 start = ALIGN(start, SECTION_SIZE);
 389                         if (end < limit) {
 390                                 limit = end & SECTION_MASK;
 391                                 memblock_set_current_limit(limit);
 392                         }
 393                 }
 394                 __map_memblock(start, end);
 395         }
 396
 397         /* Limit no longer required. */
 398         memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
 399 }
 400
 401 static void __init fixup_executable(void)
 402 {
 403 #ifdef CONFIG_DEBUG_RODATA
 404         /* now that we are actually fully mapped, make the start/end more fine grained */
 405         if (!IS_ALIGNED((unsigned long)_stext, SWAPPER_BLOCK_SIZE)) {
 406                 unsigned long aligned_start = round_down(__pa(_stext),
 407                                                          SWAPPER_BLOCK_SIZE);
 408
 409                 create_mapping(aligned_start, __phys_to_virt(aligned_start),
 410                                 __pa(_stext) - aligned_start,
 411                                 PAGE_KERNEL);
 412         }
 413
 414         if (!IS_ALIGNED((unsigned long)__init_end, SWAPPER_BLOCK_SIZE)) {
 415                 unsigned long aligned_end = round_up(__pa(__init_end),
 416                                                           SWAPPER_BLOCK_SIZE);
 417                 create_mapping(__pa(__init_end), (unsigned long)__init_end,
 418                                 aligned_end - __pa(__init_end),
 419                                 PAGE_KERNEL);
 420         }
 421 #endif
 422 }
 423
 424 #ifdef CONFIG_DEBUG_RODATA
 425 void mark_rodata_ro(void)
 426 {
 427         create_mapping_late(__pa(_stext), (unsigned long)_stext,
 428                                 (unsigned long)_etext - (unsigned long)_stext,
 429                                 PAGE_KERNEL_ROX);
 430
 431 }
 432 #endif
 433
 434 void fixup_init(void)
 435 {
 436         create_mapping_late(__pa(__init_begin), (unsigned long)__init_begin,
 437                         (unsigned long)__init_end - (unsigned long)__init_begin,
 438                         PAGE_KERNEL);
 439 }
 440
 441 /*
 442  * paging_init() sets up the page tables, initialises the zone memory
 443  * maps and sets up the zero page.
 444  */
 445 void __init paging_init(void)
 446 {
 447         void *zero_page;
 448
 449         map_mem();
 450         fixup_executable();
 451
 452         /* allocate the zero page. */
 453         zero_page = early_alloc(PAGE_SIZE);
 454
 455         bootmem_init();
 456
 457         empty_zero_page = virt_to_page(zero_page);
 458
 459         /* Ensure the zero page is visible to the page table walker */
 460         dsb(ishst);
 461
 462         /*
 463          * TTBR0 is only used for the identity mapping at this stage. Make it
 464          * point to zero page to avoid speculatively fetching new entries.
 465          */
 466         cpu_set_reserved_ttbr0();
 467         local_flush_tlb_all();
 468         cpu_set_default_tcr_t0sz();
 469 }
 470
 471 /*
 472  * Check whether a kernel address is valid (derived from arch/x86/).
 473  */
 474 int kern_addr_valid(unsigned long addr)
 475 {
 476         pgd_t *pgd;
 477         pud_t *pud;
 478         pmd_t *pmd;
 479         pte_t *pte;
 480
 481         if ((((long)addr) >> VA_BITS) != -1UL)
 482                 return 0;
 483
 484         pgd = pgd_offset_k(addr);
 485         if (pgd_none(*pgd))
 486                 return 0;
 487
 488         pud = pud_offset(pgd, addr);
 489         if (pud_none(*pud))
 490                 return 0;
 491
 492         if (pud_sect(*pud))
 493                 return pfn_valid(pud_pfn(*pud));
 494
 495         pmd = pmd_offset(pud, addr);
 496         if (pmd_none(*pmd))
 497                 return 0;
 498
 499         if (pmd_sect(*pmd))
 500                 return pfn_valid(pmd_pfn(*pmd));
 501
 502         pte = pte_offset_kernel(pmd, addr);
 503         if (pte_none(*pte))
 504                 return 0;
 505
 506         return pfn_valid(pte_pfn(*pte));
 507 }
 508 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 509 #if !ARM64_SWAPPER_USES_SECTION_MAPS
 510 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 511 {
 512         return vmemmap_populate_basepages(start, end, node);
 513 }
 514 #else   /* !ARM64_SWAPPER_USES_SECTION_MAPS */
 515 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 516 {
 517         unsigned long addr = start;
 518         unsigned long next;
 519         pgd_t *pgd;
 520         pud_t *pud;
 521         pmd_t *pmd;
 522
 523         do {
 524                 next = pmd_addr_end(addr, end);
 525
 526                 pgd = vmemmap_pgd_populate(addr, node);
 527                 if (!pgd)
 528                         return -ENOMEM;
 529
 530                 pud = vmemmap_pud_populate(pgd, addr, node);
 531                 if (!pud)
 532                         return -ENOMEM;
 533
 534                 pmd = pmd_offset(pud, addr);
 535                 if (pmd_none(*pmd)) {
 536                         void *p = NULL;
 537
 538                         p = vmemmap_alloc_block_buf(PMD_SIZE, node);
 539                         if (!p)
 540                                 return -ENOMEM;
 541
 542                         set_pmd(pmd, __pmd(__pa(p) | PROT_SECT_NORMAL));
 543                 } else
 544                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 545         } while (addr = next, addr != end);
 546
 547         return 0;
 548 }
 549 #endif  /* CONFIG_ARM64_64K_PAGES */
 550 void vmemmap_free(unsigned long start, unsigned long end)
 551 {
 552 }
 553 #endif  /* CONFIG_SPARSEMEM_VMEMMAP */
 554
 555 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
 556 #if CONFIG_PGTABLE_LEVELS > 2
 557 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss;
 558 #endif
 559 #if CONFIG_PGTABLE_LEVELS > 3
 560 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss;
 561 #endif
 562
 563 static inline pud_t * fixmap_pud(unsigned long addr)
 564 {
 565         pgd_t *pgd = pgd_offset_k(addr);
 566
 567         BUG_ON(pgd_none(*pgd) || pgd_bad(*pgd));
 568
 569         return pud_offset(pgd, addr);
 570 }
 571
 572 static inline pmd_t * fixmap_pmd(unsigned long addr)
 573 {
 574         pud_t *pud = fixmap_pud(addr);
 575
 576         BUG_ON(pud_none(*pud) || pud_bad(*pud));
 577
 578         return pmd_offset(pud, addr);
 579 }
 580
 581 static inline pte_t * fixmap_pte(unsigned long addr)
 582 {
 583         pmd_t *pmd = fixmap_pmd(addr);
 584
 585         BUG_ON(pmd_none(*pmd) || pmd_bad(*pmd));
 586
 587         return pte_offset_kernel(pmd, addr);
 588 }
 589
 590 void __init early_fixmap_init(void)
 591 {
 592         pgd_t *pgd;
 593         pud_t *pud;
 594         pmd_t *pmd;
 595         unsigned long addr = FIXADDR_START;
 596
 597         pgd = pgd_offset_k(addr);
 598         pgd_populate(&init_mm, pgd, bm_pud);
 599         pud = pud_offset(pgd, addr);
 600         pud_populate(&init_mm, pud, bm_pmd);
 601         pmd = pmd_offset(pud, addr);
 602         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 603
 604         /*
 605          * The boot-ioremap range spans multiple pmds, for which
 606          * we are not preparted:
 607          */
 608         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 609                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 610
 611         if ((pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
 612              || pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 613                 WARN_ON(1);
 614                 pr_warn("pmd %p != %p, %p\n",
 615                         pmd, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
 616                         fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
 617                 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 618                         fix_to_virt(FIX_BTMAP_BEGIN));
 619                 pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 620                         fix_to_virt(FIX_BTMAP_END));
 621
 622                 pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 623                 pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
 624         }
 625 }
 626
 627 void __set_fixmap(enum fixed_addresses idx,
 628                                phys_addr_t phys, pgprot_t flags)
 629 {
 630         unsigned long addr = __fix_to_virt(idx);
 631         pte_t *pte;
 632
 633         BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
 634
 635         pte = fixmap_pte(addr);
 636
 637         if (pgprot_val(flags)) {
 638                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 639         } else {
 640                 pte_clear(&init_mm, addr, pte);
 641                 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
 642         }
 643 }
 644
 645 void *__init fixmap_remap_fdt(phys_addr_t dt_phys)
 646 {
 647         const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
 648         pgprot_t prot = PAGE_KERNEL_RO;
 649         int size, offset;
 650         void *dt_virt;
 651
 652         /*
 653          * Check whether the physical FDT address is set and meets the minimum
 654          * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
 655          * at least 8 bytes so that we can always access the size field of the
 656          * FDT header after mapping the first chunk, double check here if that
 657          * is indeed the case.
 658          */
 659         BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
 660         if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
 661                 return NULL;
 662
 663         /*
 664          * Make sure that the FDT region can be mapped without the need to
 665          * allocate additional translation table pages, so that it is safe
 666          * to call create_mapping() this early.
 667          *
 668          * On 64k pages, the FDT will be mapped using PTEs, so we need to
 669          * be in the same PMD as the rest of the fixmap.
 670          * On 4k pages, we'll use section mappings for the FDT so we only
 671          * have to be in the same PUD.
 672          */
 673         BUILD_BUG_ON(dt_virt_base % SZ_2M);
 674
 675         BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
 676                      __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
 677
 678         offset = dt_phys % SWAPPER_BLOCK_SIZE;
 679         dt_virt = (void *)dt_virt_base + offset;
 680
 681         /* map the first chunk so we can read the size from the header */
 682         create_mapping(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
 683                        SWAPPER_BLOCK_SIZE, prot);
 684
 685         if (fdt_check_header(dt_virt) != 0)
 686                 return NULL;
 687
 688         size = fdt_totalsize(dt_virt);
 689         if (size > MAX_FDT_SIZE)
 690                 return NULL;
 691
 692         if (offset + size > SWAPPER_BLOCK_SIZE)
 693                 create_mapping(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
 694                                round_up(offset + size, SWAPPER_BLOCK_SIZE), prot);
 695
 696         memblock_reserve(dt_phys, size);
 697
 698         return dt_virt;
 699 }