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/barrier.h>
  34 #include <asm/cputype.h>
  35 #include <asm/fixmap.h>
  36 #include <asm/kasan.h>
  37 #include <asm/kernel-pgtable.h>
  38 #include <asm/sections.h>
  39 #include <asm/setup.h>
  40 #include <asm/sizes.h>
  41 #include <asm/tlb.h>
  42 #include <asm/memblock.h>
  43 #include <asm/mmu_context.h>
  44
  45 #include "mm.h"
  46
  47 u64 idmap_t0sz = TCR_T0SZ(VA_BITS);
  48
  49 u64 kimage_voffset __read_mostly;
  50 EXPORT_SYMBOL(kimage_voffset);
  51
  52 /*
  53  * Empty_zero_page is a special page that is used for zero-initialized data
  54  * and COW.
  55  */
  56 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
  57 EXPORT_SYMBOL(empty_zero_page);
  58
  59 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
  60 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused;
  61 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused;
  62
  63 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
  64                               unsigned long size, pgprot_t vma_prot)
  65 {
  66         if (!pfn_valid(pfn))
  67                 return pgprot_noncached(vma_prot);
  68         else if (file->f_flags & O_SYNC)
  69                 return pgprot_writecombine(vma_prot);
  70         return vma_prot;
  71 }
  72 EXPORT_SYMBOL(phys_mem_access_prot);
  73
  74 static phys_addr_t __init early_pgtable_alloc(void)
  75 {
  76         phys_addr_t phys;
  77         void *ptr;
  78
  79         phys = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
  80
  81         /*
  82          * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
  83          * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
  84          * any level of table.
  85          */
  86         ptr = pte_set_fixmap(phys);
  87
  88         memset(ptr, 0, PAGE_SIZE);
  89
  90         /*
  91          * Implicit barriers also ensure the zeroed page is visible to the page
  92          * table walker
  93          */
  94         pte_clear_fixmap();
  95
  96         return phys;
  97 }
  98
  99 static void alloc_init_pte(pmd_t *pmd, unsigned long addr,
 100                                   unsigned long end, unsigned long pfn,
 101                                   pgprot_t prot,
 102                                   phys_addr_t (*pgtable_alloc)(void))
 103 {
 104         pte_t *pte;
 105
 106         BUG_ON(pmd_sect(*pmd));
 107         if (pmd_none(*pmd)) {
 108                 phys_addr_t pte_phys;
 109                 BUG_ON(!pgtable_alloc);
 110                 pte_phys = pgtable_alloc();
 111                 pte = pte_set_fixmap(pte_phys);
 112                 __pmd_populate(pmd, pte_phys, PMD_TYPE_TABLE);
 113                 pte_clear_fixmap();
 114         }
 115         BUG_ON(pmd_bad(*pmd));
 116
 117         pte = pte_set_fixmap_offset(pmd, addr);
 118         do {
 119                 set_pte(pte, pfn_pte(pfn, prot));
 120                 pfn++;
 121         } while (pte++, addr += PAGE_SIZE, addr != end);
 122
 123         pte_clear_fixmap();
 124 }
 125
 126 static void alloc_init_pmd(pud_t *pud, unsigned long addr, unsigned long end,
 127                                   phys_addr_t phys, pgprot_t prot,
 128                                   phys_addr_t (*pgtable_alloc)(void),
 129                                   bool allow_block_mappings)
 130 {
 131         pmd_t *pmd;
 132         unsigned long next;
 133
 134         /*
 135          * Check for initial section mappings in the pgd/pud and remove them.
 136          */
 137         BUG_ON(pud_sect(*pud));
 138         if (pud_none(*pud)) {
 139                 phys_addr_t pmd_phys;
 140                 BUG_ON(!pgtable_alloc);
 141                 pmd_phys = pgtable_alloc();
 142                 pmd = pmd_set_fixmap(pmd_phys);
 143                 __pud_populate(pud, pmd_phys, PUD_TYPE_TABLE);
 144                 pmd_clear_fixmap();
 145         }
 146         BUG_ON(pud_bad(*pud));
 147
 148         pmd = pmd_set_fixmap_offset(pud, addr);
 149         do {
 150                 next = pmd_addr_end(addr, end);
 151                 /* try section mapping first */
 152                 if (((addr | next | phys) & ~SECTION_MASK) == 0 &&
 153                       allow_block_mappings) {
 154                         pmd_t old_pmd =*pmd;
 155                         pmd_set_huge(pmd, phys, prot);
 156                         /*
 157                          * Check for previous table entries created during
 158                          * boot (__create_page_tables) and flush them.
 159                          */
 160                         if (!pmd_none(old_pmd)) {
 161                                 flush_tlb_all();
 162                                 if (pmd_table(old_pmd)) {
 163                                         phys_addr_t table = pmd_page_paddr(old_pmd);
 164                                         if (!WARN_ON_ONCE(slab_is_available()))
 165                                                 memblock_free(table, PAGE_SIZE);
 166                                 }
 167                         }
 168                 } else {
 169                         alloc_init_pte(pmd, addr, next, __phys_to_pfn(phys),
 170                                        prot, pgtable_alloc);
 171                 }
 172                 phys += next - addr;
 173         } while (pmd++, addr = next, addr != end);
 174
 175         pmd_clear_fixmap();
 176 }
 177
 178 static inline bool use_1G_block(unsigned long addr, unsigned long next,
 179                         unsigned long phys)
 180 {
 181         if (PAGE_SHIFT != 12)
 182                 return false;
 183
 184         if (((addr | next | phys) & ~PUD_MASK) != 0)
 185                 return false;
 186
 187         return true;
 188 }
 189
 190 static void alloc_init_pud(pgd_t *pgd, unsigned long addr, unsigned long end,
 191                                   phys_addr_t phys, pgprot_t prot,
 192                                   phys_addr_t (*pgtable_alloc)(void),
 193                                   bool allow_block_mappings)
 194 {
 195         pud_t *pud;
 196         unsigned long next;
 197
 198         if (pgd_none(*pgd)) {
 199                 phys_addr_t pud_phys;
 200                 BUG_ON(!pgtable_alloc);
 201                 pud_phys = pgtable_alloc();
 202                 __pgd_populate(pgd, pud_phys, PUD_TYPE_TABLE);
 203         }
 204         BUG_ON(pgd_bad(*pgd));
 205
 206         pud = pud_set_fixmap_offset(pgd, addr);
 207         do {
 208                 next = pud_addr_end(addr, end);
 209
 210                 /*
 211                  * For 4K granule only, attempt to put down a 1GB block
 212                  */
 213                 if (use_1G_block(addr, next, phys) && allow_block_mappings) {
 214                         pud_t old_pud = *pud;
 215                         pud_set_huge(pud, phys, prot);
 216
 217                         /*
 218                          * If we have an old value for a pud, it will
 219                          * be pointing to a pmd table that we no longer
 220                          * need (from swapper_pg_dir).
 221                          *
 222                          * Look up the old pmd table and free it.
 223                          */
 224                         if (!pud_none(old_pud)) {
 225                                 flush_tlb_all();
 226                                 if (pud_table(old_pud)) {
 227                                         phys_addr_t table = pud_page_paddr(old_pud);
 228                                         if (!WARN_ON_ONCE(slab_is_available()))
 229                                                 memblock_free(table, PAGE_SIZE);
 230                                 }
 231                         }
 232                 } else {
 233                         alloc_init_pmd(pud, addr, next, phys, prot,
 234                                        pgtable_alloc, allow_block_mappings);
 235                 }
 236                 phys += next - addr;
 237         } while (pud++, addr = next, addr != end);
 238
 239         pud_clear_fixmap();
 240 }
 241
 242 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
 243                                  unsigned long virt, phys_addr_t size,
 244                                  pgprot_t prot,
 245                                  phys_addr_t (*pgtable_alloc)(void),
 246                                  bool allow_block_mappings)
 247 {
 248         unsigned long addr, length, end, next;
 249         pgd_t *pgd = pgd_offset_raw(pgdir, virt);
 250
 251         /*
 252          * If the virtual and physical address don't have the same offset
 253          * within a page, we cannot map the region as the caller expects.
 254          */
 255         if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
 256                 return;
 257
 258         phys &= PAGE_MASK;
 259         addr = virt & PAGE_MASK;
 260         length = PAGE_ALIGN(size + (virt & ~PAGE_MASK));
 261
 262         end = addr + length;
 263         do {
 264                 next = pgd_addr_end(addr, end);
 265                 alloc_init_pud(pgd, addr, next, phys, prot, pgtable_alloc,
 266                                allow_block_mappings);
 267                 phys += next - addr;
 268         } while (pgd++, addr = next, addr != end);
 269 }
 270
 271 static phys_addr_t pgd_pgtable_alloc(void)
 272 {
 273         void *ptr = (void *)__get_free_page(PGALLOC_GFP);
 274         if (!ptr || !pgtable_page_ctor(virt_to_page(ptr)))
 275                 BUG();
 276
 277         /* Ensure the zeroed page is visible to the page table walker */
 278         dsb(ishst);
 279         return __pa(ptr);
 280 }
 281
 282 /*
 283  * This function can only be used to modify existing table entries,
 284  * without allocating new levels of table. Note that this permits the
 285  * creation of new section or page entries.
 286  */
 287 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
 288                                   phys_addr_t size, pgprot_t prot)
 289 {
 290         if (virt < VMALLOC_START) {
 291                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 292                         &phys, virt);
 293                 return;
 294         }
 295         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, true);
 296 }
 297
 298 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
 299                                unsigned long virt, phys_addr_t size,
 300                                pgprot_t prot, bool allow_block_mappings)
 301 {
 302         BUG_ON(mm == &init_mm);
 303
 304         __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
 305                              pgd_pgtable_alloc, allow_block_mappings);
 306 }
 307
 308 static void create_mapping_late(phys_addr_t phys, unsigned long virt,
 309                                   phys_addr_t size, pgprot_t prot)
 310 {
 311         if (virt < VMALLOC_START) {
 312                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 313                         &phys, virt);
 314                 return;
 315         }
 316
 317         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot,
 318                              NULL, !debug_pagealloc_enabled());
 319 }
 320
 321 static void __init __map_memblock(pgd_t *pgd, phys_addr_t start, phys_addr_t end)
 322 {
 323         unsigned long kernel_start = __pa(_text);
 324         unsigned long kernel_end = __pa(__init_begin);
 325
 326         /*
 327          * Take care not to create a writable alias for the
 328          * read-only text and rodata sections of the kernel image.
 329          */
 330
 331         /* No overlap with the kernel text/rodata */
 332         if (end < kernel_start || start >= kernel_end) {
 333                 __create_pgd_mapping(pgd, start, __phys_to_virt(start),
 334                                      end - start, PAGE_KERNEL,
 335                                      early_pgtable_alloc,
 336                                      !debug_pagealloc_enabled());
 337                 return;
 338         }
 339
 340         /*
 341          * This block overlaps the kernel text/rodata mappings.
 342          * Map the portion(s) which don't overlap.
 343          */
 344         if (start < kernel_start)
 345                 __create_pgd_mapping(pgd, start,
 346                                      __phys_to_virt(start),
 347                                      kernel_start - start, PAGE_KERNEL,
 348                                      early_pgtable_alloc,
 349                                      !debug_pagealloc_enabled());
 350         if (kernel_end < end)
 351                 __create_pgd_mapping(pgd, kernel_end,
 352                                      __phys_to_virt(kernel_end),
 353                                      end - kernel_end, PAGE_KERNEL,
 354                                      early_pgtable_alloc,
 355                                      !debug_pagealloc_enabled());
 356
 357         /*
 358          * Map the linear alias of the [_text, __init_begin) interval as
 359          * read-only/non-executable. This makes the contents of the
 360          * region accessible to subsystems such as hibernate, but
 361          * protects it from inadvertent modification or execution.
 362          */
 363         __create_pgd_mapping(pgd, kernel_start, __phys_to_virt(kernel_start),
 364                              kernel_end - kernel_start, PAGE_KERNEL_RO,
 365                              early_pgtable_alloc, !debug_pagealloc_enabled());
 366 }
 367
 368 static void __init map_mem(pgd_t *pgd)
 369 {
 370         struct memblock_region *reg;
 371
 372         /* map all the memory banks */
 373         for_each_memblock(memory, reg) {
 374                 phys_addr_t start = reg->base;
 375                 phys_addr_t end = start + reg->size;
 376
 377                 if (start >= end)
 378                         break;
 379                 if (memblock_is_nomap(reg))
 380                         continue;
 381
 382                 __map_memblock(pgd, start, end);
 383         }
 384 }
 385
 386 void mark_rodata_ro(void)
 387 {
 388         unsigned long section_size;
 389
 390         section_size = (unsigned long)_etext - (unsigned long)_text;
 391         create_mapping_late(__pa(_text), (unsigned long)_text,
 392                             section_size, PAGE_KERNEL_ROX);
 393         /*
 394          * mark .rodata as read only. Use __init_begin rather than __end_rodata
 395          * to cover NOTES and EXCEPTION_TABLE.
 396          */
 397         section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
 398         create_mapping_late(__pa(__start_rodata), (unsigned long)__start_rodata,
 399                             section_size, PAGE_KERNEL_RO);
 400 }
 401
 402 void fixup_init(void)
 403 {
 404         /*
 405          * Unmap the __init region but leave the VM area in place. This
 406          * prevents the region from being reused for kernel modules, which
 407          * is not supported by kallsyms.
 408          */
 409         unmap_kernel_range((u64)__init_begin, (u64)(__init_end - __init_begin));
 410 }
 411
 412 static void __init map_kernel_segment(pgd_t *pgd, void *va_start, void *va_end,
 413                                       pgprot_t prot, struct vm_struct *vma)
 414 {
 415         phys_addr_t pa_start = __pa(va_start);
 416         unsigned long size = va_end - va_start;
 417
 418         BUG_ON(!PAGE_ALIGNED(pa_start));
 419         BUG_ON(!PAGE_ALIGNED(size));
 420
 421         __create_pgd_mapping(pgd, pa_start, (unsigned long)va_start, size, prot,
 422                              early_pgtable_alloc, !debug_pagealloc_enabled());
 423
 424         vma->addr       = va_start;
 425         vma->phys_addr  = pa_start;
 426         vma->size       = size;
 427         vma->flags      = VM_MAP;
 428         vma->caller     = __builtin_return_address(0);
 429
 430         vm_area_add_early(vma);
 431 }
 432
 433 /*
 434  * Create fine-grained mappings for the kernel.
 435  */
 436 static void __init map_kernel(pgd_t *pgd)
 437 {
 438         static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_init, vmlinux_data;
 439
 440         map_kernel_segment(pgd, _text, _etext, PAGE_KERNEL_EXEC, &vmlinux_text);
 441         map_kernel_segment(pgd, __start_rodata, __init_begin, PAGE_KERNEL, &vmlinux_rodata);
 442         map_kernel_segment(pgd, __init_begin, __init_end, PAGE_KERNEL_EXEC,
 443                            &vmlinux_init);
 444         map_kernel_segment(pgd, _data, _end, PAGE_KERNEL, &vmlinux_data);
 445
 446         if (!pgd_val(*pgd_offset_raw(pgd, FIXADDR_START))) {
 447                 /*
 448                  * The fixmap falls in a separate pgd to the kernel, and doesn't
 449                  * live in the carveout for the swapper_pg_dir. We can simply
 450                  * re-use the existing dir for the fixmap.
 451                  */
 452                 set_pgd(pgd_offset_raw(pgd, FIXADDR_START),
 453                         *pgd_offset_k(FIXADDR_START));
 454         } else if (CONFIG_PGTABLE_LEVELS > 3) {
 455                 /*
 456                  * The fixmap shares its top level pgd entry with the kernel
 457                  * mapping. This can really only occur when we are running
 458                  * with 16k/4 levels, so we can simply reuse the pud level
 459                  * entry instead.
 460                  */
 461                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 462                 set_pud(pud_set_fixmap_offset(pgd, FIXADDR_START),
 463                         __pud(__pa(bm_pmd) | PUD_TYPE_TABLE));
 464                 pud_clear_fixmap();
 465         } else {
 466                 BUG();
 467         }
 468
 469         kasan_copy_shadow(pgd);
 470 }
 471
 472 /*
 473  * paging_init() sets up the page tables, initialises the zone memory
 474  * maps and sets up the zero page.
 475  */
 476 void __init paging_init(void)
 477 {
 478         phys_addr_t pgd_phys = early_pgtable_alloc();
 479         pgd_t *pgd = pgd_set_fixmap(pgd_phys);
 480
 481         map_kernel(pgd);
 482         map_mem(pgd);
 483
 484         /*
 485          * We want to reuse the original swapper_pg_dir so we don't have to
 486          * communicate the new address to non-coherent secondaries in
 487          * secondary_entry, and so cpu_switch_mm can generate the address with
 488          * adrp+add rather than a load from some global variable.
 489          *
 490          * To do this we need to go via a temporary pgd.
 491          */
 492         cpu_replace_ttbr1(__va(pgd_phys));
 493         memcpy(swapper_pg_dir, pgd, PAGE_SIZE);
 494         cpu_replace_ttbr1(swapper_pg_dir);
 495
 496         pgd_clear_fixmap();
 497         memblock_free(pgd_phys, PAGE_SIZE);
 498
 499         /*
 500          * We only reuse the PGD from the swapper_pg_dir, not the pud + pmd
 501          * allocated with it.
 502          */
 503         memblock_free(__pa(swapper_pg_dir) + PAGE_SIZE,
 504                       SWAPPER_DIR_SIZE - PAGE_SIZE);
 505 }
 506
 507 /*
 508  * Check whether a kernel address is valid (derived from arch/x86/).
 509  */
 510 int kern_addr_valid(unsigned long addr)
 511 {
 512         pgd_t *pgd;
 513         pud_t *pud;
 514         pmd_t *pmd;
 515         pte_t *pte;
 516
 517         if ((((long)addr) >> VA_BITS) != -1UL)
 518                 return 0;
 519
 520         pgd = pgd_offset_k(addr);
 521         if (pgd_none(*pgd))
 522                 return 0;
 523
 524         pud = pud_offset(pgd, addr);
 525         if (pud_none(*pud))
 526                 return 0;
 527
 528         if (pud_sect(*pud))
 529                 return pfn_valid(pud_pfn(*pud));
 530
 531         pmd = pmd_offset(pud, addr);
 532         if (pmd_none(*pmd))
 533                 return 0;
 534
 535         if (pmd_sect(*pmd))
 536                 return pfn_valid(pmd_pfn(*pmd));
 537
 538         pte = pte_offset_kernel(pmd, addr);
 539         if (pte_none(*pte))
 540                 return 0;
 541
 542         return pfn_valid(pte_pfn(*pte));
 543 }
 544 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 545 #if !ARM64_SWAPPER_USES_SECTION_MAPS
 546 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 547 {
 548         return vmemmap_populate_basepages(start, end, node);
 549 }
 550 #else   /* !ARM64_SWAPPER_USES_SECTION_MAPS */
 551 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 552 {
 553         unsigned long addr = start;
 554         unsigned long next;
 555         pgd_t *pgd;
 556         pud_t *pud;
 557         pmd_t *pmd;
 558
 559         do {
 560                 next = pmd_addr_end(addr, end);
 561
 562                 pgd = vmemmap_pgd_populate(addr, node);
 563                 if (!pgd)
 564                         return -ENOMEM;
 565
 566                 pud = vmemmap_pud_populate(pgd, addr, node);
 567                 if (!pud)
 568                         return -ENOMEM;
 569
 570                 pmd = pmd_offset(pud, addr);
 571                 if (pmd_none(*pmd)) {
 572                         void *p = NULL;
 573
 574                         p = vmemmap_alloc_block_buf(PMD_SIZE, node);
 575                         if (!p)
 576                                 return -ENOMEM;
 577
 578                         set_pmd(pmd, __pmd(__pa(p) | PROT_SECT_NORMAL));
 579                 } else
 580                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 581         } while (addr = next, addr != end);
 582
 583         return 0;
 584 }
 585 #endif  /* CONFIG_ARM64_64K_PAGES */
 586 void vmemmap_free(unsigned long start, unsigned long end)
 587 {
 588 }
 589 #endif  /* CONFIG_SPARSEMEM_VMEMMAP */
 590
 591 static inline pud_t * fixmap_pud(unsigned long addr)
 592 {
 593         pgd_t *pgd = pgd_offset_k(addr);
 594
 595         BUG_ON(pgd_none(*pgd) || pgd_bad(*pgd));
 596
 597         return pud_offset_kimg(pgd, addr);
 598 }
 599
 600 static inline pmd_t * fixmap_pmd(unsigned long addr)
 601 {
 602         pud_t *pud = fixmap_pud(addr);
 603
 604         BUG_ON(pud_none(*pud) || pud_bad(*pud));
 605
 606         return pmd_offset_kimg(pud, addr);
 607 }
 608
 609 static inline pte_t * fixmap_pte(unsigned long addr)
 610 {
 611         return &bm_pte[pte_index(addr)];
 612 }
 613
 614 void __init early_fixmap_init(void)
 615 {
 616         pgd_t *pgd;
 617         pud_t *pud;
 618         pmd_t *pmd;
 619         unsigned long addr = FIXADDR_START;
 620
 621         pgd = pgd_offset_k(addr);
 622         if (CONFIG_PGTABLE_LEVELS > 3 &&
 623             !(pgd_none(*pgd) || pgd_page_paddr(*pgd) == __pa(bm_pud))) {
 624                 /*
 625                  * We only end up here if the kernel mapping and the fixmap
 626                  * share the top level pgd entry, which should only happen on
 627                  * 16k/4 levels configurations.
 628                  */
 629                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 630                 pud = pud_offset_kimg(pgd, addr);
 631         } else {
 632                 pgd_populate(&init_mm, pgd, bm_pud);
 633                 pud = fixmap_pud(addr);
 634         }
 635         pud_populate(&init_mm, pud, bm_pmd);
 636         pmd = fixmap_pmd(addr);
 637         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 638
 639         /*
 640          * The boot-ioremap range spans multiple pmds, for which
 641          * we are not prepared:
 642          */
 643         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 644                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 645
 646         if ((pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
 647              || pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 648                 WARN_ON(1);
 649                 pr_warn("pmd %p != %p, %p\n",
 650                         pmd, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
 651                         fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
 652                 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 653                         fix_to_virt(FIX_BTMAP_BEGIN));
 654                 pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 655                         fix_to_virt(FIX_BTMAP_END));
 656
 657                 pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 658                 pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
 659         }
 660 }
 661
 662 void __set_fixmap(enum fixed_addresses idx,
 663                                phys_addr_t phys, pgprot_t flags)
 664 {
 665         unsigned long addr = __fix_to_virt(idx);
 666         pte_t *pte;
 667
 668         BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
 669
 670         pte = fixmap_pte(addr);
 671
 672         if (pgprot_val(flags)) {
 673                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 674         } else {
 675                 pte_clear(&init_mm, addr, pte);
 676                 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
 677         }
 678 }
 679
 680 void *__init __fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
 681 {
 682         const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
 683         int offset;
 684         void *dt_virt;
 685
 686         /*
 687          * Check whether the physical FDT address is set and meets the minimum
 688          * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
 689          * at least 8 bytes so that we can always access the size field of the
 690          * FDT header after mapping the first chunk, double check here if that
 691          * is indeed the case.
 692          */
 693         BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
 694         if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
 695                 return NULL;
 696
 697         /*
 698          * Make sure that the FDT region can be mapped without the need to
 699          * allocate additional translation table pages, so that it is safe
 700          * to call create_mapping_noalloc() this early.
 701          *
 702          * On 64k pages, the FDT will be mapped using PTEs, so we need to
 703          * be in the same PMD as the rest of the fixmap.
 704          * On 4k pages, we'll use section mappings for the FDT so we only
 705          * have to be in the same PUD.
 706          */
 707         BUILD_BUG_ON(dt_virt_base % SZ_2M);
 708
 709         BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
 710                      __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
 711
 712         offset = dt_phys % SWAPPER_BLOCK_SIZE;
 713         dt_virt = (void *)dt_virt_base + offset;
 714
 715         /* map the first chunk so we can read the size from the header */
 716         create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
 717                         dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
 718
 719         if (fdt_check_header(dt_virt) != 0)
 720                 return NULL;
 721
 722         *size = fdt_totalsize(dt_virt);
 723         if (*size > MAX_FDT_SIZE)
 724                 return NULL;
 725
 726         if (offset + *size > SWAPPER_BLOCK_SIZE)
 727                 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
 728                                round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
 729
 730         return dt_virt;
 731 }
 732
 733 void *__init fixmap_remap_fdt(phys_addr_t dt_phys)
 734 {
 735         void *dt_virt;
 736         int size;
 737
 738         dt_virt = __fixmap_remap_fdt(dt_phys, &size, PAGE_KERNEL_RO);
 739         if (!dt_virt)
 740                 return NULL;
 741
 742         memblock_reserve(dt_phys, size);
 743         return dt_virt;
 744 }
 745
 746 int __init arch_ioremap_pud_supported(void)
 747 {
 748         /* only 4k granule supports level 1 block mappings */
 749         return IS_ENABLED(CONFIG_ARM64_4K_PAGES);
 750 }
 751
 752 int __init arch_ioremap_pmd_supported(void)
 753 {
 754         return 1;
 755 }
 756
 757 int pud_set_huge(pud_t *pud, phys_addr_t phys, pgprot_t prot)
 758 {
 759         BUG_ON(phys & ~PUD_MASK);
 760         set_pud(pud, __pud(phys | PUD_TYPE_SECT | pgprot_val(mk_sect_prot(prot))));
 761         return 1;
 762 }
 763
 764 int pmd_set_huge(pmd_t *pmd, phys_addr_t phys, pgprot_t prot)
 765 {
 766         BUG_ON(phys & ~PMD_MASK);
 767         set_pmd(pmd, __pmd(phys | PMD_TYPE_SECT | pgprot_val(mk_sect_prot(prot))));
 768         return 1;
 769 }
 770
 771 int pud_clear_huge(pud_t *pud)
 772 {
 773         if (!pud_sect(*pud))
 774                 return 0;
 775         pud_clear(pud);
 776         return 1;
 777 }
 778
 779 int pmd_clear_huge(pmd_t *pmd)
 780 {
 781         if (!pmd_sect(*pmd))
 782                 return 0;
 783         pmd_clear(pmd);
 784         return 1;
 785 }