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         BUG_ON(!phys);
  81
  82         /*
  83          * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
  84          * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
  85          * any level of table.
  86          */
  87         ptr = pte_set_fixmap(phys);
  88
  89         memset(ptr, 0, PAGE_SIZE);
  90
  91         /*
  92          * Implicit barriers also ensure the zeroed page is visible to the page
  93          * table walker
  94          */
  95         pte_clear_fixmap();
  96
  97         return phys;
  98 }
  99
 100 /*
 101  * remap a PMD into pages
 102  */
 103 static void split_pmd(pmd_t *pmd, pte_t *pte)
 104 {
 105         unsigned long pfn = pmd_pfn(*pmd);
 106         int i = 0;
 107
 108         do {
 109                 /*
 110                  * Need to have the least restrictive permissions available
 111                  * permissions will be fixed up later
 112                  */
 113                 set_pte(pte, pfn_pte(pfn, PAGE_KERNEL_EXEC));
 114                 pfn++;
 115         } while (pte++, i++, i < PTRS_PER_PTE);
 116 }
 117
 118 static void alloc_init_pte(pmd_t *pmd, unsigned long addr,
 119                                   unsigned long end, unsigned long pfn,
 120                                   pgprot_t prot,
 121                                   phys_addr_t (*pgtable_alloc)(void))
 122 {
 123         pte_t *pte;
 124
 125         if (pmd_none(*pmd) || pmd_sect(*pmd)) {
 126                 phys_addr_t pte_phys;
 127                 BUG_ON(!pgtable_alloc);
 128                 pte_phys = pgtable_alloc();
 129                 pte = pte_set_fixmap(pte_phys);
 130                 if (pmd_sect(*pmd))
 131                         split_pmd(pmd, pte);
 132                 __pmd_populate(pmd, pte_phys, PMD_TYPE_TABLE);
 133                 flush_tlb_all();
 134                 pte_clear_fixmap();
 135         }
 136         BUG_ON(pmd_bad(*pmd));
 137
 138         pte = pte_set_fixmap_offset(pmd, addr);
 139         do {
 140                 set_pte(pte, pfn_pte(pfn, prot));
 141                 pfn++;
 142         } while (pte++, addr += PAGE_SIZE, addr != end);
 143
 144         pte_clear_fixmap();
 145 }
 146
 147 static void split_pud(pud_t *old_pud, pmd_t *pmd)
 148 {
 149         unsigned long addr = pud_pfn(*old_pud) << PAGE_SHIFT;
 150         pgprot_t prot = __pgprot(pud_val(*old_pud) ^ addr);
 151         int i = 0;
 152
 153         do {
 154                 set_pmd(pmd, __pmd(addr | pgprot_val(prot)));
 155                 addr += PMD_SIZE;
 156         } while (pmd++, i++, i < PTRS_PER_PMD);
 157 }
 158
 159 #ifdef CONFIG_DEBUG_PAGEALLOC
 160 static bool block_mappings_allowed(phys_addr_t (*pgtable_alloc)(void))
 161 {
 162
 163         /*
 164          * If debug_page_alloc is enabled we must map the linear map
 165          * using pages. However, other mappings created by
 166          * create_mapping_noalloc must use sections in some cases. Allow
 167          * sections to be used in those cases, where no pgtable_alloc
 168          * function is provided.
 169          */
 170         return !pgtable_alloc || !debug_pagealloc_enabled();
 171 }
 172 #else
 173 static bool block_mappings_allowed(phys_addr_t (*pgtable_alloc)(void))
 174 {
 175         return true;
 176 }
 177 #endif
 178
 179 static void alloc_init_pmd(pud_t *pud, unsigned long addr, unsigned long end,
 180                                   phys_addr_t phys, pgprot_t prot,
 181                                   phys_addr_t (*pgtable_alloc)(void))
 182 {
 183         pmd_t *pmd;
 184         unsigned long next;
 185
 186         /*
 187          * Check for initial section mappings in the pgd/pud and remove them.
 188          */
 189         if (pud_none(*pud) || pud_sect(*pud)) {
 190                 phys_addr_t pmd_phys;
 191                 BUG_ON(!pgtable_alloc);
 192                 pmd_phys = pgtable_alloc();
 193                 pmd = pmd_set_fixmap(pmd_phys);
 194                 if (pud_sect(*pud)) {
 195                         /*
 196                          * need to have the 1G of mappings continue to be
 197                          * present
 198                          */
 199                         split_pud(pud, pmd);
 200                 }
 201                 __pud_populate(pud, pmd_phys, PUD_TYPE_TABLE);
 202                 flush_tlb_all();
 203                 pmd_clear_fixmap();
 204         }
 205         BUG_ON(pud_bad(*pud));
 206
 207         pmd = pmd_set_fixmap_offset(pud, addr);
 208         do {
 209                 next = pmd_addr_end(addr, end);
 210                 /* try section mapping first */
 211                 if (((addr | next | phys) & ~SECTION_MASK) == 0 &&
 212                       block_mappings_allowed(pgtable_alloc)) {
 213                         pmd_t old_pmd =*pmd;
 214                         pmd_set_huge(pmd, phys, prot);
 215                         /*
 216                          * Check for previous table entries created during
 217                          * boot (__create_page_tables) and flush them.
 218                          */
 219                         if (!pmd_none(old_pmd)) {
 220                                 flush_tlb_all();
 221                                 if (pmd_table(old_pmd)) {
 222                                         phys_addr_t table = pmd_page_paddr(old_pmd);
 223                                         if (!WARN_ON_ONCE(slab_is_available()))
 224                                                 memblock_free(table, PAGE_SIZE);
 225                                 }
 226                         }
 227                 } else {
 228                         alloc_init_pte(pmd, addr, next, __phys_to_pfn(phys),
 229                                        prot, pgtable_alloc);
 230                 }
 231                 phys += next - addr;
 232         } while (pmd++, addr = next, addr != end);
 233
 234         pmd_clear_fixmap();
 235 }
 236
 237 static inline bool use_1G_block(unsigned long addr, unsigned long next,
 238                         unsigned long phys)
 239 {
 240         if (PAGE_SHIFT != 12)
 241                 return false;
 242
 243         if (((addr | next | phys) & ~PUD_MASK) != 0)
 244                 return false;
 245
 246         return true;
 247 }
 248
 249 static void alloc_init_pud(pgd_t *pgd, unsigned long addr, unsigned long end,
 250                                   phys_addr_t phys, pgprot_t prot,
 251                                   phys_addr_t (*pgtable_alloc)(void))
 252 {
 253         pud_t *pud;
 254         unsigned long next;
 255
 256         if (pgd_none(*pgd)) {
 257                 phys_addr_t pud_phys;
 258                 BUG_ON(!pgtable_alloc);
 259                 pud_phys = pgtable_alloc();
 260                 __pgd_populate(pgd, pud_phys, PUD_TYPE_TABLE);
 261         }
 262         BUG_ON(pgd_bad(*pgd));
 263
 264         pud = pud_set_fixmap_offset(pgd, addr);
 265         do {
 266                 next = pud_addr_end(addr, end);
 267
 268                 /*
 269                  * For 4K granule only, attempt to put down a 1GB block
 270                  */
 271                 if (use_1G_block(addr, next, phys) &&
 272                     block_mappings_allowed(pgtable_alloc)) {
 273                         pud_t old_pud = *pud;
 274                         pud_set_huge(pud, phys, prot);
 275
 276                         /*
 277                          * If we have an old value for a pud, it will
 278                          * be pointing to a pmd table that we no longer
 279                          * need (from swapper_pg_dir).
 280                          *
 281                          * Look up the old pmd table and free it.
 282                          */
 283                         if (!pud_none(old_pud)) {
 284                                 flush_tlb_all();
 285                                 if (pud_table(old_pud)) {
 286                                         phys_addr_t table = pud_page_paddr(old_pud);
 287                                         if (!WARN_ON_ONCE(slab_is_available()))
 288                                                 memblock_free(table, PAGE_SIZE);
 289                                 }
 290                         }
 291                 } else {
 292                         alloc_init_pmd(pud, addr, next, phys, prot,
 293                                        pgtable_alloc);
 294                 }
 295                 phys += next - addr;
 296         } while (pud++, addr = next, addr != end);
 297
 298         pud_clear_fixmap();
 299 }
 300
 301 /*
 302  * Create the page directory entries and any necessary page tables for the
 303  * mapping specified by 'md'.
 304  */
 305 static void init_pgd(pgd_t *pgd, phys_addr_t phys, unsigned long virt,
 306                                     phys_addr_t size, pgprot_t prot,
 307                                     phys_addr_t (*pgtable_alloc)(void))
 308 {
 309         unsigned long addr, length, end, next;
 310
 311         /*
 312          * If the virtual and physical address don't have the same offset
 313          * within a page, we cannot map the region as the caller expects.
 314          */
 315         if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
 316                 return;
 317
 318         phys &= PAGE_MASK;
 319         addr = virt & PAGE_MASK;
 320         length = PAGE_ALIGN(size + (virt & ~PAGE_MASK));
 321
 322         end = addr + length;
 323         do {
 324                 next = pgd_addr_end(addr, end);
 325                 alloc_init_pud(pgd, addr, next, phys, prot, pgtable_alloc);
 326                 phys += next - addr;
 327         } while (pgd++, addr = next, addr != end);
 328 }
 329
 330 static phys_addr_t late_pgtable_alloc(void)
 331 {
 332         void *ptr = (void *)__get_free_page(PGALLOC_GFP);
 333         BUG_ON(!ptr);
 334
 335         /* Ensure the zeroed page is visible to the page table walker */
 336         dsb(ishst);
 337         return __pa(ptr);
 338 }
 339
 340 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
 341                                  unsigned long virt, phys_addr_t size,
 342                                  pgprot_t prot,
 343                                  phys_addr_t (*alloc)(void))
 344 {
 345         init_pgd(pgd_offset_raw(pgdir, virt), phys, virt, size, prot, alloc);
 346 }
 347
 348 /*
 349  * This function can only be used to modify existing table entries,
 350  * without allocating new levels of table. Note that this permits the
 351  * creation of new section or page entries.
 352  */
 353 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
 354                                   phys_addr_t size, pgprot_t prot)
 355 {
 356         if (virt < VMALLOC_START) {
 357                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 358                         &phys, virt);
 359                 return;
 360         }
 361         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot,
 362                              NULL);
 363 }
 364
 365 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
 366                                unsigned long virt, phys_addr_t size,
 367                                pgprot_t prot)
 368 {
 369         __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
 370                              late_pgtable_alloc);
 371 }
 372
 373 static void create_mapping_late(phys_addr_t phys, unsigned long virt,
 374                                   phys_addr_t size, pgprot_t prot)
 375 {
 376         if (virt < VMALLOC_START) {
 377                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 378                         &phys, virt);
 379                 return;
 380         }
 381
 382         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot,
 383                              late_pgtable_alloc);
 384 }
 385
 386 static void __init __map_memblock(pgd_t *pgd, phys_addr_t start, phys_addr_t end)
 387 {
 388         unsigned long kernel_start = __pa(_text);
 389         unsigned long kernel_end = __pa(__init_begin);
 390
 391         /*
 392          * Take care not to create a writable alias for the
 393          * read-only text and rodata sections of the kernel image.
 394          */
 395
 396         /* No overlap with the kernel text/rodata */
 397         if (end < kernel_start || start >= kernel_end) {
 398                 __create_pgd_mapping(pgd, start, __phys_to_virt(start),
 399                                      end - start, PAGE_KERNEL,
 400                                      early_pgtable_alloc);
 401                 return;
 402         }
 403
 404         /*
 405          * This block overlaps the kernel text/rodata mappings.
 406          * Map the portion(s) which don't overlap.
 407          */
 408         if (start < kernel_start)
 409                 __create_pgd_mapping(pgd, start,
 410                                      __phys_to_virt(start),
 411                                      kernel_start - start, PAGE_KERNEL,
 412                                      early_pgtable_alloc);
 413         if (kernel_end < end)
 414                 __create_pgd_mapping(pgd, kernel_end,
 415                                      __phys_to_virt(kernel_end),
 416                                      end - kernel_end, PAGE_KERNEL,
 417                                      early_pgtable_alloc);
 418
 419         /*
 420          * Map the linear alias of the [_text, __init_begin) interval as
 421          * read-only/non-executable. This makes the contents of the
 422          * region accessible to subsystems such as hibernate, but
 423          * protects it from inadvertent modification or execution.
 424          */
 425         __create_pgd_mapping(pgd, kernel_start, __phys_to_virt(kernel_start),
 426                              kernel_end - kernel_start, PAGE_KERNEL_RO,
 427                              early_pgtable_alloc);
 428 }
 429
 430 static void __init map_mem(pgd_t *pgd)
 431 {
 432         struct memblock_region *reg;
 433
 434         /* map all the memory banks */
 435         for_each_memblock(memory, reg) {
 436                 phys_addr_t start = reg->base;
 437                 phys_addr_t end = start + reg->size;
 438
 439                 if (start >= end)
 440                         break;
 441                 if (memblock_is_nomap(reg))
 442                         continue;
 443
 444                 __map_memblock(pgd, start, end);
 445         }
 446 }
 447
 448 void mark_rodata_ro(void)
 449 {
 450         unsigned long section_size;
 451
 452         section_size = (unsigned long)_etext - (unsigned long)_text;
 453         create_mapping_late(__pa(_text), (unsigned long)_text,
 454                             section_size, PAGE_KERNEL_ROX);
 455         /*
 456          * mark .rodata as read only. Use __init_begin rather than __end_rodata
 457          * to cover NOTES and EXCEPTION_TABLE.
 458          */
 459         section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
 460         create_mapping_late(__pa(__start_rodata), (unsigned long)__start_rodata,
 461                             section_size, PAGE_KERNEL_RO);
 462 }
 463
 464 void fixup_init(void)
 465 {
 466         /*
 467          * Unmap the __init region but leave the VM area in place. This
 468          * prevents the region from being reused for kernel modules, which
 469          * is not supported by kallsyms.
 470          */
 471         unmap_kernel_range((u64)__init_begin, (u64)(__init_end - __init_begin));
 472 }
 473
 474 static void __init map_kernel_segment(pgd_t *pgd, void *va_start, void *va_end,
 475                                       pgprot_t prot, struct vm_struct *vma)
 476 {
 477         phys_addr_t pa_start = __pa(va_start);
 478         unsigned long size = va_end - va_start;
 479
 480         BUG_ON(!PAGE_ALIGNED(pa_start));
 481         BUG_ON(!PAGE_ALIGNED(size));
 482
 483         __create_pgd_mapping(pgd, pa_start, (unsigned long)va_start, size, prot,
 484                              early_pgtable_alloc);
 485
 486         vma->addr       = va_start;
 487         vma->phys_addr  = pa_start;
 488         vma->size       = size;
 489         vma->flags      = VM_MAP;
 490         vma->caller     = __builtin_return_address(0);
 491
 492         vm_area_add_early(vma);
 493 }
 494
 495 /*
 496  * Create fine-grained mappings for the kernel.
 497  */
 498 static void __init map_kernel(pgd_t *pgd)
 499 {
 500         static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_init, vmlinux_data;
 501
 502         map_kernel_segment(pgd, _text, _etext, PAGE_KERNEL_EXEC, &vmlinux_text);
 503         map_kernel_segment(pgd, __start_rodata, __init_begin, PAGE_KERNEL, &vmlinux_rodata);
 504         map_kernel_segment(pgd, __init_begin, __init_end, PAGE_KERNEL_EXEC,
 505                            &vmlinux_init);
 506         map_kernel_segment(pgd, _data, _end, PAGE_KERNEL, &vmlinux_data);
 507
 508         if (!pgd_val(*pgd_offset_raw(pgd, FIXADDR_START))) {
 509                 /*
 510                  * The fixmap falls in a separate pgd to the kernel, and doesn't
 511                  * live in the carveout for the swapper_pg_dir. We can simply
 512                  * re-use the existing dir for the fixmap.
 513                  */
 514                 set_pgd(pgd_offset_raw(pgd, FIXADDR_START),
 515                         *pgd_offset_k(FIXADDR_START));
 516         } else if (CONFIG_PGTABLE_LEVELS > 3) {
 517                 /*
 518                  * The fixmap shares its top level pgd entry with the kernel
 519                  * mapping. This can really only occur when we are running
 520                  * with 16k/4 levels, so we can simply reuse the pud level
 521                  * entry instead.
 522                  */
 523                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 524                 set_pud(pud_set_fixmap_offset(pgd, FIXADDR_START),
 525                         __pud(__pa(bm_pmd) | PUD_TYPE_TABLE));
 526                 pud_clear_fixmap();
 527         } else {
 528                 BUG();
 529         }
 530
 531         kasan_copy_shadow(pgd);
 532 }
 533
 534 /*
 535  * paging_init() sets up the page tables, initialises the zone memory
 536  * maps and sets up the zero page.
 537  */
 538 void __init paging_init(void)
 539 {
 540         phys_addr_t pgd_phys = early_pgtable_alloc();
 541         pgd_t *pgd = pgd_set_fixmap(pgd_phys);
 542
 543         map_kernel(pgd);
 544         map_mem(pgd);
 545
 546         /*
 547          * We want to reuse the original swapper_pg_dir so we don't have to
 548          * communicate the new address to non-coherent secondaries in
 549          * secondary_entry, and so cpu_switch_mm can generate the address with
 550          * adrp+add rather than a load from some global variable.
 551          *
 552          * To do this we need to go via a temporary pgd.
 553          */
 554         cpu_replace_ttbr1(__va(pgd_phys));
 555         memcpy(swapper_pg_dir, pgd, PAGE_SIZE);
 556         cpu_replace_ttbr1(swapper_pg_dir);
 557
 558         pgd_clear_fixmap();
 559         memblock_free(pgd_phys, PAGE_SIZE);
 560
 561         /*
 562          * We only reuse the PGD from the swapper_pg_dir, not the pud + pmd
 563          * allocated with it.
 564          */
 565         memblock_free(__pa(swapper_pg_dir) + PAGE_SIZE,
 566                       SWAPPER_DIR_SIZE - PAGE_SIZE);
 567
 568         bootmem_init();
 569 }
 570
 571 /*
 572  * Check whether a kernel address is valid (derived from arch/x86/).
 573  */
 574 int kern_addr_valid(unsigned long addr)
 575 {
 576         pgd_t *pgd;
 577         pud_t *pud;
 578         pmd_t *pmd;
 579         pte_t *pte;
 580
 581         if ((((long)addr) >> VA_BITS) != -1UL)
 582                 return 0;
 583
 584         pgd = pgd_offset_k(addr);
 585         if (pgd_none(*pgd))
 586                 return 0;
 587
 588         pud = pud_offset(pgd, addr);
 589         if (pud_none(*pud))
 590                 return 0;
 591
 592         if (pud_sect(*pud))
 593                 return pfn_valid(pud_pfn(*pud));
 594
 595         pmd = pmd_offset(pud, addr);
 596         if (pmd_none(*pmd))
 597                 return 0;
 598
 599         if (pmd_sect(*pmd))
 600                 return pfn_valid(pmd_pfn(*pmd));
 601
 602         pte = pte_offset_kernel(pmd, addr);
 603         if (pte_none(*pte))
 604                 return 0;
 605
 606         return pfn_valid(pte_pfn(*pte));
 607 }
 608 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 609 #if !ARM64_SWAPPER_USES_SECTION_MAPS
 610 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 611 {
 612         return vmemmap_populate_basepages(start, end, node);
 613 }
 614 #else   /* !ARM64_SWAPPER_USES_SECTION_MAPS */
 615 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 616 {
 617         unsigned long addr = start;
 618         unsigned long next;
 619         pgd_t *pgd;
 620         pud_t *pud;
 621         pmd_t *pmd;
 622
 623         do {
 624                 next = pmd_addr_end(addr, end);
 625
 626                 pgd = vmemmap_pgd_populate(addr, node);
 627                 if (!pgd)
 628                         return -ENOMEM;
 629
 630                 pud = vmemmap_pud_populate(pgd, addr, node);
 631                 if (!pud)
 632                         return -ENOMEM;
 633
 634                 pmd = pmd_offset(pud, addr);
 635                 if (pmd_none(*pmd)) {
 636                         void *p = NULL;
 637
 638                         p = vmemmap_alloc_block_buf(PMD_SIZE, node);
 639                         if (!p)
 640                                 return -ENOMEM;
 641
 642                         set_pmd(pmd, __pmd(__pa(p) | PROT_SECT_NORMAL));
 643                 } else
 644                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 645         } while (addr = next, addr != end);
 646
 647         return 0;
 648 }
 649 #endif  /* CONFIG_ARM64_64K_PAGES */
 650 void vmemmap_free(unsigned long start, unsigned long end)
 651 {
 652 }
 653 #endif  /* CONFIG_SPARSEMEM_VMEMMAP */
 654
 655 static inline pud_t * fixmap_pud(unsigned long addr)
 656 {
 657         pgd_t *pgd = pgd_offset_k(addr);
 658
 659         BUG_ON(pgd_none(*pgd) || pgd_bad(*pgd));
 660
 661         return pud_offset_kimg(pgd, addr);
 662 }
 663
 664 static inline pmd_t * fixmap_pmd(unsigned long addr)
 665 {
 666         pud_t *pud = fixmap_pud(addr);
 667
 668         BUG_ON(pud_none(*pud) || pud_bad(*pud));
 669
 670         return pmd_offset_kimg(pud, addr);
 671 }
 672
 673 static inline pte_t * fixmap_pte(unsigned long addr)
 674 {
 675         return &bm_pte[pte_index(addr)];
 676 }
 677
 678 void __init early_fixmap_init(void)
 679 {
 680         pgd_t *pgd;
 681         pud_t *pud;
 682         pmd_t *pmd;
 683         unsigned long addr = FIXADDR_START;
 684
 685         pgd = pgd_offset_k(addr);
 686         if (CONFIG_PGTABLE_LEVELS > 3 &&
 687             !(pgd_none(*pgd) || pgd_page_paddr(*pgd) == __pa(bm_pud))) {
 688                 /*
 689                  * We only end up here if the kernel mapping and the fixmap
 690                  * share the top level pgd entry, which should only happen on
 691                  * 16k/4 levels configurations.
 692                  */
 693                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 694                 pud = pud_offset_kimg(pgd, addr);
 695         } else {
 696                 pgd_populate(&init_mm, pgd, bm_pud);
 697                 pud = fixmap_pud(addr);
 698         }
 699         pud_populate(&init_mm, pud, bm_pmd);
 700         pmd = fixmap_pmd(addr);
 701         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 702
 703         /*
 704          * The boot-ioremap range spans multiple pmds, for which
 705          * we are not prepared:
 706          */
 707         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 708                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 709
 710         if ((pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
 711              || pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 712                 WARN_ON(1);
 713                 pr_warn("pmd %p != %p, %p\n",
 714                         pmd, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
 715                         fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
 716                 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 717                         fix_to_virt(FIX_BTMAP_BEGIN));
 718                 pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 719                         fix_to_virt(FIX_BTMAP_END));
 720
 721                 pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 722                 pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
 723         }
 724 }
 725
 726 void __set_fixmap(enum fixed_addresses idx,
 727                                phys_addr_t phys, pgprot_t flags)
 728 {
 729         unsigned long addr = __fix_to_virt(idx);
 730         pte_t *pte;
 731
 732         BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
 733
 734         pte = fixmap_pte(addr);
 735
 736         if (pgprot_val(flags)) {
 737                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 738         } else {
 739                 pte_clear(&init_mm, addr, pte);
 740                 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
 741         }
 742 }
 743
 744 void *__init __fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
 745 {
 746         const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
 747         int offset;
 748         void *dt_virt;
 749
 750         /*
 751          * Check whether the physical FDT address is set and meets the minimum
 752          * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
 753          * at least 8 bytes so that we can always access the magic and size
 754          * fields of the FDT header after mapping the first chunk, double check
 755          * here if that is indeed the case.
 756          */
 757         BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
 758         if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
 759                 return NULL;
 760
 761         /*
 762          * Make sure that the FDT region can be mapped without the need to
 763          * allocate additional translation table pages, so that it is safe
 764          * to call create_mapping_noalloc() this early.
 765          *
 766          * On 64k pages, the FDT will be mapped using PTEs, so we need to
 767          * be in the same PMD as the rest of the fixmap.
 768          * On 4k pages, we'll use section mappings for the FDT so we only
 769          * have to be in the same PUD.
 770          */
 771         BUILD_BUG_ON(dt_virt_base % SZ_2M);
 772
 773         BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
 774                      __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
 775
 776         offset = dt_phys % SWAPPER_BLOCK_SIZE;
 777         dt_virt = (void *)dt_virt_base + offset;
 778
 779         /* map the first chunk so we can read the size from the header */
 780         create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
 781                         dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
 782
 783         if (fdt_magic(dt_virt) != FDT_MAGIC)
 784                 return NULL;
 785
 786         *size = fdt_totalsize(dt_virt);
 787         if (*size > MAX_FDT_SIZE)
 788                 return NULL;
 789
 790         if (offset + *size > SWAPPER_BLOCK_SIZE)
 791                 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
 792                                round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
 793
 794         return dt_virt;
 795 }
 796
 797 void *__init fixmap_remap_fdt(phys_addr_t dt_phys)
 798 {
 799         void *dt_virt;
 800         int size;
 801
 802         dt_virt = __fixmap_remap_fdt(dt_phys, &size, PAGE_KERNEL_RO);
 803         if (!dt_virt)
 804                 return NULL;
 805
 806         memblock_reserve(dt_phys, size);
 807         return dt_virt;
 808 }
 809
 810 int __init arch_ioremap_pud_supported(void)
 811 {
 812         /* only 4k granule supports level 1 block mappings */
 813         return IS_ENABLED(CONFIG_ARM64_4K_PAGES);
 814 }
 815
 816 int __init arch_ioremap_pmd_supported(void)
 817 {
 818         return 1;
 819 }
 820
 821 int pud_set_huge(pud_t *pud, phys_addr_t phys, pgprot_t prot)
 822 {
 823         BUG_ON(phys & ~PUD_MASK);
 824         set_pud(pud, __pud(phys | PUD_TYPE_SECT | pgprot_val(mk_sect_prot(prot))));
 825         return 1;
 826 }
 827
 828 int pmd_set_huge(pmd_t *pmd, phys_addr_t phys, pgprot_t prot)
 829 {
 830         BUG_ON(phys & ~PMD_MASK);
 831         set_pmd(pmd, __pmd(phys | PMD_TYPE_SECT | pgprot_val(mk_sect_prot(prot))));
 832         return 1;
 833 }
 834
 835 int pud_clear_huge(pud_t *pud)
 836 {
 837         if (!pud_sect(*pud))
 838                 return 0;
 839         pud_clear(pud);
 840         return 1;
 841 }
 842
 843 int pmd_clear_huge(pmd_t *pmd)
 844 {
 845         if (!pmd_sect(*pmd))
 846                 return 0;
 847         pmd_clear(pmd);
 848         return 1;
 849 }