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/cache.h>
  21 #include <linux/export.h>
  22 #include <linux/kernel.h>
  23 #include <linux/errno.h>
  24 #include <linux/init.h>
  25 #include <linux/libfdt.h>
  26 #include <linux/mman.h>
  27 #include <linux/nodemask.h>
  28 #include <linux/memblock.h>
  29 #include <linux/fs.h>
  30 #include <linux/io.h>
  31 #include <linux/mm.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 #include <asm/ptdump.h>
  45
  46 u64 idmap_t0sz = TCR_T0SZ(VA_BITS);
  47
  48 u64 kimage_voffset __ro_after_init;
  49 EXPORT_SYMBOL(kimage_voffset);
  50
  51 /*
  52  * Empty_zero_page is a special page that is used for zero-initialized data
  53  * and COW.
  54  */
  55 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
  56 EXPORT_SYMBOL(empty_zero_page);
  57
  58 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
  59 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused;
  60 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused;
  61
  62 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
  63                               unsigned long size, pgprot_t vma_prot)
  64 {
  65         if (!pfn_valid(pfn))
  66                 return pgprot_noncached(vma_prot);
  67         else if (file->f_flags & O_SYNC)
  68                 return pgprot_writecombine(vma_prot);
  69         return vma_prot;
  70 }
  71 EXPORT_SYMBOL(phys_mem_access_prot);
  72
  73 static phys_addr_t __init early_pgtable_alloc(void)
  74 {
  75         phys_addr_t phys;
  76         void *ptr;
  77
  78         phys = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
  79
  80         /*
  81          * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
  82          * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
  83          * any level of table.
  84          */
  85         ptr = pte_set_fixmap(phys);
  86
  87         memset(ptr, 0, PAGE_SIZE);
  88
  89         /*
  90          * Implicit barriers also ensure the zeroed page is visible to the page
  91          * table walker
  92          */
  93         pte_clear_fixmap();
  94
  95         return phys;
  96 }
  97
  98 static bool pgattr_change_is_safe(u64 old, u64 new)
  99 {
 100         /*
 101          * The following mapping attributes may be updated in live
 102          * kernel mappings without the need for break-before-make.
 103          */
 104         static const pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE;
 105
 106         return old  == 0 || new  == 0 || ((old ^ new) & ~mask) == 0;
 107 }
 108
 109 static void alloc_init_pte(pmd_t *pmd, unsigned long addr,
 110                                   unsigned long end, phys_addr_t phys,
 111                                   pgprot_t prot,
 112                                   phys_addr_t (*pgtable_alloc)(void))
 113 {
 114         pte_t *pte;
 115
 116         BUG_ON(pmd_sect(*pmd));
 117         if (pmd_none(*pmd)) {
 118                 phys_addr_t pte_phys;
 119                 BUG_ON(!pgtable_alloc);
 120                 pte_phys = pgtable_alloc();
 121                 pte = pte_set_fixmap(pte_phys);
 122                 __pmd_populate(pmd, pte_phys, PMD_TYPE_TABLE);
 123                 pte_clear_fixmap();
 124         }
 125         BUG_ON(pmd_bad(*pmd));
 126
 127         pte = pte_set_fixmap_offset(pmd, addr);
 128         do {
 129                 pte_t old_pte = *pte;
 130
 131                 set_pte(pte, pfn_pte(__phys_to_pfn(phys), prot));
 132
 133                 /*
 134                  * After the PTE entry has been populated once, we
 135                  * only allow updates to the permission attributes.
 136                  */
 137                 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte), pte_val(*pte)));
 138
 139                 phys += PAGE_SIZE;
 140         } while (pte++, addr += PAGE_SIZE, addr != end);
 141
 142         pte_clear_fixmap();
 143 }
 144
 145 static void alloc_init_pmd(pud_t *pud, unsigned long addr, unsigned long end,
 146                                   phys_addr_t phys, pgprot_t prot,
 147                                   phys_addr_t (*pgtable_alloc)(void),
 148                                   bool page_mappings_only)
 149 {
 150         pmd_t *pmd;
 151         unsigned long next;
 152
 153         /*
 154          * Check for initial section mappings in the pgd/pud and remove them.
 155          */
 156         BUG_ON(pud_sect(*pud));
 157         if (pud_none(*pud)) {
 158                 phys_addr_t pmd_phys;
 159                 BUG_ON(!pgtable_alloc);
 160                 pmd_phys = pgtable_alloc();
 161                 pmd = pmd_set_fixmap(pmd_phys);
 162                 __pud_populate(pud, pmd_phys, PUD_TYPE_TABLE);
 163                 pmd_clear_fixmap();
 164         }
 165         BUG_ON(pud_bad(*pud));
 166
 167         pmd = pmd_set_fixmap_offset(pud, addr);
 168         do {
 169                 pmd_t old_pmd = *pmd;
 170
 171                 next = pmd_addr_end(addr, end);
 172
 173                 /* try section mapping first */
 174                 if (((addr | next | phys) & ~SECTION_MASK) == 0 &&
 175                       !page_mappings_only) {
 176                         pmd_set_huge(pmd, phys, prot);
 177
 178                         /*
 179                          * After the PMD entry has been populated once, we
 180                          * only allow updates to the permission attributes.
 181                          */
 182                         BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
 183                                                       pmd_val(*pmd)));
 184                 } else {
 185                         alloc_init_pte(pmd, addr, next, phys,
 186                                        prot, pgtable_alloc);
 187
 188                         BUG_ON(pmd_val(old_pmd) != 0 &&
 189                                pmd_val(old_pmd) != pmd_val(*pmd));
 190                 }
 191                 phys += next - addr;
 192         } while (pmd++, addr = next, addr != end);
 193
 194         pmd_clear_fixmap();
 195 }
 196
 197 static inline bool use_1G_block(unsigned long addr, unsigned long next,
 198                         unsigned long phys)
 199 {
 200         if (PAGE_SHIFT != 12)
 201                 return false;
 202
 203         if (((addr | next | phys) & ~PUD_MASK) != 0)
 204                 return false;
 205
 206         return true;
 207 }
 208
 209 static void alloc_init_pud(pgd_t *pgd, unsigned long addr, unsigned long end,
 210                                   phys_addr_t phys, pgprot_t prot,
 211                                   phys_addr_t (*pgtable_alloc)(void),
 212                                   bool page_mappings_only)
 213 {
 214         pud_t *pud;
 215         unsigned long next;
 216
 217         if (pgd_none(*pgd)) {
 218                 phys_addr_t pud_phys;
 219                 BUG_ON(!pgtable_alloc);
 220                 pud_phys = pgtable_alloc();
 221                 __pgd_populate(pgd, pud_phys, PUD_TYPE_TABLE);
 222         }
 223         BUG_ON(pgd_bad(*pgd));
 224
 225         pud = pud_set_fixmap_offset(pgd, addr);
 226         do {
 227                 pud_t old_pud = *pud;
 228
 229                 next = pud_addr_end(addr, end);
 230
 231                 /*
 232                  * For 4K granule only, attempt to put down a 1GB block
 233                  */
 234                 if (use_1G_block(addr, next, phys) && !page_mappings_only) {
 235                         pud_set_huge(pud, phys, prot);
 236
 237                         /*
 238                          * After the PUD entry has been populated once, we
 239                          * only allow updates to the permission attributes.
 240                          */
 241                         BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
 242                                                       pud_val(*pud)));
 243                 } else {
 244                         alloc_init_pmd(pud, addr, next, phys, prot,
 245                                        pgtable_alloc, page_mappings_only);
 246
 247                         BUG_ON(pud_val(old_pud) != 0 &&
 248                                pud_val(old_pud) != pud_val(*pud));
 249                 }
 250                 phys += next - addr;
 251         } while (pud++, addr = next, addr != end);
 252
 253         pud_clear_fixmap();
 254 }
 255
 256 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
 257                                  unsigned long virt, phys_addr_t size,
 258                                  pgprot_t prot,
 259                                  phys_addr_t (*pgtable_alloc)(void),
 260                                  bool page_mappings_only)
 261 {
 262         unsigned long addr, length, end, next;
 263         pgd_t *pgd = pgd_offset_raw(pgdir, virt);
 264
 265         /*
 266          * If the virtual and physical address don't have the same offset
 267          * within a page, we cannot map the region as the caller expects.
 268          */
 269         if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
 270                 return;
 271
 272         phys &= PAGE_MASK;
 273         addr = virt & PAGE_MASK;
 274         length = PAGE_ALIGN(size + (virt & ~PAGE_MASK));
 275
 276         end = addr + length;
 277         do {
 278                 next = pgd_addr_end(addr, end);
 279                 alloc_init_pud(pgd, addr, next, phys, prot, pgtable_alloc,
 280                                page_mappings_only);
 281                 phys += next - addr;
 282         } while (pgd++, addr = next, addr != end);
 283 }
 284
 285 static phys_addr_t pgd_pgtable_alloc(void)
 286 {
 287         void *ptr = (void *)__get_free_page(PGALLOC_GFP);
 288         if (!ptr || !pgtable_page_ctor(virt_to_page(ptr)))
 289                 BUG();
 290
 291         /* Ensure the zeroed page is visible to the page table walker */
 292         dsb(ishst);
 293         return __pa(ptr);
 294 }
 295
 296 /*
 297  * This function can only be used to modify existing table entries,
 298  * without allocating new levels of table. Note that this permits the
 299  * creation of new section or page entries.
 300  */
 301 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
 302                                   phys_addr_t size, pgprot_t prot)
 303 {
 304         if (virt < VMALLOC_START) {
 305                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 306                         &phys, virt);
 307                 return;
 308         }
 309         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, false);
 310 }
 311
 312 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
 313                                unsigned long virt, phys_addr_t size,
 314                                pgprot_t prot, bool page_mappings_only)
 315 {
 316         BUG_ON(mm == &init_mm);
 317
 318         __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
 319                              pgd_pgtable_alloc, page_mappings_only);
 320 }
 321
 322 static void update_mapping_prot(phys_addr_t phys, unsigned long virt,
 323                                 phys_addr_t size, pgprot_t prot)
 324 {
 325         if (virt < VMALLOC_START) {
 326                 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n",
 327                         &phys, virt);
 328                 return;
 329         }
 330
 331         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot,
 332                              NULL, debug_pagealloc_enabled());
 333
 334         /* flush the TLBs after updating live kernel mappings */
 335         flush_tlb_kernel_range(virt, virt + size);
 336 }
 337
 338 static void __init __map_memblock(pgd_t *pgd, phys_addr_t start, phys_addr_t end)
 339 {
 340         phys_addr_t kernel_start = __pa_symbol(_text);
 341         phys_addr_t kernel_end = __pa_symbol(__init_begin);
 342
 343         /*
 344          * Take care not to create a writable alias for the
 345          * read-only text and rodata sections of the kernel image.
 346          */
 347
 348         /* No overlap with the kernel text/rodata */
 349         if (end < kernel_start || start >= kernel_end) {
 350                 __create_pgd_mapping(pgd, start, __phys_to_virt(start),
 351                                      end - start, PAGE_KERNEL,
 352                                      early_pgtable_alloc,
 353                                      debug_pagealloc_enabled());
 354                 return;
 355         }
 356
 357         /*
 358          * This block overlaps the kernel text/rodata mappings.
 359          * Map the portion(s) which don't overlap.
 360          */
 361         if (start < kernel_start)
 362                 __create_pgd_mapping(pgd, start,
 363                                      __phys_to_virt(start),
 364                                      kernel_start - start, PAGE_KERNEL,
 365                                      early_pgtable_alloc,
 366                                      debug_pagealloc_enabled());
 367         if (kernel_end < end)
 368                 __create_pgd_mapping(pgd, kernel_end,
 369                                      __phys_to_virt(kernel_end),
 370                                      end - kernel_end, PAGE_KERNEL,
 371                                      early_pgtable_alloc,
 372                                      debug_pagealloc_enabled());
 373
 374         /*
 375          * Map the linear alias of the [_text, __init_begin) interval
 376          * as non-executable now, and remove the write permission in
 377          * mark_linear_text_alias_ro() below (which will be called after
 378          * alternative patching has completed). This makes the contents
 379          * of the region accessible to subsystems such as hibernate,
 380          * but protects it from inadvertent modification or execution.
 381          */
 382         __create_pgd_mapping(pgd, kernel_start, __phys_to_virt(kernel_start),
 383                              kernel_end - kernel_start, PAGE_KERNEL,
 384                              early_pgtable_alloc, debug_pagealloc_enabled());
 385 }
 386
 387 void __init mark_linear_text_alias_ro(void)
 388 {
 389         /*
 390          * Remove the write permissions from the linear alias of .text/.rodata
 391          */
 392         update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text),
 393                             (unsigned long)__init_begin - (unsigned long)_text,
 394                             PAGE_KERNEL_RO);
 395 }
 396
 397 static void __init map_mem(pgd_t *pgd)
 398 {
 399         struct memblock_region *reg;
 400
 401         /* map all the memory banks */
 402         for_each_memblock(memory, reg) {
 403                 phys_addr_t start = reg->base;
 404                 phys_addr_t end = start + reg->size;
 405
 406                 if (start >= end)
 407                         break;
 408                 if (memblock_is_nomap(reg))
 409                         continue;
 410
 411                 __map_memblock(pgd, start, end);
 412         }
 413 }
 414
 415 void mark_rodata_ro(void)
 416 {
 417         unsigned long section_size;
 418
 419         /*
 420          * mark .rodata as read only. Use __init_begin rather than __end_rodata
 421          * to cover NOTES and EXCEPTION_TABLE.
 422          */
 423         section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
 424         update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata,
 425                             section_size, PAGE_KERNEL_RO);
 426
 427         debug_checkwx();
 428 }
 429
 430 static void __init map_kernel_segment(pgd_t *pgd, void *va_start, void *va_end,
 431                                       pgprot_t prot, struct vm_struct *vma)
 432 {
 433         phys_addr_t pa_start = __pa_symbol(va_start);
 434         unsigned long size = va_end - va_start;
 435
 436         BUG_ON(!PAGE_ALIGNED(pa_start));
 437         BUG_ON(!PAGE_ALIGNED(size));
 438
 439         __create_pgd_mapping(pgd, pa_start, (unsigned long)va_start, size, prot,
 440                              early_pgtable_alloc, debug_pagealloc_enabled());
 441
 442         vma->addr       = va_start;
 443         vma->phys_addr  = pa_start;
 444         vma->size       = size;
 445         vma->flags      = VM_MAP;
 446         vma->caller     = __builtin_return_address(0);
 447
 448         vm_area_add_early(vma);
 449 }
 450
 451 static int __init parse_rodata(char *arg)
 452 {
 453         return strtobool(arg, &rodata_enabled);
 454 }
 455 early_param("rodata", parse_rodata);
 456
 457 /*
 458  * Create fine-grained mappings for the kernel.
 459  */
 460 static void __init map_kernel(pgd_t *pgd)
 461 {
 462         static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext,
 463                                 vmlinux_initdata, vmlinux_data;
 464
 465         /*
 466          * External debuggers may need to write directly to the text
 467          * mapping to install SW breakpoints. Allow this (only) when
 468          * explicitly requested with rodata=off.
 469          */
 470         pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
 471
 472         map_kernel_segment(pgd, _text, _etext, text_prot, &vmlinux_text);
 473         map_kernel_segment(pgd, __start_rodata, __inittext_begin, PAGE_KERNEL,
 474                            &vmlinux_rodata);
 475         map_kernel_segment(pgd, __inittext_begin, __inittext_end, text_prot,
 476                            &vmlinux_inittext);
 477         map_kernel_segment(pgd, __initdata_begin, __initdata_end, PAGE_KERNEL,
 478                            &vmlinux_initdata);
 479         map_kernel_segment(pgd, _data, _end, PAGE_KERNEL, &vmlinux_data);
 480
 481         if (!pgd_val(*pgd_offset_raw(pgd, FIXADDR_START))) {
 482                 /*
 483                  * The fixmap falls in a separate pgd to the kernel, and doesn't
 484                  * live in the carveout for the swapper_pg_dir. We can simply
 485                  * re-use the existing dir for the fixmap.
 486                  */
 487                 set_pgd(pgd_offset_raw(pgd, FIXADDR_START),
 488                         *pgd_offset_k(FIXADDR_START));
 489         } else if (CONFIG_PGTABLE_LEVELS > 3) {
 490                 /*
 491                  * The fixmap shares its top level pgd entry with the kernel
 492                  * mapping. This can really only occur when we are running
 493                  * with 16k/4 levels, so we can simply reuse the pud level
 494                  * entry instead.
 495                  */
 496                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 497                 set_pud(pud_set_fixmap_offset(pgd, FIXADDR_START),
 498                         __pud(__pa_symbol(bm_pmd) | PUD_TYPE_TABLE));
 499                 pud_clear_fixmap();
 500         } else {
 501                 BUG();
 502         }
 503
 504         kasan_copy_shadow(pgd);
 505 }
 506
 507 /*
 508  * paging_init() sets up the page tables, initialises the zone memory
 509  * maps and sets up the zero page.
 510  */
 511 void __init paging_init(void)
 512 {
 513         phys_addr_t pgd_phys = early_pgtable_alloc();
 514         pgd_t *pgd = pgd_set_fixmap(pgd_phys);
 515
 516         map_kernel(pgd);
 517         map_mem(pgd);
 518
 519         /*
 520          * We want to reuse the original swapper_pg_dir so we don't have to
 521          * communicate the new address to non-coherent secondaries in
 522          * secondary_entry, and so cpu_switch_mm can generate the address with
 523          * adrp+add rather than a load from some global variable.
 524          *
 525          * To do this we need to go via a temporary pgd.
 526          */
 527         cpu_replace_ttbr1(__va(pgd_phys));
 528         memcpy(swapper_pg_dir, pgd, PGD_SIZE);
 529         cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
 530
 531         pgd_clear_fixmap();
 532         memblock_free(pgd_phys, PAGE_SIZE);
 533
 534         /*
 535          * We only reuse the PGD from the swapper_pg_dir, not the pud + pmd
 536          * allocated with it.
 537          */
 538         memblock_free(__pa_symbol(swapper_pg_dir) + PAGE_SIZE,
 539                       SWAPPER_DIR_SIZE - PAGE_SIZE);
 540 }
 541
 542 /*
 543  * Check whether a kernel address is valid (derived from arch/x86/).
 544  */
 545 int kern_addr_valid(unsigned long addr)
 546 {
 547         pgd_t *pgd;
 548         pud_t *pud;
 549         pmd_t *pmd;
 550         pte_t *pte;
 551
 552         if ((((long)addr) >> VA_BITS) != -1UL)
 553                 return 0;
 554
 555         pgd = pgd_offset_k(addr);
 556         if (pgd_none(*pgd))
 557                 return 0;
 558
 559         pud = pud_offset(pgd, addr);
 560         if (pud_none(*pud))
 561                 return 0;
 562
 563         if (pud_sect(*pud))
 564                 return pfn_valid(pud_pfn(*pud));
 565
 566         pmd = pmd_offset(pud, addr);
 567         if (pmd_none(*pmd))
 568                 return 0;
 569
 570         if (pmd_sect(*pmd))
 571                 return pfn_valid(pmd_pfn(*pmd));
 572
 573         pte = pte_offset_kernel(pmd, addr);
 574         if (pte_none(*pte))
 575                 return 0;
 576
 577         return pfn_valid(pte_pfn(*pte));
 578 }
 579 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 580 #if !ARM64_SWAPPER_USES_SECTION_MAPS
 581 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 582 {
 583         return vmemmap_populate_basepages(start, end, node);
 584 }
 585 #else   /* !ARM64_SWAPPER_USES_SECTION_MAPS */
 586 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 587 {
 588         unsigned long addr = start;
 589         unsigned long next;
 590         pgd_t *pgd;
 591         pud_t *pud;
 592         pmd_t *pmd;
 593
 594         do {
 595                 next = pmd_addr_end(addr, end);
 596
 597                 pgd = vmemmap_pgd_populate(addr, node);
 598                 if (!pgd)
 599                         return -ENOMEM;
 600
 601                 pud = vmemmap_pud_populate(pgd, addr, node);
 602                 if (!pud)
 603                         return -ENOMEM;
 604
 605                 pmd = pmd_offset(pud, addr);
 606                 if (pmd_none(*pmd)) {
 607                         void *p = NULL;
 608
 609                         p = vmemmap_alloc_block_buf(PMD_SIZE, node);
 610                         if (!p)
 611                                 return -ENOMEM;
 612
 613                         set_pmd(pmd, __pmd(__pa(p) | PROT_SECT_NORMAL));
 614                 } else
 615                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 616         } while (addr = next, addr != end);
 617
 618         return 0;
 619 }
 620 #endif  /* CONFIG_ARM64_64K_PAGES */
 621 void vmemmap_free(unsigned long start, unsigned long end)
 622 {
 623 }
 624 #endif  /* CONFIG_SPARSEMEM_VMEMMAP */
 625
 626 static inline pud_t * fixmap_pud(unsigned long addr)
 627 {
 628         pgd_t *pgd = pgd_offset_k(addr);
 629
 630         BUG_ON(pgd_none(*pgd) || pgd_bad(*pgd));
 631
 632         return pud_offset_kimg(pgd, addr);
 633 }
 634
 635 static inline pmd_t * fixmap_pmd(unsigned long addr)
 636 {
 637         pud_t *pud = fixmap_pud(addr);
 638
 639         BUG_ON(pud_none(*pud) || pud_bad(*pud));
 640
 641         return pmd_offset_kimg(pud, addr);
 642 }
 643
 644 static inline pte_t * fixmap_pte(unsigned long addr)
 645 {
 646         return &bm_pte[pte_index(addr)];
 647 }
 648
 649 /*
 650  * The p*d_populate functions call virt_to_phys implicitly so they can't be used
 651  * directly on kernel symbols (bm_p*d). This function is called too early to use
 652  * lm_alias so __p*d_populate functions must be used to populate with the
 653  * physical address from __pa_symbol.
 654  */
 655 void __init early_fixmap_init(void)
 656 {
 657         pgd_t *pgd;
 658         pud_t *pud;
 659         pmd_t *pmd;
 660         unsigned long addr = FIXADDR_START;
 661
 662         pgd = pgd_offset_k(addr);
 663         if (CONFIG_PGTABLE_LEVELS > 3 &&
 664             !(pgd_none(*pgd) || pgd_page_paddr(*pgd) == __pa_symbol(bm_pud))) {
 665                 /*
 666                  * We only end up here if the kernel mapping and the fixmap
 667                  * share the top level pgd entry, which should only happen on
 668                  * 16k/4 levels configurations.
 669                  */
 670                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 671                 pud = pud_offset_kimg(pgd, addr);
 672         } else {
 673                 if (pgd_none(*pgd))
 674                         __pgd_populate(pgd, __pa_symbol(bm_pud), PUD_TYPE_TABLE);
 675                 pud = fixmap_pud(addr);
 676         }
 677         if (pud_none(*pud))
 678                 __pud_populate(pud, __pa_symbol(bm_pmd), PMD_TYPE_TABLE);
 679         pmd = fixmap_pmd(addr);
 680         __pmd_populate(pmd, __pa_symbol(bm_pte), PMD_TYPE_TABLE);
 681
 682         /*
 683          * The boot-ioremap range spans multiple pmds, for which
 684          * we are not prepared:
 685          */
 686         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 687                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 688
 689         if ((pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
 690              || pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 691                 WARN_ON(1);
 692                 pr_warn("pmd %p != %p, %p\n",
 693                         pmd, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
 694                         fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
 695                 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 696                         fix_to_virt(FIX_BTMAP_BEGIN));
 697                 pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 698                         fix_to_virt(FIX_BTMAP_END));
 699
 700                 pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 701                 pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
 702         }
 703 }
 704
 705 void __set_fixmap(enum fixed_addresses idx,
 706                                phys_addr_t phys, pgprot_t flags)
 707 {
 708         unsigned long addr = __fix_to_virt(idx);
 709         pte_t *pte;
 710
 711         BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
 712
 713         pte = fixmap_pte(addr);
 714
 715         if (pgprot_val(flags)) {
 716                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 717         } else {
 718                 pte_clear(&init_mm, addr, pte);
 719                 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
 720         }
 721 }
 722
 723 void *__init __fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
 724 {
 725         const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
 726         int offset;
 727         void *dt_virt;
 728
 729         /*
 730          * Check whether the physical FDT address is set and meets the minimum
 731          * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
 732          * at least 8 bytes so that we can always access the magic and size
 733          * fields of the FDT header after mapping the first chunk, double check
 734          * here if that is indeed the case.
 735          */
 736         BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
 737         if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
 738                 return NULL;
 739
 740         /*
 741          * Make sure that the FDT region can be mapped without the need to
 742          * allocate additional translation table pages, so that it is safe
 743          * to call create_mapping_noalloc() this early.
 744          *
 745          * On 64k pages, the FDT will be mapped using PTEs, so we need to
 746          * be in the same PMD as the rest of the fixmap.
 747          * On 4k pages, we'll use section mappings for the FDT so we only
 748          * have to be in the same PUD.
 749          */
 750         BUILD_BUG_ON(dt_virt_base % SZ_2M);
 751
 752         BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
 753                      __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
 754
 755         offset = dt_phys % SWAPPER_BLOCK_SIZE;
 756         dt_virt = (void *)dt_virt_base + offset;
 757
 758         /* map the first chunk so we can read the size from the header */
 759         create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
 760                         dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
 761
 762         if (fdt_magic(dt_virt) != FDT_MAGIC)
 763                 return NULL;
 764
 765         *size = fdt_totalsize(dt_virt);
 766         if (*size > MAX_FDT_SIZE)
 767                 return NULL;
 768
 769         if (offset + *size > SWAPPER_BLOCK_SIZE)
 770                 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
 771                                round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
 772
 773         return dt_virt;
 774 }
 775
 776 void *__init fixmap_remap_fdt(phys_addr_t dt_phys)
 777 {
 778         void *dt_virt;
 779         int size;
 780
 781         dt_virt = __fixmap_remap_fdt(dt_phys, &size, PAGE_KERNEL_RO);
 782         if (!dt_virt)
 783                 return NULL;
 784
 785         memblock_reserve(dt_phys, size);
 786         return dt_virt;
 787 }
 788
 789 int __init arch_ioremap_pud_supported(void)
 790 {
 791         /* only 4k granule supports level 1 block mappings */
 792         return IS_ENABLED(CONFIG_ARM64_4K_PAGES);
 793 }
 794
 795 int __init arch_ioremap_pmd_supported(void)
 796 {
 797         return 1;
 798 }
 799
 800 int pud_set_huge(pud_t *pud, phys_addr_t phys, pgprot_t prot)
 801 {
 802         BUG_ON(phys & ~PUD_MASK);
 803         set_pud(pud, __pud(phys | PUD_TYPE_SECT | pgprot_val(mk_sect_prot(prot))));
 804         return 1;
 805 }
 806
 807 int pmd_set_huge(pmd_t *pmd, phys_addr_t phys, pgprot_t prot)
 808 {
 809         BUG_ON(phys & ~PMD_MASK);
 810         set_pmd(pmd, __pmd(phys | PMD_TYPE_SECT | pgprot_val(mk_sect_prot(prot))));
 811         return 1;
 812 }
 813
 814 int pud_clear_huge(pud_t *pud)
 815 {
 816         if (!pud_sect(*pud))
 817                 return 0;
 818         pud_clear(pud);
 819         return 1;
 820 }
 821
 822 int pmd_clear_huge(pmd_t *pmd)
 823 {
 824         if (!pmd_sect(*pmd))
 825                 return 0;
 826         pmd_clear(pmd);
 827         return 1;
 828 }