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
  32 #include <asm/barrier.h>
  33 #include <asm/cputype.h>
  34 #include <asm/fixmap.h>
  35 #include <asm/kasan.h>
  36 #include <asm/kernel-pgtable.h>
  37 #include <asm/sections.h>
  38 #include <asm/setup.h>
  39 #include <asm/sizes.h>
  40 #include <asm/tlb.h>
  41 #include <asm/memblock.h>
  42 #include <asm/mmu_context.h>
  43
  44 u64 idmap_t0sz = TCR_T0SZ(VA_BITS);
  45
  46 u64 kimage_voffset __ro_after_init;
  47 EXPORT_SYMBOL(kimage_voffset);
  48
  49 /*
  50  * Empty_zero_page is a special page that is used for zero-initialized data
  51  * and COW.
  52  */
  53 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
  54 EXPORT_SYMBOL(empty_zero_page);
  55
  56 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
  57 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused;
  58 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused;
  59
  60 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
  61                               unsigned long size, pgprot_t vma_prot)
  62 {
  63         if (!pfn_valid(pfn))
  64                 return pgprot_noncached(vma_prot);
  65         else if (file->f_flags & O_SYNC)
  66                 return pgprot_writecombine(vma_prot);
  67         return vma_prot;
  68 }
  69 EXPORT_SYMBOL(phys_mem_access_prot);
  70
  71 static phys_addr_t __init early_pgtable_alloc(void)
  72 {
  73         phys_addr_t phys;
  74         void *ptr;
  75
  76         phys = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
  77
  78         /*
  79          * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
  80          * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
  81          * any level of table.
  82          */
  83         ptr = pte_set_fixmap(phys);
  84
  85         memset(ptr, 0, PAGE_SIZE);
  86
  87         /*
  88          * Implicit barriers also ensure the zeroed page is visible to the page
  89          * table walker
  90          */
  91         pte_clear_fixmap();
  92
  93         return phys;
  94 }
  95
  96 static bool pgattr_change_is_safe(u64 old, u64 new)
  97 {
  98         /*
  99          * The following mapping attributes may be updated in live
 100          * kernel mappings without the need for break-before-make.
 101          */
 102         static const pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE;
 103
 104         return old  == 0 || new  == 0 || ((old ^ new) & ~mask) == 0;
 105 }
 106
 107 static void alloc_init_pte(pmd_t *pmd, unsigned long addr,
 108                                   unsigned long end, unsigned long pfn,
 109                                   pgprot_t prot,
 110                                   phys_addr_t (*pgtable_alloc)(void))
 111 {
 112         pte_t *pte;
 113
 114         BUG_ON(pmd_sect(*pmd));
 115         if (pmd_none(*pmd)) {
 116                 phys_addr_t pte_phys;
 117                 BUG_ON(!pgtable_alloc);
 118                 pte_phys = pgtable_alloc();
 119                 pte = pte_set_fixmap(pte_phys);
 120                 __pmd_populate(pmd, pte_phys, PMD_TYPE_TABLE);
 121                 pte_clear_fixmap();
 122         }
 123         BUG_ON(pmd_bad(*pmd));
 124
 125         pte = pte_set_fixmap_offset(pmd, addr);
 126         do {
 127                 pte_t old_pte = *pte;
 128
 129                 set_pte(pte, pfn_pte(pfn, prot));
 130                 pfn++;
 131
 132                 /*
 133                  * After the PTE entry has been populated once, we
 134                  * only allow updates to the permission attributes.
 135                  */
 136                 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte), pte_val(*pte)));
 137
 138         } while (pte++, addr += PAGE_SIZE, addr != end);
 139
 140         pte_clear_fixmap();
 141 }
 142
 143 static void alloc_init_pmd(pud_t *pud, unsigned long addr, unsigned long end,
 144                                   phys_addr_t phys, pgprot_t prot,
 145                                   phys_addr_t (*pgtable_alloc)(void),
 146                                   bool allow_block_mappings)
 147 {
 148         pmd_t *pmd;
 149         unsigned long next;
 150
 151         /*
 152          * Check for initial section mappings in the pgd/pud and remove them.
 153          */
 154         BUG_ON(pud_sect(*pud));
 155         if (pud_none(*pud)) {
 156                 phys_addr_t pmd_phys;
 157                 BUG_ON(!pgtable_alloc);
 158                 pmd_phys = pgtable_alloc();
 159                 pmd = pmd_set_fixmap(pmd_phys);
 160                 __pud_populate(pud, pmd_phys, PUD_TYPE_TABLE);
 161                 pmd_clear_fixmap();
 162         }
 163         BUG_ON(pud_bad(*pud));
 164
 165         pmd = pmd_set_fixmap_offset(pud, addr);
 166         do {
 167                 pmd_t old_pmd = *pmd;
 168
 169                 next = pmd_addr_end(addr, end);
 170
 171                 /* try section mapping first */
 172                 if (((addr | next | phys) & ~SECTION_MASK) == 0 &&
 173                       allow_block_mappings) {
 174                         pmd_set_huge(pmd, phys, prot);
 175
 176                         /*
 177                          * After the PMD entry has been populated once, we
 178                          * only allow updates to the permission attributes.
 179                          */
 180                         BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
 181                                                       pmd_val(*pmd)));
 182                 } else {
 183                         alloc_init_pte(pmd, addr, next, __phys_to_pfn(phys),
 184                                        prot, pgtable_alloc);
 185
 186                         BUG_ON(pmd_val(old_pmd) != 0 &&
 187                                pmd_val(old_pmd) != pmd_val(*pmd));
 188                 }
 189                 phys += next - addr;
 190         } while (pmd++, addr = next, addr != end);
 191
 192         pmd_clear_fixmap();
 193 }
 194
 195 static inline bool use_1G_block(unsigned long addr, unsigned long next,
 196                         unsigned long phys)
 197 {
 198         if (PAGE_SHIFT != 12)
 199                 return false;
 200
 201         if (((addr | next | phys) & ~PUD_MASK) != 0)
 202                 return false;
 203
 204         return true;
 205 }
 206
 207 static void alloc_init_pud(pgd_t *pgd, unsigned long addr, unsigned long end,
 208                                   phys_addr_t phys, pgprot_t prot,
 209                                   phys_addr_t (*pgtable_alloc)(void),
 210                                   bool allow_block_mappings)
 211 {
 212         pud_t *pud;
 213         unsigned long next;
 214
 215         if (pgd_none(*pgd)) {
 216                 phys_addr_t pud_phys;
 217                 BUG_ON(!pgtable_alloc);
 218                 pud_phys = pgtable_alloc();
 219                 __pgd_populate(pgd, pud_phys, PUD_TYPE_TABLE);
 220         }
 221         BUG_ON(pgd_bad(*pgd));
 222
 223         pud = pud_set_fixmap_offset(pgd, addr);
 224         do {
 225                 pud_t old_pud = *pud;
 226
 227                 next = pud_addr_end(addr, end);
 228
 229                 /*
 230                  * For 4K granule only, attempt to put down a 1GB block
 231                  */
 232                 if (use_1G_block(addr, next, phys) && allow_block_mappings) {
 233                         pud_set_huge(pud, phys, prot);
 234
 235                         /*
 236                          * After the PUD entry has been populated once, we
 237                          * only allow updates to the permission attributes.
 238                          */
 239                         BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
 240                                                       pud_val(*pud)));
 241                 } else {
 242                         alloc_init_pmd(pud, addr, next, phys, prot,
 243                                        pgtable_alloc, allow_block_mappings);
 244
 245                         BUG_ON(pud_val(old_pud) != 0 &&
 246                                pud_val(old_pud) != pud_val(*pud));
 247                 }
 248                 phys += next - addr;
 249         } while (pud++, addr = next, addr != end);
 250
 251         pud_clear_fixmap();
 252 }
 253
 254 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
 255                                  unsigned long virt, phys_addr_t size,
 256                                  pgprot_t prot,
 257                                  phys_addr_t (*pgtable_alloc)(void),
 258                                  bool allow_block_mappings)
 259 {
 260         unsigned long addr, length, end, next;
 261         pgd_t *pgd = pgd_offset_raw(pgdir, virt);
 262
 263         /*
 264          * If the virtual and physical address don't have the same offset
 265          * within a page, we cannot map the region as the caller expects.
 266          */
 267         if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
 268                 return;
 269
 270         phys &= PAGE_MASK;
 271         addr = virt & PAGE_MASK;
 272         length = PAGE_ALIGN(size + (virt & ~PAGE_MASK));
 273
 274         end = addr + length;
 275         do {
 276                 next = pgd_addr_end(addr, end);
 277                 alloc_init_pud(pgd, addr, next, phys, prot, pgtable_alloc,
 278                                allow_block_mappings);
 279                 phys += next - addr;
 280         } while (pgd++, addr = next, addr != end);
 281 }
 282
 283 static phys_addr_t pgd_pgtable_alloc(void)
 284 {
 285         void *ptr = (void *)__get_free_page(PGALLOC_GFP);
 286         if (!ptr || !pgtable_page_ctor(virt_to_page(ptr)))
 287                 BUG();
 288
 289         /* Ensure the zeroed page is visible to the page table walker */
 290         dsb(ishst);
 291         return __pa(ptr);
 292 }
 293
 294 /*
 295  * This function can only be used to modify existing table entries,
 296  * without allocating new levels of table. Note that this permits the
 297  * creation of new section or page entries.
 298  */
 299 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
 300                                   phys_addr_t size, pgprot_t prot)
 301 {
 302         if (virt < VMALLOC_START) {
 303                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 304                         &phys, virt);
 305                 return;
 306         }
 307         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, true);
 308 }
 309
 310 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
 311                                unsigned long virt, phys_addr_t size,
 312                                pgprot_t prot, bool allow_block_mappings)
 313 {
 314         BUG_ON(mm == &init_mm);
 315
 316         __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
 317                              pgd_pgtable_alloc, allow_block_mappings);
 318 }
 319
 320 static void create_mapping_late(phys_addr_t phys, unsigned long virt,
 321                                   phys_addr_t size, pgprot_t prot)
 322 {
 323         if (virt < VMALLOC_START) {
 324                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 325                         &phys, virt);
 326                 return;
 327         }
 328
 329         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot,
 330                              NULL, !debug_pagealloc_enabled());
 331 }
 332
 333 static void __init __map_memblock(pgd_t *pgd, phys_addr_t start, phys_addr_t end)
 334 {
 335         unsigned long kernel_start = __pa(_text);
 336         unsigned long kernel_end = __pa(__init_begin);
 337
 338         /*
 339          * Take care not to create a writable alias for the
 340          * read-only text and rodata sections of the kernel image.
 341          */
 342
 343         /* No overlap with the kernel text/rodata */
 344         if (end < kernel_start || start >= kernel_end) {
 345                 __create_pgd_mapping(pgd, start, __phys_to_virt(start),
 346                                      end - start, PAGE_KERNEL,
 347                                      early_pgtable_alloc,
 348                                      !debug_pagealloc_enabled());
 349                 return;
 350         }
 351
 352         /*
 353          * This block overlaps the kernel text/rodata mappings.
 354          * Map the portion(s) which don't overlap.
 355          */
 356         if (start < kernel_start)
 357                 __create_pgd_mapping(pgd, start,
 358                                      __phys_to_virt(start),
 359                                      kernel_start - start, PAGE_KERNEL,
 360                                      early_pgtable_alloc,
 361                                      !debug_pagealloc_enabled());
 362         if (kernel_end < end)
 363                 __create_pgd_mapping(pgd, kernel_end,
 364                                      __phys_to_virt(kernel_end),
 365                                      end - kernel_end, PAGE_KERNEL,
 366                                      early_pgtable_alloc,
 367                                      !debug_pagealloc_enabled());
 368
 369         /*
 370          * Map the linear alias of the [_text, __init_begin) interval as
 371          * read-only/non-executable. This makes the contents of the
 372          * region accessible to subsystems such as hibernate, but
 373          * protects it from inadvertent modification or execution.
 374          */
 375         __create_pgd_mapping(pgd, kernel_start, __phys_to_virt(kernel_start),
 376                              kernel_end - kernel_start, PAGE_KERNEL_RO,
 377                              early_pgtable_alloc, !debug_pagealloc_enabled());
 378 }
 379
 380 static void __init map_mem(pgd_t *pgd)
 381 {
 382         struct memblock_region *reg;
 383
 384         /* map all the memory banks */
 385         for_each_memblock(memory, reg) {
 386                 phys_addr_t start = reg->base;
 387                 phys_addr_t end = start + reg->size;
 388
 389                 if (start >= end)
 390                         break;
 391                 if (memblock_is_nomap(reg))
 392                         continue;
 393
 394                 __map_memblock(pgd, start, end);
 395         }
 396 }
 397
 398 void mark_rodata_ro(void)
 399 {
 400         unsigned long section_size;
 401
 402         section_size = (unsigned long)_etext - (unsigned long)_text;
 403         create_mapping_late(__pa(_text), (unsigned long)_text,
 404                             section_size, PAGE_KERNEL_ROX);
 405         /*
 406          * mark .rodata as read only. Use __init_begin rather than __end_rodata
 407          * to cover NOTES and EXCEPTION_TABLE.
 408          */
 409         section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
 410         create_mapping_late(__pa(__start_rodata), (unsigned long)__start_rodata,
 411                             section_size, PAGE_KERNEL_RO);
 412
 413         /* flush the TLBs after updating live kernel mappings */
 414         flush_tlb_all();
 415 }
 416
 417 static void __init map_kernel_segment(pgd_t *pgd, void *va_start, void *va_end,
 418                                       pgprot_t prot, struct vm_struct *vma)
 419 {
 420         phys_addr_t pa_start = __pa(va_start);
 421         unsigned long size = va_end - va_start;
 422
 423         BUG_ON(!PAGE_ALIGNED(pa_start));
 424         BUG_ON(!PAGE_ALIGNED(size));
 425
 426         __create_pgd_mapping(pgd, pa_start, (unsigned long)va_start, size, prot,
 427                              early_pgtable_alloc, !debug_pagealloc_enabled());
 428
 429         vma->addr       = va_start;
 430         vma->phys_addr  = pa_start;
 431         vma->size       = size;
 432         vma->flags      = VM_MAP;
 433         vma->caller     = __builtin_return_address(0);
 434
 435         vm_area_add_early(vma);
 436 }
 437
 438 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
 439 static int __init map_entry_trampoline(void)
 440 {
 441         extern char __entry_tramp_text_start[];
 442
 443         pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
 444         phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start);
 445
 446         /* The trampoline is always mapped and can therefore be global */
 447         pgprot_val(prot) &= ~PTE_NG;
 448
 449         /* Map only the text into the trampoline page table */
 450         memset(tramp_pg_dir, 0, PGD_SIZE);
 451         __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, PAGE_SIZE,
 452                              prot, pgd_pgtable_alloc, 0);
 453
 454         /* Map both the text and data into the kernel page table */
 455         __set_fixmap(FIX_ENTRY_TRAMP_TEXT, pa_start, prot);
 456         if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
 457                 extern char __entry_tramp_data_start[];
 458
 459                 __set_fixmap(FIX_ENTRY_TRAMP_DATA,
 460                              __pa_symbol(__entry_tramp_data_start),
 461                              PAGE_KERNEL_RO);
 462         }
 463
 464         return 0;
 465 }
 466 core_initcall(map_entry_trampoline);
 467 #endif
 468
 469 /*
 470  * Create fine-grained mappings for the kernel.
 471  */
 472 static void __init map_kernel(pgd_t *pgd)
 473 {
 474         static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_init, vmlinux_data;
 475
 476         map_kernel_segment(pgd, _text, _etext, PAGE_KERNEL_EXEC, &vmlinux_text);
 477         map_kernel_segment(pgd, __start_rodata, __init_begin, PAGE_KERNEL, &vmlinux_rodata);
 478         map_kernel_segment(pgd, __init_begin, __init_end, PAGE_KERNEL_EXEC,
 479                            &vmlinux_init);
 480         map_kernel_segment(pgd, _data, _end, PAGE_KERNEL, &vmlinux_data);
 481
 482         if (!pgd_val(*pgd_offset_raw(pgd, FIXADDR_START))) {
 483                 /*
 484                  * The fixmap falls in a separate pgd to the kernel, and doesn't
 485                  * live in the carveout for the swapper_pg_dir. We can simply
 486                  * re-use the existing dir for the fixmap.
 487                  */
 488                 set_pgd(pgd_offset_raw(pgd, FIXADDR_START),
 489                         *pgd_offset_k(FIXADDR_START));
 490         } else if (CONFIG_PGTABLE_LEVELS > 3) {
 491                 /*
 492                  * The fixmap shares its top level pgd entry with the kernel
 493                  * mapping. This can really only occur when we are running
 494                  * with 16k/4 levels, so we can simply reuse the pud level
 495                  * entry instead.
 496                  */
 497                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 498                 pud_populate(&init_mm, pud_set_fixmap_offset(pgd, FIXADDR_START),
 499                              lm_alias(bm_pmd));
 500                 pud_clear_fixmap();
 501         } else {
 502                 BUG();
 503         }
 504
 505         kasan_copy_shadow(pgd);
 506 }
 507
 508 /*
 509  * paging_init() sets up the page tables, initialises the zone memory
 510  * maps and sets up the zero page.
 511  */
 512 void __init paging_init(void)
 513 {
 514         phys_addr_t pgd_phys = early_pgtable_alloc();
 515         pgd_t *pgd = pgd_set_fixmap(pgd_phys);
 516
 517         map_kernel(pgd);
 518         map_mem(pgd);
 519
 520         /*
 521          * We want to reuse the original swapper_pg_dir so we don't have to
 522          * communicate the new address to non-coherent secondaries in
 523          * secondary_entry, and so cpu_switch_mm can generate the address with
 524          * adrp+add rather than a load from some global variable.
 525          *
 526          * To do this we need to go via a temporary pgd.
 527          */
 528         cpu_replace_ttbr1(__va(pgd_phys));
 529         memcpy(swapper_pg_dir, pgd, PGD_SIZE);
 530         cpu_replace_ttbr1(swapper_pg_dir);
 531
 532         pgd_clear_fixmap();
 533         memblock_free(pgd_phys, PAGE_SIZE);
 534
 535         /*
 536          * We only reuse the PGD from the swapper_pg_dir, not the pud + pmd
 537          * allocated with it.
 538          */
 539         memblock_free(__pa(swapper_pg_dir) + PAGE_SIZE,
 540                       SWAPPER_DIR_SIZE - PAGE_SIZE);
 541 }
 542
 543 /*
 544  * Check whether a kernel address is valid (derived from arch/x86/).
 545  */
 546 int kern_addr_valid(unsigned long addr)
 547 {
 548         pgd_t *pgd;
 549         pud_t *pud;
 550         pmd_t *pmd;
 551         pte_t *pte;
 552
 553         if ((((long)addr) >> VA_BITS) != -1UL)
 554                 return 0;
 555
 556         pgd = pgd_offset_k(addr);
 557         if (pgd_none(*pgd))
 558                 return 0;
 559
 560         pud = pud_offset(pgd, addr);
 561         if (pud_none(*pud))
 562                 return 0;
 563
 564         if (pud_sect(*pud))
 565                 return pfn_valid(pud_pfn(*pud));
 566
 567         pmd = pmd_offset(pud, addr);
 568         if (pmd_none(*pmd))
 569                 return 0;
 570
 571         if (pmd_sect(*pmd))
 572                 return pfn_valid(pmd_pfn(*pmd));
 573
 574         pte = pte_offset_kernel(pmd, addr);
 575         if (pte_none(*pte))
 576                 return 0;
 577
 578         return pfn_valid(pte_pfn(*pte));
 579 }
 580 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 581 #if !ARM64_SWAPPER_USES_SECTION_MAPS
 582 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 583 {
 584         return vmemmap_populate_basepages(start, end, node);
 585 }
 586 #else   /* !ARM64_SWAPPER_USES_SECTION_MAPS */
 587 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 588 {
 589         unsigned long addr = start;
 590         unsigned long next;
 591         pgd_t *pgd;
 592         pud_t *pud;
 593         pmd_t *pmd;
 594
 595         do {
 596                 next = pmd_addr_end(addr, end);
 597
 598                 pgd = vmemmap_pgd_populate(addr, node);
 599                 if (!pgd)
 600                         return -ENOMEM;
 601
 602                 pud = vmemmap_pud_populate(pgd, addr, node);
 603                 if (!pud)
 604                         return -ENOMEM;
 605
 606                 pmd = pmd_offset(pud, addr);
 607                 if (pmd_none(*pmd)) {
 608                         void *p = NULL;
 609
 610                         p = vmemmap_alloc_block_buf(PMD_SIZE, node);
 611                         if (!p)
 612                                 return -ENOMEM;
 613
 614                         pmd_set_huge(pmd, __pa(p), __pgprot(PROT_SECT_NORMAL));
 615                 } else
 616                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 617         } while (addr = next, addr != end);
 618
 619         return 0;
 620 }
 621 #endif  /* CONFIG_ARM64_64K_PAGES */
 622 void vmemmap_free(unsigned long start, unsigned long end)
 623 {
 624 }
 625 #endif  /* CONFIG_SPARSEMEM_VMEMMAP */
 626
 627 static inline pud_t * fixmap_pud(unsigned long addr)
 628 {
 629         pgd_t *pgd = pgd_offset_k(addr);
 630
 631         BUG_ON(pgd_none(*pgd) || pgd_bad(*pgd));
 632
 633         return pud_offset_kimg(pgd, addr);
 634 }
 635
 636 static inline pmd_t * fixmap_pmd(unsigned long addr)
 637 {
 638         pud_t *pud = fixmap_pud(addr);
 639
 640         BUG_ON(pud_none(*pud) || pud_bad(*pud));
 641
 642         return pmd_offset_kimg(pud, addr);
 643 }
 644
 645 static inline pte_t * fixmap_pte(unsigned long addr)
 646 {
 647         return &bm_pte[pte_index(addr)];
 648 }
 649
 650 void __init early_fixmap_init(void)
 651 {
 652         pgd_t *pgd;
 653         pud_t *pud;
 654         pmd_t *pmd;
 655         unsigned long addr = FIXADDR_START;
 656
 657         pgd = pgd_offset_k(addr);
 658         if (CONFIG_PGTABLE_LEVELS > 3 &&
 659             !(pgd_none(*pgd) || pgd_page_paddr(*pgd) == __pa(bm_pud))) {
 660                 /*
 661                  * We only end up here if the kernel mapping and the fixmap
 662                  * share the top level pgd entry, which should only happen on
 663                  * 16k/4 levels configurations.
 664                  */
 665                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 666                 pud = pud_offset_kimg(pgd, addr);
 667         } else {
 668                 pgd_populate(&init_mm, pgd, bm_pud);
 669                 pud = fixmap_pud(addr);
 670         }
 671         pud_populate(&init_mm, pud, bm_pmd);
 672         pmd = fixmap_pmd(addr);
 673         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 674
 675         /*
 676          * The boot-ioremap range spans multiple pmds, for which
 677          * we are not prepared:
 678          */
 679         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 680                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 681
 682         if ((pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
 683              || pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 684                 WARN_ON(1);
 685                 pr_warn("pmd %p != %p, %p\n",
 686                         pmd, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
 687                         fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
 688                 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 689                         fix_to_virt(FIX_BTMAP_BEGIN));
 690                 pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 691                         fix_to_virt(FIX_BTMAP_END));
 692
 693                 pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 694                 pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
 695         }
 696 }
 697
 698 void __set_fixmap(enum fixed_addresses idx,
 699                                phys_addr_t phys, pgprot_t flags)
 700 {
 701         unsigned long addr = __fix_to_virt(idx);
 702         pte_t *pte;
 703
 704         BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
 705
 706         pte = fixmap_pte(addr);
 707
 708         if (pgprot_val(flags)) {
 709                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 710         } else {
 711                 pte_clear(&init_mm, addr, pte);
 712                 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
 713         }
 714 }
 715
 716 void *__init __fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
 717 {
 718         const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
 719         int offset;
 720         void *dt_virt;
 721
 722         /*
 723          * Check whether the physical FDT address is set and meets the minimum
 724          * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
 725          * at least 8 bytes so that we can always access the magic and size
 726          * fields of the FDT header after mapping the first chunk, double check
 727          * here if that is indeed the case.
 728          */
 729         BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
 730         if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
 731                 return NULL;
 732
 733         /*
 734          * Make sure that the FDT region can be mapped without the need to
 735          * allocate additional translation table pages, so that it is safe
 736          * to call create_mapping_noalloc() this early.
 737          *
 738          * On 64k pages, the FDT will be mapped using PTEs, so we need to
 739          * be in the same PMD as the rest of the fixmap.
 740          * On 4k pages, we'll use section mappings for the FDT so we only
 741          * have to be in the same PUD.
 742          */
 743         BUILD_BUG_ON(dt_virt_base % SZ_2M);
 744
 745         BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
 746                      __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
 747
 748         offset = dt_phys % SWAPPER_BLOCK_SIZE;
 749         dt_virt = (void *)dt_virt_base + offset;
 750
 751         /* map the first chunk so we can read the size from the header */
 752         create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
 753                         dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
 754
 755         if (fdt_magic(dt_virt) != FDT_MAGIC)
 756                 return NULL;
 757
 758         *size = fdt_totalsize(dt_virt);
 759         if (*size > MAX_FDT_SIZE)
 760                 return NULL;
 761
 762         if (offset + *size > SWAPPER_BLOCK_SIZE)
 763                 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
 764                                round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
 765
 766         return dt_virt;
 767 }
 768
 769 void *__init fixmap_remap_fdt(phys_addr_t dt_phys)
 770 {
 771         void *dt_virt;
 772         int size;
 773
 774         dt_virt = __fixmap_remap_fdt(dt_phys, &size, PAGE_KERNEL_RO);
 775         if (!dt_virt)
 776                 return NULL;
 777
 778         memblock_reserve(dt_phys, size);
 779         return dt_virt;
 780 }
 781
 782 int __init arch_ioremap_pud_supported(void)
 783 {
 784         /*
 785          * Only 4k granule supports level 1 block mappings.
 786          * SW table walks can't handle removal of intermediate entries.
 787          */
 788         return IS_ENABLED(CONFIG_ARM64_4K_PAGES) &&
 789                !IS_ENABLED(CONFIG_ARM64_PTDUMP_DEBUGFS);
 790 }
 791
 792 int __init arch_ioremap_pmd_supported(void)
 793 {
 794         /* See arch_ioremap_pud_supported() */
 795         return !IS_ENABLED(CONFIG_ARM64_PTDUMP_DEBUGFS);
 796 }
 797
 798 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
 799 {
 800         pgprot_t sect_prot = __pgprot(PUD_TYPE_SECT |
 801                                         pgprot_val(mk_sect_prot(prot)));
 802         pud_t new_pud = pfn_pud(__phys_to_pfn(phys), sect_prot);
 803
 804         /* Only allow permission changes for now */
 805         if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)),
 806                                    pud_val(new_pud)))
 807                 return 0;
 808
 809         BUG_ON(phys & ~PUD_MASK);
 810         set_pud(pudp, new_pud);
 811         return 1;
 812 }
 813
 814 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
 815 {
 816         pgprot_t sect_prot = __pgprot(PMD_TYPE_SECT |
 817                                         pgprot_val(mk_sect_prot(prot)));
 818         pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), sect_prot);
 819
 820         /* Only allow permission changes for now */
 821         if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)),
 822                                    pmd_val(new_pmd)))
 823                 return 0;
 824
 825         BUG_ON(phys & ~PMD_MASK);
 826         set_pmd(pmdp, new_pmd);
 827         return 1;
 828 }
 829
 830 int pud_clear_huge(pud_t *pud)
 831 {
 832         if (!pud_sect(*pud))
 833                 return 0;
 834         pud_clear(pud);
 835         return 1;
 836 }
 837
 838 int pmd_clear_huge(pmd_t *pmd)
 839 {
 840         if (!pmd_sect(*pmd))
 841                 return 0;
 842         pmd_clear(pmd);
 843         return 1;
 844 }
 845
 846 int pud_free_pmd_page(pud_t *pud, unsigned long addr)
 847 {
 848         return pud_none(*pud);
 849 }
 850
 851 int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
 852 {
 853         return pmd_none(*pmd);
 854 }