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