arch/arm64/mm/mmu.c

   1 /*
   2  * Based on arch/arm/mm/mmu.c
   3  *
   4  * Copyright (C) 1995-2005 Russell King
   5  * Copyright (C) 2012 ARM Ltd.
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
  18  */
  19
  20 #include <linux/export.h>
  21 #include <linux/kernel.h>
  22 #include <linux/errno.h>
  23 #include <linux/init.h>
  24 #include <linux/libfdt.h>
  25 #include <linux/mman.h>
  26 #include <linux/nodemask.h>
  27 #include <linux/memblock.h>
  28 #include <linux/fs.h>
  29 #include <linux/io.h>
  30 #include <linux/slab.h>
  31 #include <linux/stop_machine.h>
  32
  33 #include <asm/barrier.h>
  34 #include <asm/cputype.h>
  35 #include <asm/fixmap.h>
  36 #include <asm/kasan.h>
  37 #include <asm/kernel-pgtable.h>
  38 #include <asm/sections.h>
  39 #include <asm/setup.h>
  40 #include <asm/sizes.h>
  41 #include <asm/tlb.h>
  42 #include <asm/memblock.h>
  43 #include <asm/mmu_context.h>
  44
  45 #include "mm.h"
  46
  47 u64 idmap_t0sz = TCR_T0SZ(VA_BITS);
  48
  49 u64 kimage_voffset __read_mostly;
  50 EXPORT_SYMBOL(kimage_voffset);
  51
  52 /*
  53  * Empty_zero_page is a special page that is used for zero-initialized data
  54  * and COW.
  55  */
  56 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
  57 EXPORT_SYMBOL(empty_zero_page);
  58
  59 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
  60 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused;
  61 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused;
  62
  63 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
  64                               unsigned long size, pgprot_t vma_prot)
  65 {
  66         if (!pfn_valid(pfn))
  67                 return pgprot_noncached(vma_prot);
  68         else if (file->f_flags & O_SYNC)
  69                 return pgprot_writecombine(vma_prot);
  70         return vma_prot;
  71 }
  72 EXPORT_SYMBOL(phys_mem_access_prot);
  73
  74 static phys_addr_t __init early_pgtable_alloc(void)
  75 {
  76         phys_addr_t phys;
  77         void *ptr;
  78
  79         phys = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
  80         BUG_ON(!phys);
  81
  82         /*
  83          * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
  84          * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
  85          * any level of table.
  86          */
  87         ptr = pte_set_fixmap(phys);
  88
  89         memset(ptr, 0, PAGE_SIZE);
  90
  91         /*
  92          * Implicit barriers also ensure the zeroed page is visible to the page
  93          * table walker
  94          */
  95         pte_clear_fixmap();
  96
  97         return phys;
  98 }
  99
 100 /*
 101  * remap a PMD into pages
 102  */
 103 static void split_pmd(pmd_t *pmd, pte_t *pte)
 104 {
 105         unsigned long pfn = pmd_pfn(*pmd);
 106         int i = 0;
 107
 108         do {
 109                 /*
 110                  * Need to have the least restrictive permissions available
 111                  * permissions will be fixed up later
 112                  */
 113                 set_pte(pte, pfn_pte(pfn, PAGE_KERNEL_EXEC));
 114                 pfn++;
 115         } while (pte++, i++, i < PTRS_PER_PTE);
 116 }
 117
 118 static void alloc_init_pte(pmd_t *pmd, unsigned long addr,
 119                                   unsigned long end, unsigned long pfn,
 120                                   pgprot_t prot,
 121                                   phys_addr_t (*pgtable_alloc)(void))
 122 {
 123         pte_t *pte;
 124
 125         if (pmd_none(*pmd) || pmd_sect(*pmd)) {
 126                 phys_addr_t pte_phys;
 127                 BUG_ON(!pgtable_alloc);
 128                 pte_phys = pgtable_alloc();
 129                 pte = pte_set_fixmap(pte_phys);
 130                 if (pmd_sect(*pmd))
 131                         split_pmd(pmd, pte);
 132                 __pmd_populate(pmd, pte_phys, PMD_TYPE_TABLE);
 133                 flush_tlb_all();
 134                 pte_clear_fixmap();
 135         }
 136         BUG_ON(pmd_bad(*pmd));
 137
 138         pte = pte_set_fixmap_offset(pmd, addr);
 139         do {
 140                 set_pte(pte, pfn_pte(pfn, prot));
 141                 pfn++;
 142         } while (pte++, addr += PAGE_SIZE, addr != end);
 143
 144         pte_clear_fixmap();
 145 }
 146
 147 static void split_pud(pud_t *old_pud, pmd_t *pmd)
 148 {
 149         unsigned long addr = pud_pfn(*old_pud) << PAGE_SHIFT;
 150         pgprot_t prot = __pgprot(pud_val(*old_pud) ^ addr);
 151         int i = 0;
 152
 153         do {
 154                 set_pmd(pmd, __pmd(addr | pgprot_val(prot)));
 155                 addr += PMD_SIZE;
 156         } while (pmd++, i++, i < PTRS_PER_PMD);
 157 }
 158
 159 #ifdef CONFIG_DEBUG_PAGEALLOC
 160 static bool block_mappings_allowed(phys_addr_t (*pgtable_alloc)(void))
 161 {
 162
 163         /*
 164          * If debug_page_alloc is enabled we must map the linear map
 165          * using pages. However, other mappings created by
 166          * create_mapping_noalloc must use sections in some cases. Allow
 167          * sections to be used in those cases, where no pgtable_alloc
 168          * function is provided.
 169          */
 170         return !pgtable_alloc || !debug_pagealloc_enabled();
 171 }
 172 #else
 173 static bool block_mappings_allowed(phys_addr_t (*pgtable_alloc)(void))
 174 {
 175         return true;
 176 }
 177 #endif
 178
 179 static void alloc_init_pmd(pud_t *pud, unsigned long addr, unsigned long end,
 180                                   phys_addr_t phys, pgprot_t prot,
 181                                   phys_addr_t (*pgtable_alloc)(void))
 182 {
 183         pmd_t *pmd;
 184         unsigned long next;
 185
 186         /*
 187          * Check for initial section mappings in the pgd/pud and remove them.
 188          */
 189         if (pud_none(*pud) || pud_sect(*pud)) {
 190                 phys_addr_t pmd_phys;
 191                 BUG_ON(!pgtable_alloc);
 192                 pmd_phys = pgtable_alloc();
 193                 pmd = pmd_set_fixmap(pmd_phys);
 194                 if (pud_sect(*pud)) {
 195                         /*
 196                          * need to have the 1G of mappings continue to be
 197                          * present
 198                          */
 199                         split_pud(pud, pmd);
 200                 }
 201                 __pud_populate(pud, pmd_phys, PUD_TYPE_TABLE);
 202                 flush_tlb_all();
 203                 pmd_clear_fixmap();
 204         }
 205         BUG_ON(pud_bad(*pud));
 206
 207         pmd = pmd_set_fixmap_offset(pud, addr);
 208         do {
 209                 next = pmd_addr_end(addr, end);
 210                 /* try section mapping first */
 211                 if (((addr | next | phys) & ~SECTION_MASK) == 0 &&
 212                       block_mappings_allowed(pgtable_alloc)) {
 213                         pmd_t old_pmd =*pmd;
 214                         set_pmd(pmd, __pmd(phys |
 215                                            pgprot_val(mk_sect_prot(prot))));
 216                         /*
 217                          * Check for previous table entries created during
 218                          * boot (__create_page_tables) and flush them.
 219                          */
 220                         if (!pmd_none(old_pmd)) {
 221                                 flush_tlb_all();
 222                                 if (pmd_table(old_pmd)) {
 223                                         phys_addr_t table = pmd_page_paddr(old_pmd);
 224                                         if (!WARN_ON_ONCE(slab_is_available()))
 225                                                 memblock_free(table, PAGE_SIZE);
 226                                 }
 227                         }
 228                 } else {
 229                         alloc_init_pte(pmd, addr, next, __phys_to_pfn(phys),
 230                                        prot, pgtable_alloc);
 231                 }
 232                 phys += next - addr;
 233         } while (pmd++, addr = next, addr != end);
 234
 235         pmd_clear_fixmap();
 236 }
 237
 238 static inline bool use_1G_block(unsigned long addr, unsigned long next,
 239                         unsigned long phys)
 240 {
 241         if (PAGE_SHIFT != 12)
 242                 return false;
 243
 244         if (((addr | next | phys) & ~PUD_MASK) != 0)
 245                 return false;
 246
 247         return true;
 248 }
 249
 250 static void alloc_init_pud(pgd_t *pgd, unsigned long addr, unsigned long end,
 251                                   phys_addr_t phys, pgprot_t prot,
 252                                   phys_addr_t (*pgtable_alloc)(void))
 253 {
 254         pud_t *pud;
 255         unsigned long next;
 256
 257         if (pgd_none(*pgd)) {
 258                 phys_addr_t pud_phys;
 259                 BUG_ON(!pgtable_alloc);
 260                 pud_phys = pgtable_alloc();
 261                 __pgd_populate(pgd, pud_phys, PUD_TYPE_TABLE);
 262         }
 263         BUG_ON(pgd_bad(*pgd));
 264
 265         pud = pud_set_fixmap_offset(pgd, addr);
 266         do {
 267                 next = pud_addr_end(addr, end);
 268
 269                 /*
 270                  * For 4K granule only, attempt to put down a 1GB block
 271                  */
 272                 if (use_1G_block(addr, next, phys) &&
 273                     block_mappings_allowed(pgtable_alloc)) {
 274                         pud_t old_pud = *pud;
 275                         set_pud(pud, __pud(phys |
 276                                            pgprot_val(mk_sect_prot(prot))));
 277
 278                         /*
 279                          * If we have an old value for a pud, it will
 280                          * be pointing to a pmd table that we no longer
 281                          * need (from swapper_pg_dir).
 282                          *
 283                          * Look up the old pmd table and free it.
 284                          */
 285                         if (!pud_none(old_pud)) {
 286                                 flush_tlb_all();
 287                                 if (pud_table(old_pud)) {
 288                                         phys_addr_t table = pud_page_paddr(old_pud);
 289                                         if (!WARN_ON_ONCE(slab_is_available()))
 290                                                 memblock_free(table, PAGE_SIZE);
 291                                 }
 292                         }
 293                 } else {
 294                         alloc_init_pmd(pud, addr, next, phys, prot,
 295                                        pgtable_alloc);
 296                 }
 297                 phys += next - addr;
 298         } while (pud++, addr = next, addr != end);
 299
 300         pud_clear_fixmap();
 301 }
 302
 303 /*
 304  * Create the page directory entries and any necessary page tables for the
 305  * mapping specified by 'md'.
 306  */
 307 static void init_pgd(pgd_t *pgd, phys_addr_t phys, unsigned long virt,
 308                                     phys_addr_t size, pgprot_t prot,
 309                                     phys_addr_t (*pgtable_alloc)(void))
 310 {
 311         unsigned long addr, length, end, next;
 312
 313         /*
 314          * If the virtual and physical address don't have the same offset
 315          * within a page, we cannot map the region as the caller expects.
 316          */
 317         if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
 318                 return;
 319
 320         phys &= PAGE_MASK;
 321         addr = virt & PAGE_MASK;
 322         length = PAGE_ALIGN(size + (virt & ~PAGE_MASK));
 323
 324         end = addr + length;
 325         do {
 326                 next = pgd_addr_end(addr, end);
 327                 alloc_init_pud(pgd, addr, next, phys, prot, pgtable_alloc);
 328                 phys += next - addr;
 329         } while (pgd++, addr = next, addr != end);
 330 }
 331
 332 static phys_addr_t late_pgtable_alloc(void)
 333 {
 334         void *ptr = (void *)__get_free_page(PGALLOC_GFP);
 335         BUG_ON(!ptr);
 336
 337         /* Ensure the zeroed page is visible to the page table walker */
 338         dsb(ishst);
 339         return __pa(ptr);
 340 }
 341
 342 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
 343                                  unsigned long virt, phys_addr_t size,
 344                                  pgprot_t prot,
 345                                  phys_addr_t (*alloc)(void))
 346 {
 347         init_pgd(pgd_offset_raw(pgdir, virt), phys, virt, size, prot, alloc);
 348 }
 349
 350 /*
 351  * This function can only be used to modify existing table entries,
 352  * without allocating new levels of table. Note that this permits the
 353  * creation of new section or page entries.
 354  */
 355 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
 356                                   phys_addr_t size, pgprot_t prot)
 357 {
 358         if (virt < VMALLOC_START) {
 359                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 360                         &phys, virt);
 361                 return;
 362         }
 363         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot,
 364                              NULL);
 365 }
 366
 367 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
 368                                unsigned long virt, phys_addr_t size,
 369                                pgprot_t prot)
 370 {
 371         __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
 372                              late_pgtable_alloc);
 373 }
 374
 375 static void create_mapping_late(phys_addr_t phys, unsigned long virt,
 376                                   phys_addr_t size, pgprot_t prot)
 377 {
 378         if (virt < VMALLOC_START) {
 379                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 380                         &phys, virt);
 381                 return;
 382         }
 383
 384         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot,
 385                              late_pgtable_alloc);
 386 }
 387
 388 static void __init __map_memblock(pgd_t *pgd, phys_addr_t start, phys_addr_t end)
 389 {
 390
 391         unsigned long kernel_start = __pa(_stext);
 392         unsigned long kernel_end = __pa(_etext);
 393
 394         /*
 395          * Take care not to create a writable alias for the
 396          * read-only text and rodata sections of the kernel image.
 397          */
 398
 399         /* No overlap with the kernel text */
 400         if (end < kernel_start || start >= kernel_end) {
 401                 __create_pgd_mapping(pgd, start, __phys_to_virt(start),
 402                                      end - start, PAGE_KERNEL,
 403                                      early_pgtable_alloc);
 404                 return;
 405         }
 406
 407         /*
 408          * This block overlaps the kernel text mapping. Map the portion(s) which
 409          * don't overlap.
 410          */
 411         if (start < kernel_start)
 412                 __create_pgd_mapping(pgd, start,
 413                                      __phys_to_virt(start),
 414                                      kernel_start - start, PAGE_KERNEL,
 415                                      early_pgtable_alloc);
 416         if (kernel_end < end)
 417                 __create_pgd_mapping(pgd, kernel_end,
 418                                      __phys_to_virt(kernel_end),
 419                                      end - kernel_end, PAGE_KERNEL,
 420                                      early_pgtable_alloc);
 421 }
 422
 423 static void __init map_mem(pgd_t *pgd)
 424 {
 425         struct memblock_region *reg;
 426
 427         /* map all the memory banks */
 428         for_each_memblock(memory, reg) {
 429                 phys_addr_t start = reg->base;
 430                 phys_addr_t end = start + reg->size;
 431
 432                 if (start >= end)
 433                         break;
 434                 if (memblock_is_nomap(reg))
 435                         continue;
 436
 437                 __map_memblock(pgd, start, end);
 438         }
 439 }
 440
 441 void mark_rodata_ro(void)
 442 {
 443         if (!IS_ENABLED(CONFIG_DEBUG_RODATA))
 444                 return;
 445
 446         create_mapping_late(__pa(_stext), (unsigned long)_stext,
 447                                 (unsigned long)_etext - (unsigned long)_stext,
 448                                 PAGE_KERNEL_ROX);
 449 }
 450
 451 void fixup_init(void)
 452 {
 453         /*
 454          * Unmap the __init region but leave the VM area in place. This
 455          * prevents the region from being reused for kernel modules, which
 456          * is not supported by kallsyms.
 457          */
 458         unmap_kernel_range((u64)__init_begin, (u64)(__init_end - __init_begin));
 459 }
 460
 461 static void __init map_kernel_chunk(pgd_t *pgd, void *va_start, void *va_end,
 462                                     pgprot_t prot, struct vm_struct *vma)
 463 {
 464         phys_addr_t pa_start = __pa(va_start);
 465         unsigned long size = va_end - va_start;
 466
 467         BUG_ON(!PAGE_ALIGNED(pa_start));
 468         BUG_ON(!PAGE_ALIGNED(size));
 469
 470         __create_pgd_mapping(pgd, pa_start, (unsigned long)va_start, size, prot,
 471                              early_pgtable_alloc);
 472
 473         vma->addr       = va_start;
 474         vma->phys_addr  = pa_start;
 475         vma->size       = size;
 476         vma->flags      = VM_MAP;
 477         vma->caller     = map_kernel_chunk;
 478
 479         vm_area_add_early(vma);
 480 }
 481
 482 /*
 483  * Create fine-grained mappings for the kernel.
 484  */
 485 static void __init map_kernel(pgd_t *pgd)
 486 {
 487         static struct vm_struct vmlinux_text, vmlinux_init, vmlinux_data;
 488
 489         map_kernel_chunk(pgd, _stext, _etext, PAGE_KERNEL_EXEC, &vmlinux_text);
 490         map_kernel_chunk(pgd, __init_begin, __init_end, PAGE_KERNEL_EXEC,
 491                          &vmlinux_init);
 492         map_kernel_chunk(pgd, _data, _end, PAGE_KERNEL, &vmlinux_data);
 493
 494         if (!pgd_val(*pgd_offset_raw(pgd, FIXADDR_START))) {
 495                 /*
 496                  * The fixmap falls in a separate pgd to the kernel, and doesn't
 497                  * live in the carveout for the swapper_pg_dir. We can simply
 498                  * re-use the existing dir for the fixmap.
 499                  */
 500                 set_pgd(pgd_offset_raw(pgd, FIXADDR_START),
 501                         *pgd_offset_k(FIXADDR_START));
 502         } else if (CONFIG_PGTABLE_LEVELS > 3) {
 503                 /*
 504                  * The fixmap shares its top level pgd entry with the kernel
 505                  * mapping. This can really only occur when we are running
 506                  * with 16k/4 levels, so we can simply reuse the pud level
 507                  * entry instead.
 508                  */
 509                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 510                 set_pud(pud_set_fixmap_offset(pgd, FIXADDR_START),
 511                         __pud(__pa(bm_pmd) | PUD_TYPE_TABLE));
 512                 pud_clear_fixmap();
 513         } else {
 514                 BUG();
 515         }
 516
 517         kasan_copy_shadow(pgd);
 518 }
 519
 520 /*
 521  * paging_init() sets up the page tables, initialises the zone memory
 522  * maps and sets up the zero page.
 523  */
 524 void __init paging_init(void)
 525 {
 526         phys_addr_t pgd_phys = early_pgtable_alloc();
 527         pgd_t *pgd = pgd_set_fixmap(pgd_phys);
 528
 529         map_kernel(pgd);
 530         map_mem(pgd);
 531
 532         /*
 533          * We want to reuse the original swapper_pg_dir so we don't have to
 534          * communicate the new address to non-coherent secondaries in
 535          * secondary_entry, and so cpu_switch_mm can generate the address with
 536          * adrp+add rather than a load from some global variable.
 537          *
 538          * To do this we need to go via a temporary pgd.
 539          */
 540         cpu_replace_ttbr1(__va(pgd_phys));
 541         memcpy(swapper_pg_dir, pgd, PAGE_SIZE);
 542         cpu_replace_ttbr1(swapper_pg_dir);
 543
 544         pgd_clear_fixmap();
 545         memblock_free(pgd_phys, PAGE_SIZE);
 546
 547         /*
 548          * We only reuse the PGD from the swapper_pg_dir, not the pud + pmd
 549          * allocated with it.
 550          */
 551         memblock_free(__pa(swapper_pg_dir) + PAGE_SIZE,
 552                       SWAPPER_DIR_SIZE - PAGE_SIZE);
 553
 554         bootmem_init();
 555 }
 556
 557 /*
 558  * Check whether a kernel address is valid (derived from arch/x86/).
 559  */
 560 int kern_addr_valid(unsigned long addr)
 561 {
 562         pgd_t *pgd;
 563         pud_t *pud;
 564         pmd_t *pmd;
 565         pte_t *pte;
 566
 567         if ((((long)addr) >> VA_BITS) != -1UL)
 568                 return 0;
 569
 570         pgd = pgd_offset_k(addr);
 571         if (pgd_none(*pgd))
 572                 return 0;
 573
 574         pud = pud_offset(pgd, addr);
 575         if (pud_none(*pud))
 576                 return 0;
 577
 578         if (pud_sect(*pud))
 579                 return pfn_valid(pud_pfn(*pud));
 580
 581         pmd = pmd_offset(pud, addr);
 582         if (pmd_none(*pmd))
 583                 return 0;
 584
 585         if (pmd_sect(*pmd))
 586                 return pfn_valid(pmd_pfn(*pmd));
 587
 588         pte = pte_offset_kernel(pmd, addr);
 589         if (pte_none(*pte))
 590                 return 0;
 591
 592         return pfn_valid(pte_pfn(*pte));
 593 }
 594 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 595 #if !ARM64_SWAPPER_USES_SECTION_MAPS
 596 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 597 {
 598         return vmemmap_populate_basepages(start, end, node);
 599 }
 600 #else   /* !ARM64_SWAPPER_USES_SECTION_MAPS */
 601 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 602 {
 603         unsigned long addr = start;
 604         unsigned long next;
 605         pgd_t *pgd;
 606         pud_t *pud;
 607         pmd_t *pmd;
 608
 609         do {
 610                 next = pmd_addr_end(addr, end);
 611
 612                 pgd = vmemmap_pgd_populate(addr, node);
 613                 if (!pgd)
 614                         return -ENOMEM;
 615
 616                 pud = vmemmap_pud_populate(pgd, addr, node);
 617                 if (!pud)
 618                         return -ENOMEM;
 619
 620                 pmd = pmd_offset(pud, addr);
 621                 if (pmd_none(*pmd)) {
 622                         void *p = NULL;
 623
 624                         p = vmemmap_alloc_block_buf(PMD_SIZE, node);
 625                         if (!p)
 626                                 return -ENOMEM;
 627
 628                         set_pmd(pmd, __pmd(__pa(p) | PROT_SECT_NORMAL));
 629                 } else
 630                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 631         } while (addr = next, addr != end);
 632
 633         return 0;
 634 }
 635 #endif  /* CONFIG_ARM64_64K_PAGES */
 636 void vmemmap_free(unsigned long start, unsigned long end)
 637 {
 638 }
 639 #endif  /* CONFIG_SPARSEMEM_VMEMMAP */
 640
 641 static inline pud_t * fixmap_pud(unsigned long addr)
 642 {
 643         pgd_t *pgd = pgd_offset_k(addr);
 644
 645         BUG_ON(pgd_none(*pgd) || pgd_bad(*pgd));
 646
 647         return pud_offset_kimg(pgd, addr);
 648 }
 649
 650 static inline pmd_t * fixmap_pmd(unsigned long addr)
 651 {
 652         pud_t *pud = fixmap_pud(addr);
 653
 654         BUG_ON(pud_none(*pud) || pud_bad(*pud));
 655
 656         return pmd_offset_kimg(pud, addr);
 657 }
 658
 659 static inline pte_t * fixmap_pte(unsigned long addr)
 660 {
 661         return &bm_pte[pte_index(addr)];
 662 }
 663
 664 void __init early_fixmap_init(void)
 665 {
 666         pgd_t *pgd;
 667         pud_t *pud;
 668         pmd_t *pmd;
 669         unsigned long addr = FIXADDR_START;
 670
 671         pgd = pgd_offset_k(addr);
 672         if (CONFIG_PGTABLE_LEVELS > 3 && !pgd_none(*pgd)) {
 673                 /*
 674                  * We only end up here if the kernel mapping and the fixmap
 675                  * share the top level pgd entry, which should only happen on
 676                  * 16k/4 levels configurations.
 677                  */
 678                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 679                 pud = pud_offset_kimg(pgd, addr);
 680         } else {
 681                 pgd_populate(&init_mm, pgd, bm_pud);
 682                 pud = fixmap_pud(addr);
 683         }
 684         pud_populate(&init_mm, pud, bm_pmd);
 685         pmd = fixmap_pmd(addr);
 686         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 687
 688         /*
 689          * The boot-ioremap range spans multiple pmds, for which
 690          * we are not prepared:
 691          */
 692         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 693                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 694
 695         if ((pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
 696              || pmd != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 697                 WARN_ON(1);
 698                 pr_warn("pmd %p != %p, %p\n",
 699                         pmd, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
 700                         fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
 701                 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 702                         fix_to_virt(FIX_BTMAP_BEGIN));
 703                 pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 704                         fix_to_virt(FIX_BTMAP_END));
 705
 706                 pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 707                 pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
 708         }
 709 }
 710
 711 void __set_fixmap(enum fixed_addresses idx,
 712                                phys_addr_t phys, pgprot_t flags)
 713 {
 714         unsigned long addr = __fix_to_virt(idx);
 715         pte_t *pte;
 716
 717         BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
 718
 719         pte = fixmap_pte(addr);
 720
 721         if (pgprot_val(flags)) {
 722                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 723         } else {
 724                 pte_clear(&init_mm, addr, pte);
 725                 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
 726         }
 727 }
 728
 729 void *__init fixmap_remap_fdt(phys_addr_t dt_phys)
 730 {
 731         const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
 732         pgprot_t prot = PAGE_KERNEL_RO;
 733         int size, offset;
 734         void *dt_virt;
 735
 736         /*
 737          * Check whether the physical FDT address is set and meets the minimum
 738          * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
 739          * at least 8 bytes so that we can always access the size field of the
 740          * FDT header after mapping the first chunk, double check here if that
 741          * is indeed the case.
 742          */
 743         BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
 744         if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
 745                 return NULL;
 746
 747         /*
 748          * Make sure that the FDT region can be mapped without the need to
 749          * allocate additional translation table pages, so that it is safe
 750          * to call create_mapping_noalloc() this early.
 751          *
 752          * On 64k pages, the FDT will be mapped using PTEs, so we need to
 753          * be in the same PMD as the rest of the fixmap.
 754          * On 4k pages, we'll use section mappings for the FDT so we only
 755          * have to be in the same PUD.
 756          */
 757         BUILD_BUG_ON(dt_virt_base % SZ_2M);
 758
 759         BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
 760                      __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
 761
 762         offset = dt_phys % SWAPPER_BLOCK_SIZE;
 763         dt_virt = (void *)dt_virt_base + offset;
 764
 765         /* map the first chunk so we can read the size from the header */
 766         create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
 767                         dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
 768
 769         if (fdt_check_header(dt_virt) != 0)
 770                 return NULL;
 771
 772         size = fdt_totalsize(dt_virt);
 773         if (size > MAX_FDT_SIZE)
 774                 return NULL;
 775
 776         if (offset + size > SWAPPER_BLOCK_SIZE)
 777                 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
 778                                round_up(offset + size, SWAPPER_BLOCK_SIZE), prot);
 779
 780         memblock_reserve(dt_phys, size);
 781
 782         return dt_virt;
 783 }
 784
 785 int __init arch_ioremap_pud_supported(void)
 786 {
 787         /* only 4k granule supports level 1 block mappings */
 788         return IS_ENABLED(CONFIG_ARM64_4K_PAGES);
 789 }
 790
 791 int __init arch_ioremap_pmd_supported(void)
 792 {
 793         return 1;
 794 }
 795
 796 int pud_set_huge(pud_t *pud, phys_addr_t phys, pgprot_t prot)
 797 {
 798         BUG_ON(phys & ~PUD_MASK);
 799         set_pud(pud, __pud(phys | PUD_TYPE_SECT | pgprot_val(mk_sect_prot(prot))));
 800         return 1;
 801 }
 802
 803 int pmd_set_huge(pmd_t *pmd, phys_addr_t phys, pgprot_t prot)
 804 {
 805         BUG_ON(phys & ~PMD_MASK);
 806         set_pmd(pmd, __pmd(phys | PMD_TYPE_SECT | pgprot_val(mk_sect_prot(prot))));
 807         return 1;
 808 }
 809
 810 int pud_clear_huge(pud_t *pud)
 811 {
 812         if (!pud_sect(*pud))
 813                 return 0;
 814         pud_clear(pud);
 815         return 1;
 816 }
 817
 818 int pmd_clear_huge(pmd_t *pmd)
 819 {
 820         if (!pmd_sect(*pmd))
 821                 return 0;
 822         pmd_clear(pmd);
 823         return 1;
 824 }