arch/arm64/mm/mmu.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Based on arch/arm/mm/mmu.c
   4  *
   5  * Copyright (C) 1995-2005 Russell King
   6  * Copyright (C) 2012 ARM Ltd.
   7  */
   8
   9 #include <linux/cache.h>
  10 #include <linux/export.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/init.h>
  14 #include <linux/ioport.h>
  15 #include <linux/kexec.h>
  16 #include <linux/libfdt.h>
  17 #include <linux/mman.h>
  18 #include <linux/nodemask.h>
  19 #include <linux/memblock.h>
  20 #include <linux/fs.h>
  21 #include <linux/io.h>
  22 #include <linux/mm.h>
  23 #include <linux/vmalloc.h>
  24
  25 #include <asm/barrier.h>
  26 #include <asm/cputype.h>
  27 #include <asm/fixmap.h>
  28 #include <asm/kasan.h>
  29 #include <asm/kernel-pgtable.h>
  30 #include <asm/sections.h>
  31 #include <asm/setup.h>
  32 #include <linux/sizes.h>
  33 #include <asm/tlb.h>
  34 #include <asm/mmu_context.h>
  35 #include <asm/ptdump.h>
  36 #include <asm/tlbflush.h>
  37
  38 #define NO_BLOCK_MAPPINGS       BIT(0)
  39 #define NO_CONT_MAPPINGS        BIT(1)
  40
  41 u64 idmap_t0sz = TCR_T0SZ(VA_BITS);
  42 u64 idmap_ptrs_per_pgd = PTRS_PER_PGD;
  43 u64 vabits_user __ro_after_init;
  44 EXPORT_SYMBOL(vabits_user);
  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 static DEFINE_SPINLOCK(swapper_pgdir_lock);
  61
  62 void set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
  63 {
  64         pgd_t *fixmap_pgdp;
  65
  66         spin_lock(&swapper_pgdir_lock);
  67         fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp));
  68         WRITE_ONCE(*fixmap_pgdp, pgd);
  69         /*
  70          * We need dsb(ishst) here to ensure the page-table-walker sees
  71          * our new entry before set_p?d() returns. The fixmap's
  72          * flush_tlb_kernel_range() via clear_fixmap() does this for us.
  73          */
  74         pgd_clear_fixmap();
  75         spin_unlock(&swapper_pgdir_lock);
  76 }
  77
  78 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
  79                               unsigned long size, pgprot_t vma_prot)
  80 {
  81         if (!pfn_valid(pfn))
  82                 return pgprot_noncached(vma_prot);
  83         else if (file->f_flags & O_SYNC)
  84                 return pgprot_writecombine(vma_prot);
  85         return vma_prot;
  86 }
  87 EXPORT_SYMBOL(phys_mem_access_prot);
  88
  89 static phys_addr_t __init early_pgtable_alloc(int shift)
  90 {
  91         phys_addr_t phys;
  92         void *ptr;
  93
  94         phys = memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
  95         if (!phys)
  96                 panic("Failed to allocate page table page\n");
  97
  98         /*
  99          * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
 100          * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
 101          * any level of table.
 102          */
 103         ptr = pte_set_fixmap(phys);
 104
 105         memset(ptr, 0, PAGE_SIZE);
 106
 107         /*
 108          * Implicit barriers also ensure the zeroed page is visible to the page
 109          * table walker
 110          */
 111         pte_clear_fixmap();
 112
 113         return phys;
 114 }
 115
 116 static bool pgattr_change_is_safe(u64 old, u64 new)
 117 {
 118         /*
 119          * The following mapping attributes may be updated in live
 120          * kernel mappings without the need for break-before-make.
 121          */
 122         static const pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG;
 123
 124         /* creating or taking down mappings is always safe */
 125         if (old == 0 || new == 0)
 126                 return true;
 127
 128         /* live contiguous mappings may not be manipulated at all */
 129         if ((old | new) & PTE_CONT)
 130                 return false;
 131
 132         /* Transitioning from Non-Global to Global is unsafe */
 133         if (old & ~new & PTE_NG)
 134                 return false;
 135
 136         return ((old ^ new) & ~mask) == 0;
 137 }
 138
 139 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end,
 140                      phys_addr_t phys, pgprot_t prot)
 141 {
 142         pte_t *ptep;
 143
 144         ptep = pte_set_fixmap_offset(pmdp, addr);
 145         do {
 146                 pte_t old_pte = READ_ONCE(*ptep);
 147
 148                 set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot));
 149
 150                 /*
 151                  * After the PTE entry has been populated once, we
 152                  * only allow updates to the permission attributes.
 153                  */
 154                 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte),
 155                                               READ_ONCE(pte_val(*ptep))));
 156
 157                 phys += PAGE_SIZE;
 158         } while (ptep++, addr += PAGE_SIZE, addr != end);
 159
 160         pte_clear_fixmap();
 161 }
 162
 163 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
 164                                 unsigned long end, phys_addr_t phys,
 165                                 pgprot_t prot,
 166                                 phys_addr_t (*pgtable_alloc)(int),
 167                                 int flags)
 168 {
 169         unsigned long next;
 170         pmd_t pmd = READ_ONCE(*pmdp);
 171
 172         BUG_ON(pmd_sect(pmd));
 173         if (pmd_none(pmd)) {
 174                 phys_addr_t pte_phys;
 175                 BUG_ON(!pgtable_alloc);
 176                 pte_phys = pgtable_alloc(PAGE_SHIFT);
 177                 __pmd_populate(pmdp, pte_phys, PMD_TYPE_TABLE);
 178                 pmd = READ_ONCE(*pmdp);
 179         }
 180         BUG_ON(pmd_bad(pmd));
 181
 182         do {
 183                 pgprot_t __prot = prot;
 184
 185                 next = pte_cont_addr_end(addr, end);
 186
 187                 /* use a contiguous mapping if the range is suitably aligned */
 188                 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) &&
 189                     (flags & NO_CONT_MAPPINGS) == 0)
 190                         __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
 191
 192                 init_pte(pmdp, addr, next, phys, __prot);
 193
 194                 phys += next - addr;
 195         } while (addr = next, addr != end);
 196 }
 197
 198 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end,
 199                      phys_addr_t phys, pgprot_t prot,
 200                      phys_addr_t (*pgtable_alloc)(int), int flags)
 201 {
 202         unsigned long next;
 203         pmd_t *pmdp;
 204
 205         pmdp = pmd_set_fixmap_offset(pudp, addr);
 206         do {
 207                 pmd_t old_pmd = READ_ONCE(*pmdp);
 208
 209                 next = pmd_addr_end(addr, end);
 210
 211                 /* try section mapping first */
 212                 if (((addr | next | phys) & ~SECTION_MASK) == 0 &&
 213                     (flags & NO_BLOCK_MAPPINGS) == 0) {
 214                         pmd_set_huge(pmdp, phys, prot);
 215
 216                         /*
 217                          * After the PMD entry has been populated once, we
 218                          * only allow updates to the permission attributes.
 219                          */
 220                         BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
 221                                                       READ_ONCE(pmd_val(*pmdp))));
 222                 } else {
 223                         alloc_init_cont_pte(pmdp, addr, next, phys, prot,
 224                                             pgtable_alloc, flags);
 225
 226                         BUG_ON(pmd_val(old_pmd) != 0 &&
 227                                pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp)));
 228                 }
 229                 phys += next - addr;
 230         } while (pmdp++, addr = next, addr != end);
 231
 232         pmd_clear_fixmap();
 233 }
 234
 235 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
 236                                 unsigned long end, phys_addr_t phys,
 237                                 pgprot_t prot,
 238                                 phys_addr_t (*pgtable_alloc)(int), int flags)
 239 {
 240         unsigned long next;
 241         pud_t pud = READ_ONCE(*pudp);
 242
 243         /*
 244          * Check for initial section mappings in the pgd/pud.
 245          */
 246         BUG_ON(pud_sect(pud));
 247         if (pud_none(pud)) {
 248                 phys_addr_t pmd_phys;
 249                 BUG_ON(!pgtable_alloc);
 250                 pmd_phys = pgtable_alloc(PMD_SHIFT);
 251                 __pud_populate(pudp, pmd_phys, PUD_TYPE_TABLE);
 252                 pud = READ_ONCE(*pudp);
 253         }
 254         BUG_ON(pud_bad(pud));
 255
 256         do {
 257                 pgprot_t __prot = prot;
 258
 259                 next = pmd_cont_addr_end(addr, end);
 260
 261                 /* use a contiguous mapping if the range is suitably aligned */
 262                 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) &&
 263                     (flags & NO_CONT_MAPPINGS) == 0)
 264                         __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
 265
 266                 init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags);
 267
 268                 phys += next - addr;
 269         } while (addr = next, addr != end);
 270 }
 271
 272 static inline bool use_1G_block(unsigned long addr, unsigned long next,
 273                         unsigned long phys)
 274 {
 275         if (PAGE_SHIFT != 12)
 276                 return false;
 277
 278         if (((addr | next | phys) & ~PUD_MASK) != 0)
 279                 return false;
 280
 281         return true;
 282 }
 283
 284 static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end,
 285                            phys_addr_t phys, pgprot_t prot,
 286                            phys_addr_t (*pgtable_alloc)(int),
 287                            int flags)
 288 {
 289         unsigned long next;
 290         pud_t *pudp;
 291         pgd_t pgd = READ_ONCE(*pgdp);
 292
 293         if (pgd_none(pgd)) {
 294                 phys_addr_t pud_phys;
 295                 BUG_ON(!pgtable_alloc);
 296                 pud_phys = pgtable_alloc(PUD_SHIFT);
 297                 __pgd_populate(pgdp, pud_phys, PUD_TYPE_TABLE);
 298                 pgd = READ_ONCE(*pgdp);
 299         }
 300         BUG_ON(pgd_bad(pgd));
 301
 302         pudp = pud_set_fixmap_offset(pgdp, addr);
 303         do {
 304                 pud_t old_pud = READ_ONCE(*pudp);
 305
 306                 next = pud_addr_end(addr, end);
 307
 308                 /*
 309                  * For 4K granule only, attempt to put down a 1GB block
 310                  */
 311                 if (use_1G_block(addr, next, phys) &&
 312                     (flags & NO_BLOCK_MAPPINGS) == 0) {
 313                         pud_set_huge(pudp, phys, prot);
 314
 315                         /*
 316                          * After the PUD entry has been populated once, we
 317                          * only allow updates to the permission attributes.
 318                          */
 319                         BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
 320                                                       READ_ONCE(pud_val(*pudp))));
 321                 } else {
 322                         alloc_init_cont_pmd(pudp, addr, next, phys, prot,
 323                                             pgtable_alloc, flags);
 324
 325                         BUG_ON(pud_val(old_pud) != 0 &&
 326                                pud_val(old_pud) != READ_ONCE(pud_val(*pudp)));
 327                 }
 328                 phys += next - addr;
 329         } while (pudp++, addr = next, addr != end);
 330
 331         pud_clear_fixmap();
 332 }
 333
 334 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
 335                                  unsigned long virt, phys_addr_t size,
 336                                  pgprot_t prot,
 337                                  phys_addr_t (*pgtable_alloc)(int),
 338                                  int flags)
 339 {
 340         unsigned long addr, length, end, next;
 341         pgd_t *pgdp = pgd_offset_raw(pgdir, virt);
 342
 343         /*
 344          * If the virtual and physical address don't have the same offset
 345          * within a page, we cannot map the region as the caller expects.
 346          */
 347         if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
 348                 return;
 349
 350         phys &= PAGE_MASK;
 351         addr = virt & PAGE_MASK;
 352         length = PAGE_ALIGN(size + (virt & ~PAGE_MASK));
 353
 354         end = addr + length;
 355         do {
 356                 next = pgd_addr_end(addr, end);
 357                 alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc,
 358                                flags);
 359                 phys += next - addr;
 360         } while (pgdp++, addr = next, addr != end);
 361 }
 362
 363 static phys_addr_t __pgd_pgtable_alloc(int shift)
 364 {
 365         void *ptr = (void *)__get_free_page(PGALLOC_GFP);
 366         BUG_ON(!ptr);
 367
 368         /* Ensure the zeroed page is visible to the page table walker */
 369         dsb(ishst);
 370         return __pa(ptr);
 371 }
 372
 373 static phys_addr_t pgd_pgtable_alloc(int shift)
 374 {
 375         phys_addr_t pa = __pgd_pgtable_alloc(shift);
 376
 377         /*
 378          * Call proper page table ctor in case later we need to
 379          * call core mm functions like apply_to_page_range() on
 380          * this pre-allocated page table.
 381          *
 382          * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is
 383          * folded, and if so pgtable_pmd_page_ctor() becomes nop.
 384          */
 385         if (shift == PAGE_SHIFT)
 386                 BUG_ON(!pgtable_page_ctor(phys_to_page(pa)));
 387         else if (shift == PMD_SHIFT)
 388                 BUG_ON(!pgtable_pmd_page_ctor(phys_to_page(pa)));
 389
 390         return pa;
 391 }
 392
 393 /*
 394  * This function can only be used to modify existing table entries,
 395  * without allocating new levels of table. Note that this permits the
 396  * creation of new section or page entries.
 397  */
 398 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
 399                                   phys_addr_t size, pgprot_t prot)
 400 {
 401         if (virt < VMALLOC_START) {
 402                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 403                         &phys, virt);
 404                 return;
 405         }
 406         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
 407                              NO_CONT_MAPPINGS);
 408 }
 409
 410 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
 411                                unsigned long virt, phys_addr_t size,
 412                                pgprot_t prot, bool page_mappings_only)
 413 {
 414         int flags = 0;
 415
 416         BUG_ON(mm == &init_mm);
 417
 418         if (page_mappings_only)
 419                 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
 420
 421         __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
 422                              pgd_pgtable_alloc, flags);
 423 }
 424
 425 static void update_mapping_prot(phys_addr_t phys, unsigned long virt,
 426                                 phys_addr_t size, pgprot_t prot)
 427 {
 428         if (virt < VMALLOC_START) {
 429                 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n",
 430                         &phys, virt);
 431                 return;
 432         }
 433
 434         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
 435                              NO_CONT_MAPPINGS);
 436
 437         /* flush the TLBs after updating live kernel mappings */
 438         flush_tlb_kernel_range(virt, virt + size);
 439 }
 440
 441 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start,
 442                                   phys_addr_t end, pgprot_t prot, int flags)
 443 {
 444         __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start,
 445                              prot, early_pgtable_alloc, flags);
 446 }
 447
 448 void __init mark_linear_text_alias_ro(void)
 449 {
 450         /*
 451          * Remove the write permissions from the linear alias of .text/.rodata
 452          */
 453         update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text),
 454                             (unsigned long)__init_begin - (unsigned long)_text,
 455                             PAGE_KERNEL_RO);
 456 }
 457
 458 static void __init map_mem(pgd_t *pgdp)
 459 {
 460         phys_addr_t kernel_start = __pa_symbol(_text);
 461         phys_addr_t kernel_end = __pa_symbol(__init_begin);
 462         struct memblock_region *reg;
 463         int flags = 0;
 464
 465         if (rodata_full || debug_pagealloc_enabled())
 466                 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
 467
 468         /*
 469          * Take care not to create a writable alias for the
 470          * read-only text and rodata sections of the kernel image.
 471          * So temporarily mark them as NOMAP to skip mappings in
 472          * the following for-loop
 473          */
 474         memblock_mark_nomap(kernel_start, kernel_end - kernel_start);
 475 #ifdef CONFIG_KEXEC_CORE
 476         if (crashk_res.end)
 477                 memblock_mark_nomap(crashk_res.start,
 478                                     resource_size(&crashk_res));
 479 #endif
 480
 481         /* map all the memory banks */
 482         for_each_memblock(memory, reg) {
 483                 phys_addr_t start = reg->base;
 484                 phys_addr_t end = start + reg->size;
 485
 486                 if (start >= end)
 487                         break;
 488                 if (memblock_is_nomap(reg))
 489                         continue;
 490
 491                 __map_memblock(pgdp, start, end, PAGE_KERNEL, flags);
 492         }
 493
 494         /*
 495          * Map the linear alias of the [_text, __init_begin) interval
 496          * as non-executable now, and remove the write permission in
 497          * mark_linear_text_alias_ro() below (which will be called after
 498          * alternative patching has completed). This makes the contents
 499          * of the region accessible to subsystems such as hibernate,
 500          * but protects it from inadvertent modification or execution.
 501          * Note that contiguous mappings cannot be remapped in this way,
 502          * so we should avoid them here.
 503          */
 504         __map_memblock(pgdp, kernel_start, kernel_end,
 505                        PAGE_KERNEL, NO_CONT_MAPPINGS);
 506         memblock_clear_nomap(kernel_start, kernel_end - kernel_start);
 507
 508 #ifdef CONFIG_KEXEC_CORE
 509         /*
 510          * Use page-level mappings here so that we can shrink the region
 511          * in page granularity and put back unused memory to buddy system
 512          * through /sys/kernel/kexec_crash_size interface.
 513          */
 514         if (crashk_res.end) {
 515                 __map_memblock(pgdp, crashk_res.start, crashk_res.end + 1,
 516                                PAGE_KERNEL,
 517                                NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
 518                 memblock_clear_nomap(crashk_res.start,
 519                                      resource_size(&crashk_res));
 520         }
 521 #endif
 522 }
 523
 524 void mark_rodata_ro(void)
 525 {
 526         unsigned long section_size;
 527
 528         /*
 529          * mark .rodata as read only. Use __init_begin rather than __end_rodata
 530          * to cover NOTES and EXCEPTION_TABLE.
 531          */
 532         section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
 533         update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata,
 534                             section_size, PAGE_KERNEL_RO);
 535
 536         debug_checkwx();
 537 }
 538
 539 static void __init map_kernel_segment(pgd_t *pgdp, void *va_start, void *va_end,
 540                                       pgprot_t prot, struct vm_struct *vma,
 541                                       int flags, unsigned long vm_flags)
 542 {
 543         phys_addr_t pa_start = __pa_symbol(va_start);
 544         unsigned long size = va_end - va_start;
 545
 546         BUG_ON(!PAGE_ALIGNED(pa_start));
 547         BUG_ON(!PAGE_ALIGNED(size));
 548
 549         __create_pgd_mapping(pgdp, pa_start, (unsigned long)va_start, size, prot,
 550                              early_pgtable_alloc, flags);
 551
 552         if (!(vm_flags & VM_NO_GUARD))
 553                 size += PAGE_SIZE;
 554
 555         vma->addr       = va_start;
 556         vma->phys_addr  = pa_start;
 557         vma->size       = size;
 558         vma->flags      = VM_MAP | vm_flags;
 559         vma->caller     = __builtin_return_address(0);
 560
 561         vm_area_add_early(vma);
 562 }
 563
 564 static int __init parse_rodata(char *arg)
 565 {
 566         int ret = strtobool(arg, &rodata_enabled);
 567         if (!ret) {
 568                 rodata_full = false;
 569                 return 0;
 570         }
 571
 572         /* permit 'full' in addition to boolean options */
 573         if (strcmp(arg, "full"))
 574                 return -EINVAL;
 575
 576         rodata_enabled = true;
 577         rodata_full = true;
 578         return 0;
 579 }
 580 early_param("rodata", parse_rodata);
 581
 582 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
 583 static int __init map_entry_trampoline(void)
 584 {
 585         pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
 586         phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start);
 587
 588         /* The trampoline is always mapped and can therefore be global */
 589         pgprot_val(prot) &= ~PTE_NG;
 590
 591         /* Map only the text into the trampoline page table */
 592         memset(tramp_pg_dir, 0, PGD_SIZE);
 593         __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, PAGE_SIZE,
 594                              prot, __pgd_pgtable_alloc, 0);
 595
 596         /* Map both the text and data into the kernel page table */
 597         __set_fixmap(FIX_ENTRY_TRAMP_TEXT, pa_start, prot);
 598         if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
 599                 extern char __entry_tramp_data_start[];
 600
 601                 __set_fixmap(FIX_ENTRY_TRAMP_DATA,
 602                              __pa_symbol(__entry_tramp_data_start),
 603                              PAGE_KERNEL_RO);
 604         }
 605
 606         return 0;
 607 }
 608 core_initcall(map_entry_trampoline);
 609 #endif
 610
 611 /*
 612  * Create fine-grained mappings for the kernel.
 613  */
 614 static void __init map_kernel(pgd_t *pgdp)
 615 {
 616         static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext,
 617                                 vmlinux_initdata, vmlinux_data;
 618
 619         /*
 620          * External debuggers may need to write directly to the text
 621          * mapping to install SW breakpoints. Allow this (only) when
 622          * explicitly requested with rodata=off.
 623          */
 624         pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
 625
 626         /*
 627          * Only rodata will be remapped with different permissions later on,
 628          * all other segments are allowed to use contiguous mappings.
 629          */
 630         map_kernel_segment(pgdp, _text, _etext, text_prot, &vmlinux_text, 0,
 631                            VM_NO_GUARD);
 632         map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL,
 633                            &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD);
 634         map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot,
 635                            &vmlinux_inittext, 0, VM_NO_GUARD);
 636         map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL,
 637                            &vmlinux_initdata, 0, VM_NO_GUARD);
 638         map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0);
 639
 640         if (!READ_ONCE(pgd_val(*pgd_offset_raw(pgdp, FIXADDR_START)))) {
 641                 /*
 642                  * The fixmap falls in a separate pgd to the kernel, and doesn't
 643                  * live in the carveout for the swapper_pg_dir. We can simply
 644                  * re-use the existing dir for the fixmap.
 645                  */
 646                 set_pgd(pgd_offset_raw(pgdp, FIXADDR_START),
 647                         READ_ONCE(*pgd_offset_k(FIXADDR_START)));
 648         } else if (CONFIG_PGTABLE_LEVELS > 3) {
 649                 /*
 650                  * The fixmap shares its top level pgd entry with the kernel
 651                  * mapping. This can really only occur when we are running
 652                  * with 16k/4 levels, so we can simply reuse the pud level
 653                  * entry instead.
 654                  */
 655                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 656                 pud_populate(&init_mm,
 657                              pud_set_fixmap_offset(pgdp, FIXADDR_START),
 658                              lm_alias(bm_pmd));
 659                 pud_clear_fixmap();
 660         } else {
 661                 BUG();
 662         }
 663
 664         kasan_copy_shadow(pgdp);
 665 }
 666
 667 void __init paging_init(void)
 668 {
 669         pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir));
 670
 671         map_kernel(pgdp);
 672         map_mem(pgdp);
 673
 674         pgd_clear_fixmap();
 675
 676         cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
 677         init_mm.pgd = swapper_pg_dir;
 678
 679         memblock_free(__pa_symbol(init_pg_dir),
 680                       __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir));
 681
 682         memblock_allow_resize();
 683 }
 684
 685 /*
 686  * Check whether a kernel address is valid (derived from arch/x86/).
 687  */
 688 int kern_addr_valid(unsigned long addr)
 689 {
 690         pgd_t *pgdp;
 691         pud_t *pudp, pud;
 692         pmd_t *pmdp, pmd;
 693         pte_t *ptep, pte;
 694
 695         if ((((long)addr) >> VA_BITS) != -1UL)
 696                 return 0;
 697
 698         pgdp = pgd_offset_k(addr);
 699         if (pgd_none(READ_ONCE(*pgdp)))
 700                 return 0;
 701
 702         pudp = pud_offset(pgdp, addr);
 703         pud = READ_ONCE(*pudp);
 704         if (pud_none(pud))
 705                 return 0;
 706
 707         if (pud_sect(pud))
 708                 return pfn_valid(pud_pfn(pud));
 709
 710         pmdp = pmd_offset(pudp, addr);
 711         pmd = READ_ONCE(*pmdp);
 712         if (pmd_none(pmd))
 713                 return 0;
 714
 715         if (pmd_sect(pmd))
 716                 return pfn_valid(pmd_pfn(pmd));
 717
 718         ptep = pte_offset_kernel(pmdp, addr);
 719         pte = READ_ONCE(*ptep);
 720         if (pte_none(pte))
 721                 return 0;
 722
 723         return pfn_valid(pte_pfn(pte));
 724 }
 725 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 726 #if !ARM64_SWAPPER_USES_SECTION_MAPS
 727 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
 728                 struct vmem_altmap *altmap)
 729 {
 730         return vmemmap_populate_basepages(start, end, node);
 731 }
 732 #else   /* !ARM64_SWAPPER_USES_SECTION_MAPS */
 733 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
 734                 struct vmem_altmap *altmap)
 735 {
 736         unsigned long addr = start;
 737         unsigned long next;
 738         pgd_t *pgdp;
 739         pud_t *pudp;
 740         pmd_t *pmdp;
 741
 742         do {
 743                 next = pmd_addr_end(addr, end);
 744
 745                 pgdp = vmemmap_pgd_populate(addr, node);
 746                 if (!pgdp)
 747                         return -ENOMEM;
 748
 749                 pudp = vmemmap_pud_populate(pgdp, addr, node);
 750                 if (!pudp)
 751                         return -ENOMEM;
 752
 753                 pmdp = pmd_offset(pudp, addr);
 754                 if (pmd_none(READ_ONCE(*pmdp))) {
 755                         void *p = NULL;
 756
 757                         p = vmemmap_alloc_block_buf(PMD_SIZE, node);
 758                         if (!p)
 759                                 return -ENOMEM;
 760
 761                         pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL));
 762                 } else
 763                         vmemmap_verify((pte_t *)pmdp, node, addr, next);
 764         } while (addr = next, addr != end);
 765
 766         return 0;
 767 }
 768 #endif  /* !ARM64_SWAPPER_USES_SECTION_MAPS */
 769 void vmemmap_free(unsigned long start, unsigned long end,
 770                 struct vmem_altmap *altmap)
 771 {
 772 }
 773 #endif  /* CONFIG_SPARSEMEM_VMEMMAP */
 774
 775 static inline pud_t * fixmap_pud(unsigned long addr)
 776 {
 777         pgd_t *pgdp = pgd_offset_k(addr);
 778         pgd_t pgd = READ_ONCE(*pgdp);
 779
 780         BUG_ON(pgd_none(pgd) || pgd_bad(pgd));
 781
 782         return pud_offset_kimg(pgdp, addr);
 783 }
 784
 785 static inline pmd_t * fixmap_pmd(unsigned long addr)
 786 {
 787         pud_t *pudp = fixmap_pud(addr);
 788         pud_t pud = READ_ONCE(*pudp);
 789
 790         BUG_ON(pud_none(pud) || pud_bad(pud));
 791
 792         return pmd_offset_kimg(pudp, addr);
 793 }
 794
 795 static inline pte_t * fixmap_pte(unsigned long addr)
 796 {
 797         return &bm_pte[pte_index(addr)];
 798 }
 799
 800 /*
 801  * The p*d_populate functions call virt_to_phys implicitly so they can't be used
 802  * directly on kernel symbols (bm_p*d). This function is called too early to use
 803  * lm_alias so __p*d_populate functions must be used to populate with the
 804  * physical address from __pa_symbol.
 805  */
 806 void __init early_fixmap_init(void)
 807 {
 808         pgd_t *pgdp, pgd;
 809         pud_t *pudp;
 810         pmd_t *pmdp;
 811         unsigned long addr = FIXADDR_START;
 812
 813         pgdp = pgd_offset_k(addr);
 814         pgd = READ_ONCE(*pgdp);
 815         if (CONFIG_PGTABLE_LEVELS > 3 &&
 816             !(pgd_none(pgd) || pgd_page_paddr(pgd) == __pa_symbol(bm_pud))) {
 817                 /*
 818                  * We only end up here if the kernel mapping and the fixmap
 819                  * share the top level pgd entry, which should only happen on
 820                  * 16k/4 levels configurations.
 821                  */
 822                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 823                 pudp = pud_offset_kimg(pgdp, addr);
 824         } else {
 825                 if (pgd_none(pgd))
 826                         __pgd_populate(pgdp, __pa_symbol(bm_pud), PUD_TYPE_TABLE);
 827                 pudp = fixmap_pud(addr);
 828         }
 829         if (pud_none(READ_ONCE(*pudp)))
 830                 __pud_populate(pudp, __pa_symbol(bm_pmd), PMD_TYPE_TABLE);
 831         pmdp = fixmap_pmd(addr);
 832         __pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE);
 833
 834         /*
 835          * The boot-ioremap range spans multiple pmds, for which
 836          * we are not prepared:
 837          */
 838         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 839                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 840
 841         if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
 842              || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 843                 WARN_ON(1);
 844                 pr_warn("pmdp %p != %p, %p\n",
 845                         pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
 846                         fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
 847                 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 848                         fix_to_virt(FIX_BTMAP_BEGIN));
 849                 pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 850                         fix_to_virt(FIX_BTMAP_END));
 851
 852                 pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 853                 pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
 854         }
 855 }
 856
 857 /*
 858  * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we
 859  * ever need to use IPIs for TLB broadcasting, then we're in trouble here.
 860  */
 861 void __set_fixmap(enum fixed_addresses idx,
 862                                phys_addr_t phys, pgprot_t flags)
 863 {
 864         unsigned long addr = __fix_to_virt(idx);
 865         pte_t *ptep;
 866
 867         BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
 868
 869         ptep = fixmap_pte(addr);
 870
 871         if (pgprot_val(flags)) {
 872                 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags));
 873         } else {
 874                 pte_clear(&init_mm, addr, ptep);
 875                 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
 876         }
 877 }
 878
 879 void *__init __fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
 880 {
 881         const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
 882         int offset;
 883         void *dt_virt;
 884
 885         /*
 886          * Check whether the physical FDT address is set and meets the minimum
 887          * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
 888          * at least 8 bytes so that we can always access the magic and size
 889          * fields of the FDT header after mapping the first chunk, double check
 890          * here if that is indeed the case.
 891          */
 892         BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
 893         if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
 894                 return NULL;
 895
 896         /*
 897          * Make sure that the FDT region can be mapped without the need to
 898          * allocate additional translation table pages, so that it is safe
 899          * to call create_mapping_noalloc() this early.
 900          *
 901          * On 64k pages, the FDT will be mapped using PTEs, so we need to
 902          * be in the same PMD as the rest of the fixmap.
 903          * On 4k pages, we'll use section mappings for the FDT so we only
 904          * have to be in the same PUD.
 905          */
 906         BUILD_BUG_ON(dt_virt_base % SZ_2M);
 907
 908         BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
 909                      __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
 910
 911         offset = dt_phys % SWAPPER_BLOCK_SIZE;
 912         dt_virt = (void *)dt_virt_base + offset;
 913
 914         /* map the first chunk so we can read the size from the header */
 915         create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
 916                         dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
 917
 918         if (fdt_magic(dt_virt) != FDT_MAGIC)
 919                 return NULL;
 920
 921         *size = fdt_totalsize(dt_virt);
 922         if (*size > MAX_FDT_SIZE)
 923                 return NULL;
 924
 925         if (offset + *size > SWAPPER_BLOCK_SIZE)
 926                 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
 927                                round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
 928
 929         return dt_virt;
 930 }
 931
 932 void *__init fixmap_remap_fdt(phys_addr_t dt_phys)
 933 {
 934         void *dt_virt;
 935         int size;
 936
 937         dt_virt = __fixmap_remap_fdt(dt_phys, &size, PAGE_KERNEL_RO);
 938         if (!dt_virt)
 939                 return NULL;
 940
 941         memblock_reserve(dt_phys, size);
 942         return dt_virt;
 943 }
 944
 945 int __init arch_ioremap_pud_supported(void)
 946 {
 947         /*
 948          * Only 4k granule supports level 1 block mappings.
 949          * SW table walks can't handle removal of intermediate entries.
 950          */
 951         return IS_ENABLED(CONFIG_ARM64_4K_PAGES) &&
 952                !IS_ENABLED(CONFIG_ARM64_PTDUMP_DEBUGFS);
 953 }
 954
 955 int __init arch_ioremap_pmd_supported(void)
 956 {
 957         /* See arch_ioremap_pud_supported() */
 958         return !IS_ENABLED(CONFIG_ARM64_PTDUMP_DEBUGFS);
 959 }
 960
 961 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
 962 {
 963         pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot));
 964
 965         /* Only allow permission changes for now */
 966         if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)),
 967                                    pud_val(new_pud)))
 968                 return 0;
 969
 970         VM_BUG_ON(phys & ~PUD_MASK);
 971         set_pud(pudp, new_pud);
 972         return 1;
 973 }
 974
 975 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
 976 {
 977         pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot));
 978
 979         /* Only allow permission changes for now */
 980         if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)),
 981                                    pmd_val(new_pmd)))
 982                 return 0;
 983
 984         VM_BUG_ON(phys & ~PMD_MASK);
 985         set_pmd(pmdp, new_pmd);
 986         return 1;
 987 }
 988
 989 int pud_clear_huge(pud_t *pudp)
 990 {
 991         if (!pud_sect(READ_ONCE(*pudp)))
 992                 return 0;
 993         pud_clear(pudp);
 994         return 1;
 995 }
 996
 997 int pmd_clear_huge(pmd_t *pmdp)
 998 {
 999         if (!pmd_sect(READ_ONCE(*pmdp)))
1000                 return 0;
1001         pmd_clear(pmdp);
1002         return 1;
1003 }
1004
1005 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
1006 {
1007         pte_t *table;
1008         pmd_t pmd;
1009
1010         pmd = READ_ONCE(*pmdp);
1011
1012         if (!pmd_table(pmd)) {
1013                 VM_WARN_ON(1);
1014                 return 1;
1015         }
1016
1017         table = pte_offset_kernel(pmdp, addr);
1018         pmd_clear(pmdp);
1019         __flush_tlb_kernel_pgtable(addr);
1020         pte_free_kernel(NULL, table);
1021         return 1;
1022 }
1023
1024 int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
1025 {
1026         pmd_t *table;
1027         pmd_t *pmdp;
1028         pud_t pud;
1029         unsigned long next, end;
1030
1031         pud = READ_ONCE(*pudp);
1032
1033         if (!pud_table(pud)) {
1034                 VM_WARN_ON(1);
1035                 return 1;
1036         }
1037
1038         table = pmd_offset(pudp, addr);
1039         pmdp = table;
1040         next = addr;
1041         end = addr + PUD_SIZE;
1042         do {
1043                 pmd_free_pte_page(pmdp, next);
1044         } while (pmdp++, next += PMD_SIZE, next != end);
1045
1046         pud_clear(pudp);
1047         __flush_tlb_kernel_pgtable(addr);
1048         pmd_free(NULL, table);
1049         return 1;
1050 }
1051
1052 int p4d_free_pud_page(p4d_t *p4d, unsigned long addr)
1053 {
1054         return 0;       /* Don't attempt a block mapping */
1055 }
1056
1057 #ifdef CONFIG_MEMORY_HOTPLUG
1058 int arch_add_memory(int nid, u64 start, u64 size,
1059                         struct mhp_restrictions *restrictions)
1060 {
1061         int flags = 0;
1062
1063         if (rodata_full || debug_pagealloc_enabled())
1064                 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
1065
1066         __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
1067                              size, PAGE_KERNEL, __pgd_pgtable_alloc, flags);
1068
1069         return __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT,
1070                            restrictions);
1071 }
1072 #endif