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/memremap.h>
  21 #include <linux/memory.h>
  22 #include <linux/fs.h>
  23 #include <linux/io.h>
  24 #include <linux/mm.h>
  25 #include <linux/vmalloc.h>
  26 #include <linux/set_memory.h>
  27 #include <linux/kfence.h>
  28
  29 #include <asm/barrier.h>
  30 #include <asm/cputype.h>
  31 #include <asm/fixmap.h>
  32 #include <asm/kasan.h>
  33 #include <asm/kernel-pgtable.h>
  34 #include <asm/sections.h>
  35 #include <asm/setup.h>
  36 #include <linux/sizes.h>
  37 #include <asm/tlb.h>
  38 #include <asm/mmu_context.h>
  39 #include <asm/ptdump.h>
  40 #include <asm/tlbflush.h>
  41 #include <asm/pgalloc.h>
  42 #include <asm/kfence.h>
  43
  44 #define NO_BLOCK_MAPPINGS       BIT(0)
  45 #define NO_CONT_MAPPINGS        BIT(1)
  46 #define NO_EXEC_MAPPINGS        BIT(2)  /* assumes FEAT_HPDS is not used */
  47
  48 int idmap_t0sz __ro_after_init;
  49
  50 #if VA_BITS > 48
  51 u64 vabits_actual __ro_after_init = VA_BITS_MIN;
  52 EXPORT_SYMBOL(vabits_actual);
  53 #endif
  54
  55 u64 kimage_vaddr __ro_after_init = (u64)&_text;
  56 EXPORT_SYMBOL(kimage_vaddr);
  57
  58 u64 kimage_voffset __ro_after_init;
  59 EXPORT_SYMBOL(kimage_voffset);
  60
  61 u32 __boot_cpu_mode[] = { BOOT_CPU_MODE_EL2, BOOT_CPU_MODE_EL1 };
  62
  63 /*
  64  * The booting CPU updates the failed status @__early_cpu_boot_status,
  65  * with MMU turned off.
  66  */
  67 long __section(".mmuoff.data.write") __early_cpu_boot_status;
  68
  69 /*
  70  * Empty_zero_page is a special page that is used for zero-initialized data
  71  * and COW.
  72  */
  73 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
  74 EXPORT_SYMBOL(empty_zero_page);
  75
  76 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
  77 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused;
  78 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused;
  79
  80 static DEFINE_SPINLOCK(swapper_pgdir_lock);
  81 static DEFINE_MUTEX(fixmap_lock);
  82
  83 void set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
  84 {
  85         pgd_t *fixmap_pgdp;
  86
  87         spin_lock(&swapper_pgdir_lock);
  88         fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp));
  89         WRITE_ONCE(*fixmap_pgdp, pgd);
  90         /*
  91          * We need dsb(ishst) here to ensure the page-table-walker sees
  92          * our new entry before set_p?d() returns. The fixmap's
  93          * flush_tlb_kernel_range() via clear_fixmap() does this for us.
  94          */
  95         pgd_clear_fixmap();
  96         spin_unlock(&swapper_pgdir_lock);
  97 }
  98
  99 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
 100                               unsigned long size, pgprot_t vma_prot)
 101 {
 102         if (!pfn_is_map_memory(pfn))
 103                 return pgprot_noncached(vma_prot);
 104         else if (file->f_flags & O_SYNC)
 105                 return pgprot_writecombine(vma_prot);
 106         return vma_prot;
 107 }
 108 EXPORT_SYMBOL(phys_mem_access_prot);
 109
 110 static phys_addr_t __init early_pgtable_alloc(int shift)
 111 {
 112         phys_addr_t phys;
 113         void *ptr;
 114
 115         phys = memblock_phys_alloc_range(PAGE_SIZE, PAGE_SIZE, 0,
 116                                          MEMBLOCK_ALLOC_NOLEAKTRACE);
 117         if (!phys)
 118                 panic("Failed to allocate page table page\n");
 119
 120         /*
 121          * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
 122          * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
 123          * any level of table.
 124          */
 125         ptr = pte_set_fixmap(phys);
 126
 127         memset(ptr, 0, PAGE_SIZE);
 128
 129         /*
 130          * Implicit barriers also ensure the zeroed page is visible to the page
 131          * table walker
 132          */
 133         pte_clear_fixmap();
 134
 135         return phys;
 136 }
 137
 138 static bool pgattr_change_is_safe(u64 old, u64 new)
 139 {
 140         /*
 141          * The following mapping attributes may be updated in live
 142          * kernel mappings without the need for break-before-make.
 143          */
 144         pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG;
 145
 146         /* creating or taking down mappings is always safe */
 147         if (old == 0 || new == 0)
 148                 return true;
 149
 150         /* live contiguous mappings may not be manipulated at all */
 151         if ((old | new) & PTE_CONT)
 152                 return false;
 153
 154         /* Transitioning from Non-Global to Global is unsafe */
 155         if (old & ~new & PTE_NG)
 156                 return false;
 157
 158         /*
 159          * Changing the memory type between Normal and Normal-Tagged is safe
 160          * since Tagged is considered a permission attribute from the
 161          * mismatched attribute aliases perspective.
 162          */
 163         if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
 164              (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) &&
 165             ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
 166              (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)))
 167                 mask |= PTE_ATTRINDX_MASK;
 168
 169         return ((old ^ new) & ~mask) == 0;
 170 }
 171
 172 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end,
 173                      phys_addr_t phys, pgprot_t prot)
 174 {
 175         pte_t *ptep;
 176
 177         ptep = pte_set_fixmap_offset(pmdp, addr);
 178         do {
 179                 pte_t old_pte = READ_ONCE(*ptep);
 180
 181                 set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot));
 182
 183                 /*
 184                  * After the PTE entry has been populated once, we
 185                  * only allow updates to the permission attributes.
 186                  */
 187                 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte),
 188                                               READ_ONCE(pte_val(*ptep))));
 189
 190                 phys += PAGE_SIZE;
 191         } while (ptep++, addr += PAGE_SIZE, addr != end);
 192
 193         pte_clear_fixmap();
 194 }
 195
 196 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
 197                                 unsigned long end, phys_addr_t phys,
 198                                 pgprot_t prot,
 199                                 phys_addr_t (*pgtable_alloc)(int),
 200                                 int flags)
 201 {
 202         unsigned long next;
 203         pmd_t pmd = READ_ONCE(*pmdp);
 204
 205         BUG_ON(pmd_sect(pmd));
 206         if (pmd_none(pmd)) {
 207                 pmdval_t pmdval = PMD_TYPE_TABLE | PMD_TABLE_UXN;
 208                 phys_addr_t pte_phys;
 209
 210                 if (flags & NO_EXEC_MAPPINGS)
 211                         pmdval |= PMD_TABLE_PXN;
 212                 BUG_ON(!pgtable_alloc);
 213                 pte_phys = pgtable_alloc(PAGE_SHIFT);
 214                 __pmd_populate(pmdp, pte_phys, pmdval);
 215                 pmd = READ_ONCE(*pmdp);
 216         }
 217         BUG_ON(pmd_bad(pmd));
 218
 219         do {
 220                 pgprot_t __prot = prot;
 221
 222                 next = pte_cont_addr_end(addr, end);
 223
 224                 /* use a contiguous mapping if the range is suitably aligned */
 225                 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) &&
 226                     (flags & NO_CONT_MAPPINGS) == 0)
 227                         __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
 228
 229                 init_pte(pmdp, addr, next, phys, __prot);
 230
 231                 phys += next - addr;
 232         } while (addr = next, addr != end);
 233 }
 234
 235 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end,
 236                      phys_addr_t phys, pgprot_t prot,
 237                      phys_addr_t (*pgtable_alloc)(int), int flags)
 238 {
 239         unsigned long next;
 240         pmd_t *pmdp;
 241
 242         pmdp = pmd_set_fixmap_offset(pudp, addr);
 243         do {
 244                 pmd_t old_pmd = READ_ONCE(*pmdp);
 245
 246                 next = pmd_addr_end(addr, end);
 247
 248                 /* try section mapping first */
 249                 if (((addr | next | phys) & ~PMD_MASK) == 0 &&
 250                     (flags & NO_BLOCK_MAPPINGS) == 0) {
 251                         pmd_set_huge(pmdp, phys, prot);
 252
 253                         /*
 254                          * After the PMD entry has been populated once, we
 255                          * only allow updates to the permission attributes.
 256                          */
 257                         BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
 258                                                       READ_ONCE(pmd_val(*pmdp))));
 259                 } else {
 260                         alloc_init_cont_pte(pmdp, addr, next, phys, prot,
 261                                             pgtable_alloc, flags);
 262
 263                         BUG_ON(pmd_val(old_pmd) != 0 &&
 264                                pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp)));
 265                 }
 266                 phys += next - addr;
 267         } while (pmdp++, addr = next, addr != end);
 268
 269         pmd_clear_fixmap();
 270 }
 271
 272 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
 273                                 unsigned long end, phys_addr_t phys,
 274                                 pgprot_t prot,
 275                                 phys_addr_t (*pgtable_alloc)(int), int flags)
 276 {
 277         unsigned long next;
 278         pud_t pud = READ_ONCE(*pudp);
 279
 280         /*
 281          * Check for initial section mappings in the pgd/pud.
 282          */
 283         BUG_ON(pud_sect(pud));
 284         if (pud_none(pud)) {
 285                 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN;
 286                 phys_addr_t pmd_phys;
 287
 288                 if (flags & NO_EXEC_MAPPINGS)
 289                         pudval |= PUD_TABLE_PXN;
 290                 BUG_ON(!pgtable_alloc);
 291                 pmd_phys = pgtable_alloc(PMD_SHIFT);
 292                 __pud_populate(pudp, pmd_phys, pudval);
 293                 pud = READ_ONCE(*pudp);
 294         }
 295         BUG_ON(pud_bad(pud));
 296
 297         do {
 298                 pgprot_t __prot = prot;
 299
 300                 next = pmd_cont_addr_end(addr, end);
 301
 302                 /* use a contiguous mapping if the range is suitably aligned */
 303                 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) &&
 304                     (flags & NO_CONT_MAPPINGS) == 0)
 305                         __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
 306
 307                 init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags);
 308
 309                 phys += next - addr;
 310         } while (addr = next, addr != end);
 311 }
 312
 313 static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end,
 314                            phys_addr_t phys, pgprot_t prot,
 315                            phys_addr_t (*pgtable_alloc)(int),
 316                            int flags)
 317 {
 318         unsigned long next;
 319         pud_t *pudp;
 320         p4d_t *p4dp = p4d_offset(pgdp, addr);
 321         p4d_t p4d = READ_ONCE(*p4dp);
 322
 323         if (p4d_none(p4d)) {
 324                 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN;
 325                 phys_addr_t pud_phys;
 326
 327                 if (flags & NO_EXEC_MAPPINGS)
 328                         p4dval |= P4D_TABLE_PXN;
 329                 BUG_ON(!pgtable_alloc);
 330                 pud_phys = pgtable_alloc(PUD_SHIFT);
 331                 __p4d_populate(p4dp, pud_phys, p4dval);
 332                 p4d = READ_ONCE(*p4dp);
 333         }
 334         BUG_ON(p4d_bad(p4d));
 335
 336         pudp = pud_set_fixmap_offset(p4dp, addr);
 337         do {
 338                 pud_t old_pud = READ_ONCE(*pudp);
 339
 340                 next = pud_addr_end(addr, end);
 341
 342                 /*
 343                  * For 4K granule only, attempt to put down a 1GB block
 344                  */
 345                 if (pud_sect_supported() &&
 346                    ((addr | next | phys) & ~PUD_MASK) == 0 &&
 347                     (flags & NO_BLOCK_MAPPINGS) == 0) {
 348                         pud_set_huge(pudp, phys, prot);
 349
 350                         /*
 351                          * After the PUD entry has been populated once, we
 352                          * only allow updates to the permission attributes.
 353                          */
 354                         BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
 355                                                       READ_ONCE(pud_val(*pudp))));
 356                 } else {
 357                         alloc_init_cont_pmd(pudp, addr, next, phys, prot,
 358                                             pgtable_alloc, flags);
 359
 360                         BUG_ON(pud_val(old_pud) != 0 &&
 361                                pud_val(old_pud) != READ_ONCE(pud_val(*pudp)));
 362                 }
 363                 phys += next - addr;
 364         } while (pudp++, addr = next, addr != end);
 365
 366         pud_clear_fixmap();
 367 }
 368
 369 static void __create_pgd_mapping_locked(pgd_t *pgdir, phys_addr_t phys,
 370                                         unsigned long virt, phys_addr_t size,
 371                                         pgprot_t prot,
 372                                         phys_addr_t (*pgtable_alloc)(int),
 373                                         int flags)
 374 {
 375         unsigned long addr, end, next;
 376         pgd_t *pgdp = pgd_offset_pgd(pgdir, virt);
 377
 378         /*
 379          * If the virtual and physical address don't have the same offset
 380          * within a page, we cannot map the region as the caller expects.
 381          */
 382         if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
 383                 return;
 384
 385         phys &= PAGE_MASK;
 386         addr = virt & PAGE_MASK;
 387         end = PAGE_ALIGN(virt + size);
 388
 389         do {
 390                 next = pgd_addr_end(addr, end);
 391                 alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc,
 392                                flags);
 393                 phys += next - addr;
 394         } while (pgdp++, addr = next, addr != end);
 395 }
 396
 397 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
 398                                  unsigned long virt, phys_addr_t size,
 399                                  pgprot_t prot,
 400                                  phys_addr_t (*pgtable_alloc)(int),
 401                                  int flags)
 402 {
 403         mutex_lock(&fixmap_lock);
 404         __create_pgd_mapping_locked(pgdir, phys, virt, size, prot,
 405                                     pgtable_alloc, flags);
 406         mutex_unlock(&fixmap_lock);
 407 }
 408
 409 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
 410 extern __alias(__create_pgd_mapping_locked)
 411 void create_kpti_ng_temp_pgd(pgd_t *pgdir, phys_addr_t phys, unsigned long virt,
 412                              phys_addr_t size, pgprot_t prot,
 413                              phys_addr_t (*pgtable_alloc)(int), int flags);
 414 #endif
 415
 416 static phys_addr_t __pgd_pgtable_alloc(int shift)
 417 {
 418         void *ptr = (void *)__get_free_page(GFP_PGTABLE_KERNEL);
 419         BUG_ON(!ptr);
 420
 421         /* Ensure the zeroed page is visible to the page table walker */
 422         dsb(ishst);
 423         return __pa(ptr);
 424 }
 425
 426 static phys_addr_t pgd_pgtable_alloc(int shift)
 427 {
 428         phys_addr_t pa = __pgd_pgtable_alloc(shift);
 429
 430         /*
 431          * Call proper page table ctor in case later we need to
 432          * call core mm functions like apply_to_page_range() on
 433          * this pre-allocated page table.
 434          *
 435          * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is
 436          * folded, and if so pgtable_pmd_page_ctor() becomes nop.
 437          */
 438         if (shift == PAGE_SHIFT)
 439                 BUG_ON(!pgtable_pte_page_ctor(phys_to_page(pa)));
 440         else if (shift == PMD_SHIFT)
 441                 BUG_ON(!pgtable_pmd_page_ctor(phys_to_page(pa)));
 442
 443         return pa;
 444 }
 445
 446 /*
 447  * This function can only be used to modify existing table entries,
 448  * without allocating new levels of table. Note that this permits the
 449  * creation of new section or page entries.
 450  */
 451 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
 452                                   phys_addr_t size, pgprot_t prot)
 453 {
 454         if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
 455                 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
 456                         &phys, virt);
 457                 return;
 458         }
 459         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
 460                              NO_CONT_MAPPINGS);
 461 }
 462
 463 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
 464                                unsigned long virt, phys_addr_t size,
 465                                pgprot_t prot, bool page_mappings_only)
 466 {
 467         int flags = 0;
 468
 469         BUG_ON(mm == &init_mm);
 470
 471         if (page_mappings_only)
 472                 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
 473
 474         __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
 475                              pgd_pgtable_alloc, flags);
 476 }
 477
 478 static void update_mapping_prot(phys_addr_t phys, unsigned long virt,
 479                                 phys_addr_t size, pgprot_t prot)
 480 {
 481         if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
 482                 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n",
 483                         &phys, virt);
 484                 return;
 485         }
 486
 487         __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
 488                              NO_CONT_MAPPINGS);
 489
 490         /* flush the TLBs after updating live kernel mappings */
 491         flush_tlb_kernel_range(virt, virt + size);
 492 }
 493
 494 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start,
 495                                   phys_addr_t end, pgprot_t prot, int flags)
 496 {
 497         __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start,
 498                              prot, early_pgtable_alloc, flags);
 499 }
 500
 501 void __init mark_linear_text_alias_ro(void)
 502 {
 503         /*
 504          * Remove the write permissions from the linear alias of .text/.rodata
 505          */
 506         update_mapping_prot(__pa_symbol(_stext), (unsigned long)lm_alias(_stext),
 507                             (unsigned long)__init_begin - (unsigned long)_stext,
 508                             PAGE_KERNEL_RO);
 509 }
 510
 511 static bool crash_mem_map __initdata;
 512
 513 static int __init enable_crash_mem_map(char *arg)
 514 {
 515         /*
 516          * Proper parameter parsing is done by reserve_crashkernel(). We only
 517          * need to know if the linear map has to avoid block mappings so that
 518          * the crashkernel reservations can be unmapped later.
 519          */
 520         crash_mem_map = true;
 521
 522         return 0;
 523 }
 524 early_param("crashkernel", enable_crash_mem_map);
 525
 526 #ifdef CONFIG_KFENCE
 527
 528 bool __ro_after_init kfence_early_init = !!CONFIG_KFENCE_SAMPLE_INTERVAL;
 529
 530 /* early_param() will be parsed before map_mem() below. */
 531 static int __init parse_kfence_early_init(char *arg)
 532 {
 533         int val;
 534
 535         if (get_option(&arg, &val))
 536                 kfence_early_init = !!val;
 537         return 0;
 538 }
 539 early_param("kfence.sample_interval", parse_kfence_early_init);
 540
 541 static phys_addr_t __init arm64_kfence_alloc_pool(void)
 542 {
 543         phys_addr_t kfence_pool;
 544
 545         if (!kfence_early_init)
 546                 return 0;
 547
 548         kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
 549         if (!kfence_pool) {
 550                 pr_err("failed to allocate kfence pool\n");
 551                 kfence_early_init = false;
 552                 return 0;
 553         }
 554
 555         /* Temporarily mark as NOMAP. */
 556         memblock_mark_nomap(kfence_pool, KFENCE_POOL_SIZE);
 557
 558         return kfence_pool;
 559 }
 560
 561 static void __init arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp)
 562 {
 563         if (!kfence_pool)
 564                 return;
 565
 566         /* KFENCE pool needs page-level mapping. */
 567         __map_memblock(pgdp, kfence_pool, kfence_pool + KFENCE_POOL_SIZE,
 568                         pgprot_tagged(PAGE_KERNEL),
 569                         NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
 570         memblock_clear_nomap(kfence_pool, KFENCE_POOL_SIZE);
 571         __kfence_pool = phys_to_virt(kfence_pool);
 572 }
 573 #else /* CONFIG_KFENCE */
 574
 575 static inline phys_addr_t arm64_kfence_alloc_pool(void) { return 0; }
 576 static inline void arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) { }
 577
 578 #endif /* CONFIG_KFENCE */
 579
 580 static void __init map_mem(pgd_t *pgdp)
 581 {
 582         static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN);
 583         phys_addr_t kernel_start = __pa_symbol(_stext);
 584         phys_addr_t kernel_end = __pa_symbol(__init_begin);
 585         phys_addr_t start, end;
 586         phys_addr_t early_kfence_pool;
 587         int flags = NO_EXEC_MAPPINGS;
 588         u64 i;
 589
 590         /*
 591          * Setting hierarchical PXNTable attributes on table entries covering
 592          * the linear region is only possible if it is guaranteed that no table
 593          * entries at any level are being shared between the linear region and
 594          * the vmalloc region. Check whether this is true for the PGD level, in
 595          * which case it is guaranteed to be true for all other levels as well.
 596          */
 597         BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end));
 598
 599         early_kfence_pool = arm64_kfence_alloc_pool();
 600
 601         if (can_set_direct_map())
 602                 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
 603
 604         /*
 605          * Take care not to create a writable alias for the
 606          * read-only text and rodata sections of the kernel image.
 607          * So temporarily mark them as NOMAP to skip mappings in
 608          * the following for-loop
 609          */
 610         memblock_mark_nomap(kernel_start, kernel_end - kernel_start);
 611
 612 #ifdef CONFIG_KEXEC_CORE
 613         if (crash_mem_map) {
 614                 if (defer_reserve_crashkernel())
 615                         flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
 616                 else if (crashk_res.end)
 617                         memblock_mark_nomap(crashk_res.start,
 618                             resource_size(&crashk_res));
 619         }
 620 #endif
 621
 622         /* map all the memory banks */
 623         for_each_mem_range(i, &start, &end) {
 624                 if (start >= end)
 625                         break;
 626                 /*
 627                  * The linear map must allow allocation tags reading/writing
 628                  * if MTE is present. Otherwise, it has the same attributes as
 629                  * PAGE_KERNEL.
 630                  */
 631                 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL),
 632                                flags);
 633         }
 634
 635         /*
 636          * Map the linear alias of the [_stext, __init_begin) interval
 637          * as non-executable now, and remove the write permission in
 638          * mark_linear_text_alias_ro() below (which will be called after
 639          * alternative patching has completed). This makes the contents
 640          * of the region accessible to subsystems such as hibernate,
 641          * but protects it from inadvertent modification or execution.
 642          * Note that contiguous mappings cannot be remapped in this way,
 643          * so we should avoid them here.
 644          */
 645         __map_memblock(pgdp, kernel_start, kernel_end,
 646                        PAGE_KERNEL, NO_CONT_MAPPINGS);
 647         memblock_clear_nomap(kernel_start, kernel_end - kernel_start);
 648
 649         /*
 650          * Use page-level mappings here so that we can shrink the region
 651          * in page granularity and put back unused memory to buddy system
 652          * through /sys/kernel/kexec_crash_size interface.
 653          */
 654 #ifdef CONFIG_KEXEC_CORE
 655         if (crash_mem_map && !defer_reserve_crashkernel()) {
 656                 if (crashk_res.end) {
 657                         __map_memblock(pgdp, crashk_res.start,
 658                                        crashk_res.end + 1,
 659                                        PAGE_KERNEL,
 660                                        NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
 661                         memblock_clear_nomap(crashk_res.start,
 662                                              resource_size(&crashk_res));
 663                 }
 664         }
 665 #endif
 666
 667         arm64_kfence_map_pool(early_kfence_pool, pgdp);
 668 }
 669
 670 void mark_rodata_ro(void)
 671 {
 672         unsigned long section_size;
 673
 674         /*
 675          * mark .rodata as read only. Use __init_begin rather than __end_rodata
 676          * to cover NOTES and EXCEPTION_TABLE.
 677          */
 678         section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
 679         update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata,
 680                             section_size, PAGE_KERNEL_RO);
 681
 682         debug_checkwx();
 683 }
 684
 685 static void __init map_kernel_segment(pgd_t *pgdp, void *va_start, void *va_end,
 686                                       pgprot_t prot, struct vm_struct *vma,
 687                                       int flags, unsigned long vm_flags)
 688 {
 689         phys_addr_t pa_start = __pa_symbol(va_start);
 690         unsigned long size = va_end - va_start;
 691
 692         BUG_ON(!PAGE_ALIGNED(pa_start));
 693         BUG_ON(!PAGE_ALIGNED(size));
 694
 695         __create_pgd_mapping(pgdp, pa_start, (unsigned long)va_start, size, prot,
 696                              early_pgtable_alloc, flags);
 697
 698         if (!(vm_flags & VM_NO_GUARD))
 699                 size += PAGE_SIZE;
 700
 701         vma->addr       = va_start;
 702         vma->phys_addr  = pa_start;
 703         vma->size       = size;
 704         vma->flags      = VM_MAP | vm_flags;
 705         vma->caller     = __builtin_return_address(0);
 706
 707         vm_area_add_early(vma);
 708 }
 709
 710 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
 711 static int __init map_entry_trampoline(void)
 712 {
 713         int i;
 714
 715         pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
 716         phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start);
 717
 718         /* The trampoline is always mapped and can therefore be global */
 719         pgprot_val(prot) &= ~PTE_NG;
 720
 721         /* Map only the text into the trampoline page table */
 722         memset(tramp_pg_dir, 0, PGD_SIZE);
 723         __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS,
 724                              entry_tramp_text_size(), prot,
 725                              __pgd_pgtable_alloc, NO_BLOCK_MAPPINGS);
 726
 727         /* Map both the text and data into the kernel page table */
 728         for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++)
 729                 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i,
 730                              pa_start + i * PAGE_SIZE, prot);
 731
 732         if (IS_ENABLED(CONFIG_RELOCATABLE))
 733                 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i,
 734                              pa_start + i * PAGE_SIZE, PAGE_KERNEL_RO);
 735
 736         return 0;
 737 }
 738 core_initcall(map_entry_trampoline);
 739 #endif
 740
 741 /*
 742  * Open coded check for BTI, only for use to determine configuration
 743  * for early mappings for before the cpufeature code has run.
 744  */
 745 static bool arm64_early_this_cpu_has_bti(void)
 746 {
 747         u64 pfr1;
 748
 749         if (!IS_ENABLED(CONFIG_ARM64_BTI_KERNEL))
 750                 return false;
 751
 752         pfr1 = __read_sysreg_by_encoding(SYS_ID_AA64PFR1_EL1);
 753         return cpuid_feature_extract_unsigned_field(pfr1,
 754                                                     ID_AA64PFR1_EL1_BT_SHIFT);
 755 }
 756
 757 /*
 758  * Create fine-grained mappings for the kernel.
 759  */
 760 static void __init map_kernel(pgd_t *pgdp)
 761 {
 762         static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext,
 763                                 vmlinux_initdata, vmlinux_data;
 764
 765         /*
 766          * External debuggers may need to write directly to the text
 767          * mapping to install SW breakpoints. Allow this (only) when
 768          * explicitly requested with rodata=off.
 769          */
 770         pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
 771
 772         /*
 773          * If we have a CPU that supports BTI and a kernel built for
 774          * BTI then mark the kernel executable text as guarded pages
 775          * now so we don't have to rewrite the page tables later.
 776          */
 777         if (arm64_early_this_cpu_has_bti())
 778                 text_prot = __pgprot_modify(text_prot, PTE_GP, PTE_GP);
 779
 780         /*
 781          * Only rodata will be remapped with different permissions later on,
 782          * all other segments are allowed to use contiguous mappings.
 783          */
 784         map_kernel_segment(pgdp, _stext, _etext, text_prot, &vmlinux_text, 0,
 785                            VM_NO_GUARD);
 786         map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL,
 787                            &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD);
 788         map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot,
 789                            &vmlinux_inittext, 0, VM_NO_GUARD);
 790         map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL,
 791                            &vmlinux_initdata, 0, VM_NO_GUARD);
 792         map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0);
 793
 794         if (!READ_ONCE(pgd_val(*pgd_offset_pgd(pgdp, FIXADDR_START)))) {
 795                 /*
 796                  * The fixmap falls in a separate pgd to the kernel, and doesn't
 797                  * live in the carveout for the swapper_pg_dir. We can simply
 798                  * re-use the existing dir for the fixmap.
 799                  */
 800                 set_pgd(pgd_offset_pgd(pgdp, FIXADDR_START),
 801                         READ_ONCE(*pgd_offset_k(FIXADDR_START)));
 802         } else if (CONFIG_PGTABLE_LEVELS > 3) {
 803                 pgd_t *bm_pgdp;
 804                 p4d_t *bm_p4dp;
 805                 pud_t *bm_pudp;
 806                 /*
 807                  * The fixmap shares its top level pgd entry with the kernel
 808                  * mapping. This can really only occur when we are running
 809                  * with 16k/4 levels, so we can simply reuse the pud level
 810                  * entry instead.
 811                  */
 812                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
 813                 bm_pgdp = pgd_offset_pgd(pgdp, FIXADDR_START);
 814                 bm_p4dp = p4d_offset(bm_pgdp, FIXADDR_START);
 815                 bm_pudp = pud_set_fixmap_offset(bm_p4dp, FIXADDR_START);
 816                 pud_populate(&init_mm, bm_pudp, lm_alias(bm_pmd));
 817                 pud_clear_fixmap();
 818         } else {
 819                 BUG();
 820         }
 821
 822         kasan_copy_shadow(pgdp);
 823 }
 824
 825 static void __init create_idmap(void)
 826 {
 827         u64 start = __pa_symbol(__idmap_text_start);
 828         u64 size = __pa_symbol(__idmap_text_end) - start;
 829         pgd_t *pgd = idmap_pg_dir;
 830         u64 pgd_phys;
 831
 832         /* check if we need an additional level of translation */
 833         if (VA_BITS < 48 && idmap_t0sz < (64 - VA_BITS_MIN)) {
 834                 pgd_phys = early_pgtable_alloc(PAGE_SHIFT);
 835                 set_pgd(&idmap_pg_dir[start >> VA_BITS],
 836                         __pgd(pgd_phys | P4D_TYPE_TABLE));
 837                 pgd = __va(pgd_phys);
 838         }
 839         __create_pgd_mapping(pgd, start, start, size, PAGE_KERNEL_ROX,
 840                              early_pgtable_alloc, 0);
 841
 842         if (IS_ENABLED(CONFIG_UNMAP_KERNEL_AT_EL0)) {
 843                 extern u32 __idmap_kpti_flag;
 844                 u64 pa = __pa_symbol(&__idmap_kpti_flag);
 845
 846                 /*
 847                  * The KPTI G-to-nG conversion code needs a read-write mapping
 848                  * of its synchronization flag in the ID map.
 849                  */
 850                 __create_pgd_mapping(pgd, pa, pa, sizeof(u32), PAGE_KERNEL,
 851                                      early_pgtable_alloc, 0);
 852         }
 853 }
 854
 855 void __init paging_init(void)
 856 {
 857         pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir));
 858         extern pgd_t init_idmap_pg_dir[];
 859
 860         idmap_t0sz = 63UL - __fls(__pa_symbol(_end) | GENMASK(VA_BITS_MIN - 1, 0));
 861
 862         map_kernel(pgdp);
 863         map_mem(pgdp);
 864
 865         pgd_clear_fixmap();
 866
 867         cpu_replace_ttbr1(lm_alias(swapper_pg_dir), init_idmap_pg_dir);
 868         init_mm.pgd = swapper_pg_dir;
 869
 870         memblock_phys_free(__pa_symbol(init_pg_dir),
 871                            __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir));
 872
 873         memblock_allow_resize();
 874
 875         create_idmap();
 876 }
 877
 878 /*
 879  * Check whether a kernel address is valid (derived from arch/x86/).
 880  */
 881 int kern_addr_valid(unsigned long addr)
 882 {
 883         pgd_t *pgdp;
 884         p4d_t *p4dp;
 885         pud_t *pudp, pud;
 886         pmd_t *pmdp, pmd;
 887         pte_t *ptep, pte;
 888
 889         addr = arch_kasan_reset_tag(addr);
 890         if ((((long)addr) >> VA_BITS) != -1UL)
 891                 return 0;
 892
 893         pgdp = pgd_offset_k(addr);
 894         if (pgd_none(READ_ONCE(*pgdp)))
 895                 return 0;
 896
 897         p4dp = p4d_offset(pgdp, addr);
 898         if (p4d_none(READ_ONCE(*p4dp)))
 899                 return 0;
 900
 901         pudp = pud_offset(p4dp, addr);
 902         pud = READ_ONCE(*pudp);
 903         if (pud_none(pud))
 904                 return 0;
 905
 906         if (pud_sect(pud))
 907                 return pfn_valid(pud_pfn(pud));
 908
 909         pmdp = pmd_offset(pudp, addr);
 910         pmd = READ_ONCE(*pmdp);
 911         if (pmd_none(pmd))
 912                 return 0;
 913
 914         if (pmd_sect(pmd))
 915                 return pfn_valid(pmd_pfn(pmd));
 916
 917         ptep = pte_offset_kernel(pmdp, addr);
 918         pte = READ_ONCE(*ptep);
 919         if (pte_none(pte))
 920                 return 0;
 921
 922         return pfn_valid(pte_pfn(pte));
 923 }
 924
 925 #ifdef CONFIG_MEMORY_HOTPLUG
 926 static void free_hotplug_page_range(struct page *page, size_t size,
 927                                     struct vmem_altmap *altmap)
 928 {
 929         if (altmap) {
 930                 vmem_altmap_free(altmap, size >> PAGE_SHIFT);
 931         } else {
 932                 WARN_ON(PageReserved(page));
 933                 free_pages((unsigned long)page_address(page), get_order(size));
 934         }
 935 }
 936
 937 static void free_hotplug_pgtable_page(struct page *page)
 938 {
 939         free_hotplug_page_range(page, PAGE_SIZE, NULL);
 940 }
 941
 942 static bool pgtable_range_aligned(unsigned long start, unsigned long end,
 943                                   unsigned long floor, unsigned long ceiling,
 944                                   unsigned long mask)
 945 {
 946         start &= mask;
 947         if (start < floor)
 948                 return false;
 949
 950         if (ceiling) {
 951                 ceiling &= mask;
 952                 if (!ceiling)
 953                         return false;
 954         }
 955
 956         if (end - 1 > ceiling - 1)
 957                 return false;
 958         return true;
 959 }
 960
 961 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr,
 962                                     unsigned long end, bool free_mapped,
 963                                     struct vmem_altmap *altmap)
 964 {
 965         pte_t *ptep, pte;
 966
 967         do {
 968                 ptep = pte_offset_kernel(pmdp, addr);
 969                 pte = READ_ONCE(*ptep);
 970                 if (pte_none(pte))
 971                         continue;
 972
 973                 WARN_ON(!pte_present(pte));
 974                 pte_clear(&init_mm, addr, ptep);
 975                 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
 976                 if (free_mapped)
 977                         free_hotplug_page_range(pte_page(pte),
 978                                                 PAGE_SIZE, altmap);
 979         } while (addr += PAGE_SIZE, addr < end);
 980 }
 981
 982 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr,
 983                                     unsigned long end, bool free_mapped,
 984                                     struct vmem_altmap *altmap)
 985 {
 986         unsigned long next;
 987         pmd_t *pmdp, pmd;
 988
 989         do {
 990                 next = pmd_addr_end(addr, end);
 991                 pmdp = pmd_offset(pudp, addr);
 992                 pmd = READ_ONCE(*pmdp);
 993                 if (pmd_none(pmd))
 994                         continue;
 995
 996                 WARN_ON(!pmd_present(pmd));
 997                 if (pmd_sect(pmd)) {
 998                         pmd_clear(pmdp);
 999
1000                         /*
1001                          * One TLBI should be sufficient here as the PMD_SIZE
1002                          * range is mapped with a single block entry.
1003                          */
1004                         flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
1005                         if (free_mapped)
1006                                 free_hotplug_page_range(pmd_page(pmd),
1007                                                         PMD_SIZE, altmap);
1008                         continue;
1009                 }
1010                 WARN_ON(!pmd_table(pmd));
1011                 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap);
1012         } while (addr = next, addr < end);
1013 }
1014
1015 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr,
1016                                     unsigned long end, bool free_mapped,
1017                                     struct vmem_altmap *altmap)
1018 {
1019         unsigned long next;
1020         pud_t *pudp, pud;
1021
1022         do {
1023                 next = pud_addr_end(addr, end);
1024                 pudp = pud_offset(p4dp, addr);
1025                 pud = READ_ONCE(*pudp);
1026                 if (pud_none(pud))
1027                         continue;
1028
1029                 WARN_ON(!pud_present(pud));
1030                 if (pud_sect(pud)) {
1031                         pud_clear(pudp);
1032
1033                         /*
1034                          * One TLBI should be sufficient here as the PUD_SIZE
1035                          * range is mapped with a single block entry.
1036                          */
1037                         flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
1038                         if (free_mapped)
1039                                 free_hotplug_page_range(pud_page(pud),
1040                                                         PUD_SIZE, altmap);
1041                         continue;
1042                 }
1043                 WARN_ON(!pud_table(pud));
1044                 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap);
1045         } while (addr = next, addr < end);
1046 }
1047
1048 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr,
1049                                     unsigned long end, bool free_mapped,
1050                                     struct vmem_altmap *altmap)
1051 {
1052         unsigned long next;
1053         p4d_t *p4dp, p4d;
1054
1055         do {
1056                 next = p4d_addr_end(addr, end);
1057                 p4dp = p4d_offset(pgdp, addr);
1058                 p4d = READ_ONCE(*p4dp);
1059                 if (p4d_none(p4d))
1060                         continue;
1061
1062                 WARN_ON(!p4d_present(p4d));
1063                 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap);
1064         } while (addr = next, addr < end);
1065 }
1066
1067 static void unmap_hotplug_range(unsigned long addr, unsigned long end,
1068                                 bool free_mapped, struct vmem_altmap *altmap)
1069 {
1070         unsigned long next;
1071         pgd_t *pgdp, pgd;
1072
1073         /*
1074          * altmap can only be used as vmemmap mapping backing memory.
1075          * In case the backing memory itself is not being freed, then
1076          * altmap is irrelevant. Warn about this inconsistency when
1077          * encountered.
1078          */
1079         WARN_ON(!free_mapped && altmap);
1080
1081         do {
1082                 next = pgd_addr_end(addr, end);
1083                 pgdp = pgd_offset_k(addr);
1084                 pgd = READ_ONCE(*pgdp);
1085                 if (pgd_none(pgd))
1086                         continue;
1087
1088                 WARN_ON(!pgd_present(pgd));
1089                 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap);
1090         } while (addr = next, addr < end);
1091 }
1092
1093 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
1094                                  unsigned long end, unsigned long floor,
1095                                  unsigned long ceiling)
1096 {
1097         pte_t *ptep, pte;
1098         unsigned long i, start = addr;
1099
1100         do {
1101                 ptep = pte_offset_kernel(pmdp, addr);
1102                 pte = READ_ONCE(*ptep);
1103
1104                 /*
1105                  * This is just a sanity check here which verifies that
1106                  * pte clearing has been done by earlier unmap loops.
1107                  */
1108                 WARN_ON(!pte_none(pte));
1109         } while (addr += PAGE_SIZE, addr < end);
1110
1111         if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK))
1112                 return;
1113
1114         /*
1115          * Check whether we can free the pte page if the rest of the
1116          * entries are empty. Overlap with other regions have been
1117          * handled by the floor/ceiling check.
1118          */
1119         ptep = pte_offset_kernel(pmdp, 0UL);
1120         for (i = 0; i < PTRS_PER_PTE; i++) {
1121                 if (!pte_none(READ_ONCE(ptep[i])))
1122                         return;
1123         }
1124
1125         pmd_clear(pmdp);
1126         __flush_tlb_kernel_pgtable(start);
1127         free_hotplug_pgtable_page(virt_to_page(ptep));
1128 }
1129
1130 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
1131                                  unsigned long end, unsigned long floor,
1132                                  unsigned long ceiling)
1133 {
1134         pmd_t *pmdp, pmd;
1135         unsigned long i, next, start = addr;
1136
1137         do {
1138                 next = pmd_addr_end(addr, end);
1139                 pmdp = pmd_offset(pudp, addr);
1140                 pmd = READ_ONCE(*pmdp);
1141                 if (pmd_none(pmd))
1142                         continue;
1143
1144                 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd));
1145                 free_empty_pte_table(pmdp, addr, next, floor, ceiling);
1146         } while (addr = next, addr < end);
1147
1148         if (CONFIG_PGTABLE_LEVELS <= 2)
1149                 return;
1150
1151         if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK))
1152                 return;
1153
1154         /*
1155          * Check whether we can free the pmd page if the rest of the
1156          * entries are empty. Overlap with other regions have been
1157          * handled by the floor/ceiling check.
1158          */
1159         pmdp = pmd_offset(pudp, 0UL);
1160         for (i = 0; i < PTRS_PER_PMD; i++) {
1161                 if (!pmd_none(READ_ONCE(pmdp[i])))
1162                         return;
1163         }
1164
1165         pud_clear(pudp);
1166         __flush_tlb_kernel_pgtable(start);
1167         free_hotplug_pgtable_page(virt_to_page(pmdp));
1168 }
1169
1170 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr,
1171                                  unsigned long end, unsigned long floor,
1172                                  unsigned long ceiling)
1173 {
1174         pud_t *pudp, pud;
1175         unsigned long i, next, start = addr;
1176
1177         do {
1178                 next = pud_addr_end(addr, end);
1179                 pudp = pud_offset(p4dp, addr);
1180                 pud = READ_ONCE(*pudp);
1181                 if (pud_none(pud))
1182                         continue;
1183
1184                 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud));
1185                 free_empty_pmd_table(pudp, addr, next, floor, ceiling);
1186         } while (addr = next, addr < end);
1187
1188         if (CONFIG_PGTABLE_LEVELS <= 3)
1189                 return;
1190
1191         if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK))
1192                 return;
1193
1194         /*
1195          * Check whether we can free the pud page if the rest of the
1196          * entries are empty. Overlap with other regions have been
1197          * handled by the floor/ceiling check.
1198          */
1199         pudp = pud_offset(p4dp, 0UL);
1200         for (i = 0; i < PTRS_PER_PUD; i++) {
1201                 if (!pud_none(READ_ONCE(pudp[i])))
1202                         return;
1203         }
1204
1205         p4d_clear(p4dp);
1206         __flush_tlb_kernel_pgtable(start);
1207         free_hotplug_pgtable_page(virt_to_page(pudp));
1208 }
1209
1210 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
1211                                  unsigned long end, unsigned long floor,
1212                                  unsigned long ceiling)
1213 {
1214         unsigned long next;
1215         p4d_t *p4dp, p4d;
1216
1217         do {
1218                 next = p4d_addr_end(addr, end);
1219                 p4dp = p4d_offset(pgdp, addr);
1220                 p4d = READ_ONCE(*p4dp);
1221                 if (p4d_none(p4d))
1222                         continue;
1223
1224                 WARN_ON(!p4d_present(p4d));
1225                 free_empty_pud_table(p4dp, addr, next, floor, ceiling);
1226         } while (addr = next, addr < end);
1227 }
1228
1229 static void free_empty_tables(unsigned long addr, unsigned long end,
1230                               unsigned long floor, unsigned long ceiling)
1231 {
1232         unsigned long next;
1233         pgd_t *pgdp, pgd;
1234
1235         do {
1236                 next = pgd_addr_end(addr, end);
1237                 pgdp = pgd_offset_k(addr);
1238                 pgd = READ_ONCE(*pgdp);
1239                 if (pgd_none(pgd))
1240                         continue;
1241
1242                 WARN_ON(!pgd_present(pgd));
1243                 free_empty_p4d_table(pgdp, addr, next, floor, ceiling);
1244         } while (addr = next, addr < end);
1245 }
1246 #endif
1247
1248 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1249                 struct vmem_altmap *altmap)
1250 {
1251         unsigned long addr = start;
1252         unsigned long next;
1253         pgd_t *pgdp;
1254         p4d_t *p4dp;
1255         pud_t *pudp;
1256         pmd_t *pmdp;
1257
1258         WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1259
1260         if (!ARM64_KERNEL_USES_PMD_MAPS)
1261                 return vmemmap_populate_basepages(start, end, node, altmap);
1262
1263         do {
1264                 next = pmd_addr_end(addr, end);
1265
1266                 pgdp = vmemmap_pgd_populate(addr, node);
1267                 if (!pgdp)
1268                         return -ENOMEM;
1269
1270                 p4dp = vmemmap_p4d_populate(pgdp, addr, node);
1271                 if (!p4dp)
1272                         return -ENOMEM;
1273
1274                 pudp = vmemmap_pud_populate(p4dp, addr, node);
1275                 if (!pudp)
1276                         return -ENOMEM;
1277
1278                 pmdp = pmd_offset(pudp, addr);
1279                 if (pmd_none(READ_ONCE(*pmdp))) {
1280                         void *p = NULL;
1281
1282                         p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
1283                         if (!p) {
1284                                 if (vmemmap_populate_basepages(addr, next, node, altmap))
1285                                         return -ENOMEM;
1286                                 continue;
1287                         }
1288
1289                         pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL));
1290                 } else
1291                         vmemmap_verify((pte_t *)pmdp, node, addr, next);
1292         } while (addr = next, addr != end);
1293
1294         return 0;
1295 }
1296
1297 #ifdef CONFIG_MEMORY_HOTPLUG
1298 void vmemmap_free(unsigned long start, unsigned long end,
1299                 struct vmem_altmap *altmap)
1300 {
1301         WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1302
1303         unmap_hotplug_range(start, end, true, altmap);
1304         free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END);
1305 }
1306 #endif /* CONFIG_MEMORY_HOTPLUG */
1307
1308 static inline pud_t *fixmap_pud(unsigned long addr)
1309 {
1310         pgd_t *pgdp = pgd_offset_k(addr);
1311         p4d_t *p4dp = p4d_offset(pgdp, addr);
1312         p4d_t p4d = READ_ONCE(*p4dp);
1313
1314         BUG_ON(p4d_none(p4d) || p4d_bad(p4d));
1315
1316         return pud_offset_kimg(p4dp, addr);
1317 }
1318
1319 static inline pmd_t *fixmap_pmd(unsigned long addr)
1320 {
1321         pud_t *pudp = fixmap_pud(addr);
1322         pud_t pud = READ_ONCE(*pudp);
1323
1324         BUG_ON(pud_none(pud) || pud_bad(pud));
1325
1326         return pmd_offset_kimg(pudp, addr);
1327 }
1328
1329 static inline pte_t *fixmap_pte(unsigned long addr)
1330 {
1331         return &bm_pte[pte_index(addr)];
1332 }
1333
1334 /*
1335  * The p*d_populate functions call virt_to_phys implicitly so they can't be used
1336  * directly on kernel symbols (bm_p*d). This function is called too early to use
1337  * lm_alias so __p*d_populate functions must be used to populate with the
1338  * physical address from __pa_symbol.
1339  */
1340 void __init early_fixmap_init(void)
1341 {
1342         pgd_t *pgdp;
1343         p4d_t *p4dp, p4d;
1344         pud_t *pudp;
1345         pmd_t *pmdp;
1346         unsigned long addr = FIXADDR_START;
1347
1348         pgdp = pgd_offset_k(addr);
1349         p4dp = p4d_offset(pgdp, addr);
1350         p4d = READ_ONCE(*p4dp);
1351         if (CONFIG_PGTABLE_LEVELS > 3 &&
1352             !(p4d_none(p4d) || p4d_page_paddr(p4d) == __pa_symbol(bm_pud))) {
1353                 /*
1354                  * We only end up here if the kernel mapping and the fixmap
1355                  * share the top level pgd entry, which should only happen on
1356                  * 16k/4 levels configurations.
1357                  */
1358                 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
1359                 pudp = pud_offset_kimg(p4dp, addr);
1360         } else {
1361                 if (p4d_none(p4d))
1362                         __p4d_populate(p4dp, __pa_symbol(bm_pud), P4D_TYPE_TABLE);
1363                 pudp = fixmap_pud(addr);
1364         }
1365         if (pud_none(READ_ONCE(*pudp)))
1366                 __pud_populate(pudp, __pa_symbol(bm_pmd), PUD_TYPE_TABLE);
1367         pmdp = fixmap_pmd(addr);
1368         __pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE);
1369
1370         /*
1371          * The boot-ioremap range spans multiple pmds, for which
1372          * we are not prepared:
1373          */
1374         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
1375                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
1376
1377         if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
1378              || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
1379                 WARN_ON(1);
1380                 pr_warn("pmdp %p != %p, %p\n",
1381                         pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
1382                         fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
1383                 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
1384                         fix_to_virt(FIX_BTMAP_BEGIN));
1385                 pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
1386                         fix_to_virt(FIX_BTMAP_END));
1387
1388                 pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
1389                 pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
1390         }
1391 }
1392
1393 /*
1394  * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we
1395  * ever need to use IPIs for TLB broadcasting, then we're in trouble here.
1396  */
1397 void __set_fixmap(enum fixed_addresses idx,
1398                                phys_addr_t phys, pgprot_t flags)
1399 {
1400         unsigned long addr = __fix_to_virt(idx);
1401         pte_t *ptep;
1402
1403         BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
1404
1405         ptep = fixmap_pte(addr);
1406
1407         if (pgprot_val(flags)) {
1408                 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags));
1409         } else {
1410                 pte_clear(&init_mm, addr, ptep);
1411                 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
1412         }
1413 }
1414
1415 void *__init fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
1416 {
1417         const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
1418         int offset;
1419         void *dt_virt;
1420
1421         /*
1422          * Check whether the physical FDT address is set and meets the minimum
1423          * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
1424          * at least 8 bytes so that we can always access the magic and size
1425          * fields of the FDT header after mapping the first chunk, double check
1426          * here if that is indeed the case.
1427          */
1428         BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
1429         if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
1430                 return NULL;
1431
1432         /*
1433          * Make sure that the FDT region can be mapped without the need to
1434          * allocate additional translation table pages, so that it is safe
1435          * to call create_mapping_noalloc() this early.
1436          *
1437          * On 64k pages, the FDT will be mapped using PTEs, so we need to
1438          * be in the same PMD as the rest of the fixmap.
1439          * On 4k pages, we'll use section mappings for the FDT so we only
1440          * have to be in the same PUD.
1441          */
1442         BUILD_BUG_ON(dt_virt_base % SZ_2M);
1443
1444         BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
1445                      __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
1446
1447         offset = dt_phys % SWAPPER_BLOCK_SIZE;
1448         dt_virt = (void *)dt_virt_base + offset;
1449
1450         /* map the first chunk so we can read the size from the header */
1451         create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
1452                         dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
1453
1454         if (fdt_magic(dt_virt) != FDT_MAGIC)
1455                 return NULL;
1456
1457         *size = fdt_totalsize(dt_virt);
1458         if (*size > MAX_FDT_SIZE)
1459                 return NULL;
1460
1461         if (offset + *size > SWAPPER_BLOCK_SIZE)
1462                 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
1463                                round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
1464
1465         return dt_virt;
1466 }
1467
1468 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
1469 {
1470         pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot));
1471
1472         /* Only allow permission changes for now */
1473         if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)),
1474                                    pud_val(new_pud)))
1475                 return 0;
1476
1477         VM_BUG_ON(phys & ~PUD_MASK);
1478         set_pud(pudp, new_pud);
1479         return 1;
1480 }
1481
1482 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
1483 {
1484         pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot));
1485
1486         /* Only allow permission changes for now */
1487         if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)),
1488                                    pmd_val(new_pmd)))
1489                 return 0;
1490
1491         VM_BUG_ON(phys & ~PMD_MASK);
1492         set_pmd(pmdp, new_pmd);
1493         return 1;
1494 }
1495
1496 int pud_clear_huge(pud_t *pudp)
1497 {
1498         if (!pud_sect(READ_ONCE(*pudp)))
1499                 return 0;
1500         pud_clear(pudp);
1501         return 1;
1502 }
1503
1504 int pmd_clear_huge(pmd_t *pmdp)
1505 {
1506         if (!pmd_sect(READ_ONCE(*pmdp)))
1507                 return 0;
1508         pmd_clear(pmdp);
1509         return 1;
1510 }
1511
1512 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
1513 {
1514         pte_t *table;
1515         pmd_t pmd;
1516
1517         pmd = READ_ONCE(*pmdp);
1518
1519         if (!pmd_table(pmd)) {
1520                 VM_WARN_ON(1);
1521                 return 1;
1522         }
1523
1524         table = pte_offset_kernel(pmdp, addr);
1525         pmd_clear(pmdp);
1526         __flush_tlb_kernel_pgtable(addr);
1527         pte_free_kernel(NULL, table);
1528         return 1;
1529 }
1530
1531 int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
1532 {
1533         pmd_t *table;
1534         pmd_t *pmdp;
1535         pud_t pud;
1536         unsigned long next, end;
1537
1538         pud = READ_ONCE(*pudp);
1539
1540         if (!pud_table(pud)) {
1541                 VM_WARN_ON(1);
1542                 return 1;
1543         }
1544
1545         table = pmd_offset(pudp, addr);
1546         pmdp = table;
1547         next = addr;
1548         end = addr + PUD_SIZE;
1549         do {
1550                 pmd_free_pte_page(pmdp, next);
1551         } while (pmdp++, next += PMD_SIZE, next != end);
1552
1553         pud_clear(pudp);
1554         __flush_tlb_kernel_pgtable(addr);
1555         pmd_free(NULL, table);
1556         return 1;
1557 }
1558
1559 #ifdef CONFIG_MEMORY_HOTPLUG
1560 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size)
1561 {
1562         unsigned long end = start + size;
1563
1564         WARN_ON(pgdir != init_mm.pgd);
1565         WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END));
1566
1567         unmap_hotplug_range(start, end, false, NULL);
1568         free_empty_tables(start, end, PAGE_OFFSET, PAGE_END);
1569 }
1570
1571 struct range arch_get_mappable_range(void)
1572 {
1573         struct range mhp_range;
1574         u64 start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual));
1575         u64 end_linear_pa = __pa(PAGE_END - 1);
1576
1577         if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
1578                 /*
1579                  * Check for a wrap, it is possible because of randomized linear
1580                  * mapping the start physical address is actually bigger than
1581                  * the end physical address. In this case set start to zero
1582                  * because [0, end_linear_pa] range must still be able to cover
1583                  * all addressable physical addresses.
1584                  */
1585                 if (start_linear_pa > end_linear_pa)
1586                         start_linear_pa = 0;
1587         }
1588
1589         WARN_ON(start_linear_pa > end_linear_pa);
1590
1591         /*
1592          * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)]
1593          * accommodating both its ends but excluding PAGE_END. Max physical
1594          * range which can be mapped inside this linear mapping range, must
1595          * also be derived from its end points.
1596          */
1597         mhp_range.start = start_linear_pa;
1598         mhp_range.end =  end_linear_pa;
1599
1600         return mhp_range;
1601 }
1602
1603 int arch_add_memory(int nid, u64 start, u64 size,
1604                     struct mhp_params *params)
1605 {
1606         int ret, flags = NO_EXEC_MAPPINGS;
1607
1608         VM_BUG_ON(!mhp_range_allowed(start, size, true));
1609
1610         if (can_set_direct_map())
1611                 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
1612
1613         __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
1614                              size, params->pgprot, __pgd_pgtable_alloc,
1615                              flags);
1616
1617         memblock_clear_nomap(start, size);
1618
1619         ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT,
1620                            params);
1621         if (ret)
1622                 __remove_pgd_mapping(swapper_pg_dir,
1623                                      __phys_to_virt(start), size);
1624         else {
1625                 max_pfn = PFN_UP(start + size);
1626                 max_low_pfn = max_pfn;
1627         }
1628
1629         return ret;
1630 }
1631
1632 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
1633 {
1634         unsigned long start_pfn = start >> PAGE_SHIFT;
1635         unsigned long nr_pages = size >> PAGE_SHIFT;
1636
1637         __remove_pages(start_pfn, nr_pages, altmap);
1638         __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size);
1639 }
1640
1641 /*
1642  * This memory hotplug notifier helps prevent boot memory from being
1643  * inadvertently removed as it blocks pfn range offlining process in
1644  * __offline_pages(). Hence this prevents both offlining as well as
1645  * removal process for boot memory which is initially always online.
1646  * In future if and when boot memory could be removed, this notifier
1647  * should be dropped and free_hotplug_page_range() should handle any
1648  * reserved pages allocated during boot.
1649  */
1650 static int prevent_bootmem_remove_notifier(struct notifier_block *nb,
1651                                            unsigned long action, void *data)
1652 {
1653         struct mem_section *ms;
1654         struct memory_notify *arg = data;
1655         unsigned long end_pfn = arg->start_pfn + arg->nr_pages;
1656         unsigned long pfn = arg->start_pfn;
1657
1658         if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE))
1659                 return NOTIFY_OK;
1660
1661         for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
1662                 unsigned long start = PFN_PHYS(pfn);
1663                 unsigned long end = start + (1UL << PA_SECTION_SHIFT);
1664
1665                 ms = __pfn_to_section(pfn);
1666                 if (!early_section(ms))
1667                         continue;
1668
1669                 if (action == MEM_GOING_OFFLINE) {
1670                         /*
1671                          * Boot memory removal is not supported. Prevent
1672                          * it via blocking any attempted offline request
1673                          * for the boot memory and just report it.
1674                          */
1675                         pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end);
1676                         return NOTIFY_BAD;
1677                 } else if (action == MEM_OFFLINE) {
1678                         /*
1679                          * This should have never happened. Boot memory
1680                          * offlining should have been prevented by this
1681                          * very notifier. Probably some memory removal
1682                          * procedure might have changed which would then
1683                          * require further debug.
1684                          */
1685                         pr_err("Boot memory [%lx %lx] offlined\n", start, end);
1686
1687                         /*
1688                          * Core memory hotplug does not process a return
1689                          * code from the notifier for MEM_OFFLINE events.
1690                          * The error condition has been reported. Return
1691                          * from here as if ignored.
1692                          */
1693                         return NOTIFY_DONE;
1694                 }
1695         }
1696         return NOTIFY_OK;
1697 }
1698
1699 static struct notifier_block prevent_bootmem_remove_nb = {
1700         .notifier_call = prevent_bootmem_remove_notifier,
1701 };
1702
1703 /*
1704  * This ensures that boot memory sections on the platform are online
1705  * from early boot. Memory sections could not be prevented from being
1706  * offlined, unless for some reason they are not online to begin with.
1707  * This helps validate the basic assumption on which the above memory
1708  * event notifier works to prevent boot memory section offlining and
1709  * its possible removal.
1710  */
1711 static void validate_bootmem_online(void)
1712 {
1713         phys_addr_t start, end, addr;
1714         struct mem_section *ms;
1715         u64 i;
1716
1717         /*
1718          * Scanning across all memblock might be expensive
1719          * on some big memory systems. Hence enable this
1720          * validation only with DEBUG_VM.
1721          */
1722         if (!IS_ENABLED(CONFIG_DEBUG_VM))
1723                 return;
1724
1725         for_each_mem_range(i, &start, &end) {
1726                 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) {
1727                         ms = __pfn_to_section(PHYS_PFN(addr));
1728
1729                         /*
1730                          * All memory ranges in the system at this point
1731                          * should have been marked as early sections.
1732                          */
1733                         WARN_ON(!early_section(ms));
1734
1735                         /*
1736                          * Memory notifier mechanism here to prevent boot
1737                          * memory offlining depends on the fact that each
1738                          * early section memory on the system is initially
1739                          * online. Otherwise a given memory section which
1740                          * is already offline will be overlooked and can
1741                          * be removed completely. Call out such sections.
1742                          */
1743                         if (!online_section(ms))
1744                                 pr_err("Boot memory [%llx %llx] is offline, can be removed\n",
1745                                         addr, addr + (1UL << PA_SECTION_SHIFT));
1746                 }
1747         }
1748 }
1749
1750 static int __init prevent_bootmem_remove_init(void)
1751 {
1752         int ret = 0;
1753
1754         if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
1755                 return ret;
1756
1757         validate_bootmem_online();
1758         ret = register_memory_notifier(&prevent_bootmem_remove_nb);
1759         if (ret)
1760                 pr_err("%s: Notifier registration failed %d\n", __func__, ret);
1761
1762         return ret;
1763 }
1764 early_initcall(prevent_bootmem_remove_init);
1765 #endif