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/memory.h> 21 #include <linux/fs.h> 22 #include <linux/io.h> 23 #include <linux/mm.h> 24 #include <linux/vmalloc.h> 25 #include <linux/set_memory.h> 26 27 #include <asm/barrier.h> 28 #include <asm/cputype.h> 29 #include <asm/fixmap.h> 30 #include <asm/kasan.h> 31 #include <asm/kernel-pgtable.h> 32 #include <asm/sections.h> 33 #include <asm/setup.h> 34 #include <linux/sizes.h> 35 #include <asm/tlb.h> 36 #include <asm/mmu_context.h> 37 #include <asm/ptdump.h> 38 #include <asm/tlbflush.h> 39 #include <asm/pgalloc.h> 40 41 #define NO_BLOCK_MAPPINGS BIT(0) 42 #define NO_CONT_MAPPINGS BIT(1) 43 #define NO_EXEC_MAPPINGS BIT(2) /* assumes FEAT_HPDS is not used */ 44 45 u64 idmap_t0sz = TCR_T0SZ(VA_BITS_MIN); 46 u64 idmap_ptrs_per_pgd = PTRS_PER_PGD; 47 48 u64 __section(".mmuoff.data.write") vabits_actual; 49 EXPORT_SYMBOL(vabits_actual); 50 51 u64 kimage_voffset __ro_after_init; 52 EXPORT_SYMBOL(kimage_voffset); 53 54 /* 55 * Empty_zero_page is a special page that is used for zero-initialized data 56 * and COW. 57 */ 58 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss; 59 EXPORT_SYMBOL(empty_zero_page); 60 61 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss; 62 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused; 63 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused; 64 65 static DEFINE_SPINLOCK(swapper_pgdir_lock); 66 67 void set_swapper_pgd(pgd_t *pgdp, pgd_t pgd) 68 { 69 pgd_t *fixmap_pgdp; 70 71 spin_lock(&swapper_pgdir_lock); 72 fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp)); 73 WRITE_ONCE(*fixmap_pgdp, pgd); 74 /* 75 * We need dsb(ishst) here to ensure the page-table-walker sees 76 * our new entry before set_p?d() returns. The fixmap's 77 * flush_tlb_kernel_range() via clear_fixmap() does this for us. 78 */ 79 pgd_clear_fixmap(); 80 spin_unlock(&swapper_pgdir_lock); 81 } 82 83 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn, 84 unsigned long size, pgprot_t vma_prot) 85 { 86 if (!pfn_is_map_memory(pfn)) 87 return pgprot_noncached(vma_prot); 88 else if (file->f_flags & O_SYNC) 89 return pgprot_writecombine(vma_prot); 90 return vma_prot; 91 } 92 EXPORT_SYMBOL(phys_mem_access_prot); 93 94 static phys_addr_t __init early_pgtable_alloc(int shift) 95 { 96 phys_addr_t phys; 97 void *ptr; 98 99 phys = memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE); 100 if (!phys) 101 panic("Failed to allocate page table page\n"); 102 103 /* 104 * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE 105 * slot will be free, so we can (ab)use the FIX_PTE slot to initialise 106 * any level of table. 107 */ 108 ptr = pte_set_fixmap(phys); 109 110 memset(ptr, 0, PAGE_SIZE); 111 112 /* 113 * Implicit barriers also ensure the zeroed page is visible to the page 114 * table walker 115 */ 116 pte_clear_fixmap(); 117 118 return phys; 119 } 120 121 static bool pgattr_change_is_safe(u64 old, u64 new) 122 { 123 /* 124 * The following mapping attributes may be updated in live 125 * kernel mappings without the need for break-before-make. 126 */ 127 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG; 128 129 /* creating or taking down mappings is always safe */ 130 if (old == 0 || new == 0) 131 return true; 132 133 /* live contiguous mappings may not be manipulated at all */ 134 if ((old | new) & PTE_CONT) 135 return false; 136 137 /* Transitioning from Non-Global to Global is unsafe */ 138 if (old & ~new & PTE_NG) 139 return false; 140 141 /* 142 * Changing the memory type between Normal and Normal-Tagged is safe 143 * since Tagged is considered a permission attribute from the 144 * mismatched attribute aliases perspective. 145 */ 146 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 147 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) && 148 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 149 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED))) 150 mask |= PTE_ATTRINDX_MASK; 151 152 return ((old ^ new) & ~mask) == 0; 153 } 154 155 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end, 156 phys_addr_t phys, pgprot_t prot) 157 { 158 pte_t *ptep; 159 160 ptep = pte_set_fixmap_offset(pmdp, addr); 161 do { 162 pte_t old_pte = READ_ONCE(*ptep); 163 164 set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot)); 165 166 /* 167 * After the PTE entry has been populated once, we 168 * only allow updates to the permission attributes. 169 */ 170 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte), 171 READ_ONCE(pte_val(*ptep)))); 172 173 phys += PAGE_SIZE; 174 } while (ptep++, addr += PAGE_SIZE, addr != end); 175 176 pte_clear_fixmap(); 177 } 178 179 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr, 180 unsigned long end, phys_addr_t phys, 181 pgprot_t prot, 182 phys_addr_t (*pgtable_alloc)(int), 183 int flags) 184 { 185 unsigned long next; 186 pmd_t pmd = READ_ONCE(*pmdp); 187 188 BUG_ON(pmd_sect(pmd)); 189 if (pmd_none(pmd)) { 190 pmdval_t pmdval = PMD_TYPE_TABLE | PMD_TABLE_UXN; 191 phys_addr_t pte_phys; 192 193 if (flags & NO_EXEC_MAPPINGS) 194 pmdval |= PMD_TABLE_PXN; 195 BUG_ON(!pgtable_alloc); 196 pte_phys = pgtable_alloc(PAGE_SHIFT); 197 __pmd_populate(pmdp, pte_phys, pmdval); 198 pmd = READ_ONCE(*pmdp); 199 } 200 BUG_ON(pmd_bad(pmd)); 201 202 do { 203 pgprot_t __prot = prot; 204 205 next = pte_cont_addr_end(addr, end); 206 207 /* use a contiguous mapping if the range is suitably aligned */ 208 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) && 209 (flags & NO_CONT_MAPPINGS) == 0) 210 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 211 212 init_pte(pmdp, addr, next, phys, __prot); 213 214 phys += next - addr; 215 } while (addr = next, addr != end); 216 } 217 218 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end, 219 phys_addr_t phys, pgprot_t prot, 220 phys_addr_t (*pgtable_alloc)(int), int flags) 221 { 222 unsigned long next; 223 pmd_t *pmdp; 224 225 pmdp = pmd_set_fixmap_offset(pudp, addr); 226 do { 227 pmd_t old_pmd = READ_ONCE(*pmdp); 228 229 next = pmd_addr_end(addr, end); 230 231 /* try section mapping first */ 232 if (((addr | next | phys) & ~PMD_MASK) == 0 && 233 (flags & NO_BLOCK_MAPPINGS) == 0) { 234 pmd_set_huge(pmdp, phys, prot); 235 236 /* 237 * After the PMD entry has been populated once, we 238 * only allow updates to the permission attributes. 239 */ 240 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd), 241 READ_ONCE(pmd_val(*pmdp)))); 242 } else { 243 alloc_init_cont_pte(pmdp, addr, next, phys, prot, 244 pgtable_alloc, flags); 245 246 BUG_ON(pmd_val(old_pmd) != 0 && 247 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp))); 248 } 249 phys += next - addr; 250 } while (pmdp++, addr = next, addr != end); 251 252 pmd_clear_fixmap(); 253 } 254 255 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr, 256 unsigned long end, phys_addr_t phys, 257 pgprot_t prot, 258 phys_addr_t (*pgtable_alloc)(int), int flags) 259 { 260 unsigned long next; 261 pud_t pud = READ_ONCE(*pudp); 262 263 /* 264 * Check for initial section mappings in the pgd/pud. 265 */ 266 BUG_ON(pud_sect(pud)); 267 if (pud_none(pud)) { 268 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN; 269 phys_addr_t pmd_phys; 270 271 if (flags & NO_EXEC_MAPPINGS) 272 pudval |= PUD_TABLE_PXN; 273 BUG_ON(!pgtable_alloc); 274 pmd_phys = pgtable_alloc(PMD_SHIFT); 275 __pud_populate(pudp, pmd_phys, pudval); 276 pud = READ_ONCE(*pudp); 277 } 278 BUG_ON(pud_bad(pud)); 279 280 do { 281 pgprot_t __prot = prot; 282 283 next = pmd_cont_addr_end(addr, end); 284 285 /* use a contiguous mapping if the range is suitably aligned */ 286 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) && 287 (flags & NO_CONT_MAPPINGS) == 0) 288 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 289 290 init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags); 291 292 phys += next - addr; 293 } while (addr = next, addr != end); 294 } 295 296 static inline bool use_1G_block(unsigned long addr, unsigned long next, 297 unsigned long phys) 298 { 299 if (PAGE_SHIFT != 12) 300 return false; 301 302 if (((addr | next | phys) & ~PUD_MASK) != 0) 303 return false; 304 305 return true; 306 } 307 308 static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end, 309 phys_addr_t phys, pgprot_t prot, 310 phys_addr_t (*pgtable_alloc)(int), 311 int flags) 312 { 313 unsigned long next; 314 pud_t *pudp; 315 p4d_t *p4dp = p4d_offset(pgdp, addr); 316 p4d_t p4d = READ_ONCE(*p4dp); 317 318 if (p4d_none(p4d)) { 319 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN; 320 phys_addr_t pud_phys; 321 322 if (flags & NO_EXEC_MAPPINGS) 323 p4dval |= P4D_TABLE_PXN; 324 BUG_ON(!pgtable_alloc); 325 pud_phys = pgtable_alloc(PUD_SHIFT); 326 __p4d_populate(p4dp, pud_phys, p4dval); 327 p4d = READ_ONCE(*p4dp); 328 } 329 BUG_ON(p4d_bad(p4d)); 330 331 pudp = pud_set_fixmap_offset(p4dp, addr); 332 do { 333 pud_t old_pud = READ_ONCE(*pudp); 334 335 next = pud_addr_end(addr, end); 336 337 /* 338 * For 4K granule only, attempt to put down a 1GB block 339 */ 340 if (use_1G_block(addr, next, phys) && 341 (flags & NO_BLOCK_MAPPINGS) == 0) { 342 pud_set_huge(pudp, phys, prot); 343 344 /* 345 * After the PUD entry has been populated once, we 346 * only allow updates to the permission attributes. 347 */ 348 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud), 349 READ_ONCE(pud_val(*pudp)))); 350 } else { 351 alloc_init_cont_pmd(pudp, addr, next, phys, prot, 352 pgtable_alloc, flags); 353 354 BUG_ON(pud_val(old_pud) != 0 && 355 pud_val(old_pud) != READ_ONCE(pud_val(*pudp))); 356 } 357 phys += next - addr; 358 } while (pudp++, addr = next, addr != end); 359 360 pud_clear_fixmap(); 361 } 362 363 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, 364 unsigned long virt, phys_addr_t size, 365 pgprot_t prot, 366 phys_addr_t (*pgtable_alloc)(int), 367 int flags) 368 { 369 unsigned long addr, end, next; 370 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt); 371 372 /* 373 * If the virtual and physical address don't have the same offset 374 * within a page, we cannot map the region as the caller expects. 375 */ 376 if (WARN_ON((phys ^ virt) & ~PAGE_MASK)) 377 return; 378 379 phys &= PAGE_MASK; 380 addr = virt & PAGE_MASK; 381 end = PAGE_ALIGN(virt + size); 382 383 do { 384 next = pgd_addr_end(addr, end); 385 alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc, 386 flags); 387 phys += next - addr; 388 } while (pgdp++, addr = next, addr != end); 389 } 390 391 static phys_addr_t __pgd_pgtable_alloc(int shift) 392 { 393 void *ptr = (void *)__get_free_page(GFP_PGTABLE_KERNEL); 394 BUG_ON(!ptr); 395 396 /* Ensure the zeroed page is visible to the page table walker */ 397 dsb(ishst); 398 return __pa(ptr); 399 } 400 401 static phys_addr_t pgd_pgtable_alloc(int shift) 402 { 403 phys_addr_t pa = __pgd_pgtable_alloc(shift); 404 405 /* 406 * Call proper page table ctor in case later we need to 407 * call core mm functions like apply_to_page_range() on 408 * this pre-allocated page table. 409 * 410 * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is 411 * folded, and if so pgtable_pmd_page_ctor() becomes nop. 412 */ 413 if (shift == PAGE_SHIFT) 414 BUG_ON(!pgtable_pte_page_ctor(phys_to_page(pa))); 415 else if (shift == PMD_SHIFT) 416 BUG_ON(!pgtable_pmd_page_ctor(phys_to_page(pa))); 417 418 return pa; 419 } 420 421 /* 422 * This function can only be used to modify existing table entries, 423 * without allocating new levels of table. Note that this permits the 424 * creation of new section or page entries. 425 */ 426 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt, 427 phys_addr_t size, pgprot_t prot) 428 { 429 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) { 430 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n", 431 &phys, virt); 432 return; 433 } 434 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 435 NO_CONT_MAPPINGS); 436 } 437 438 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys, 439 unsigned long virt, phys_addr_t size, 440 pgprot_t prot, bool page_mappings_only) 441 { 442 int flags = 0; 443 444 BUG_ON(mm == &init_mm); 445 446 if (page_mappings_only) 447 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 448 449 __create_pgd_mapping(mm->pgd, phys, virt, size, prot, 450 pgd_pgtable_alloc, flags); 451 } 452 453 static void update_mapping_prot(phys_addr_t phys, unsigned long virt, 454 phys_addr_t size, pgprot_t prot) 455 { 456 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) { 457 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n", 458 &phys, virt); 459 return; 460 } 461 462 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 463 NO_CONT_MAPPINGS); 464 465 /* flush the TLBs after updating live kernel mappings */ 466 flush_tlb_kernel_range(virt, virt + size); 467 } 468 469 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start, 470 phys_addr_t end, pgprot_t prot, int flags) 471 { 472 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start, 473 prot, early_pgtable_alloc, flags); 474 } 475 476 void __init mark_linear_text_alias_ro(void) 477 { 478 /* 479 * Remove the write permissions from the linear alias of .text/.rodata 480 */ 481 update_mapping_prot(__pa_symbol(_stext), (unsigned long)lm_alias(_stext), 482 (unsigned long)__init_begin - (unsigned long)_stext, 483 PAGE_KERNEL_RO); 484 } 485 486 static bool crash_mem_map __initdata; 487 488 static int __init enable_crash_mem_map(char *arg) 489 { 490 /* 491 * Proper parameter parsing is done by reserve_crashkernel(). We only 492 * need to know if the linear map has to avoid block mappings so that 493 * the crashkernel reservations can be unmapped later. 494 */ 495 crash_mem_map = true; 496 497 return 0; 498 } 499 early_param("crashkernel", enable_crash_mem_map); 500 501 static void __init map_mem(pgd_t *pgdp) 502 { 503 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN); 504 phys_addr_t kernel_start = __pa_symbol(_stext); 505 phys_addr_t kernel_end = __pa_symbol(__init_begin); 506 phys_addr_t start, end; 507 int flags = NO_EXEC_MAPPINGS; 508 u64 i; 509 510 /* 511 * Setting hierarchical PXNTable attributes on table entries covering 512 * the linear region is only possible if it is guaranteed that no table 513 * entries at any level are being shared between the linear region and 514 * the vmalloc region. Check whether this is true for the PGD level, in 515 * which case it is guaranteed to be true for all other levels as well. 516 */ 517 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end)); 518 519 if (can_set_direct_map() || crash_mem_map || IS_ENABLED(CONFIG_KFENCE)) 520 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 521 522 /* 523 * Take care not to create a writable alias for the 524 * read-only text and rodata sections of the kernel image. 525 * So temporarily mark them as NOMAP to skip mappings in 526 * the following for-loop 527 */ 528 memblock_mark_nomap(kernel_start, kernel_end - kernel_start); 529 530 /* map all the memory banks */ 531 for_each_mem_range(i, &start, &end) { 532 if (start >= end) 533 break; 534 /* 535 * The linear map must allow allocation tags reading/writing 536 * if MTE is present. Otherwise, it has the same attributes as 537 * PAGE_KERNEL. 538 */ 539 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL), 540 flags); 541 } 542 543 /* 544 * Map the linear alias of the [_stext, __init_begin) interval 545 * as non-executable now, and remove the write permission in 546 * mark_linear_text_alias_ro() below (which will be called after 547 * alternative patching has completed). This makes the contents 548 * of the region accessible to subsystems such as hibernate, 549 * but protects it from inadvertent modification or execution. 550 * Note that contiguous mappings cannot be remapped in this way, 551 * so we should avoid them here. 552 */ 553 __map_memblock(pgdp, kernel_start, kernel_end, 554 PAGE_KERNEL, NO_CONT_MAPPINGS); 555 memblock_clear_nomap(kernel_start, kernel_end - kernel_start); 556 } 557 558 void mark_rodata_ro(void) 559 { 560 unsigned long section_size; 561 562 /* 563 * mark .rodata as read only. Use __init_begin rather than __end_rodata 564 * to cover NOTES and EXCEPTION_TABLE. 565 */ 566 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata; 567 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata, 568 section_size, PAGE_KERNEL_RO); 569 570 debug_checkwx(); 571 } 572 573 static void __init map_kernel_segment(pgd_t *pgdp, void *va_start, void *va_end, 574 pgprot_t prot, struct vm_struct *vma, 575 int flags, unsigned long vm_flags) 576 { 577 phys_addr_t pa_start = __pa_symbol(va_start); 578 unsigned long size = va_end - va_start; 579 580 BUG_ON(!PAGE_ALIGNED(pa_start)); 581 BUG_ON(!PAGE_ALIGNED(size)); 582 583 __create_pgd_mapping(pgdp, pa_start, (unsigned long)va_start, size, prot, 584 early_pgtable_alloc, flags); 585 586 if (!(vm_flags & VM_NO_GUARD)) 587 size += PAGE_SIZE; 588 589 vma->addr = va_start; 590 vma->phys_addr = pa_start; 591 vma->size = size; 592 vma->flags = VM_MAP | vm_flags; 593 vma->caller = __builtin_return_address(0); 594 595 vm_area_add_early(vma); 596 } 597 598 static int __init parse_rodata(char *arg) 599 { 600 int ret = strtobool(arg, &rodata_enabled); 601 if (!ret) { 602 rodata_full = false; 603 return 0; 604 } 605 606 /* permit 'full' in addition to boolean options */ 607 if (strcmp(arg, "full")) 608 return -EINVAL; 609 610 rodata_enabled = true; 611 rodata_full = true; 612 return 0; 613 } 614 early_param("rodata", parse_rodata); 615 616 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0 617 static int __init map_entry_trampoline(void) 618 { 619 pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 620 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start); 621 622 /* The trampoline is always mapped and can therefore be global */ 623 pgprot_val(prot) &= ~PTE_NG; 624 625 /* Map only the text into the trampoline page table */ 626 memset(tramp_pg_dir, 0, PGD_SIZE); 627 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, PAGE_SIZE, 628 prot, __pgd_pgtable_alloc, 0); 629 630 /* Map both the text and data into the kernel page table */ 631 __set_fixmap(FIX_ENTRY_TRAMP_TEXT, pa_start, prot); 632 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 633 extern char __entry_tramp_data_start[]; 634 635 __set_fixmap(FIX_ENTRY_TRAMP_DATA, 636 __pa_symbol(__entry_tramp_data_start), 637 PAGE_KERNEL_RO); 638 } 639 640 return 0; 641 } 642 core_initcall(map_entry_trampoline); 643 #endif 644 645 /* 646 * Open coded check for BTI, only for use to determine configuration 647 * for early mappings for before the cpufeature code has run. 648 */ 649 static bool arm64_early_this_cpu_has_bti(void) 650 { 651 u64 pfr1; 652 653 if (!IS_ENABLED(CONFIG_ARM64_BTI_KERNEL)) 654 return false; 655 656 pfr1 = __read_sysreg_by_encoding(SYS_ID_AA64PFR1_EL1); 657 return cpuid_feature_extract_unsigned_field(pfr1, 658 ID_AA64PFR1_BT_SHIFT); 659 } 660 661 /* 662 * Create fine-grained mappings for the kernel. 663 */ 664 static void __init map_kernel(pgd_t *pgdp) 665 { 666 static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext, 667 vmlinux_initdata, vmlinux_data; 668 669 /* 670 * External debuggers may need to write directly to the text 671 * mapping to install SW breakpoints. Allow this (only) when 672 * explicitly requested with rodata=off. 673 */ 674 pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 675 676 /* 677 * If we have a CPU that supports BTI and a kernel built for 678 * BTI then mark the kernel executable text as guarded pages 679 * now so we don't have to rewrite the page tables later. 680 */ 681 if (arm64_early_this_cpu_has_bti()) 682 text_prot = __pgprot_modify(text_prot, PTE_GP, PTE_GP); 683 684 /* 685 * Only rodata will be remapped with different permissions later on, 686 * all other segments are allowed to use contiguous mappings. 687 */ 688 map_kernel_segment(pgdp, _stext, _etext, text_prot, &vmlinux_text, 0, 689 VM_NO_GUARD); 690 map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL, 691 &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD); 692 map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot, 693 &vmlinux_inittext, 0, VM_NO_GUARD); 694 map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL, 695 &vmlinux_initdata, 0, VM_NO_GUARD); 696 map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0); 697 698 if (!READ_ONCE(pgd_val(*pgd_offset_pgd(pgdp, FIXADDR_START)))) { 699 /* 700 * The fixmap falls in a separate pgd to the kernel, and doesn't 701 * live in the carveout for the swapper_pg_dir. We can simply 702 * re-use the existing dir for the fixmap. 703 */ 704 set_pgd(pgd_offset_pgd(pgdp, FIXADDR_START), 705 READ_ONCE(*pgd_offset_k(FIXADDR_START))); 706 } else if (CONFIG_PGTABLE_LEVELS > 3) { 707 pgd_t *bm_pgdp; 708 p4d_t *bm_p4dp; 709 pud_t *bm_pudp; 710 /* 711 * The fixmap shares its top level pgd entry with the kernel 712 * mapping. This can really only occur when we are running 713 * with 16k/4 levels, so we can simply reuse the pud level 714 * entry instead. 715 */ 716 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES)); 717 bm_pgdp = pgd_offset_pgd(pgdp, FIXADDR_START); 718 bm_p4dp = p4d_offset(bm_pgdp, FIXADDR_START); 719 bm_pudp = pud_set_fixmap_offset(bm_p4dp, FIXADDR_START); 720 pud_populate(&init_mm, bm_pudp, lm_alias(bm_pmd)); 721 pud_clear_fixmap(); 722 } else { 723 BUG(); 724 } 725 726 kasan_copy_shadow(pgdp); 727 } 728 729 void __init paging_init(void) 730 { 731 pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir)); 732 733 map_kernel(pgdp); 734 map_mem(pgdp); 735 736 pgd_clear_fixmap(); 737 738 cpu_replace_ttbr1(lm_alias(swapper_pg_dir)); 739 init_mm.pgd = swapper_pg_dir; 740 741 memblock_free(__pa_symbol(init_pg_dir), 742 __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir)); 743 744 memblock_allow_resize(); 745 } 746 747 /* 748 * Check whether a kernel address is valid (derived from arch/x86/). 749 */ 750 int kern_addr_valid(unsigned long addr) 751 { 752 pgd_t *pgdp; 753 p4d_t *p4dp; 754 pud_t *pudp, pud; 755 pmd_t *pmdp, pmd; 756 pte_t *ptep, pte; 757 758 addr = arch_kasan_reset_tag(addr); 759 if ((((long)addr) >> VA_BITS) != -1UL) 760 return 0; 761 762 pgdp = pgd_offset_k(addr); 763 if (pgd_none(READ_ONCE(*pgdp))) 764 return 0; 765 766 p4dp = p4d_offset(pgdp, addr); 767 if (p4d_none(READ_ONCE(*p4dp))) 768 return 0; 769 770 pudp = pud_offset(p4dp, addr); 771 pud = READ_ONCE(*pudp); 772 if (pud_none(pud)) 773 return 0; 774 775 if (pud_sect(pud)) 776 return pfn_valid(pud_pfn(pud)); 777 778 pmdp = pmd_offset(pudp, addr); 779 pmd = READ_ONCE(*pmdp); 780 if (pmd_none(pmd)) 781 return 0; 782 783 if (pmd_sect(pmd)) 784 return pfn_valid(pmd_pfn(pmd)); 785 786 ptep = pte_offset_kernel(pmdp, addr); 787 pte = READ_ONCE(*ptep); 788 if (pte_none(pte)) 789 return 0; 790 791 return pfn_valid(pte_pfn(pte)); 792 } 793 794 #ifdef CONFIG_MEMORY_HOTPLUG 795 static void free_hotplug_page_range(struct page *page, size_t size, 796 struct vmem_altmap *altmap) 797 { 798 if (altmap) { 799 vmem_altmap_free(altmap, size >> PAGE_SHIFT); 800 } else { 801 WARN_ON(PageReserved(page)); 802 free_pages((unsigned long)page_address(page), get_order(size)); 803 } 804 } 805 806 static void free_hotplug_pgtable_page(struct page *page) 807 { 808 free_hotplug_page_range(page, PAGE_SIZE, NULL); 809 } 810 811 static bool pgtable_range_aligned(unsigned long start, unsigned long end, 812 unsigned long floor, unsigned long ceiling, 813 unsigned long mask) 814 { 815 start &= mask; 816 if (start < floor) 817 return false; 818 819 if (ceiling) { 820 ceiling &= mask; 821 if (!ceiling) 822 return false; 823 } 824 825 if (end - 1 > ceiling - 1) 826 return false; 827 return true; 828 } 829 830 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr, 831 unsigned long end, bool free_mapped, 832 struct vmem_altmap *altmap) 833 { 834 pte_t *ptep, pte; 835 836 do { 837 ptep = pte_offset_kernel(pmdp, addr); 838 pte = READ_ONCE(*ptep); 839 if (pte_none(pte)) 840 continue; 841 842 WARN_ON(!pte_present(pte)); 843 pte_clear(&init_mm, addr, ptep); 844 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 845 if (free_mapped) 846 free_hotplug_page_range(pte_page(pte), 847 PAGE_SIZE, altmap); 848 } while (addr += PAGE_SIZE, addr < end); 849 } 850 851 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr, 852 unsigned long end, bool free_mapped, 853 struct vmem_altmap *altmap) 854 { 855 unsigned long next; 856 pmd_t *pmdp, pmd; 857 858 do { 859 next = pmd_addr_end(addr, end); 860 pmdp = pmd_offset(pudp, addr); 861 pmd = READ_ONCE(*pmdp); 862 if (pmd_none(pmd)) 863 continue; 864 865 WARN_ON(!pmd_present(pmd)); 866 if (pmd_sect(pmd)) { 867 pmd_clear(pmdp); 868 869 /* 870 * One TLBI should be sufficient here as the PMD_SIZE 871 * range is mapped with a single block entry. 872 */ 873 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 874 if (free_mapped) 875 free_hotplug_page_range(pmd_page(pmd), 876 PMD_SIZE, altmap); 877 continue; 878 } 879 WARN_ON(!pmd_table(pmd)); 880 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap); 881 } while (addr = next, addr < end); 882 } 883 884 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr, 885 unsigned long end, bool free_mapped, 886 struct vmem_altmap *altmap) 887 { 888 unsigned long next; 889 pud_t *pudp, pud; 890 891 do { 892 next = pud_addr_end(addr, end); 893 pudp = pud_offset(p4dp, addr); 894 pud = READ_ONCE(*pudp); 895 if (pud_none(pud)) 896 continue; 897 898 WARN_ON(!pud_present(pud)); 899 if (pud_sect(pud)) { 900 pud_clear(pudp); 901 902 /* 903 * One TLBI should be sufficient here as the PUD_SIZE 904 * range is mapped with a single block entry. 905 */ 906 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 907 if (free_mapped) 908 free_hotplug_page_range(pud_page(pud), 909 PUD_SIZE, altmap); 910 continue; 911 } 912 WARN_ON(!pud_table(pud)); 913 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap); 914 } while (addr = next, addr < end); 915 } 916 917 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr, 918 unsigned long end, bool free_mapped, 919 struct vmem_altmap *altmap) 920 { 921 unsigned long next; 922 p4d_t *p4dp, p4d; 923 924 do { 925 next = p4d_addr_end(addr, end); 926 p4dp = p4d_offset(pgdp, addr); 927 p4d = READ_ONCE(*p4dp); 928 if (p4d_none(p4d)) 929 continue; 930 931 WARN_ON(!p4d_present(p4d)); 932 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap); 933 } while (addr = next, addr < end); 934 } 935 936 static void unmap_hotplug_range(unsigned long addr, unsigned long end, 937 bool free_mapped, struct vmem_altmap *altmap) 938 { 939 unsigned long next; 940 pgd_t *pgdp, pgd; 941 942 /* 943 * altmap can only be used as vmemmap mapping backing memory. 944 * In case the backing memory itself is not being freed, then 945 * altmap is irrelevant. Warn about this inconsistency when 946 * encountered. 947 */ 948 WARN_ON(!free_mapped && altmap); 949 950 do { 951 next = pgd_addr_end(addr, end); 952 pgdp = pgd_offset_k(addr); 953 pgd = READ_ONCE(*pgdp); 954 if (pgd_none(pgd)) 955 continue; 956 957 WARN_ON(!pgd_present(pgd)); 958 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap); 959 } while (addr = next, addr < end); 960 } 961 962 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr, 963 unsigned long end, unsigned long floor, 964 unsigned long ceiling) 965 { 966 pte_t *ptep, pte; 967 unsigned long i, start = addr; 968 969 do { 970 ptep = pte_offset_kernel(pmdp, addr); 971 pte = READ_ONCE(*ptep); 972 973 /* 974 * This is just a sanity check here which verifies that 975 * pte clearing has been done by earlier unmap loops. 976 */ 977 WARN_ON(!pte_none(pte)); 978 } while (addr += PAGE_SIZE, addr < end); 979 980 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK)) 981 return; 982 983 /* 984 * Check whether we can free the pte page if the rest of the 985 * entries are empty. Overlap with other regions have been 986 * handled by the floor/ceiling check. 987 */ 988 ptep = pte_offset_kernel(pmdp, 0UL); 989 for (i = 0; i < PTRS_PER_PTE; i++) { 990 if (!pte_none(READ_ONCE(ptep[i]))) 991 return; 992 } 993 994 pmd_clear(pmdp); 995 __flush_tlb_kernel_pgtable(start); 996 free_hotplug_pgtable_page(virt_to_page(ptep)); 997 } 998 999 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr, 1000 unsigned long end, unsigned long floor, 1001 unsigned long ceiling) 1002 { 1003 pmd_t *pmdp, pmd; 1004 unsigned long i, next, start = addr; 1005 1006 do { 1007 next = pmd_addr_end(addr, end); 1008 pmdp = pmd_offset(pudp, addr); 1009 pmd = READ_ONCE(*pmdp); 1010 if (pmd_none(pmd)) 1011 continue; 1012 1013 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd)); 1014 free_empty_pte_table(pmdp, addr, next, floor, ceiling); 1015 } while (addr = next, addr < end); 1016 1017 if (CONFIG_PGTABLE_LEVELS <= 2) 1018 return; 1019 1020 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK)) 1021 return; 1022 1023 /* 1024 * Check whether we can free the pmd page if the rest of the 1025 * entries are empty. Overlap with other regions have been 1026 * handled by the floor/ceiling check. 1027 */ 1028 pmdp = pmd_offset(pudp, 0UL); 1029 for (i = 0; i < PTRS_PER_PMD; i++) { 1030 if (!pmd_none(READ_ONCE(pmdp[i]))) 1031 return; 1032 } 1033 1034 pud_clear(pudp); 1035 __flush_tlb_kernel_pgtable(start); 1036 free_hotplug_pgtable_page(virt_to_page(pmdp)); 1037 } 1038 1039 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr, 1040 unsigned long end, unsigned long floor, 1041 unsigned long ceiling) 1042 { 1043 pud_t *pudp, pud; 1044 unsigned long i, next, start = addr; 1045 1046 do { 1047 next = pud_addr_end(addr, end); 1048 pudp = pud_offset(p4dp, addr); 1049 pud = READ_ONCE(*pudp); 1050 if (pud_none(pud)) 1051 continue; 1052 1053 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud)); 1054 free_empty_pmd_table(pudp, addr, next, floor, ceiling); 1055 } while (addr = next, addr < end); 1056 1057 if (CONFIG_PGTABLE_LEVELS <= 3) 1058 return; 1059 1060 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK)) 1061 return; 1062 1063 /* 1064 * Check whether we can free the pud page if the rest of the 1065 * entries are empty. Overlap with other regions have been 1066 * handled by the floor/ceiling check. 1067 */ 1068 pudp = pud_offset(p4dp, 0UL); 1069 for (i = 0; i < PTRS_PER_PUD; i++) { 1070 if (!pud_none(READ_ONCE(pudp[i]))) 1071 return; 1072 } 1073 1074 p4d_clear(p4dp); 1075 __flush_tlb_kernel_pgtable(start); 1076 free_hotplug_pgtable_page(virt_to_page(pudp)); 1077 } 1078 1079 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr, 1080 unsigned long end, unsigned long floor, 1081 unsigned long ceiling) 1082 { 1083 unsigned long next; 1084 p4d_t *p4dp, p4d; 1085 1086 do { 1087 next = p4d_addr_end(addr, end); 1088 p4dp = p4d_offset(pgdp, addr); 1089 p4d = READ_ONCE(*p4dp); 1090 if (p4d_none(p4d)) 1091 continue; 1092 1093 WARN_ON(!p4d_present(p4d)); 1094 free_empty_pud_table(p4dp, addr, next, floor, ceiling); 1095 } while (addr = next, addr < end); 1096 } 1097 1098 static void free_empty_tables(unsigned long addr, unsigned long end, 1099 unsigned long floor, unsigned long ceiling) 1100 { 1101 unsigned long next; 1102 pgd_t *pgdp, pgd; 1103 1104 do { 1105 next = pgd_addr_end(addr, end); 1106 pgdp = pgd_offset_k(addr); 1107 pgd = READ_ONCE(*pgdp); 1108 if (pgd_none(pgd)) 1109 continue; 1110 1111 WARN_ON(!pgd_present(pgd)); 1112 free_empty_p4d_table(pgdp, addr, next, floor, ceiling); 1113 } while (addr = next, addr < end); 1114 } 1115 #endif 1116 1117 #if !ARM64_KERNEL_USES_PMD_MAPS 1118 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1119 struct vmem_altmap *altmap) 1120 { 1121 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1122 return vmemmap_populate_basepages(start, end, node, altmap); 1123 } 1124 #else /* !ARM64_KERNEL_USES_PMD_MAPS */ 1125 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1126 struct vmem_altmap *altmap) 1127 { 1128 unsigned long addr = start; 1129 unsigned long next; 1130 pgd_t *pgdp; 1131 p4d_t *p4dp; 1132 pud_t *pudp; 1133 pmd_t *pmdp; 1134 1135 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1136 do { 1137 next = pmd_addr_end(addr, end); 1138 1139 pgdp = vmemmap_pgd_populate(addr, node); 1140 if (!pgdp) 1141 return -ENOMEM; 1142 1143 p4dp = vmemmap_p4d_populate(pgdp, addr, node); 1144 if (!p4dp) 1145 return -ENOMEM; 1146 1147 pudp = vmemmap_pud_populate(p4dp, addr, node); 1148 if (!pudp) 1149 return -ENOMEM; 1150 1151 pmdp = pmd_offset(pudp, addr); 1152 if (pmd_none(READ_ONCE(*pmdp))) { 1153 void *p = NULL; 1154 1155 p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap); 1156 if (!p) { 1157 if (vmemmap_populate_basepages(addr, next, node, altmap)) 1158 return -ENOMEM; 1159 continue; 1160 } 1161 1162 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL)); 1163 } else 1164 vmemmap_verify((pte_t *)pmdp, node, addr, next); 1165 } while (addr = next, addr != end); 1166 1167 return 0; 1168 } 1169 #endif /* !ARM64_KERNEL_USES_PMD_MAPS */ 1170 1171 #ifdef CONFIG_MEMORY_HOTPLUG 1172 void vmemmap_free(unsigned long start, unsigned long end, 1173 struct vmem_altmap *altmap) 1174 { 1175 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1176 1177 unmap_hotplug_range(start, end, true, altmap); 1178 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END); 1179 } 1180 #endif /* CONFIG_MEMORY_HOTPLUG */ 1181 1182 static inline pud_t *fixmap_pud(unsigned long addr) 1183 { 1184 pgd_t *pgdp = pgd_offset_k(addr); 1185 p4d_t *p4dp = p4d_offset(pgdp, addr); 1186 p4d_t p4d = READ_ONCE(*p4dp); 1187 1188 BUG_ON(p4d_none(p4d) || p4d_bad(p4d)); 1189 1190 return pud_offset_kimg(p4dp, addr); 1191 } 1192 1193 static inline pmd_t *fixmap_pmd(unsigned long addr) 1194 { 1195 pud_t *pudp = fixmap_pud(addr); 1196 pud_t pud = READ_ONCE(*pudp); 1197 1198 BUG_ON(pud_none(pud) || pud_bad(pud)); 1199 1200 return pmd_offset_kimg(pudp, addr); 1201 } 1202 1203 static inline pte_t *fixmap_pte(unsigned long addr) 1204 { 1205 return &bm_pte[pte_index(addr)]; 1206 } 1207 1208 /* 1209 * The p*d_populate functions call virt_to_phys implicitly so they can't be used 1210 * directly on kernel symbols (bm_p*d). This function is called too early to use 1211 * lm_alias so __p*d_populate functions must be used to populate with the 1212 * physical address from __pa_symbol. 1213 */ 1214 void __init early_fixmap_init(void) 1215 { 1216 pgd_t *pgdp; 1217 p4d_t *p4dp, p4d; 1218 pud_t *pudp; 1219 pmd_t *pmdp; 1220 unsigned long addr = FIXADDR_START; 1221 1222 pgdp = pgd_offset_k(addr); 1223 p4dp = p4d_offset(pgdp, addr); 1224 p4d = READ_ONCE(*p4dp); 1225 if (CONFIG_PGTABLE_LEVELS > 3 && 1226 !(p4d_none(p4d) || p4d_page_paddr(p4d) == __pa_symbol(bm_pud))) { 1227 /* 1228 * We only end up here if the kernel mapping and the fixmap 1229 * share the top level pgd entry, which should only happen on 1230 * 16k/4 levels configurations. 1231 */ 1232 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES)); 1233 pudp = pud_offset_kimg(p4dp, addr); 1234 } else { 1235 if (p4d_none(p4d)) 1236 __p4d_populate(p4dp, __pa_symbol(bm_pud), P4D_TYPE_TABLE); 1237 pudp = fixmap_pud(addr); 1238 } 1239 if (pud_none(READ_ONCE(*pudp))) 1240 __pud_populate(pudp, __pa_symbol(bm_pmd), PUD_TYPE_TABLE); 1241 pmdp = fixmap_pmd(addr); 1242 __pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE); 1243 1244 /* 1245 * The boot-ioremap range spans multiple pmds, for which 1246 * we are not prepared: 1247 */ 1248 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT) 1249 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT)); 1250 1251 if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN))) 1252 || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) { 1253 WARN_ON(1); 1254 pr_warn("pmdp %p != %p, %p\n", 1255 pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)), 1256 fixmap_pmd(fix_to_virt(FIX_BTMAP_END))); 1257 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n", 1258 fix_to_virt(FIX_BTMAP_BEGIN)); 1259 pr_warn("fix_to_virt(FIX_BTMAP_END): %08lx\n", 1260 fix_to_virt(FIX_BTMAP_END)); 1261 1262 pr_warn("FIX_BTMAP_END: %d\n", FIX_BTMAP_END); 1263 pr_warn("FIX_BTMAP_BEGIN: %d\n", FIX_BTMAP_BEGIN); 1264 } 1265 } 1266 1267 /* 1268 * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we 1269 * ever need to use IPIs for TLB broadcasting, then we're in trouble here. 1270 */ 1271 void __set_fixmap(enum fixed_addresses idx, 1272 phys_addr_t phys, pgprot_t flags) 1273 { 1274 unsigned long addr = __fix_to_virt(idx); 1275 pte_t *ptep; 1276 1277 BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses); 1278 1279 ptep = fixmap_pte(addr); 1280 1281 if (pgprot_val(flags)) { 1282 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags)); 1283 } else { 1284 pte_clear(&init_mm, addr, ptep); 1285 flush_tlb_kernel_range(addr, addr+PAGE_SIZE); 1286 } 1287 } 1288 1289 void *__init fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot) 1290 { 1291 const u64 dt_virt_base = __fix_to_virt(FIX_FDT); 1292 int offset; 1293 void *dt_virt; 1294 1295 /* 1296 * Check whether the physical FDT address is set and meets the minimum 1297 * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be 1298 * at least 8 bytes so that we can always access the magic and size 1299 * fields of the FDT header after mapping the first chunk, double check 1300 * here if that is indeed the case. 1301 */ 1302 BUILD_BUG_ON(MIN_FDT_ALIGN < 8); 1303 if (!dt_phys || dt_phys % MIN_FDT_ALIGN) 1304 return NULL; 1305 1306 /* 1307 * Make sure that the FDT region can be mapped without the need to 1308 * allocate additional translation table pages, so that it is safe 1309 * to call create_mapping_noalloc() this early. 1310 * 1311 * On 64k pages, the FDT will be mapped using PTEs, so we need to 1312 * be in the same PMD as the rest of the fixmap. 1313 * On 4k pages, we'll use section mappings for the FDT so we only 1314 * have to be in the same PUD. 1315 */ 1316 BUILD_BUG_ON(dt_virt_base % SZ_2M); 1317 1318 BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT != 1319 __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT); 1320 1321 offset = dt_phys % SWAPPER_BLOCK_SIZE; 1322 dt_virt = (void *)dt_virt_base + offset; 1323 1324 /* map the first chunk so we can read the size from the header */ 1325 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), 1326 dt_virt_base, SWAPPER_BLOCK_SIZE, prot); 1327 1328 if (fdt_magic(dt_virt) != FDT_MAGIC) 1329 return NULL; 1330 1331 *size = fdt_totalsize(dt_virt); 1332 if (*size > MAX_FDT_SIZE) 1333 return NULL; 1334 1335 if (offset + *size > SWAPPER_BLOCK_SIZE) 1336 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base, 1337 round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot); 1338 1339 return dt_virt; 1340 } 1341 1342 #if CONFIG_PGTABLE_LEVELS > 3 1343 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot) 1344 { 1345 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot)); 1346 1347 /* Only allow permission changes for now */ 1348 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)), 1349 pud_val(new_pud))) 1350 return 0; 1351 1352 VM_BUG_ON(phys & ~PUD_MASK); 1353 set_pud(pudp, new_pud); 1354 return 1; 1355 } 1356 1357 int pud_clear_huge(pud_t *pudp) 1358 { 1359 if (!pud_sect(READ_ONCE(*pudp))) 1360 return 0; 1361 pud_clear(pudp); 1362 return 1; 1363 } 1364 #endif 1365 1366 #if CONFIG_PGTABLE_LEVELS > 2 1367 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot) 1368 { 1369 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot)); 1370 1371 /* Only allow permission changes for now */ 1372 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)), 1373 pmd_val(new_pmd))) 1374 return 0; 1375 1376 VM_BUG_ON(phys & ~PMD_MASK); 1377 set_pmd(pmdp, new_pmd); 1378 return 1; 1379 } 1380 1381 int pmd_clear_huge(pmd_t *pmdp) 1382 { 1383 if (!pmd_sect(READ_ONCE(*pmdp))) 1384 return 0; 1385 pmd_clear(pmdp); 1386 return 1; 1387 } 1388 #endif 1389 1390 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr) 1391 { 1392 pte_t *table; 1393 pmd_t pmd; 1394 1395 pmd = READ_ONCE(*pmdp); 1396 1397 if (!pmd_table(pmd)) { 1398 VM_WARN_ON(1); 1399 return 1; 1400 } 1401 1402 table = pte_offset_kernel(pmdp, addr); 1403 pmd_clear(pmdp); 1404 __flush_tlb_kernel_pgtable(addr); 1405 pte_free_kernel(NULL, table); 1406 return 1; 1407 } 1408 1409 int pud_free_pmd_page(pud_t *pudp, unsigned long addr) 1410 { 1411 pmd_t *table; 1412 pmd_t *pmdp; 1413 pud_t pud; 1414 unsigned long next, end; 1415 1416 pud = READ_ONCE(*pudp); 1417 1418 if (!pud_table(pud)) { 1419 VM_WARN_ON(1); 1420 return 1; 1421 } 1422 1423 table = pmd_offset(pudp, addr); 1424 pmdp = table; 1425 next = addr; 1426 end = addr + PUD_SIZE; 1427 do { 1428 pmd_free_pte_page(pmdp, next); 1429 } while (pmdp++, next += PMD_SIZE, next != end); 1430 1431 pud_clear(pudp); 1432 __flush_tlb_kernel_pgtable(addr); 1433 pmd_free(NULL, table); 1434 return 1; 1435 } 1436 1437 #ifdef CONFIG_MEMORY_HOTPLUG 1438 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size) 1439 { 1440 unsigned long end = start + size; 1441 1442 WARN_ON(pgdir != init_mm.pgd); 1443 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END)); 1444 1445 unmap_hotplug_range(start, end, false, NULL); 1446 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END); 1447 } 1448 1449 struct range arch_get_mappable_range(void) 1450 { 1451 struct range mhp_range; 1452 u64 start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual)); 1453 u64 end_linear_pa = __pa(PAGE_END - 1); 1454 1455 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 1456 /* 1457 * Check for a wrap, it is possible because of randomized linear 1458 * mapping the start physical address is actually bigger than 1459 * the end physical address. In this case set start to zero 1460 * because [0, end_linear_pa] range must still be able to cover 1461 * all addressable physical addresses. 1462 */ 1463 if (start_linear_pa > end_linear_pa) 1464 start_linear_pa = 0; 1465 } 1466 1467 WARN_ON(start_linear_pa > end_linear_pa); 1468 1469 /* 1470 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)] 1471 * accommodating both its ends but excluding PAGE_END. Max physical 1472 * range which can be mapped inside this linear mapping range, must 1473 * also be derived from its end points. 1474 */ 1475 mhp_range.start = start_linear_pa; 1476 mhp_range.end = end_linear_pa; 1477 1478 return mhp_range; 1479 } 1480 1481 int arch_add_memory(int nid, u64 start, u64 size, 1482 struct mhp_params *params) 1483 { 1484 int ret, flags = NO_EXEC_MAPPINGS; 1485 1486 VM_BUG_ON(!mhp_range_allowed(start, size, true)); 1487 1488 /* 1489 * KFENCE requires linear map to be mapped at page granularity, so that 1490 * it is possible to protect/unprotect single pages in the KFENCE pool. 1491 */ 1492 if (can_set_direct_map() || IS_ENABLED(CONFIG_KFENCE)) 1493 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1494 1495 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1496 size, params->pgprot, __pgd_pgtable_alloc, 1497 flags); 1498 1499 memblock_clear_nomap(start, size); 1500 1501 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, 1502 params); 1503 if (ret) 1504 __remove_pgd_mapping(swapper_pg_dir, 1505 __phys_to_virt(start), size); 1506 return ret; 1507 } 1508 1509 void arch_remove_memory(int nid, u64 start, u64 size, 1510 struct vmem_altmap *altmap) 1511 { 1512 unsigned long start_pfn = start >> PAGE_SHIFT; 1513 unsigned long nr_pages = size >> PAGE_SHIFT; 1514 1515 __remove_pages(start_pfn, nr_pages, altmap); 1516 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size); 1517 } 1518 1519 /* 1520 * This memory hotplug notifier helps prevent boot memory from being 1521 * inadvertently removed as it blocks pfn range offlining process in 1522 * __offline_pages(). Hence this prevents both offlining as well as 1523 * removal process for boot memory which is initially always online. 1524 * In future if and when boot memory could be removed, this notifier 1525 * should be dropped and free_hotplug_page_range() should handle any 1526 * reserved pages allocated during boot. 1527 */ 1528 static int prevent_bootmem_remove_notifier(struct notifier_block *nb, 1529 unsigned long action, void *data) 1530 { 1531 struct mem_section *ms; 1532 struct memory_notify *arg = data; 1533 unsigned long end_pfn = arg->start_pfn + arg->nr_pages; 1534 unsigned long pfn = arg->start_pfn; 1535 1536 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE)) 1537 return NOTIFY_OK; 1538 1539 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 1540 unsigned long start = PFN_PHYS(pfn); 1541 unsigned long end = start + (1UL << PA_SECTION_SHIFT); 1542 1543 ms = __pfn_to_section(pfn); 1544 if (!early_section(ms)) 1545 continue; 1546 1547 if (action == MEM_GOING_OFFLINE) { 1548 /* 1549 * Boot memory removal is not supported. Prevent 1550 * it via blocking any attempted offline request 1551 * for the boot memory and just report it. 1552 */ 1553 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end); 1554 return NOTIFY_BAD; 1555 } else if (action == MEM_OFFLINE) { 1556 /* 1557 * This should have never happened. Boot memory 1558 * offlining should have been prevented by this 1559 * very notifier. Probably some memory removal 1560 * procedure might have changed which would then 1561 * require further debug. 1562 */ 1563 pr_err("Boot memory [%lx %lx] offlined\n", start, end); 1564 1565 /* 1566 * Core memory hotplug does not process a return 1567 * code from the notifier for MEM_OFFLINE events. 1568 * The error condition has been reported. Return 1569 * from here as if ignored. 1570 */ 1571 return NOTIFY_DONE; 1572 } 1573 } 1574 return NOTIFY_OK; 1575 } 1576 1577 static struct notifier_block prevent_bootmem_remove_nb = { 1578 .notifier_call = prevent_bootmem_remove_notifier, 1579 }; 1580 1581 /* 1582 * This ensures that boot memory sections on the platform are online 1583 * from early boot. Memory sections could not be prevented from being 1584 * offlined, unless for some reason they are not online to begin with. 1585 * This helps validate the basic assumption on which the above memory 1586 * event notifier works to prevent boot memory section offlining and 1587 * its possible removal. 1588 */ 1589 static void validate_bootmem_online(void) 1590 { 1591 phys_addr_t start, end, addr; 1592 struct mem_section *ms; 1593 u64 i; 1594 1595 /* 1596 * Scanning across all memblock might be expensive 1597 * on some big memory systems. Hence enable this 1598 * validation only with DEBUG_VM. 1599 */ 1600 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 1601 return; 1602 1603 for_each_mem_range(i, &start, &end) { 1604 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) { 1605 ms = __pfn_to_section(PHYS_PFN(addr)); 1606 1607 /* 1608 * All memory ranges in the system at this point 1609 * should have been marked as early sections. 1610 */ 1611 WARN_ON(!early_section(ms)); 1612 1613 /* 1614 * Memory notifier mechanism here to prevent boot 1615 * memory offlining depends on the fact that each 1616 * early section memory on the system is initially 1617 * online. Otherwise a given memory section which 1618 * is already offline will be overlooked and can 1619 * be removed completely. Call out such sections. 1620 */ 1621 if (!online_section(ms)) 1622 pr_err("Boot memory [%llx %llx] is offline, can be removed\n", 1623 addr, addr + (1UL << PA_SECTION_SHIFT)); 1624 } 1625 } 1626 } 1627 1628 static int __init prevent_bootmem_remove_init(void) 1629 { 1630 int ret = 0; 1631 1632 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) 1633 return ret; 1634 1635 validate_bootmem_online(); 1636 ret = register_memory_notifier(&prevent_bootmem_remove_nb); 1637 if (ret) 1638 pr_err("%s: Notifier registration failed %d\n", __func__, ret); 1639 1640 return ret; 1641 } 1642 early_initcall(prevent_bootmem_remove_init); 1643 #endif 1644