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_range(PAGE_SIZE, PAGE_SIZE, 0, 100 MEMBLOCK_ALLOC_NOLEAKTRACE); 101 if (!phys) 102 panic("Failed to allocate page table page\n"); 103 104 /* 105 * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE 106 * slot will be free, so we can (ab)use the FIX_PTE slot to initialise 107 * any level of table. 108 */ 109 ptr = pte_set_fixmap(phys); 110 111 memset(ptr, 0, PAGE_SIZE); 112 113 /* 114 * Implicit barriers also ensure the zeroed page is visible to the page 115 * table walker 116 */ 117 pte_clear_fixmap(); 118 119 return phys; 120 } 121 122 static bool pgattr_change_is_safe(u64 old, u64 new) 123 { 124 /* 125 * The following mapping attributes may be updated in live 126 * kernel mappings without the need for break-before-make. 127 */ 128 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG; 129 130 /* creating or taking down mappings is always safe */ 131 if (old == 0 || new == 0) 132 return true; 133 134 /* live contiguous mappings may not be manipulated at all */ 135 if ((old | new) & PTE_CONT) 136 return false; 137 138 /* Transitioning from Non-Global to Global is unsafe */ 139 if (old & ~new & PTE_NG) 140 return false; 141 142 /* 143 * Changing the memory type between Normal and Normal-Tagged is safe 144 * since Tagged is considered a permission attribute from the 145 * mismatched attribute aliases perspective. 146 */ 147 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 148 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) && 149 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 150 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED))) 151 mask |= PTE_ATTRINDX_MASK; 152 153 return ((old ^ new) & ~mask) == 0; 154 } 155 156 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end, 157 phys_addr_t phys, pgprot_t prot) 158 { 159 pte_t *ptep; 160 161 ptep = pte_set_fixmap_offset(pmdp, addr); 162 do { 163 pte_t old_pte = READ_ONCE(*ptep); 164 165 set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot)); 166 167 /* 168 * After the PTE entry has been populated once, we 169 * only allow updates to the permission attributes. 170 */ 171 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte), 172 READ_ONCE(pte_val(*ptep)))); 173 174 phys += PAGE_SIZE; 175 } while (ptep++, addr += PAGE_SIZE, addr != end); 176 177 pte_clear_fixmap(); 178 } 179 180 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr, 181 unsigned long end, phys_addr_t phys, 182 pgprot_t prot, 183 phys_addr_t (*pgtable_alloc)(int), 184 int flags) 185 { 186 unsigned long next; 187 pmd_t pmd = READ_ONCE(*pmdp); 188 189 BUG_ON(pmd_sect(pmd)); 190 if (pmd_none(pmd)) { 191 pmdval_t pmdval = PMD_TYPE_TABLE | PMD_TABLE_UXN; 192 phys_addr_t pte_phys; 193 194 if (flags & NO_EXEC_MAPPINGS) 195 pmdval |= PMD_TABLE_PXN; 196 BUG_ON(!pgtable_alloc); 197 pte_phys = pgtable_alloc(PAGE_SHIFT); 198 __pmd_populate(pmdp, pte_phys, pmdval); 199 pmd = READ_ONCE(*pmdp); 200 } 201 BUG_ON(pmd_bad(pmd)); 202 203 do { 204 pgprot_t __prot = prot; 205 206 next = pte_cont_addr_end(addr, end); 207 208 /* use a contiguous mapping if the range is suitably aligned */ 209 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) && 210 (flags & NO_CONT_MAPPINGS) == 0) 211 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 212 213 init_pte(pmdp, addr, next, phys, __prot); 214 215 phys += next - addr; 216 } while (addr = next, addr != end); 217 } 218 219 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end, 220 phys_addr_t phys, pgprot_t prot, 221 phys_addr_t (*pgtable_alloc)(int), int flags) 222 { 223 unsigned long next; 224 pmd_t *pmdp; 225 226 pmdp = pmd_set_fixmap_offset(pudp, addr); 227 do { 228 pmd_t old_pmd = READ_ONCE(*pmdp); 229 230 next = pmd_addr_end(addr, end); 231 232 /* try section mapping first */ 233 if (((addr | next | phys) & ~PMD_MASK) == 0 && 234 (flags & NO_BLOCK_MAPPINGS) == 0) { 235 pmd_set_huge(pmdp, phys, prot); 236 237 /* 238 * After the PMD entry has been populated once, we 239 * only allow updates to the permission attributes. 240 */ 241 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd), 242 READ_ONCE(pmd_val(*pmdp)))); 243 } else { 244 alloc_init_cont_pte(pmdp, addr, next, phys, prot, 245 pgtable_alloc, flags); 246 247 BUG_ON(pmd_val(old_pmd) != 0 && 248 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp))); 249 } 250 phys += next - addr; 251 } while (pmdp++, addr = next, addr != end); 252 253 pmd_clear_fixmap(); 254 } 255 256 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr, 257 unsigned long end, phys_addr_t phys, 258 pgprot_t prot, 259 phys_addr_t (*pgtable_alloc)(int), int flags) 260 { 261 unsigned long next; 262 pud_t pud = READ_ONCE(*pudp); 263 264 /* 265 * Check for initial section mappings in the pgd/pud. 266 */ 267 BUG_ON(pud_sect(pud)); 268 if (pud_none(pud)) { 269 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN; 270 phys_addr_t pmd_phys; 271 272 if (flags & NO_EXEC_MAPPINGS) 273 pudval |= PUD_TABLE_PXN; 274 BUG_ON(!pgtable_alloc); 275 pmd_phys = pgtable_alloc(PMD_SHIFT); 276 __pud_populate(pudp, pmd_phys, pudval); 277 pud = READ_ONCE(*pudp); 278 } 279 BUG_ON(pud_bad(pud)); 280 281 do { 282 pgprot_t __prot = prot; 283 284 next = pmd_cont_addr_end(addr, end); 285 286 /* use a contiguous mapping if the range is suitably aligned */ 287 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) && 288 (flags & NO_CONT_MAPPINGS) == 0) 289 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 290 291 init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags); 292 293 phys += next - addr; 294 } while (addr = next, addr != end); 295 } 296 297 static inline bool use_1G_block(unsigned long addr, unsigned long next, 298 unsigned long phys) 299 { 300 if (PAGE_SHIFT != 12) 301 return false; 302 303 if (((addr | next | phys) & ~PUD_MASK) != 0) 304 return false; 305 306 return true; 307 } 308 309 static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end, 310 phys_addr_t phys, pgprot_t prot, 311 phys_addr_t (*pgtable_alloc)(int), 312 int flags) 313 { 314 unsigned long next; 315 pud_t *pudp; 316 p4d_t *p4dp = p4d_offset(pgdp, addr); 317 p4d_t p4d = READ_ONCE(*p4dp); 318 319 if (p4d_none(p4d)) { 320 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN; 321 phys_addr_t pud_phys; 322 323 if (flags & NO_EXEC_MAPPINGS) 324 p4dval |= P4D_TABLE_PXN; 325 BUG_ON(!pgtable_alloc); 326 pud_phys = pgtable_alloc(PUD_SHIFT); 327 __p4d_populate(p4dp, pud_phys, p4dval); 328 p4d = READ_ONCE(*p4dp); 329 } 330 BUG_ON(p4d_bad(p4d)); 331 332 pudp = pud_set_fixmap_offset(p4dp, addr); 333 do { 334 pud_t old_pud = READ_ONCE(*pudp); 335 336 next = pud_addr_end(addr, end); 337 338 /* 339 * For 4K granule only, attempt to put down a 1GB block 340 */ 341 if (use_1G_block(addr, next, phys) && 342 (flags & NO_BLOCK_MAPPINGS) == 0) { 343 pud_set_huge(pudp, phys, prot); 344 345 /* 346 * After the PUD entry has been populated once, we 347 * only allow updates to the permission attributes. 348 */ 349 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud), 350 READ_ONCE(pud_val(*pudp)))); 351 } else { 352 alloc_init_cont_pmd(pudp, addr, next, phys, prot, 353 pgtable_alloc, flags); 354 355 BUG_ON(pud_val(old_pud) != 0 && 356 pud_val(old_pud) != READ_ONCE(pud_val(*pudp))); 357 } 358 phys += next - addr; 359 } while (pudp++, addr = next, addr != end); 360 361 pud_clear_fixmap(); 362 } 363 364 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, 365 unsigned long virt, phys_addr_t size, 366 pgprot_t prot, 367 phys_addr_t (*pgtable_alloc)(int), 368 int flags) 369 { 370 unsigned long addr, end, next; 371 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt); 372 373 /* 374 * If the virtual and physical address don't have the same offset 375 * within a page, we cannot map the region as the caller expects. 376 */ 377 if (WARN_ON((phys ^ virt) & ~PAGE_MASK)) 378 return; 379 380 phys &= PAGE_MASK; 381 addr = virt & PAGE_MASK; 382 end = PAGE_ALIGN(virt + size); 383 384 do { 385 next = pgd_addr_end(addr, end); 386 alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc, 387 flags); 388 phys += next - addr; 389 } while (pgdp++, addr = next, addr != end); 390 } 391 392 static phys_addr_t __pgd_pgtable_alloc(int shift) 393 { 394 void *ptr = (void *)__get_free_page(GFP_PGTABLE_KERNEL); 395 BUG_ON(!ptr); 396 397 /* Ensure the zeroed page is visible to the page table walker */ 398 dsb(ishst); 399 return __pa(ptr); 400 } 401 402 static phys_addr_t pgd_pgtable_alloc(int shift) 403 { 404 phys_addr_t pa = __pgd_pgtable_alloc(shift); 405 406 /* 407 * Call proper page table ctor in case later we need to 408 * call core mm functions like apply_to_page_range() on 409 * this pre-allocated page table. 410 * 411 * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is 412 * folded, and if so pgtable_pmd_page_ctor() becomes nop. 413 */ 414 if (shift == PAGE_SHIFT) 415 BUG_ON(!pgtable_pte_page_ctor(phys_to_page(pa))); 416 else if (shift == PMD_SHIFT) 417 BUG_ON(!pgtable_pmd_page_ctor(phys_to_page(pa))); 418 419 return pa; 420 } 421 422 /* 423 * This function can only be used to modify existing table entries, 424 * without allocating new levels of table. Note that this permits the 425 * creation of new section or page entries. 426 */ 427 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt, 428 phys_addr_t size, pgprot_t prot) 429 { 430 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) { 431 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n", 432 &phys, virt); 433 return; 434 } 435 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 436 NO_CONT_MAPPINGS); 437 } 438 439 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys, 440 unsigned long virt, phys_addr_t size, 441 pgprot_t prot, bool page_mappings_only) 442 { 443 int flags = 0; 444 445 BUG_ON(mm == &init_mm); 446 447 if (page_mappings_only) 448 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 449 450 __create_pgd_mapping(mm->pgd, phys, virt, size, prot, 451 pgd_pgtable_alloc, flags); 452 } 453 454 static void update_mapping_prot(phys_addr_t phys, unsigned long virt, 455 phys_addr_t size, pgprot_t prot) 456 { 457 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) { 458 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n", 459 &phys, virt); 460 return; 461 } 462 463 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 464 NO_CONT_MAPPINGS); 465 466 /* flush the TLBs after updating live kernel mappings */ 467 flush_tlb_kernel_range(virt, virt + size); 468 } 469 470 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start, 471 phys_addr_t end, pgprot_t prot, int flags) 472 { 473 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start, 474 prot, early_pgtable_alloc, flags); 475 } 476 477 void __init mark_linear_text_alias_ro(void) 478 { 479 /* 480 * Remove the write permissions from the linear alias of .text/.rodata 481 */ 482 update_mapping_prot(__pa_symbol(_stext), (unsigned long)lm_alias(_stext), 483 (unsigned long)__init_begin - (unsigned long)_stext, 484 PAGE_KERNEL_RO); 485 } 486 487 static bool crash_mem_map __initdata; 488 489 static int __init enable_crash_mem_map(char *arg) 490 { 491 /* 492 * Proper parameter parsing is done by reserve_crashkernel(). We only 493 * need to know if the linear map has to avoid block mappings so that 494 * the crashkernel reservations can be unmapped later. 495 */ 496 crash_mem_map = true; 497 498 return 0; 499 } 500 early_param("crashkernel", enable_crash_mem_map); 501 502 static void __init map_mem(pgd_t *pgdp) 503 { 504 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN); 505 phys_addr_t kernel_start = __pa_symbol(_stext); 506 phys_addr_t kernel_end = __pa_symbol(__init_begin); 507 phys_addr_t start, end; 508 int flags = NO_EXEC_MAPPINGS; 509 u64 i; 510 511 /* 512 * Setting hierarchical PXNTable attributes on table entries covering 513 * the linear region is only possible if it is guaranteed that no table 514 * entries at any level are being shared between the linear region and 515 * the vmalloc region. Check whether this is true for the PGD level, in 516 * which case it is guaranteed to be true for all other levels as well. 517 */ 518 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end)); 519 520 if (can_set_direct_map() || crash_mem_map || IS_ENABLED(CONFIG_KFENCE)) 521 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 522 523 /* 524 * Take care not to create a writable alias for the 525 * read-only text and rodata sections of the kernel image. 526 * So temporarily mark them as NOMAP to skip mappings in 527 * the following for-loop 528 */ 529 memblock_mark_nomap(kernel_start, kernel_end - kernel_start); 530 531 /* map all the memory banks */ 532 for_each_mem_range(i, &start, &end) { 533 if (start >= end) 534 break; 535 /* 536 * The linear map must allow allocation tags reading/writing 537 * if MTE is present. Otherwise, it has the same attributes as 538 * PAGE_KERNEL. 539 */ 540 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL), 541 flags); 542 } 543 544 /* 545 * Map the linear alias of the [_stext, __init_begin) interval 546 * as non-executable now, and remove the write permission in 547 * mark_linear_text_alias_ro() below (which will be called after 548 * alternative patching has completed). This makes the contents 549 * of the region accessible to subsystems such as hibernate, 550 * but protects it from inadvertent modification or execution. 551 * Note that contiguous mappings cannot be remapped in this way, 552 * so we should avoid them here. 553 */ 554 __map_memblock(pgdp, kernel_start, kernel_end, 555 PAGE_KERNEL, NO_CONT_MAPPINGS); 556 memblock_clear_nomap(kernel_start, kernel_end - kernel_start); 557 } 558 559 void mark_rodata_ro(void) 560 { 561 unsigned long section_size; 562 563 /* 564 * mark .rodata as read only. Use __init_begin rather than __end_rodata 565 * to cover NOTES and EXCEPTION_TABLE. 566 */ 567 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata; 568 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata, 569 section_size, PAGE_KERNEL_RO); 570 571 debug_checkwx(); 572 } 573 574 static void __init map_kernel_segment(pgd_t *pgdp, void *va_start, void *va_end, 575 pgprot_t prot, struct vm_struct *vma, 576 int flags, unsigned long vm_flags) 577 { 578 phys_addr_t pa_start = __pa_symbol(va_start); 579 unsigned long size = va_end - va_start; 580 581 BUG_ON(!PAGE_ALIGNED(pa_start)); 582 BUG_ON(!PAGE_ALIGNED(size)); 583 584 __create_pgd_mapping(pgdp, pa_start, (unsigned long)va_start, size, prot, 585 early_pgtable_alloc, flags); 586 587 if (!(vm_flags & VM_NO_GUARD)) 588 size += PAGE_SIZE; 589 590 vma->addr = va_start; 591 vma->phys_addr = pa_start; 592 vma->size = size; 593 vma->flags = VM_MAP | vm_flags; 594 vma->caller = __builtin_return_address(0); 595 596 vm_area_add_early(vma); 597 } 598 599 static int __init parse_rodata(char *arg) 600 { 601 int ret = strtobool(arg, &rodata_enabled); 602 if (!ret) { 603 rodata_full = false; 604 return 0; 605 } 606 607 /* permit 'full' in addition to boolean options */ 608 if (strcmp(arg, "full")) 609 return -EINVAL; 610 611 rodata_enabled = true; 612 rodata_full = true; 613 return 0; 614 } 615 early_param("rodata", parse_rodata); 616 617 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0 618 static int __init map_entry_trampoline(void) 619 { 620 pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 621 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start); 622 623 /* The trampoline is always mapped and can therefore be global */ 624 pgprot_val(prot) &= ~PTE_NG; 625 626 /* Map only the text into the trampoline page table */ 627 memset(tramp_pg_dir, 0, PGD_SIZE); 628 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, PAGE_SIZE, 629 prot, __pgd_pgtable_alloc, 0); 630 631 /* Map both the text and data into the kernel page table */ 632 __set_fixmap(FIX_ENTRY_TRAMP_TEXT, pa_start, prot); 633 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 634 extern char __entry_tramp_data_start[]; 635 636 __set_fixmap(FIX_ENTRY_TRAMP_DATA, 637 __pa_symbol(__entry_tramp_data_start), 638 PAGE_KERNEL_RO); 639 } 640 641 return 0; 642 } 643 core_initcall(map_entry_trampoline); 644 #endif 645 646 /* 647 * Open coded check for BTI, only for use to determine configuration 648 * for early mappings for before the cpufeature code has run. 649 */ 650 static bool arm64_early_this_cpu_has_bti(void) 651 { 652 u64 pfr1; 653 654 if (!IS_ENABLED(CONFIG_ARM64_BTI_KERNEL)) 655 return false; 656 657 pfr1 = __read_sysreg_by_encoding(SYS_ID_AA64PFR1_EL1); 658 return cpuid_feature_extract_unsigned_field(pfr1, 659 ID_AA64PFR1_BT_SHIFT); 660 } 661 662 /* 663 * Create fine-grained mappings for the kernel. 664 */ 665 static void __init map_kernel(pgd_t *pgdp) 666 { 667 static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext, 668 vmlinux_initdata, vmlinux_data; 669 670 /* 671 * External debuggers may need to write directly to the text 672 * mapping to install SW breakpoints. Allow this (only) when 673 * explicitly requested with rodata=off. 674 */ 675 pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 676 677 /* 678 * If we have a CPU that supports BTI and a kernel built for 679 * BTI then mark the kernel executable text as guarded pages 680 * now so we don't have to rewrite the page tables later. 681 */ 682 if (arm64_early_this_cpu_has_bti()) 683 text_prot = __pgprot_modify(text_prot, PTE_GP, PTE_GP); 684 685 /* 686 * Only rodata will be remapped with different permissions later on, 687 * all other segments are allowed to use contiguous mappings. 688 */ 689 map_kernel_segment(pgdp, _stext, _etext, text_prot, &vmlinux_text, 0, 690 VM_NO_GUARD); 691 map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL, 692 &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD); 693 map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot, 694 &vmlinux_inittext, 0, VM_NO_GUARD); 695 map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL, 696 &vmlinux_initdata, 0, VM_NO_GUARD); 697 map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0); 698 699 if (!READ_ONCE(pgd_val(*pgd_offset_pgd(pgdp, FIXADDR_START)))) { 700 /* 701 * The fixmap falls in a separate pgd to the kernel, and doesn't 702 * live in the carveout for the swapper_pg_dir. We can simply 703 * re-use the existing dir for the fixmap. 704 */ 705 set_pgd(pgd_offset_pgd(pgdp, FIXADDR_START), 706 READ_ONCE(*pgd_offset_k(FIXADDR_START))); 707 } else if (CONFIG_PGTABLE_LEVELS > 3) { 708 pgd_t *bm_pgdp; 709 p4d_t *bm_p4dp; 710 pud_t *bm_pudp; 711 /* 712 * The fixmap shares its top level pgd entry with the kernel 713 * mapping. This can really only occur when we are running 714 * with 16k/4 levels, so we can simply reuse the pud level 715 * entry instead. 716 */ 717 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES)); 718 bm_pgdp = pgd_offset_pgd(pgdp, FIXADDR_START); 719 bm_p4dp = p4d_offset(bm_pgdp, FIXADDR_START); 720 bm_pudp = pud_set_fixmap_offset(bm_p4dp, FIXADDR_START); 721 pud_populate(&init_mm, bm_pudp, lm_alias(bm_pmd)); 722 pud_clear_fixmap(); 723 } else { 724 BUG(); 725 } 726 727 kasan_copy_shadow(pgdp); 728 } 729 730 void __init paging_init(void) 731 { 732 pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir)); 733 734 map_kernel(pgdp); 735 map_mem(pgdp); 736 737 pgd_clear_fixmap(); 738 739 cpu_replace_ttbr1(lm_alias(swapper_pg_dir)); 740 init_mm.pgd = swapper_pg_dir; 741 742 memblock_phys_free(__pa_symbol(init_pg_dir), 743 __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir)); 744 745 memblock_allow_resize(); 746 } 747 748 /* 749 * Check whether a kernel address is valid (derived from arch/x86/). 750 */ 751 int kern_addr_valid(unsigned long addr) 752 { 753 pgd_t *pgdp; 754 p4d_t *p4dp; 755 pud_t *pudp, pud; 756 pmd_t *pmdp, pmd; 757 pte_t *ptep, pte; 758 759 addr = arch_kasan_reset_tag(addr); 760 if ((((long)addr) >> VA_BITS) != -1UL) 761 return 0; 762 763 pgdp = pgd_offset_k(addr); 764 if (pgd_none(READ_ONCE(*pgdp))) 765 return 0; 766 767 p4dp = p4d_offset(pgdp, addr); 768 if (p4d_none(READ_ONCE(*p4dp))) 769 return 0; 770 771 pudp = pud_offset(p4dp, addr); 772 pud = READ_ONCE(*pudp); 773 if (pud_none(pud)) 774 return 0; 775 776 if (pud_sect(pud)) 777 return pfn_valid(pud_pfn(pud)); 778 779 pmdp = pmd_offset(pudp, addr); 780 pmd = READ_ONCE(*pmdp); 781 if (pmd_none(pmd)) 782 return 0; 783 784 if (pmd_sect(pmd)) 785 return pfn_valid(pmd_pfn(pmd)); 786 787 ptep = pte_offset_kernel(pmdp, addr); 788 pte = READ_ONCE(*ptep); 789 if (pte_none(pte)) 790 return 0; 791 792 return pfn_valid(pte_pfn(pte)); 793 } 794 795 #ifdef CONFIG_MEMORY_HOTPLUG 796 static void free_hotplug_page_range(struct page *page, size_t size, 797 struct vmem_altmap *altmap) 798 { 799 if (altmap) { 800 vmem_altmap_free(altmap, size >> PAGE_SHIFT); 801 } else { 802 WARN_ON(PageReserved(page)); 803 free_pages((unsigned long)page_address(page), get_order(size)); 804 } 805 } 806 807 static void free_hotplug_pgtable_page(struct page *page) 808 { 809 free_hotplug_page_range(page, PAGE_SIZE, NULL); 810 } 811 812 static bool pgtable_range_aligned(unsigned long start, unsigned long end, 813 unsigned long floor, unsigned long ceiling, 814 unsigned long mask) 815 { 816 start &= mask; 817 if (start < floor) 818 return false; 819 820 if (ceiling) { 821 ceiling &= mask; 822 if (!ceiling) 823 return false; 824 } 825 826 if (end - 1 > ceiling - 1) 827 return false; 828 return true; 829 } 830 831 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr, 832 unsigned long end, bool free_mapped, 833 struct vmem_altmap *altmap) 834 { 835 pte_t *ptep, pte; 836 837 do { 838 ptep = pte_offset_kernel(pmdp, addr); 839 pte = READ_ONCE(*ptep); 840 if (pte_none(pte)) 841 continue; 842 843 WARN_ON(!pte_present(pte)); 844 pte_clear(&init_mm, addr, ptep); 845 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 846 if (free_mapped) 847 free_hotplug_page_range(pte_page(pte), 848 PAGE_SIZE, altmap); 849 } while (addr += PAGE_SIZE, addr < end); 850 } 851 852 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr, 853 unsigned long end, bool free_mapped, 854 struct vmem_altmap *altmap) 855 { 856 unsigned long next; 857 pmd_t *pmdp, pmd; 858 859 do { 860 next = pmd_addr_end(addr, end); 861 pmdp = pmd_offset(pudp, addr); 862 pmd = READ_ONCE(*pmdp); 863 if (pmd_none(pmd)) 864 continue; 865 866 WARN_ON(!pmd_present(pmd)); 867 if (pmd_sect(pmd)) { 868 pmd_clear(pmdp); 869 870 /* 871 * One TLBI should be sufficient here as the PMD_SIZE 872 * range is mapped with a single block entry. 873 */ 874 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 875 if (free_mapped) 876 free_hotplug_page_range(pmd_page(pmd), 877 PMD_SIZE, altmap); 878 continue; 879 } 880 WARN_ON(!pmd_table(pmd)); 881 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap); 882 } while (addr = next, addr < end); 883 } 884 885 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr, 886 unsigned long end, bool free_mapped, 887 struct vmem_altmap *altmap) 888 { 889 unsigned long next; 890 pud_t *pudp, pud; 891 892 do { 893 next = pud_addr_end(addr, end); 894 pudp = pud_offset(p4dp, addr); 895 pud = READ_ONCE(*pudp); 896 if (pud_none(pud)) 897 continue; 898 899 WARN_ON(!pud_present(pud)); 900 if (pud_sect(pud)) { 901 pud_clear(pudp); 902 903 /* 904 * One TLBI should be sufficient here as the PUD_SIZE 905 * range is mapped with a single block entry. 906 */ 907 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 908 if (free_mapped) 909 free_hotplug_page_range(pud_page(pud), 910 PUD_SIZE, altmap); 911 continue; 912 } 913 WARN_ON(!pud_table(pud)); 914 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap); 915 } while (addr = next, addr < end); 916 } 917 918 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr, 919 unsigned long end, bool free_mapped, 920 struct vmem_altmap *altmap) 921 { 922 unsigned long next; 923 p4d_t *p4dp, p4d; 924 925 do { 926 next = p4d_addr_end(addr, end); 927 p4dp = p4d_offset(pgdp, addr); 928 p4d = READ_ONCE(*p4dp); 929 if (p4d_none(p4d)) 930 continue; 931 932 WARN_ON(!p4d_present(p4d)); 933 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap); 934 } while (addr = next, addr < end); 935 } 936 937 static void unmap_hotplug_range(unsigned long addr, unsigned long end, 938 bool free_mapped, struct vmem_altmap *altmap) 939 { 940 unsigned long next; 941 pgd_t *pgdp, pgd; 942 943 /* 944 * altmap can only be used as vmemmap mapping backing memory. 945 * In case the backing memory itself is not being freed, then 946 * altmap is irrelevant. Warn about this inconsistency when 947 * encountered. 948 */ 949 WARN_ON(!free_mapped && altmap); 950 951 do { 952 next = pgd_addr_end(addr, end); 953 pgdp = pgd_offset_k(addr); 954 pgd = READ_ONCE(*pgdp); 955 if (pgd_none(pgd)) 956 continue; 957 958 WARN_ON(!pgd_present(pgd)); 959 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap); 960 } while (addr = next, addr < end); 961 } 962 963 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr, 964 unsigned long end, unsigned long floor, 965 unsigned long ceiling) 966 { 967 pte_t *ptep, pte; 968 unsigned long i, start = addr; 969 970 do { 971 ptep = pte_offset_kernel(pmdp, addr); 972 pte = READ_ONCE(*ptep); 973 974 /* 975 * This is just a sanity check here which verifies that 976 * pte clearing has been done by earlier unmap loops. 977 */ 978 WARN_ON(!pte_none(pte)); 979 } while (addr += PAGE_SIZE, addr < end); 980 981 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK)) 982 return; 983 984 /* 985 * Check whether we can free the pte page if the rest of the 986 * entries are empty. Overlap with other regions have been 987 * handled by the floor/ceiling check. 988 */ 989 ptep = pte_offset_kernel(pmdp, 0UL); 990 for (i = 0; i < PTRS_PER_PTE; i++) { 991 if (!pte_none(READ_ONCE(ptep[i]))) 992 return; 993 } 994 995 pmd_clear(pmdp); 996 __flush_tlb_kernel_pgtable(start); 997 free_hotplug_pgtable_page(virt_to_page(ptep)); 998 } 999 1000 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr, 1001 unsigned long end, unsigned long floor, 1002 unsigned long ceiling) 1003 { 1004 pmd_t *pmdp, pmd; 1005 unsigned long i, next, start = addr; 1006 1007 do { 1008 next = pmd_addr_end(addr, end); 1009 pmdp = pmd_offset(pudp, addr); 1010 pmd = READ_ONCE(*pmdp); 1011 if (pmd_none(pmd)) 1012 continue; 1013 1014 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd)); 1015 free_empty_pte_table(pmdp, addr, next, floor, ceiling); 1016 } while (addr = next, addr < end); 1017 1018 if (CONFIG_PGTABLE_LEVELS <= 2) 1019 return; 1020 1021 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK)) 1022 return; 1023 1024 /* 1025 * Check whether we can free the pmd page if the rest of the 1026 * entries are empty. Overlap with other regions have been 1027 * handled by the floor/ceiling check. 1028 */ 1029 pmdp = pmd_offset(pudp, 0UL); 1030 for (i = 0; i < PTRS_PER_PMD; i++) { 1031 if (!pmd_none(READ_ONCE(pmdp[i]))) 1032 return; 1033 } 1034 1035 pud_clear(pudp); 1036 __flush_tlb_kernel_pgtable(start); 1037 free_hotplug_pgtable_page(virt_to_page(pmdp)); 1038 } 1039 1040 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr, 1041 unsigned long end, unsigned long floor, 1042 unsigned long ceiling) 1043 { 1044 pud_t *pudp, pud; 1045 unsigned long i, next, start = addr; 1046 1047 do { 1048 next = pud_addr_end(addr, end); 1049 pudp = pud_offset(p4dp, addr); 1050 pud = READ_ONCE(*pudp); 1051 if (pud_none(pud)) 1052 continue; 1053 1054 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud)); 1055 free_empty_pmd_table(pudp, addr, next, floor, ceiling); 1056 } while (addr = next, addr < end); 1057 1058 if (CONFIG_PGTABLE_LEVELS <= 3) 1059 return; 1060 1061 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK)) 1062 return; 1063 1064 /* 1065 * Check whether we can free the pud page if the rest of the 1066 * entries are empty. Overlap with other regions have been 1067 * handled by the floor/ceiling check. 1068 */ 1069 pudp = pud_offset(p4dp, 0UL); 1070 for (i = 0; i < PTRS_PER_PUD; i++) { 1071 if (!pud_none(READ_ONCE(pudp[i]))) 1072 return; 1073 } 1074 1075 p4d_clear(p4dp); 1076 __flush_tlb_kernel_pgtable(start); 1077 free_hotplug_pgtable_page(virt_to_page(pudp)); 1078 } 1079 1080 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr, 1081 unsigned long end, unsigned long floor, 1082 unsigned long ceiling) 1083 { 1084 unsigned long next; 1085 p4d_t *p4dp, p4d; 1086 1087 do { 1088 next = p4d_addr_end(addr, end); 1089 p4dp = p4d_offset(pgdp, addr); 1090 p4d = READ_ONCE(*p4dp); 1091 if (p4d_none(p4d)) 1092 continue; 1093 1094 WARN_ON(!p4d_present(p4d)); 1095 free_empty_pud_table(p4dp, addr, next, floor, ceiling); 1096 } while (addr = next, addr < end); 1097 } 1098 1099 static void free_empty_tables(unsigned long addr, unsigned long end, 1100 unsigned long floor, unsigned long ceiling) 1101 { 1102 unsigned long next; 1103 pgd_t *pgdp, pgd; 1104 1105 do { 1106 next = pgd_addr_end(addr, end); 1107 pgdp = pgd_offset_k(addr); 1108 pgd = READ_ONCE(*pgdp); 1109 if (pgd_none(pgd)) 1110 continue; 1111 1112 WARN_ON(!pgd_present(pgd)); 1113 free_empty_p4d_table(pgdp, addr, next, floor, ceiling); 1114 } while (addr = next, addr < end); 1115 } 1116 #endif 1117 1118 #if !ARM64_KERNEL_USES_PMD_MAPS 1119 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1120 struct vmem_altmap *altmap) 1121 { 1122 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1123 return vmemmap_populate_basepages(start, end, node, altmap); 1124 } 1125 #else /* !ARM64_KERNEL_USES_PMD_MAPS */ 1126 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1127 struct vmem_altmap *altmap) 1128 { 1129 unsigned long addr = start; 1130 unsigned long next; 1131 pgd_t *pgdp; 1132 p4d_t *p4dp; 1133 pud_t *pudp; 1134 pmd_t *pmdp; 1135 1136 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1137 do { 1138 next = pmd_addr_end(addr, end); 1139 1140 pgdp = vmemmap_pgd_populate(addr, node); 1141 if (!pgdp) 1142 return -ENOMEM; 1143 1144 p4dp = vmemmap_p4d_populate(pgdp, addr, node); 1145 if (!p4dp) 1146 return -ENOMEM; 1147 1148 pudp = vmemmap_pud_populate(p4dp, addr, node); 1149 if (!pudp) 1150 return -ENOMEM; 1151 1152 pmdp = pmd_offset(pudp, addr); 1153 if (pmd_none(READ_ONCE(*pmdp))) { 1154 void *p = NULL; 1155 1156 p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap); 1157 if (!p) { 1158 if (vmemmap_populate_basepages(addr, next, node, altmap)) 1159 return -ENOMEM; 1160 continue; 1161 } 1162 1163 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL)); 1164 } else 1165 vmemmap_verify((pte_t *)pmdp, node, addr, next); 1166 } while (addr = next, addr != end); 1167 1168 return 0; 1169 } 1170 #endif /* !ARM64_KERNEL_USES_PMD_MAPS */ 1171 1172 #ifdef CONFIG_MEMORY_HOTPLUG 1173 void vmemmap_free(unsigned long start, unsigned long end, 1174 struct vmem_altmap *altmap) 1175 { 1176 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1177 1178 unmap_hotplug_range(start, end, true, altmap); 1179 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END); 1180 } 1181 #endif /* CONFIG_MEMORY_HOTPLUG */ 1182 1183 static inline pud_t *fixmap_pud(unsigned long addr) 1184 { 1185 pgd_t *pgdp = pgd_offset_k(addr); 1186 p4d_t *p4dp = p4d_offset(pgdp, addr); 1187 p4d_t p4d = READ_ONCE(*p4dp); 1188 1189 BUG_ON(p4d_none(p4d) || p4d_bad(p4d)); 1190 1191 return pud_offset_kimg(p4dp, addr); 1192 } 1193 1194 static inline pmd_t *fixmap_pmd(unsigned long addr) 1195 { 1196 pud_t *pudp = fixmap_pud(addr); 1197 pud_t pud = READ_ONCE(*pudp); 1198 1199 BUG_ON(pud_none(pud) || pud_bad(pud)); 1200 1201 return pmd_offset_kimg(pudp, addr); 1202 } 1203 1204 static inline pte_t *fixmap_pte(unsigned long addr) 1205 { 1206 return &bm_pte[pte_index(addr)]; 1207 } 1208 1209 /* 1210 * The p*d_populate functions call virt_to_phys implicitly so they can't be used 1211 * directly on kernel symbols (bm_p*d). This function is called too early to use 1212 * lm_alias so __p*d_populate functions must be used to populate with the 1213 * physical address from __pa_symbol. 1214 */ 1215 void __init early_fixmap_init(void) 1216 { 1217 pgd_t *pgdp; 1218 p4d_t *p4dp, p4d; 1219 pud_t *pudp; 1220 pmd_t *pmdp; 1221 unsigned long addr = FIXADDR_START; 1222 1223 pgdp = pgd_offset_k(addr); 1224 p4dp = p4d_offset(pgdp, addr); 1225 p4d = READ_ONCE(*p4dp); 1226 if (CONFIG_PGTABLE_LEVELS > 3 && 1227 !(p4d_none(p4d) || p4d_page_paddr(p4d) == __pa_symbol(bm_pud))) { 1228 /* 1229 * We only end up here if the kernel mapping and the fixmap 1230 * share the top level pgd entry, which should only happen on 1231 * 16k/4 levels configurations. 1232 */ 1233 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES)); 1234 pudp = pud_offset_kimg(p4dp, addr); 1235 } else { 1236 if (p4d_none(p4d)) 1237 __p4d_populate(p4dp, __pa_symbol(bm_pud), P4D_TYPE_TABLE); 1238 pudp = fixmap_pud(addr); 1239 } 1240 if (pud_none(READ_ONCE(*pudp))) 1241 __pud_populate(pudp, __pa_symbol(bm_pmd), PUD_TYPE_TABLE); 1242 pmdp = fixmap_pmd(addr); 1243 __pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE); 1244 1245 /* 1246 * The boot-ioremap range spans multiple pmds, for which 1247 * we are not prepared: 1248 */ 1249 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT) 1250 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT)); 1251 1252 if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN))) 1253 || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) { 1254 WARN_ON(1); 1255 pr_warn("pmdp %p != %p, %p\n", 1256 pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)), 1257 fixmap_pmd(fix_to_virt(FIX_BTMAP_END))); 1258 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n", 1259 fix_to_virt(FIX_BTMAP_BEGIN)); 1260 pr_warn("fix_to_virt(FIX_BTMAP_END): %08lx\n", 1261 fix_to_virt(FIX_BTMAP_END)); 1262 1263 pr_warn("FIX_BTMAP_END: %d\n", FIX_BTMAP_END); 1264 pr_warn("FIX_BTMAP_BEGIN: %d\n", FIX_BTMAP_BEGIN); 1265 } 1266 } 1267 1268 /* 1269 * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we 1270 * ever need to use IPIs for TLB broadcasting, then we're in trouble here. 1271 */ 1272 void __set_fixmap(enum fixed_addresses idx, 1273 phys_addr_t phys, pgprot_t flags) 1274 { 1275 unsigned long addr = __fix_to_virt(idx); 1276 pte_t *ptep; 1277 1278 BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses); 1279 1280 ptep = fixmap_pte(addr); 1281 1282 if (pgprot_val(flags)) { 1283 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags)); 1284 } else { 1285 pte_clear(&init_mm, addr, ptep); 1286 flush_tlb_kernel_range(addr, addr+PAGE_SIZE); 1287 } 1288 } 1289 1290 void *__init fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot) 1291 { 1292 const u64 dt_virt_base = __fix_to_virt(FIX_FDT); 1293 int offset; 1294 void *dt_virt; 1295 1296 /* 1297 * Check whether the physical FDT address is set and meets the minimum 1298 * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be 1299 * at least 8 bytes so that we can always access the magic and size 1300 * fields of the FDT header after mapping the first chunk, double check 1301 * here if that is indeed the case. 1302 */ 1303 BUILD_BUG_ON(MIN_FDT_ALIGN < 8); 1304 if (!dt_phys || dt_phys % MIN_FDT_ALIGN) 1305 return NULL; 1306 1307 /* 1308 * Make sure that the FDT region can be mapped without the need to 1309 * allocate additional translation table pages, so that it is safe 1310 * to call create_mapping_noalloc() this early. 1311 * 1312 * On 64k pages, the FDT will be mapped using PTEs, so we need to 1313 * be in the same PMD as the rest of the fixmap. 1314 * On 4k pages, we'll use section mappings for the FDT so we only 1315 * have to be in the same PUD. 1316 */ 1317 BUILD_BUG_ON(dt_virt_base % SZ_2M); 1318 1319 BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT != 1320 __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT); 1321 1322 offset = dt_phys % SWAPPER_BLOCK_SIZE; 1323 dt_virt = (void *)dt_virt_base + offset; 1324 1325 /* map the first chunk so we can read the size from the header */ 1326 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), 1327 dt_virt_base, SWAPPER_BLOCK_SIZE, prot); 1328 1329 if (fdt_magic(dt_virt) != FDT_MAGIC) 1330 return NULL; 1331 1332 *size = fdt_totalsize(dt_virt); 1333 if (*size > MAX_FDT_SIZE) 1334 return NULL; 1335 1336 if (offset + *size > SWAPPER_BLOCK_SIZE) 1337 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base, 1338 round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot); 1339 1340 return dt_virt; 1341 } 1342 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 pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot) 1358 { 1359 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot)); 1360 1361 /* Only allow permission changes for now */ 1362 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)), 1363 pmd_val(new_pmd))) 1364 return 0; 1365 1366 VM_BUG_ON(phys & ~PMD_MASK); 1367 set_pmd(pmdp, new_pmd); 1368 return 1; 1369 } 1370 1371 int pud_clear_huge(pud_t *pudp) 1372 { 1373 if (!pud_sect(READ_ONCE(*pudp))) 1374 return 0; 1375 pud_clear(pudp); 1376 return 1; 1377 } 1378 1379 int pmd_clear_huge(pmd_t *pmdp) 1380 { 1381 if (!pmd_sect(READ_ONCE(*pmdp))) 1382 return 0; 1383 pmd_clear(pmdp); 1384 return 1; 1385 } 1386 1387 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr) 1388 { 1389 pte_t *table; 1390 pmd_t pmd; 1391 1392 pmd = READ_ONCE(*pmdp); 1393 1394 if (!pmd_table(pmd)) { 1395 VM_WARN_ON(1); 1396 return 1; 1397 } 1398 1399 table = pte_offset_kernel(pmdp, addr); 1400 pmd_clear(pmdp); 1401 __flush_tlb_kernel_pgtable(addr); 1402 pte_free_kernel(NULL, table); 1403 return 1; 1404 } 1405 1406 int pud_free_pmd_page(pud_t *pudp, unsigned long addr) 1407 { 1408 pmd_t *table; 1409 pmd_t *pmdp; 1410 pud_t pud; 1411 unsigned long next, end; 1412 1413 pud = READ_ONCE(*pudp); 1414 1415 if (!pud_table(pud)) { 1416 VM_WARN_ON(1); 1417 return 1; 1418 } 1419 1420 table = pmd_offset(pudp, addr); 1421 pmdp = table; 1422 next = addr; 1423 end = addr + PUD_SIZE; 1424 do { 1425 pmd_free_pte_page(pmdp, next); 1426 } while (pmdp++, next += PMD_SIZE, next != end); 1427 1428 pud_clear(pudp); 1429 __flush_tlb_kernel_pgtable(addr); 1430 pmd_free(NULL, table); 1431 return 1; 1432 } 1433 1434 #ifdef CONFIG_MEMORY_HOTPLUG 1435 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size) 1436 { 1437 unsigned long end = start + size; 1438 1439 WARN_ON(pgdir != init_mm.pgd); 1440 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END)); 1441 1442 unmap_hotplug_range(start, end, false, NULL); 1443 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END); 1444 } 1445 1446 struct range arch_get_mappable_range(void) 1447 { 1448 struct range mhp_range; 1449 u64 start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual)); 1450 u64 end_linear_pa = __pa(PAGE_END - 1); 1451 1452 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 1453 /* 1454 * Check for a wrap, it is possible because of randomized linear 1455 * mapping the start physical address is actually bigger than 1456 * the end physical address. In this case set start to zero 1457 * because [0, end_linear_pa] range must still be able to cover 1458 * all addressable physical addresses. 1459 */ 1460 if (start_linear_pa > end_linear_pa) 1461 start_linear_pa = 0; 1462 } 1463 1464 WARN_ON(start_linear_pa > end_linear_pa); 1465 1466 /* 1467 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)] 1468 * accommodating both its ends but excluding PAGE_END. Max physical 1469 * range which can be mapped inside this linear mapping range, must 1470 * also be derived from its end points. 1471 */ 1472 mhp_range.start = start_linear_pa; 1473 mhp_range.end = end_linear_pa; 1474 1475 return mhp_range; 1476 } 1477 1478 int arch_add_memory(int nid, u64 start, u64 size, 1479 struct mhp_params *params) 1480 { 1481 int ret, flags = NO_EXEC_MAPPINGS; 1482 1483 VM_BUG_ON(!mhp_range_allowed(start, size, true)); 1484 1485 /* 1486 * KFENCE requires linear map to be mapped at page granularity, so that 1487 * it is possible to protect/unprotect single pages in the KFENCE pool. 1488 */ 1489 if (can_set_direct_map() || IS_ENABLED(CONFIG_KFENCE)) 1490 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1491 1492 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1493 size, params->pgprot, __pgd_pgtable_alloc, 1494 flags); 1495 1496 memblock_clear_nomap(start, size); 1497 1498 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, 1499 params); 1500 if (ret) 1501 __remove_pgd_mapping(swapper_pg_dir, 1502 __phys_to_virt(start), size); 1503 else { 1504 max_pfn = PFN_UP(start + size); 1505 max_low_pfn = max_pfn; 1506 } 1507 1508 return ret; 1509 } 1510 1511 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap) 1512 { 1513 unsigned long start_pfn = start >> PAGE_SHIFT; 1514 unsigned long nr_pages = size >> PAGE_SHIFT; 1515 1516 __remove_pages(start_pfn, nr_pages, altmap); 1517 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size); 1518 } 1519 1520 /* 1521 * This memory hotplug notifier helps prevent boot memory from being 1522 * inadvertently removed as it blocks pfn range offlining process in 1523 * __offline_pages(). Hence this prevents both offlining as well as 1524 * removal process for boot memory which is initially always online. 1525 * In future if and when boot memory could be removed, this notifier 1526 * should be dropped and free_hotplug_page_range() should handle any 1527 * reserved pages allocated during boot. 1528 */ 1529 static int prevent_bootmem_remove_notifier(struct notifier_block *nb, 1530 unsigned long action, void *data) 1531 { 1532 struct mem_section *ms; 1533 struct memory_notify *arg = data; 1534 unsigned long end_pfn = arg->start_pfn + arg->nr_pages; 1535 unsigned long pfn = arg->start_pfn; 1536 1537 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE)) 1538 return NOTIFY_OK; 1539 1540 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 1541 unsigned long start = PFN_PHYS(pfn); 1542 unsigned long end = start + (1UL << PA_SECTION_SHIFT); 1543 1544 ms = __pfn_to_section(pfn); 1545 if (!early_section(ms)) 1546 continue; 1547 1548 if (action == MEM_GOING_OFFLINE) { 1549 /* 1550 * Boot memory removal is not supported. Prevent 1551 * it via blocking any attempted offline request 1552 * for the boot memory and just report it. 1553 */ 1554 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end); 1555 return NOTIFY_BAD; 1556 } else if (action == MEM_OFFLINE) { 1557 /* 1558 * This should have never happened. Boot memory 1559 * offlining should have been prevented by this 1560 * very notifier. Probably some memory removal 1561 * procedure might have changed which would then 1562 * require further debug. 1563 */ 1564 pr_err("Boot memory [%lx %lx] offlined\n", start, end); 1565 1566 /* 1567 * Core memory hotplug does not process a return 1568 * code from the notifier for MEM_OFFLINE events. 1569 * The error condition has been reported. Return 1570 * from here as if ignored. 1571 */ 1572 return NOTIFY_DONE; 1573 } 1574 } 1575 return NOTIFY_OK; 1576 } 1577 1578 static struct notifier_block prevent_bootmem_remove_nb = { 1579 .notifier_call = prevent_bootmem_remove_notifier, 1580 }; 1581 1582 /* 1583 * This ensures that boot memory sections on the platform are online 1584 * from early boot. Memory sections could not be prevented from being 1585 * offlined, unless for some reason they are not online to begin with. 1586 * This helps validate the basic assumption on which the above memory 1587 * event notifier works to prevent boot memory section offlining and 1588 * its possible removal. 1589 */ 1590 static void validate_bootmem_online(void) 1591 { 1592 phys_addr_t start, end, addr; 1593 struct mem_section *ms; 1594 u64 i; 1595 1596 /* 1597 * Scanning across all memblock might be expensive 1598 * on some big memory systems. Hence enable this 1599 * validation only with DEBUG_VM. 1600 */ 1601 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 1602 return; 1603 1604 for_each_mem_range(i, &start, &end) { 1605 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) { 1606 ms = __pfn_to_section(PHYS_PFN(addr)); 1607 1608 /* 1609 * All memory ranges in the system at this point 1610 * should have been marked as early sections. 1611 */ 1612 WARN_ON(!early_section(ms)); 1613 1614 /* 1615 * Memory notifier mechanism here to prevent boot 1616 * memory offlining depends on the fact that each 1617 * early section memory on the system is initially 1618 * online. Otherwise a given memory section which 1619 * is already offline will be overlooked and can 1620 * be removed completely. Call out such sections. 1621 */ 1622 if (!online_section(ms)) 1623 pr_err("Boot memory [%llx %llx] is offline, can be removed\n", 1624 addr, addr + (1UL << PA_SECTION_SHIFT)); 1625 } 1626 } 1627 } 1628 1629 static int __init prevent_bootmem_remove_init(void) 1630 { 1631 int ret = 0; 1632 1633 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) 1634 return ret; 1635 1636 validate_bootmem_online(); 1637 ret = register_memory_notifier(&prevent_bootmem_remove_nb); 1638 if (ret) 1639 pr_err("%s: Notifier registration failed %d\n", __func__, ret); 1640 1641 return ret; 1642 } 1643 early_initcall(prevent_bootmem_remove_init); 1644 #endif 1645