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 int i; 621 622 pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 623 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start); 624 625 /* The trampoline is always mapped and can therefore be global */ 626 pgprot_val(prot) &= ~PTE_NG; 627 628 /* Map only the text into the trampoline page table */ 629 memset(tramp_pg_dir, 0, PGD_SIZE); 630 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, 631 entry_tramp_text_size(), prot, 632 __pgd_pgtable_alloc, NO_BLOCK_MAPPINGS); 633 634 /* Map both the text and data into the kernel page table */ 635 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++) 636 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 637 pa_start + i * PAGE_SIZE, prot); 638 639 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 640 extern char __entry_tramp_data_start[]; 641 642 __set_fixmap(FIX_ENTRY_TRAMP_DATA, 643 __pa_symbol(__entry_tramp_data_start), 644 PAGE_KERNEL_RO); 645 } 646 647 return 0; 648 } 649 core_initcall(map_entry_trampoline); 650 #endif 651 652 /* 653 * Open coded check for BTI, only for use to determine configuration 654 * for early mappings for before the cpufeature code has run. 655 */ 656 static bool arm64_early_this_cpu_has_bti(void) 657 { 658 u64 pfr1; 659 660 if (!IS_ENABLED(CONFIG_ARM64_BTI_KERNEL)) 661 return false; 662 663 pfr1 = __read_sysreg_by_encoding(SYS_ID_AA64PFR1_EL1); 664 return cpuid_feature_extract_unsigned_field(pfr1, 665 ID_AA64PFR1_BT_SHIFT); 666 } 667 668 /* 669 * Create fine-grained mappings for the kernel. 670 */ 671 static void __init map_kernel(pgd_t *pgdp) 672 { 673 static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext, 674 vmlinux_initdata, vmlinux_data; 675 676 /* 677 * External debuggers may need to write directly to the text 678 * mapping to install SW breakpoints. Allow this (only) when 679 * explicitly requested with rodata=off. 680 */ 681 pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 682 683 /* 684 * If we have a CPU that supports BTI and a kernel built for 685 * BTI then mark the kernel executable text as guarded pages 686 * now so we don't have to rewrite the page tables later. 687 */ 688 if (arm64_early_this_cpu_has_bti()) 689 text_prot = __pgprot_modify(text_prot, PTE_GP, PTE_GP); 690 691 /* 692 * Only rodata will be remapped with different permissions later on, 693 * all other segments are allowed to use contiguous mappings. 694 */ 695 map_kernel_segment(pgdp, _stext, _etext, text_prot, &vmlinux_text, 0, 696 VM_NO_GUARD); 697 map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL, 698 &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD); 699 map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot, 700 &vmlinux_inittext, 0, VM_NO_GUARD); 701 map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL, 702 &vmlinux_initdata, 0, VM_NO_GUARD); 703 map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0); 704 705 if (!READ_ONCE(pgd_val(*pgd_offset_pgd(pgdp, FIXADDR_START)))) { 706 /* 707 * The fixmap falls in a separate pgd to the kernel, and doesn't 708 * live in the carveout for the swapper_pg_dir. We can simply 709 * re-use the existing dir for the fixmap. 710 */ 711 set_pgd(pgd_offset_pgd(pgdp, FIXADDR_START), 712 READ_ONCE(*pgd_offset_k(FIXADDR_START))); 713 } else if (CONFIG_PGTABLE_LEVELS > 3) { 714 pgd_t *bm_pgdp; 715 p4d_t *bm_p4dp; 716 pud_t *bm_pudp; 717 /* 718 * The fixmap shares its top level pgd entry with the kernel 719 * mapping. This can really only occur when we are running 720 * with 16k/4 levels, so we can simply reuse the pud level 721 * entry instead. 722 */ 723 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES)); 724 bm_pgdp = pgd_offset_pgd(pgdp, FIXADDR_START); 725 bm_p4dp = p4d_offset(bm_pgdp, FIXADDR_START); 726 bm_pudp = pud_set_fixmap_offset(bm_p4dp, FIXADDR_START); 727 pud_populate(&init_mm, bm_pudp, lm_alias(bm_pmd)); 728 pud_clear_fixmap(); 729 } else { 730 BUG(); 731 } 732 733 kasan_copy_shadow(pgdp); 734 } 735 736 void __init paging_init(void) 737 { 738 pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir)); 739 740 map_kernel(pgdp); 741 map_mem(pgdp); 742 743 pgd_clear_fixmap(); 744 745 cpu_replace_ttbr1(lm_alias(swapper_pg_dir)); 746 init_mm.pgd = swapper_pg_dir; 747 748 memblock_phys_free(__pa_symbol(init_pg_dir), 749 __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir)); 750 751 memblock_allow_resize(); 752 } 753 754 /* 755 * Check whether a kernel address is valid (derived from arch/x86/). 756 */ 757 int kern_addr_valid(unsigned long addr) 758 { 759 pgd_t *pgdp; 760 p4d_t *p4dp; 761 pud_t *pudp, pud; 762 pmd_t *pmdp, pmd; 763 pte_t *ptep, pte; 764 765 addr = arch_kasan_reset_tag(addr); 766 if ((((long)addr) >> VA_BITS) != -1UL) 767 return 0; 768 769 pgdp = pgd_offset_k(addr); 770 if (pgd_none(READ_ONCE(*pgdp))) 771 return 0; 772 773 p4dp = p4d_offset(pgdp, addr); 774 if (p4d_none(READ_ONCE(*p4dp))) 775 return 0; 776 777 pudp = pud_offset(p4dp, addr); 778 pud = READ_ONCE(*pudp); 779 if (pud_none(pud)) 780 return 0; 781 782 if (pud_sect(pud)) 783 return pfn_valid(pud_pfn(pud)); 784 785 pmdp = pmd_offset(pudp, addr); 786 pmd = READ_ONCE(*pmdp); 787 if (pmd_none(pmd)) 788 return 0; 789 790 if (pmd_sect(pmd)) 791 return pfn_valid(pmd_pfn(pmd)); 792 793 ptep = pte_offset_kernel(pmdp, addr); 794 pte = READ_ONCE(*ptep); 795 if (pte_none(pte)) 796 return 0; 797 798 return pfn_valid(pte_pfn(pte)); 799 } 800 801 #ifdef CONFIG_MEMORY_HOTPLUG 802 static void free_hotplug_page_range(struct page *page, size_t size, 803 struct vmem_altmap *altmap) 804 { 805 if (altmap) { 806 vmem_altmap_free(altmap, size >> PAGE_SHIFT); 807 } else { 808 WARN_ON(PageReserved(page)); 809 free_pages((unsigned long)page_address(page), get_order(size)); 810 } 811 } 812 813 static void free_hotplug_pgtable_page(struct page *page) 814 { 815 free_hotplug_page_range(page, PAGE_SIZE, NULL); 816 } 817 818 static bool pgtable_range_aligned(unsigned long start, unsigned long end, 819 unsigned long floor, unsigned long ceiling, 820 unsigned long mask) 821 { 822 start &= mask; 823 if (start < floor) 824 return false; 825 826 if (ceiling) { 827 ceiling &= mask; 828 if (!ceiling) 829 return false; 830 } 831 832 if (end - 1 > ceiling - 1) 833 return false; 834 return true; 835 } 836 837 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr, 838 unsigned long end, bool free_mapped, 839 struct vmem_altmap *altmap) 840 { 841 pte_t *ptep, pte; 842 843 do { 844 ptep = pte_offset_kernel(pmdp, addr); 845 pte = READ_ONCE(*ptep); 846 if (pte_none(pte)) 847 continue; 848 849 WARN_ON(!pte_present(pte)); 850 pte_clear(&init_mm, addr, ptep); 851 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 852 if (free_mapped) 853 free_hotplug_page_range(pte_page(pte), 854 PAGE_SIZE, altmap); 855 } while (addr += PAGE_SIZE, addr < end); 856 } 857 858 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr, 859 unsigned long end, bool free_mapped, 860 struct vmem_altmap *altmap) 861 { 862 unsigned long next; 863 pmd_t *pmdp, pmd; 864 865 do { 866 next = pmd_addr_end(addr, end); 867 pmdp = pmd_offset(pudp, addr); 868 pmd = READ_ONCE(*pmdp); 869 if (pmd_none(pmd)) 870 continue; 871 872 WARN_ON(!pmd_present(pmd)); 873 if (pmd_sect(pmd)) { 874 pmd_clear(pmdp); 875 876 /* 877 * One TLBI should be sufficient here as the PMD_SIZE 878 * range is mapped with a single block entry. 879 */ 880 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 881 if (free_mapped) 882 free_hotplug_page_range(pmd_page(pmd), 883 PMD_SIZE, altmap); 884 continue; 885 } 886 WARN_ON(!pmd_table(pmd)); 887 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap); 888 } while (addr = next, addr < end); 889 } 890 891 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr, 892 unsigned long end, bool free_mapped, 893 struct vmem_altmap *altmap) 894 { 895 unsigned long next; 896 pud_t *pudp, pud; 897 898 do { 899 next = pud_addr_end(addr, end); 900 pudp = pud_offset(p4dp, addr); 901 pud = READ_ONCE(*pudp); 902 if (pud_none(pud)) 903 continue; 904 905 WARN_ON(!pud_present(pud)); 906 if (pud_sect(pud)) { 907 pud_clear(pudp); 908 909 /* 910 * One TLBI should be sufficient here as the PUD_SIZE 911 * range is mapped with a single block entry. 912 */ 913 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 914 if (free_mapped) 915 free_hotplug_page_range(pud_page(pud), 916 PUD_SIZE, altmap); 917 continue; 918 } 919 WARN_ON(!pud_table(pud)); 920 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap); 921 } while (addr = next, addr < end); 922 } 923 924 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr, 925 unsigned long end, bool free_mapped, 926 struct vmem_altmap *altmap) 927 { 928 unsigned long next; 929 p4d_t *p4dp, p4d; 930 931 do { 932 next = p4d_addr_end(addr, end); 933 p4dp = p4d_offset(pgdp, addr); 934 p4d = READ_ONCE(*p4dp); 935 if (p4d_none(p4d)) 936 continue; 937 938 WARN_ON(!p4d_present(p4d)); 939 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap); 940 } while (addr = next, addr < end); 941 } 942 943 static void unmap_hotplug_range(unsigned long addr, unsigned long end, 944 bool free_mapped, struct vmem_altmap *altmap) 945 { 946 unsigned long next; 947 pgd_t *pgdp, pgd; 948 949 /* 950 * altmap can only be used as vmemmap mapping backing memory. 951 * In case the backing memory itself is not being freed, then 952 * altmap is irrelevant. Warn about this inconsistency when 953 * encountered. 954 */ 955 WARN_ON(!free_mapped && altmap); 956 957 do { 958 next = pgd_addr_end(addr, end); 959 pgdp = pgd_offset_k(addr); 960 pgd = READ_ONCE(*pgdp); 961 if (pgd_none(pgd)) 962 continue; 963 964 WARN_ON(!pgd_present(pgd)); 965 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap); 966 } while (addr = next, addr < end); 967 } 968 969 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr, 970 unsigned long end, unsigned long floor, 971 unsigned long ceiling) 972 { 973 pte_t *ptep, pte; 974 unsigned long i, start = addr; 975 976 do { 977 ptep = pte_offset_kernel(pmdp, addr); 978 pte = READ_ONCE(*ptep); 979 980 /* 981 * This is just a sanity check here which verifies that 982 * pte clearing has been done by earlier unmap loops. 983 */ 984 WARN_ON(!pte_none(pte)); 985 } while (addr += PAGE_SIZE, addr < end); 986 987 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK)) 988 return; 989 990 /* 991 * Check whether we can free the pte page if the rest of the 992 * entries are empty. Overlap with other regions have been 993 * handled by the floor/ceiling check. 994 */ 995 ptep = pte_offset_kernel(pmdp, 0UL); 996 for (i = 0; i < PTRS_PER_PTE; i++) { 997 if (!pte_none(READ_ONCE(ptep[i]))) 998 return; 999 } 1000 1001 pmd_clear(pmdp); 1002 __flush_tlb_kernel_pgtable(start); 1003 free_hotplug_pgtable_page(virt_to_page(ptep)); 1004 } 1005 1006 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr, 1007 unsigned long end, unsigned long floor, 1008 unsigned long ceiling) 1009 { 1010 pmd_t *pmdp, pmd; 1011 unsigned long i, next, start = addr; 1012 1013 do { 1014 next = pmd_addr_end(addr, end); 1015 pmdp = pmd_offset(pudp, addr); 1016 pmd = READ_ONCE(*pmdp); 1017 if (pmd_none(pmd)) 1018 continue; 1019 1020 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd)); 1021 free_empty_pte_table(pmdp, addr, next, floor, ceiling); 1022 } while (addr = next, addr < end); 1023 1024 if (CONFIG_PGTABLE_LEVELS <= 2) 1025 return; 1026 1027 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK)) 1028 return; 1029 1030 /* 1031 * Check whether we can free the pmd page if the rest of the 1032 * entries are empty. Overlap with other regions have been 1033 * handled by the floor/ceiling check. 1034 */ 1035 pmdp = pmd_offset(pudp, 0UL); 1036 for (i = 0; i < PTRS_PER_PMD; i++) { 1037 if (!pmd_none(READ_ONCE(pmdp[i]))) 1038 return; 1039 } 1040 1041 pud_clear(pudp); 1042 __flush_tlb_kernel_pgtable(start); 1043 free_hotplug_pgtable_page(virt_to_page(pmdp)); 1044 } 1045 1046 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr, 1047 unsigned long end, unsigned long floor, 1048 unsigned long ceiling) 1049 { 1050 pud_t *pudp, pud; 1051 unsigned long i, next, start = addr; 1052 1053 do { 1054 next = pud_addr_end(addr, end); 1055 pudp = pud_offset(p4dp, addr); 1056 pud = READ_ONCE(*pudp); 1057 if (pud_none(pud)) 1058 continue; 1059 1060 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud)); 1061 free_empty_pmd_table(pudp, addr, next, floor, ceiling); 1062 } while (addr = next, addr < end); 1063 1064 if (CONFIG_PGTABLE_LEVELS <= 3) 1065 return; 1066 1067 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK)) 1068 return; 1069 1070 /* 1071 * Check whether we can free the pud page if the rest of the 1072 * entries are empty. Overlap with other regions have been 1073 * handled by the floor/ceiling check. 1074 */ 1075 pudp = pud_offset(p4dp, 0UL); 1076 for (i = 0; i < PTRS_PER_PUD; i++) { 1077 if (!pud_none(READ_ONCE(pudp[i]))) 1078 return; 1079 } 1080 1081 p4d_clear(p4dp); 1082 __flush_tlb_kernel_pgtable(start); 1083 free_hotplug_pgtable_page(virt_to_page(pudp)); 1084 } 1085 1086 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr, 1087 unsigned long end, unsigned long floor, 1088 unsigned long ceiling) 1089 { 1090 unsigned long next; 1091 p4d_t *p4dp, p4d; 1092 1093 do { 1094 next = p4d_addr_end(addr, end); 1095 p4dp = p4d_offset(pgdp, addr); 1096 p4d = READ_ONCE(*p4dp); 1097 if (p4d_none(p4d)) 1098 continue; 1099 1100 WARN_ON(!p4d_present(p4d)); 1101 free_empty_pud_table(p4dp, addr, next, floor, ceiling); 1102 } while (addr = next, addr < end); 1103 } 1104 1105 static void free_empty_tables(unsigned long addr, unsigned long end, 1106 unsigned long floor, unsigned long ceiling) 1107 { 1108 unsigned long next; 1109 pgd_t *pgdp, pgd; 1110 1111 do { 1112 next = pgd_addr_end(addr, end); 1113 pgdp = pgd_offset_k(addr); 1114 pgd = READ_ONCE(*pgdp); 1115 if (pgd_none(pgd)) 1116 continue; 1117 1118 WARN_ON(!pgd_present(pgd)); 1119 free_empty_p4d_table(pgdp, addr, next, floor, ceiling); 1120 } while (addr = next, addr < end); 1121 } 1122 #endif 1123 1124 #if !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 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1129 return vmemmap_populate_basepages(start, end, node, altmap); 1130 } 1131 #else /* !ARM64_KERNEL_USES_PMD_MAPS */ 1132 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1133 struct vmem_altmap *altmap) 1134 { 1135 unsigned long addr = start; 1136 unsigned long next; 1137 pgd_t *pgdp; 1138 p4d_t *p4dp; 1139 pud_t *pudp; 1140 pmd_t *pmdp; 1141 1142 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1143 do { 1144 next = pmd_addr_end(addr, end); 1145 1146 pgdp = vmemmap_pgd_populate(addr, node); 1147 if (!pgdp) 1148 return -ENOMEM; 1149 1150 p4dp = vmemmap_p4d_populate(pgdp, addr, node); 1151 if (!p4dp) 1152 return -ENOMEM; 1153 1154 pudp = vmemmap_pud_populate(p4dp, addr, node); 1155 if (!pudp) 1156 return -ENOMEM; 1157 1158 pmdp = pmd_offset(pudp, addr); 1159 if (pmd_none(READ_ONCE(*pmdp))) { 1160 void *p = NULL; 1161 1162 p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap); 1163 if (!p) { 1164 if (vmemmap_populate_basepages(addr, next, node, altmap)) 1165 return -ENOMEM; 1166 continue; 1167 } 1168 1169 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL)); 1170 } else 1171 vmemmap_verify((pte_t *)pmdp, node, addr, next); 1172 } while (addr = next, addr != end); 1173 1174 return 0; 1175 } 1176 #endif /* !ARM64_KERNEL_USES_PMD_MAPS */ 1177 1178 #ifdef CONFIG_MEMORY_HOTPLUG 1179 void vmemmap_free(unsigned long start, unsigned long end, 1180 struct vmem_altmap *altmap) 1181 { 1182 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1183 1184 unmap_hotplug_range(start, end, true, altmap); 1185 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END); 1186 } 1187 #endif /* CONFIG_MEMORY_HOTPLUG */ 1188 1189 static inline pud_t *fixmap_pud(unsigned long addr) 1190 { 1191 pgd_t *pgdp = pgd_offset_k(addr); 1192 p4d_t *p4dp = p4d_offset(pgdp, addr); 1193 p4d_t p4d = READ_ONCE(*p4dp); 1194 1195 BUG_ON(p4d_none(p4d) || p4d_bad(p4d)); 1196 1197 return pud_offset_kimg(p4dp, addr); 1198 } 1199 1200 static inline pmd_t *fixmap_pmd(unsigned long addr) 1201 { 1202 pud_t *pudp = fixmap_pud(addr); 1203 pud_t pud = READ_ONCE(*pudp); 1204 1205 BUG_ON(pud_none(pud) || pud_bad(pud)); 1206 1207 return pmd_offset_kimg(pudp, addr); 1208 } 1209 1210 static inline pte_t *fixmap_pte(unsigned long addr) 1211 { 1212 return &bm_pte[pte_index(addr)]; 1213 } 1214 1215 /* 1216 * The p*d_populate functions call virt_to_phys implicitly so they can't be used 1217 * directly on kernel symbols (bm_p*d). This function is called too early to use 1218 * lm_alias so __p*d_populate functions must be used to populate with the 1219 * physical address from __pa_symbol. 1220 */ 1221 void __init early_fixmap_init(void) 1222 { 1223 pgd_t *pgdp; 1224 p4d_t *p4dp, p4d; 1225 pud_t *pudp; 1226 pmd_t *pmdp; 1227 unsigned long addr = FIXADDR_START; 1228 1229 pgdp = pgd_offset_k(addr); 1230 p4dp = p4d_offset(pgdp, addr); 1231 p4d = READ_ONCE(*p4dp); 1232 if (CONFIG_PGTABLE_LEVELS > 3 && 1233 !(p4d_none(p4d) || p4d_page_paddr(p4d) == __pa_symbol(bm_pud))) { 1234 /* 1235 * We only end up here if the kernel mapping and the fixmap 1236 * share the top level pgd entry, which should only happen on 1237 * 16k/4 levels configurations. 1238 */ 1239 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES)); 1240 pudp = pud_offset_kimg(p4dp, addr); 1241 } else { 1242 if (p4d_none(p4d)) 1243 __p4d_populate(p4dp, __pa_symbol(bm_pud), P4D_TYPE_TABLE); 1244 pudp = fixmap_pud(addr); 1245 } 1246 if (pud_none(READ_ONCE(*pudp))) 1247 __pud_populate(pudp, __pa_symbol(bm_pmd), PUD_TYPE_TABLE); 1248 pmdp = fixmap_pmd(addr); 1249 __pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE); 1250 1251 /* 1252 * The boot-ioremap range spans multiple pmds, for which 1253 * we are not prepared: 1254 */ 1255 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT) 1256 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT)); 1257 1258 if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN))) 1259 || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) { 1260 WARN_ON(1); 1261 pr_warn("pmdp %p != %p, %p\n", 1262 pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)), 1263 fixmap_pmd(fix_to_virt(FIX_BTMAP_END))); 1264 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n", 1265 fix_to_virt(FIX_BTMAP_BEGIN)); 1266 pr_warn("fix_to_virt(FIX_BTMAP_END): %08lx\n", 1267 fix_to_virt(FIX_BTMAP_END)); 1268 1269 pr_warn("FIX_BTMAP_END: %d\n", FIX_BTMAP_END); 1270 pr_warn("FIX_BTMAP_BEGIN: %d\n", FIX_BTMAP_BEGIN); 1271 } 1272 } 1273 1274 /* 1275 * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we 1276 * ever need to use IPIs for TLB broadcasting, then we're in trouble here. 1277 */ 1278 void __set_fixmap(enum fixed_addresses idx, 1279 phys_addr_t phys, pgprot_t flags) 1280 { 1281 unsigned long addr = __fix_to_virt(idx); 1282 pte_t *ptep; 1283 1284 BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses); 1285 1286 ptep = fixmap_pte(addr); 1287 1288 if (pgprot_val(flags)) { 1289 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags)); 1290 } else { 1291 pte_clear(&init_mm, addr, ptep); 1292 flush_tlb_kernel_range(addr, addr+PAGE_SIZE); 1293 } 1294 } 1295 1296 void *__init fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot) 1297 { 1298 const u64 dt_virt_base = __fix_to_virt(FIX_FDT); 1299 int offset; 1300 void *dt_virt; 1301 1302 /* 1303 * Check whether the physical FDT address is set and meets the minimum 1304 * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be 1305 * at least 8 bytes so that we can always access the magic and size 1306 * fields of the FDT header after mapping the first chunk, double check 1307 * here if that is indeed the case. 1308 */ 1309 BUILD_BUG_ON(MIN_FDT_ALIGN < 8); 1310 if (!dt_phys || dt_phys % MIN_FDT_ALIGN) 1311 return NULL; 1312 1313 /* 1314 * Make sure that the FDT region can be mapped without the need to 1315 * allocate additional translation table pages, so that it is safe 1316 * to call create_mapping_noalloc() this early. 1317 * 1318 * On 64k pages, the FDT will be mapped using PTEs, so we need to 1319 * be in the same PMD as the rest of the fixmap. 1320 * On 4k pages, we'll use section mappings for the FDT so we only 1321 * have to be in the same PUD. 1322 */ 1323 BUILD_BUG_ON(dt_virt_base % SZ_2M); 1324 1325 BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT != 1326 __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT); 1327 1328 offset = dt_phys % SWAPPER_BLOCK_SIZE; 1329 dt_virt = (void *)dt_virt_base + offset; 1330 1331 /* map the first chunk so we can read the size from the header */ 1332 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), 1333 dt_virt_base, SWAPPER_BLOCK_SIZE, prot); 1334 1335 if (fdt_magic(dt_virt) != FDT_MAGIC) 1336 return NULL; 1337 1338 *size = fdt_totalsize(dt_virt); 1339 if (*size > MAX_FDT_SIZE) 1340 return NULL; 1341 1342 if (offset + *size > SWAPPER_BLOCK_SIZE) 1343 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base, 1344 round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot); 1345 1346 return dt_virt; 1347 } 1348 1349 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot) 1350 { 1351 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot)); 1352 1353 /* Only allow permission changes for now */ 1354 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)), 1355 pud_val(new_pud))) 1356 return 0; 1357 1358 VM_BUG_ON(phys & ~PUD_MASK); 1359 set_pud(pudp, new_pud); 1360 return 1; 1361 } 1362 1363 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot) 1364 { 1365 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot)); 1366 1367 /* Only allow permission changes for now */ 1368 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)), 1369 pmd_val(new_pmd))) 1370 return 0; 1371 1372 VM_BUG_ON(phys & ~PMD_MASK); 1373 set_pmd(pmdp, new_pmd); 1374 return 1; 1375 } 1376 1377 int pud_clear_huge(pud_t *pudp) 1378 { 1379 if (!pud_sect(READ_ONCE(*pudp))) 1380 return 0; 1381 pud_clear(pudp); 1382 return 1; 1383 } 1384 1385 int pmd_clear_huge(pmd_t *pmdp) 1386 { 1387 if (!pmd_sect(READ_ONCE(*pmdp))) 1388 return 0; 1389 pmd_clear(pmdp); 1390 return 1; 1391 } 1392 1393 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr) 1394 { 1395 pte_t *table; 1396 pmd_t pmd; 1397 1398 pmd = READ_ONCE(*pmdp); 1399 1400 if (!pmd_table(pmd)) { 1401 VM_WARN_ON(1); 1402 return 1; 1403 } 1404 1405 table = pte_offset_kernel(pmdp, addr); 1406 pmd_clear(pmdp); 1407 __flush_tlb_kernel_pgtable(addr); 1408 pte_free_kernel(NULL, table); 1409 return 1; 1410 } 1411 1412 int pud_free_pmd_page(pud_t *pudp, unsigned long addr) 1413 { 1414 pmd_t *table; 1415 pmd_t *pmdp; 1416 pud_t pud; 1417 unsigned long next, end; 1418 1419 pud = READ_ONCE(*pudp); 1420 1421 if (!pud_table(pud)) { 1422 VM_WARN_ON(1); 1423 return 1; 1424 } 1425 1426 table = pmd_offset(pudp, addr); 1427 pmdp = table; 1428 next = addr; 1429 end = addr + PUD_SIZE; 1430 do { 1431 pmd_free_pte_page(pmdp, next); 1432 } while (pmdp++, next += PMD_SIZE, next != end); 1433 1434 pud_clear(pudp); 1435 __flush_tlb_kernel_pgtable(addr); 1436 pmd_free(NULL, table); 1437 return 1; 1438 } 1439 1440 #ifdef CONFIG_MEMORY_HOTPLUG 1441 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size) 1442 { 1443 unsigned long end = start + size; 1444 1445 WARN_ON(pgdir != init_mm.pgd); 1446 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END)); 1447 1448 unmap_hotplug_range(start, end, false, NULL); 1449 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END); 1450 } 1451 1452 struct range arch_get_mappable_range(void) 1453 { 1454 struct range mhp_range; 1455 u64 start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual)); 1456 u64 end_linear_pa = __pa(PAGE_END - 1); 1457 1458 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 1459 /* 1460 * Check for a wrap, it is possible because of randomized linear 1461 * mapping the start physical address is actually bigger than 1462 * the end physical address. In this case set start to zero 1463 * because [0, end_linear_pa] range must still be able to cover 1464 * all addressable physical addresses. 1465 */ 1466 if (start_linear_pa > end_linear_pa) 1467 start_linear_pa = 0; 1468 } 1469 1470 WARN_ON(start_linear_pa > end_linear_pa); 1471 1472 /* 1473 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)] 1474 * accommodating both its ends but excluding PAGE_END. Max physical 1475 * range which can be mapped inside this linear mapping range, must 1476 * also be derived from its end points. 1477 */ 1478 mhp_range.start = start_linear_pa; 1479 mhp_range.end = end_linear_pa; 1480 1481 return mhp_range; 1482 } 1483 1484 int arch_add_memory(int nid, u64 start, u64 size, 1485 struct mhp_params *params) 1486 { 1487 int ret, flags = NO_EXEC_MAPPINGS; 1488 1489 VM_BUG_ON(!mhp_range_allowed(start, size, true)); 1490 1491 /* 1492 * KFENCE requires linear map to be mapped at page granularity, so that 1493 * it is possible to protect/unprotect single pages in the KFENCE pool. 1494 */ 1495 if (can_set_direct_map() || IS_ENABLED(CONFIG_KFENCE)) 1496 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1497 1498 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1499 size, params->pgprot, __pgd_pgtable_alloc, 1500 flags); 1501 1502 memblock_clear_nomap(start, size); 1503 1504 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, 1505 params); 1506 if (ret) 1507 __remove_pgd_mapping(swapper_pg_dir, 1508 __phys_to_virt(start), size); 1509 else { 1510 max_pfn = PFN_UP(start + size); 1511 max_low_pfn = max_pfn; 1512 } 1513 1514 return ret; 1515 } 1516 1517 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap) 1518 { 1519 unsigned long start_pfn = start >> PAGE_SHIFT; 1520 unsigned long nr_pages = size >> PAGE_SHIFT; 1521 1522 __remove_pages(start_pfn, nr_pages, altmap); 1523 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size); 1524 } 1525 1526 /* 1527 * This memory hotplug notifier helps prevent boot memory from being 1528 * inadvertently removed as it blocks pfn range offlining process in 1529 * __offline_pages(). Hence this prevents both offlining as well as 1530 * removal process for boot memory which is initially always online. 1531 * In future if and when boot memory could be removed, this notifier 1532 * should be dropped and free_hotplug_page_range() should handle any 1533 * reserved pages allocated during boot. 1534 */ 1535 static int prevent_bootmem_remove_notifier(struct notifier_block *nb, 1536 unsigned long action, void *data) 1537 { 1538 struct mem_section *ms; 1539 struct memory_notify *arg = data; 1540 unsigned long end_pfn = arg->start_pfn + arg->nr_pages; 1541 unsigned long pfn = arg->start_pfn; 1542 1543 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE)) 1544 return NOTIFY_OK; 1545 1546 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 1547 unsigned long start = PFN_PHYS(pfn); 1548 unsigned long end = start + (1UL << PA_SECTION_SHIFT); 1549 1550 ms = __pfn_to_section(pfn); 1551 if (!early_section(ms)) 1552 continue; 1553 1554 if (action == MEM_GOING_OFFLINE) { 1555 /* 1556 * Boot memory removal is not supported. Prevent 1557 * it via blocking any attempted offline request 1558 * for the boot memory and just report it. 1559 */ 1560 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end); 1561 return NOTIFY_BAD; 1562 } else if (action == MEM_OFFLINE) { 1563 /* 1564 * This should have never happened. Boot memory 1565 * offlining should have been prevented by this 1566 * very notifier. Probably some memory removal 1567 * procedure might have changed which would then 1568 * require further debug. 1569 */ 1570 pr_err("Boot memory [%lx %lx] offlined\n", start, end); 1571 1572 /* 1573 * Core memory hotplug does not process a return 1574 * code from the notifier for MEM_OFFLINE events. 1575 * The error condition has been reported. Return 1576 * from here as if ignored. 1577 */ 1578 return NOTIFY_DONE; 1579 } 1580 } 1581 return NOTIFY_OK; 1582 } 1583 1584 static struct notifier_block prevent_bootmem_remove_nb = { 1585 .notifier_call = prevent_bootmem_remove_notifier, 1586 }; 1587 1588 /* 1589 * This ensures that boot memory sections on the platform are online 1590 * from early boot. Memory sections could not be prevented from being 1591 * offlined, unless for some reason they are not online to begin with. 1592 * This helps validate the basic assumption on which the above memory 1593 * event notifier works to prevent boot memory section offlining and 1594 * its possible removal. 1595 */ 1596 static void validate_bootmem_online(void) 1597 { 1598 phys_addr_t start, end, addr; 1599 struct mem_section *ms; 1600 u64 i; 1601 1602 /* 1603 * Scanning across all memblock might be expensive 1604 * on some big memory systems. Hence enable this 1605 * validation only with DEBUG_VM. 1606 */ 1607 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 1608 return; 1609 1610 for_each_mem_range(i, &start, &end) { 1611 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) { 1612 ms = __pfn_to_section(PHYS_PFN(addr)); 1613 1614 /* 1615 * All memory ranges in the system at this point 1616 * should have been marked as early sections. 1617 */ 1618 WARN_ON(!early_section(ms)); 1619 1620 /* 1621 * Memory notifier mechanism here to prevent boot 1622 * memory offlining depends on the fact that each 1623 * early section memory on the system is initially 1624 * online. Otherwise a given memory section which 1625 * is already offline will be overlooked and can 1626 * be removed completely. Call out such sections. 1627 */ 1628 if (!online_section(ms)) 1629 pr_err("Boot memory [%llx %llx] is offline, can be removed\n", 1630 addr, addr + (1UL << PA_SECTION_SHIFT)); 1631 } 1632 } 1633 } 1634 1635 static int __init prevent_bootmem_remove_init(void) 1636 { 1637 int ret = 0; 1638 1639 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) 1640 return ret; 1641 1642 validate_bootmem_online(); 1643 ret = register_memory_notifier(&prevent_bootmem_remove_nb); 1644 if (ret) 1645 pr_err("%s: Notifier registration failed %d\n", __func__, ret); 1646 1647 return ret; 1648 } 1649 early_initcall(prevent_bootmem_remove_init); 1650 #endif 1651