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 26 #include <asm/barrier.h> 27 #include <asm/cputype.h> 28 #include <asm/fixmap.h> 29 #include <asm/kasan.h> 30 #include <asm/kernel-pgtable.h> 31 #include <asm/sections.h> 32 #include <asm/setup.h> 33 #include <linux/sizes.h> 34 #include <asm/tlb.h> 35 #include <asm/mmu_context.h> 36 #include <asm/ptdump.h> 37 #include <asm/tlbflush.h> 38 #include <asm/pgalloc.h> 39 40 #define NO_BLOCK_MAPPINGS BIT(0) 41 #define NO_CONT_MAPPINGS BIT(1) 42 43 u64 idmap_t0sz = TCR_T0SZ(VA_BITS_MIN); 44 u64 idmap_ptrs_per_pgd = PTRS_PER_PGD; 45 46 u64 __section(".mmuoff.data.write") vabits_actual; 47 EXPORT_SYMBOL(vabits_actual); 48 49 u64 kimage_voffset __ro_after_init; 50 EXPORT_SYMBOL(kimage_voffset); 51 52 /* 53 * Empty_zero_page is a special page that is used for zero-initialized data 54 * and COW. 55 */ 56 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss; 57 EXPORT_SYMBOL(empty_zero_page); 58 59 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss; 60 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused; 61 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused; 62 63 static DEFINE_SPINLOCK(swapper_pgdir_lock); 64 65 void set_swapper_pgd(pgd_t *pgdp, pgd_t pgd) 66 { 67 pgd_t *fixmap_pgdp; 68 69 spin_lock(&swapper_pgdir_lock); 70 fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp)); 71 WRITE_ONCE(*fixmap_pgdp, pgd); 72 /* 73 * We need dsb(ishst) here to ensure the page-table-walker sees 74 * our new entry before set_p?d() returns. The fixmap's 75 * flush_tlb_kernel_range() via clear_fixmap() does this for us. 76 */ 77 pgd_clear_fixmap(); 78 spin_unlock(&swapper_pgdir_lock); 79 } 80 81 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn, 82 unsigned long size, pgprot_t vma_prot) 83 { 84 if (!pfn_valid(pfn)) 85 return pgprot_noncached(vma_prot); 86 else if (file->f_flags & O_SYNC) 87 return pgprot_writecombine(vma_prot); 88 return vma_prot; 89 } 90 EXPORT_SYMBOL(phys_mem_access_prot); 91 92 static phys_addr_t __init early_pgtable_alloc(int shift) 93 { 94 phys_addr_t phys; 95 void *ptr; 96 97 phys = memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE); 98 if (!phys) 99 panic("Failed to allocate page table page\n"); 100 101 /* 102 * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE 103 * slot will be free, so we can (ab)use the FIX_PTE slot to initialise 104 * any level of table. 105 */ 106 ptr = pte_set_fixmap(phys); 107 108 memset(ptr, 0, PAGE_SIZE); 109 110 /* 111 * Implicit barriers also ensure the zeroed page is visible to the page 112 * table walker 113 */ 114 pte_clear_fixmap(); 115 116 return phys; 117 } 118 119 static bool pgattr_change_is_safe(u64 old, u64 new) 120 { 121 /* 122 * The following mapping attributes may be updated in live 123 * kernel mappings without the need for break-before-make. 124 */ 125 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG; 126 127 /* creating or taking down mappings is always safe */ 128 if (old == 0 || new == 0) 129 return true; 130 131 /* live contiguous mappings may not be manipulated at all */ 132 if ((old | new) & PTE_CONT) 133 return false; 134 135 /* Transitioning from Non-Global to Global is unsafe */ 136 if (old & ~new & PTE_NG) 137 return false; 138 139 /* 140 * Changing the memory type between Normal and Normal-Tagged is safe 141 * since Tagged is considered a permission attribute from the 142 * mismatched attribute aliases perspective. 143 */ 144 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 145 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) && 146 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 147 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED))) 148 mask |= PTE_ATTRINDX_MASK; 149 150 return ((old ^ new) & ~mask) == 0; 151 } 152 153 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end, 154 phys_addr_t phys, pgprot_t prot) 155 { 156 pte_t *ptep; 157 158 ptep = pte_set_fixmap_offset(pmdp, addr); 159 do { 160 pte_t old_pte = READ_ONCE(*ptep); 161 162 set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot)); 163 164 /* 165 * After the PTE entry has been populated once, we 166 * only allow updates to the permission attributes. 167 */ 168 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte), 169 READ_ONCE(pte_val(*ptep)))); 170 171 phys += PAGE_SIZE; 172 } while (ptep++, addr += PAGE_SIZE, addr != end); 173 174 pte_clear_fixmap(); 175 } 176 177 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr, 178 unsigned long end, phys_addr_t phys, 179 pgprot_t prot, 180 phys_addr_t (*pgtable_alloc)(int), 181 int flags) 182 { 183 unsigned long next; 184 pmd_t pmd = READ_ONCE(*pmdp); 185 186 BUG_ON(pmd_sect(pmd)); 187 if (pmd_none(pmd)) { 188 phys_addr_t pte_phys; 189 BUG_ON(!pgtable_alloc); 190 pte_phys = pgtable_alloc(PAGE_SHIFT); 191 __pmd_populate(pmdp, pte_phys, PMD_TYPE_TABLE); 192 pmd = READ_ONCE(*pmdp); 193 } 194 BUG_ON(pmd_bad(pmd)); 195 196 do { 197 pgprot_t __prot = prot; 198 199 next = pte_cont_addr_end(addr, end); 200 201 /* use a contiguous mapping if the range is suitably aligned */ 202 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) && 203 (flags & NO_CONT_MAPPINGS) == 0) 204 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 205 206 init_pte(pmdp, addr, next, phys, __prot); 207 208 phys += next - addr; 209 } while (addr = next, addr != end); 210 } 211 212 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end, 213 phys_addr_t phys, pgprot_t prot, 214 phys_addr_t (*pgtable_alloc)(int), int flags) 215 { 216 unsigned long next; 217 pmd_t *pmdp; 218 219 pmdp = pmd_set_fixmap_offset(pudp, addr); 220 do { 221 pmd_t old_pmd = READ_ONCE(*pmdp); 222 223 next = pmd_addr_end(addr, end); 224 225 /* try section mapping first */ 226 if (((addr | next | phys) & ~SECTION_MASK) == 0 && 227 (flags & NO_BLOCK_MAPPINGS) == 0) { 228 pmd_set_huge(pmdp, phys, prot); 229 230 /* 231 * After the PMD entry has been populated once, we 232 * only allow updates to the permission attributes. 233 */ 234 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd), 235 READ_ONCE(pmd_val(*pmdp)))); 236 } else { 237 alloc_init_cont_pte(pmdp, addr, next, phys, prot, 238 pgtable_alloc, flags); 239 240 BUG_ON(pmd_val(old_pmd) != 0 && 241 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp))); 242 } 243 phys += next - addr; 244 } while (pmdp++, addr = next, addr != end); 245 246 pmd_clear_fixmap(); 247 } 248 249 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr, 250 unsigned long end, phys_addr_t phys, 251 pgprot_t prot, 252 phys_addr_t (*pgtable_alloc)(int), int flags) 253 { 254 unsigned long next; 255 pud_t pud = READ_ONCE(*pudp); 256 257 /* 258 * Check for initial section mappings in the pgd/pud. 259 */ 260 BUG_ON(pud_sect(pud)); 261 if (pud_none(pud)) { 262 phys_addr_t pmd_phys; 263 BUG_ON(!pgtable_alloc); 264 pmd_phys = pgtable_alloc(PMD_SHIFT); 265 __pud_populate(pudp, pmd_phys, PUD_TYPE_TABLE); 266 pud = READ_ONCE(*pudp); 267 } 268 BUG_ON(pud_bad(pud)); 269 270 do { 271 pgprot_t __prot = prot; 272 273 next = pmd_cont_addr_end(addr, end); 274 275 /* use a contiguous mapping if the range is suitably aligned */ 276 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) && 277 (flags & NO_CONT_MAPPINGS) == 0) 278 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 279 280 init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags); 281 282 phys += next - addr; 283 } while (addr = next, addr != end); 284 } 285 286 static inline bool use_1G_block(unsigned long addr, unsigned long next, 287 unsigned long phys) 288 { 289 if (PAGE_SHIFT != 12) 290 return false; 291 292 if (((addr | next | phys) & ~PUD_MASK) != 0) 293 return false; 294 295 return true; 296 } 297 298 static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end, 299 phys_addr_t phys, pgprot_t prot, 300 phys_addr_t (*pgtable_alloc)(int), 301 int flags) 302 { 303 unsigned long next; 304 pud_t *pudp; 305 p4d_t *p4dp = p4d_offset(pgdp, addr); 306 p4d_t p4d = READ_ONCE(*p4dp); 307 308 if (p4d_none(p4d)) { 309 phys_addr_t pud_phys; 310 BUG_ON(!pgtable_alloc); 311 pud_phys = pgtable_alloc(PUD_SHIFT); 312 __p4d_populate(p4dp, pud_phys, PUD_TYPE_TABLE); 313 p4d = READ_ONCE(*p4dp); 314 } 315 BUG_ON(p4d_bad(p4d)); 316 317 pudp = pud_set_fixmap_offset(p4dp, addr); 318 do { 319 pud_t old_pud = READ_ONCE(*pudp); 320 321 next = pud_addr_end(addr, end); 322 323 /* 324 * For 4K granule only, attempt to put down a 1GB block 325 */ 326 if (use_1G_block(addr, next, phys) && 327 (flags & NO_BLOCK_MAPPINGS) == 0) { 328 pud_set_huge(pudp, phys, prot); 329 330 /* 331 * After the PUD entry has been populated once, we 332 * only allow updates to the permission attributes. 333 */ 334 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud), 335 READ_ONCE(pud_val(*pudp)))); 336 } else { 337 alloc_init_cont_pmd(pudp, addr, next, phys, prot, 338 pgtable_alloc, flags); 339 340 BUG_ON(pud_val(old_pud) != 0 && 341 pud_val(old_pud) != READ_ONCE(pud_val(*pudp))); 342 } 343 phys += next - addr; 344 } while (pudp++, addr = next, addr != end); 345 346 pud_clear_fixmap(); 347 } 348 349 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, 350 unsigned long virt, phys_addr_t size, 351 pgprot_t prot, 352 phys_addr_t (*pgtable_alloc)(int), 353 int flags) 354 { 355 unsigned long addr, end, next; 356 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt); 357 358 /* 359 * If the virtual and physical address don't have the same offset 360 * within a page, we cannot map the region as the caller expects. 361 */ 362 if (WARN_ON((phys ^ virt) & ~PAGE_MASK)) 363 return; 364 365 phys &= PAGE_MASK; 366 addr = virt & PAGE_MASK; 367 end = PAGE_ALIGN(virt + size); 368 369 do { 370 next = pgd_addr_end(addr, end); 371 alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc, 372 flags); 373 phys += next - addr; 374 } while (pgdp++, addr = next, addr != end); 375 } 376 377 static phys_addr_t __pgd_pgtable_alloc(int shift) 378 { 379 void *ptr = (void *)__get_free_page(GFP_PGTABLE_KERNEL); 380 BUG_ON(!ptr); 381 382 /* Ensure the zeroed page is visible to the page table walker */ 383 dsb(ishst); 384 return __pa(ptr); 385 } 386 387 static phys_addr_t pgd_pgtable_alloc(int shift) 388 { 389 phys_addr_t pa = __pgd_pgtable_alloc(shift); 390 391 /* 392 * Call proper page table ctor in case later we need to 393 * call core mm functions like apply_to_page_range() on 394 * this pre-allocated page table. 395 * 396 * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is 397 * folded, and if so pgtable_pmd_page_ctor() becomes nop. 398 */ 399 if (shift == PAGE_SHIFT) 400 BUG_ON(!pgtable_pte_page_ctor(phys_to_page(pa))); 401 else if (shift == PMD_SHIFT) 402 BUG_ON(!pgtable_pmd_page_ctor(phys_to_page(pa))); 403 404 return pa; 405 } 406 407 /* 408 * This function can only be used to modify existing table entries, 409 * without allocating new levels of table. Note that this permits the 410 * creation of new section or page entries. 411 */ 412 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt, 413 phys_addr_t size, pgprot_t prot) 414 { 415 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) { 416 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n", 417 &phys, virt); 418 return; 419 } 420 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 421 NO_CONT_MAPPINGS); 422 } 423 424 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys, 425 unsigned long virt, phys_addr_t size, 426 pgprot_t prot, bool page_mappings_only) 427 { 428 int flags = 0; 429 430 BUG_ON(mm == &init_mm); 431 432 if (page_mappings_only) 433 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 434 435 __create_pgd_mapping(mm->pgd, phys, virt, size, prot, 436 pgd_pgtable_alloc, flags); 437 } 438 439 static void update_mapping_prot(phys_addr_t phys, unsigned long virt, 440 phys_addr_t size, pgprot_t prot) 441 { 442 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) { 443 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n", 444 &phys, virt); 445 return; 446 } 447 448 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 449 NO_CONT_MAPPINGS); 450 451 /* flush the TLBs after updating live kernel mappings */ 452 flush_tlb_kernel_range(virt, virt + size); 453 } 454 455 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start, 456 phys_addr_t end, pgprot_t prot, int flags) 457 { 458 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start, 459 prot, early_pgtable_alloc, flags); 460 } 461 462 void __init mark_linear_text_alias_ro(void) 463 { 464 /* 465 * Remove the write permissions from the linear alias of .text/.rodata 466 */ 467 update_mapping_prot(__pa_symbol(_stext), (unsigned long)lm_alias(_stext), 468 (unsigned long)__init_begin - (unsigned long)_stext, 469 PAGE_KERNEL_RO); 470 } 471 472 static bool crash_mem_map __initdata; 473 474 static int __init enable_crash_mem_map(char *arg) 475 { 476 /* 477 * Proper parameter parsing is done by reserve_crashkernel(). We only 478 * need to know if the linear map has to avoid block mappings so that 479 * the crashkernel reservations can be unmapped later. 480 */ 481 crash_mem_map = true; 482 483 return 0; 484 } 485 early_param("crashkernel", enable_crash_mem_map); 486 487 static void __init map_mem(pgd_t *pgdp) 488 { 489 phys_addr_t kernel_start = __pa_symbol(_stext); 490 phys_addr_t kernel_end = __pa_symbol(__init_begin); 491 phys_addr_t start, end; 492 int flags = 0; 493 u64 i; 494 495 if (rodata_full || crash_mem_map || debug_pagealloc_enabled()) 496 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 497 498 /* 499 * Take care not to create a writable alias for the 500 * read-only text and rodata sections of the kernel image. 501 * So temporarily mark them as NOMAP to skip mappings in 502 * the following for-loop 503 */ 504 memblock_mark_nomap(kernel_start, kernel_end - kernel_start); 505 506 /* map all the memory banks */ 507 for_each_mem_range(i, &start, &end) { 508 if (start >= end) 509 break; 510 /* 511 * The linear map must allow allocation tags reading/writing 512 * if MTE is present. Otherwise, it has the same attributes as 513 * PAGE_KERNEL. 514 */ 515 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL), 516 flags); 517 } 518 519 /* 520 * Map the linear alias of the [_stext, __init_begin) interval 521 * as non-executable now, and remove the write permission in 522 * mark_linear_text_alias_ro() below (which will be called after 523 * alternative patching has completed). This makes the contents 524 * of the region accessible to subsystems such as hibernate, 525 * but protects it from inadvertent modification or execution. 526 * Note that contiguous mappings cannot be remapped in this way, 527 * so we should avoid them here. 528 */ 529 __map_memblock(pgdp, kernel_start, kernel_end, 530 PAGE_KERNEL, NO_CONT_MAPPINGS); 531 memblock_clear_nomap(kernel_start, kernel_end - kernel_start); 532 } 533 534 void mark_rodata_ro(void) 535 { 536 unsigned long section_size; 537 538 /* 539 * mark .rodata as read only. Use __init_begin rather than __end_rodata 540 * to cover NOTES and EXCEPTION_TABLE. 541 */ 542 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata; 543 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata, 544 section_size, PAGE_KERNEL_RO); 545 546 debug_checkwx(); 547 } 548 549 static void __init map_kernel_segment(pgd_t *pgdp, void *va_start, void *va_end, 550 pgprot_t prot, struct vm_struct *vma, 551 int flags, unsigned long vm_flags) 552 { 553 phys_addr_t pa_start = __pa_symbol(va_start); 554 unsigned long size = va_end - va_start; 555 556 BUG_ON(!PAGE_ALIGNED(pa_start)); 557 BUG_ON(!PAGE_ALIGNED(size)); 558 559 __create_pgd_mapping(pgdp, pa_start, (unsigned long)va_start, size, prot, 560 early_pgtable_alloc, flags); 561 562 if (!(vm_flags & VM_NO_GUARD)) 563 size += PAGE_SIZE; 564 565 vma->addr = va_start; 566 vma->phys_addr = pa_start; 567 vma->size = size; 568 vma->flags = VM_MAP | vm_flags; 569 vma->caller = __builtin_return_address(0); 570 571 vm_area_add_early(vma); 572 } 573 574 static int __init parse_rodata(char *arg) 575 { 576 int ret = strtobool(arg, &rodata_enabled); 577 if (!ret) { 578 rodata_full = false; 579 return 0; 580 } 581 582 /* permit 'full' in addition to boolean options */ 583 if (strcmp(arg, "full")) 584 return -EINVAL; 585 586 rodata_enabled = true; 587 rodata_full = true; 588 return 0; 589 } 590 early_param("rodata", parse_rodata); 591 592 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0 593 static int __init map_entry_trampoline(void) 594 { 595 pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 596 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start); 597 598 /* The trampoline is always mapped and can therefore be global */ 599 pgprot_val(prot) &= ~PTE_NG; 600 601 /* Map only the text into the trampoline page table */ 602 memset(tramp_pg_dir, 0, PGD_SIZE); 603 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, PAGE_SIZE, 604 prot, __pgd_pgtable_alloc, 0); 605 606 /* Map both the text and data into the kernel page table */ 607 __set_fixmap(FIX_ENTRY_TRAMP_TEXT, pa_start, prot); 608 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 609 extern char __entry_tramp_data_start[]; 610 611 __set_fixmap(FIX_ENTRY_TRAMP_DATA, 612 __pa_symbol(__entry_tramp_data_start), 613 PAGE_KERNEL_RO); 614 } 615 616 return 0; 617 } 618 core_initcall(map_entry_trampoline); 619 #endif 620 621 /* 622 * Open coded check for BTI, only for use to determine configuration 623 * for early mappings for before the cpufeature code has run. 624 */ 625 static bool arm64_early_this_cpu_has_bti(void) 626 { 627 u64 pfr1; 628 629 if (!IS_ENABLED(CONFIG_ARM64_BTI_KERNEL)) 630 return false; 631 632 pfr1 = __read_sysreg_by_encoding(SYS_ID_AA64PFR1_EL1); 633 return cpuid_feature_extract_unsigned_field(pfr1, 634 ID_AA64PFR1_BT_SHIFT); 635 } 636 637 /* 638 * Create fine-grained mappings for the kernel. 639 */ 640 static void __init map_kernel(pgd_t *pgdp) 641 { 642 static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext, 643 vmlinux_initdata, vmlinux_data; 644 645 /* 646 * External debuggers may need to write directly to the text 647 * mapping to install SW breakpoints. Allow this (only) when 648 * explicitly requested with rodata=off. 649 */ 650 pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 651 652 /* 653 * If we have a CPU that supports BTI and a kernel built for 654 * BTI then mark the kernel executable text as guarded pages 655 * now so we don't have to rewrite the page tables later. 656 */ 657 if (arm64_early_this_cpu_has_bti()) 658 text_prot = __pgprot_modify(text_prot, PTE_GP, PTE_GP); 659 660 /* 661 * Only rodata will be remapped with different permissions later on, 662 * all other segments are allowed to use contiguous mappings. 663 */ 664 map_kernel_segment(pgdp, _stext, _etext, text_prot, &vmlinux_text, 0, 665 VM_NO_GUARD); 666 map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL, 667 &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD); 668 map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot, 669 &vmlinux_inittext, 0, VM_NO_GUARD); 670 map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL, 671 &vmlinux_initdata, 0, VM_NO_GUARD); 672 map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0); 673 674 if (!READ_ONCE(pgd_val(*pgd_offset_pgd(pgdp, FIXADDR_START)))) { 675 /* 676 * The fixmap falls in a separate pgd to the kernel, and doesn't 677 * live in the carveout for the swapper_pg_dir. We can simply 678 * re-use the existing dir for the fixmap. 679 */ 680 set_pgd(pgd_offset_pgd(pgdp, FIXADDR_START), 681 READ_ONCE(*pgd_offset_k(FIXADDR_START))); 682 } else if (CONFIG_PGTABLE_LEVELS > 3) { 683 pgd_t *bm_pgdp; 684 p4d_t *bm_p4dp; 685 pud_t *bm_pudp; 686 /* 687 * The fixmap shares its top level pgd entry with the kernel 688 * mapping. This can really only occur when we are running 689 * with 16k/4 levels, so we can simply reuse the pud level 690 * entry instead. 691 */ 692 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES)); 693 bm_pgdp = pgd_offset_pgd(pgdp, FIXADDR_START); 694 bm_p4dp = p4d_offset(bm_pgdp, FIXADDR_START); 695 bm_pudp = pud_set_fixmap_offset(bm_p4dp, FIXADDR_START); 696 pud_populate(&init_mm, bm_pudp, lm_alias(bm_pmd)); 697 pud_clear_fixmap(); 698 } else { 699 BUG(); 700 } 701 702 kasan_copy_shadow(pgdp); 703 } 704 705 void __init paging_init(void) 706 { 707 pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir)); 708 709 map_kernel(pgdp); 710 map_mem(pgdp); 711 712 pgd_clear_fixmap(); 713 714 cpu_replace_ttbr1(lm_alias(swapper_pg_dir)); 715 init_mm.pgd = swapper_pg_dir; 716 717 memblock_free(__pa_symbol(init_pg_dir), 718 __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir)); 719 720 memblock_allow_resize(); 721 } 722 723 /* 724 * Check whether a kernel address is valid (derived from arch/x86/). 725 */ 726 int kern_addr_valid(unsigned long addr) 727 { 728 pgd_t *pgdp; 729 p4d_t *p4dp; 730 pud_t *pudp, pud; 731 pmd_t *pmdp, pmd; 732 pte_t *ptep, pte; 733 734 addr = arch_kasan_reset_tag(addr); 735 if ((((long)addr) >> VA_BITS) != -1UL) 736 return 0; 737 738 pgdp = pgd_offset_k(addr); 739 if (pgd_none(READ_ONCE(*pgdp))) 740 return 0; 741 742 p4dp = p4d_offset(pgdp, addr); 743 if (p4d_none(READ_ONCE(*p4dp))) 744 return 0; 745 746 pudp = pud_offset(p4dp, addr); 747 pud = READ_ONCE(*pudp); 748 if (pud_none(pud)) 749 return 0; 750 751 if (pud_sect(pud)) 752 return pfn_valid(pud_pfn(pud)); 753 754 pmdp = pmd_offset(pudp, addr); 755 pmd = READ_ONCE(*pmdp); 756 if (pmd_none(pmd)) 757 return 0; 758 759 if (pmd_sect(pmd)) 760 return pfn_valid(pmd_pfn(pmd)); 761 762 ptep = pte_offset_kernel(pmdp, addr); 763 pte = READ_ONCE(*ptep); 764 if (pte_none(pte)) 765 return 0; 766 767 return pfn_valid(pte_pfn(pte)); 768 } 769 770 #ifdef CONFIG_MEMORY_HOTPLUG 771 static void free_hotplug_page_range(struct page *page, size_t size, 772 struct vmem_altmap *altmap) 773 { 774 if (altmap) { 775 vmem_altmap_free(altmap, size >> PAGE_SHIFT); 776 } else { 777 WARN_ON(PageReserved(page)); 778 free_pages((unsigned long)page_address(page), get_order(size)); 779 } 780 } 781 782 static void free_hotplug_pgtable_page(struct page *page) 783 { 784 free_hotplug_page_range(page, PAGE_SIZE, NULL); 785 } 786 787 static bool pgtable_range_aligned(unsigned long start, unsigned long end, 788 unsigned long floor, unsigned long ceiling, 789 unsigned long mask) 790 { 791 start &= mask; 792 if (start < floor) 793 return false; 794 795 if (ceiling) { 796 ceiling &= mask; 797 if (!ceiling) 798 return false; 799 } 800 801 if (end - 1 > ceiling - 1) 802 return false; 803 return true; 804 } 805 806 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr, 807 unsigned long end, bool free_mapped, 808 struct vmem_altmap *altmap) 809 { 810 pte_t *ptep, pte; 811 812 do { 813 ptep = pte_offset_kernel(pmdp, addr); 814 pte = READ_ONCE(*ptep); 815 if (pte_none(pte)) 816 continue; 817 818 WARN_ON(!pte_present(pte)); 819 pte_clear(&init_mm, addr, ptep); 820 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 821 if (free_mapped) 822 free_hotplug_page_range(pte_page(pte), 823 PAGE_SIZE, altmap); 824 } while (addr += PAGE_SIZE, addr < end); 825 } 826 827 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr, 828 unsigned long end, bool free_mapped, 829 struct vmem_altmap *altmap) 830 { 831 unsigned long next; 832 pmd_t *pmdp, pmd; 833 834 do { 835 next = pmd_addr_end(addr, end); 836 pmdp = pmd_offset(pudp, addr); 837 pmd = READ_ONCE(*pmdp); 838 if (pmd_none(pmd)) 839 continue; 840 841 WARN_ON(!pmd_present(pmd)); 842 if (pmd_sect(pmd)) { 843 pmd_clear(pmdp); 844 845 /* 846 * One TLBI should be sufficient here as the PMD_SIZE 847 * range is mapped with a single block entry. 848 */ 849 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 850 if (free_mapped) 851 free_hotplug_page_range(pmd_page(pmd), 852 PMD_SIZE, altmap); 853 continue; 854 } 855 WARN_ON(!pmd_table(pmd)); 856 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap); 857 } while (addr = next, addr < end); 858 } 859 860 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr, 861 unsigned long end, bool free_mapped, 862 struct vmem_altmap *altmap) 863 { 864 unsigned long next; 865 pud_t *pudp, pud; 866 867 do { 868 next = pud_addr_end(addr, end); 869 pudp = pud_offset(p4dp, addr); 870 pud = READ_ONCE(*pudp); 871 if (pud_none(pud)) 872 continue; 873 874 WARN_ON(!pud_present(pud)); 875 if (pud_sect(pud)) { 876 pud_clear(pudp); 877 878 /* 879 * One TLBI should be sufficient here as the PUD_SIZE 880 * range is mapped with a single block entry. 881 */ 882 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 883 if (free_mapped) 884 free_hotplug_page_range(pud_page(pud), 885 PUD_SIZE, altmap); 886 continue; 887 } 888 WARN_ON(!pud_table(pud)); 889 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap); 890 } while (addr = next, addr < end); 891 } 892 893 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr, 894 unsigned long end, bool free_mapped, 895 struct vmem_altmap *altmap) 896 { 897 unsigned long next; 898 p4d_t *p4dp, p4d; 899 900 do { 901 next = p4d_addr_end(addr, end); 902 p4dp = p4d_offset(pgdp, addr); 903 p4d = READ_ONCE(*p4dp); 904 if (p4d_none(p4d)) 905 continue; 906 907 WARN_ON(!p4d_present(p4d)); 908 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap); 909 } while (addr = next, addr < end); 910 } 911 912 static void unmap_hotplug_range(unsigned long addr, unsigned long end, 913 bool free_mapped, struct vmem_altmap *altmap) 914 { 915 unsigned long next; 916 pgd_t *pgdp, pgd; 917 918 /* 919 * altmap can only be used as vmemmap mapping backing memory. 920 * In case the backing memory itself is not being freed, then 921 * altmap is irrelevant. Warn about this inconsistency when 922 * encountered. 923 */ 924 WARN_ON(!free_mapped && altmap); 925 926 do { 927 next = pgd_addr_end(addr, end); 928 pgdp = pgd_offset_k(addr); 929 pgd = READ_ONCE(*pgdp); 930 if (pgd_none(pgd)) 931 continue; 932 933 WARN_ON(!pgd_present(pgd)); 934 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap); 935 } while (addr = next, addr < end); 936 } 937 938 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr, 939 unsigned long end, unsigned long floor, 940 unsigned long ceiling) 941 { 942 pte_t *ptep, pte; 943 unsigned long i, start = addr; 944 945 do { 946 ptep = pte_offset_kernel(pmdp, addr); 947 pte = READ_ONCE(*ptep); 948 949 /* 950 * This is just a sanity check here which verifies that 951 * pte clearing has been done by earlier unmap loops. 952 */ 953 WARN_ON(!pte_none(pte)); 954 } while (addr += PAGE_SIZE, addr < end); 955 956 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK)) 957 return; 958 959 /* 960 * Check whether we can free the pte page if the rest of the 961 * entries are empty. Overlap with other regions have been 962 * handled by the floor/ceiling check. 963 */ 964 ptep = pte_offset_kernel(pmdp, 0UL); 965 for (i = 0; i < PTRS_PER_PTE; i++) { 966 if (!pte_none(READ_ONCE(ptep[i]))) 967 return; 968 } 969 970 pmd_clear(pmdp); 971 __flush_tlb_kernel_pgtable(start); 972 free_hotplug_pgtable_page(virt_to_page(ptep)); 973 } 974 975 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr, 976 unsigned long end, unsigned long floor, 977 unsigned long ceiling) 978 { 979 pmd_t *pmdp, pmd; 980 unsigned long i, next, start = addr; 981 982 do { 983 next = pmd_addr_end(addr, end); 984 pmdp = pmd_offset(pudp, addr); 985 pmd = READ_ONCE(*pmdp); 986 if (pmd_none(pmd)) 987 continue; 988 989 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd)); 990 free_empty_pte_table(pmdp, addr, next, floor, ceiling); 991 } while (addr = next, addr < end); 992 993 if (CONFIG_PGTABLE_LEVELS <= 2) 994 return; 995 996 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK)) 997 return; 998 999 /* 1000 * Check whether we can free the pmd page if the rest of the 1001 * entries are empty. Overlap with other regions have been 1002 * handled by the floor/ceiling check. 1003 */ 1004 pmdp = pmd_offset(pudp, 0UL); 1005 for (i = 0; i < PTRS_PER_PMD; i++) { 1006 if (!pmd_none(READ_ONCE(pmdp[i]))) 1007 return; 1008 } 1009 1010 pud_clear(pudp); 1011 __flush_tlb_kernel_pgtable(start); 1012 free_hotplug_pgtable_page(virt_to_page(pmdp)); 1013 } 1014 1015 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr, 1016 unsigned long end, unsigned long floor, 1017 unsigned long ceiling) 1018 { 1019 pud_t *pudp, pud; 1020 unsigned long i, next, start = addr; 1021 1022 do { 1023 next = pud_addr_end(addr, end); 1024 pudp = pud_offset(p4dp, addr); 1025 pud = READ_ONCE(*pudp); 1026 if (pud_none(pud)) 1027 continue; 1028 1029 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud)); 1030 free_empty_pmd_table(pudp, addr, next, floor, ceiling); 1031 } while (addr = next, addr < end); 1032 1033 if (CONFIG_PGTABLE_LEVELS <= 3) 1034 return; 1035 1036 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK)) 1037 return; 1038 1039 /* 1040 * Check whether we can free the pud page if the rest of the 1041 * entries are empty. Overlap with other regions have been 1042 * handled by the floor/ceiling check. 1043 */ 1044 pudp = pud_offset(p4dp, 0UL); 1045 for (i = 0; i < PTRS_PER_PUD; i++) { 1046 if (!pud_none(READ_ONCE(pudp[i]))) 1047 return; 1048 } 1049 1050 p4d_clear(p4dp); 1051 __flush_tlb_kernel_pgtable(start); 1052 free_hotplug_pgtable_page(virt_to_page(pudp)); 1053 } 1054 1055 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr, 1056 unsigned long end, unsigned long floor, 1057 unsigned long ceiling) 1058 { 1059 unsigned long next; 1060 p4d_t *p4dp, p4d; 1061 1062 do { 1063 next = p4d_addr_end(addr, end); 1064 p4dp = p4d_offset(pgdp, addr); 1065 p4d = READ_ONCE(*p4dp); 1066 if (p4d_none(p4d)) 1067 continue; 1068 1069 WARN_ON(!p4d_present(p4d)); 1070 free_empty_pud_table(p4dp, addr, next, floor, ceiling); 1071 } while (addr = next, addr < end); 1072 } 1073 1074 static void free_empty_tables(unsigned long addr, unsigned long end, 1075 unsigned long floor, unsigned long ceiling) 1076 { 1077 unsigned long next; 1078 pgd_t *pgdp, pgd; 1079 1080 do { 1081 next = pgd_addr_end(addr, end); 1082 pgdp = pgd_offset_k(addr); 1083 pgd = READ_ONCE(*pgdp); 1084 if (pgd_none(pgd)) 1085 continue; 1086 1087 WARN_ON(!pgd_present(pgd)); 1088 free_empty_p4d_table(pgdp, addr, next, floor, ceiling); 1089 } while (addr = next, addr < end); 1090 } 1091 #endif 1092 1093 #ifdef CONFIG_SPARSEMEM_VMEMMAP 1094 #if !ARM64_SWAPPER_USES_SECTION_MAPS 1095 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1096 struct vmem_altmap *altmap) 1097 { 1098 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1099 return vmemmap_populate_basepages(start, end, node, altmap); 1100 } 1101 #else /* !ARM64_SWAPPER_USES_SECTION_MAPS */ 1102 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1103 struct vmem_altmap *altmap) 1104 { 1105 unsigned long addr = start; 1106 unsigned long next; 1107 pgd_t *pgdp; 1108 p4d_t *p4dp; 1109 pud_t *pudp; 1110 pmd_t *pmdp; 1111 1112 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1113 do { 1114 next = pmd_addr_end(addr, end); 1115 1116 pgdp = vmemmap_pgd_populate(addr, node); 1117 if (!pgdp) 1118 return -ENOMEM; 1119 1120 p4dp = vmemmap_p4d_populate(pgdp, addr, node); 1121 if (!p4dp) 1122 return -ENOMEM; 1123 1124 pudp = vmemmap_pud_populate(p4dp, addr, node); 1125 if (!pudp) 1126 return -ENOMEM; 1127 1128 pmdp = pmd_offset(pudp, addr); 1129 if (pmd_none(READ_ONCE(*pmdp))) { 1130 void *p = NULL; 1131 1132 p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap); 1133 if (!p) { 1134 if (vmemmap_populate_basepages(addr, next, node, altmap)) 1135 return -ENOMEM; 1136 continue; 1137 } 1138 1139 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL)); 1140 } else 1141 vmemmap_verify((pte_t *)pmdp, node, addr, next); 1142 } while (addr = next, addr != end); 1143 1144 return 0; 1145 } 1146 #endif /* !ARM64_SWAPPER_USES_SECTION_MAPS */ 1147 void vmemmap_free(unsigned long start, unsigned long end, 1148 struct vmem_altmap *altmap) 1149 { 1150 #ifdef CONFIG_MEMORY_HOTPLUG 1151 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1152 1153 unmap_hotplug_range(start, end, true, altmap); 1154 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END); 1155 #endif 1156 } 1157 #endif /* CONFIG_SPARSEMEM_VMEMMAP */ 1158 1159 static inline pud_t *fixmap_pud(unsigned long addr) 1160 { 1161 pgd_t *pgdp = pgd_offset_k(addr); 1162 p4d_t *p4dp = p4d_offset(pgdp, addr); 1163 p4d_t p4d = READ_ONCE(*p4dp); 1164 1165 BUG_ON(p4d_none(p4d) || p4d_bad(p4d)); 1166 1167 return pud_offset_kimg(p4dp, addr); 1168 } 1169 1170 static inline pmd_t *fixmap_pmd(unsigned long addr) 1171 { 1172 pud_t *pudp = fixmap_pud(addr); 1173 pud_t pud = READ_ONCE(*pudp); 1174 1175 BUG_ON(pud_none(pud) || pud_bad(pud)); 1176 1177 return pmd_offset_kimg(pudp, addr); 1178 } 1179 1180 static inline pte_t *fixmap_pte(unsigned long addr) 1181 { 1182 return &bm_pte[pte_index(addr)]; 1183 } 1184 1185 /* 1186 * The p*d_populate functions call virt_to_phys implicitly so they can't be used 1187 * directly on kernel symbols (bm_p*d). This function is called too early to use 1188 * lm_alias so __p*d_populate functions must be used to populate with the 1189 * physical address from __pa_symbol. 1190 */ 1191 void __init early_fixmap_init(void) 1192 { 1193 pgd_t *pgdp; 1194 p4d_t *p4dp, p4d; 1195 pud_t *pudp; 1196 pmd_t *pmdp; 1197 unsigned long addr = FIXADDR_START; 1198 1199 pgdp = pgd_offset_k(addr); 1200 p4dp = p4d_offset(pgdp, addr); 1201 p4d = READ_ONCE(*p4dp); 1202 if (CONFIG_PGTABLE_LEVELS > 3 && 1203 !(p4d_none(p4d) || p4d_page_paddr(p4d) == __pa_symbol(bm_pud))) { 1204 /* 1205 * We only end up here if the kernel mapping and the fixmap 1206 * share the top level pgd entry, which should only happen on 1207 * 16k/4 levels configurations. 1208 */ 1209 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES)); 1210 pudp = pud_offset_kimg(p4dp, addr); 1211 } else { 1212 if (p4d_none(p4d)) 1213 __p4d_populate(p4dp, __pa_symbol(bm_pud), PUD_TYPE_TABLE); 1214 pudp = fixmap_pud(addr); 1215 } 1216 if (pud_none(READ_ONCE(*pudp))) 1217 __pud_populate(pudp, __pa_symbol(bm_pmd), PMD_TYPE_TABLE); 1218 pmdp = fixmap_pmd(addr); 1219 __pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE); 1220 1221 /* 1222 * The boot-ioremap range spans multiple pmds, for which 1223 * we are not prepared: 1224 */ 1225 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT) 1226 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT)); 1227 1228 if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN))) 1229 || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) { 1230 WARN_ON(1); 1231 pr_warn("pmdp %p != %p, %p\n", 1232 pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)), 1233 fixmap_pmd(fix_to_virt(FIX_BTMAP_END))); 1234 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n", 1235 fix_to_virt(FIX_BTMAP_BEGIN)); 1236 pr_warn("fix_to_virt(FIX_BTMAP_END): %08lx\n", 1237 fix_to_virt(FIX_BTMAP_END)); 1238 1239 pr_warn("FIX_BTMAP_END: %d\n", FIX_BTMAP_END); 1240 pr_warn("FIX_BTMAP_BEGIN: %d\n", FIX_BTMAP_BEGIN); 1241 } 1242 } 1243 1244 /* 1245 * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we 1246 * ever need to use IPIs for TLB broadcasting, then we're in trouble here. 1247 */ 1248 void __set_fixmap(enum fixed_addresses idx, 1249 phys_addr_t phys, pgprot_t flags) 1250 { 1251 unsigned long addr = __fix_to_virt(idx); 1252 pte_t *ptep; 1253 1254 BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses); 1255 1256 ptep = fixmap_pte(addr); 1257 1258 if (pgprot_val(flags)) { 1259 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags)); 1260 } else { 1261 pte_clear(&init_mm, addr, ptep); 1262 flush_tlb_kernel_range(addr, addr+PAGE_SIZE); 1263 } 1264 } 1265 1266 void *__init fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot) 1267 { 1268 const u64 dt_virt_base = __fix_to_virt(FIX_FDT); 1269 int offset; 1270 void *dt_virt; 1271 1272 /* 1273 * Check whether the physical FDT address is set and meets the minimum 1274 * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be 1275 * at least 8 bytes so that we can always access the magic and size 1276 * fields of the FDT header after mapping the first chunk, double check 1277 * here if that is indeed the case. 1278 */ 1279 BUILD_BUG_ON(MIN_FDT_ALIGN < 8); 1280 if (!dt_phys || dt_phys % MIN_FDT_ALIGN) 1281 return NULL; 1282 1283 /* 1284 * Make sure that the FDT region can be mapped without the need to 1285 * allocate additional translation table pages, so that it is safe 1286 * to call create_mapping_noalloc() this early. 1287 * 1288 * On 64k pages, the FDT will be mapped using PTEs, so we need to 1289 * be in the same PMD as the rest of the fixmap. 1290 * On 4k pages, we'll use section mappings for the FDT so we only 1291 * have to be in the same PUD. 1292 */ 1293 BUILD_BUG_ON(dt_virt_base % SZ_2M); 1294 1295 BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT != 1296 __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT); 1297 1298 offset = dt_phys % SWAPPER_BLOCK_SIZE; 1299 dt_virt = (void *)dt_virt_base + offset; 1300 1301 /* map the first chunk so we can read the size from the header */ 1302 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), 1303 dt_virt_base, SWAPPER_BLOCK_SIZE, prot); 1304 1305 if (fdt_magic(dt_virt) != FDT_MAGIC) 1306 return NULL; 1307 1308 *size = fdt_totalsize(dt_virt); 1309 if (*size > MAX_FDT_SIZE) 1310 return NULL; 1311 1312 if (offset + *size > SWAPPER_BLOCK_SIZE) 1313 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base, 1314 round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot); 1315 1316 return dt_virt; 1317 } 1318 1319 int __init arch_ioremap_p4d_supported(void) 1320 { 1321 return 0; 1322 } 1323 1324 int __init arch_ioremap_pud_supported(void) 1325 { 1326 /* 1327 * Only 4k granule supports level 1 block mappings. 1328 * SW table walks can't handle removal of intermediate entries. 1329 */ 1330 return IS_ENABLED(CONFIG_ARM64_4K_PAGES) && 1331 !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS); 1332 } 1333 1334 int __init arch_ioremap_pmd_supported(void) 1335 { 1336 /* See arch_ioremap_pud_supported() */ 1337 return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS); 1338 } 1339 1340 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot) 1341 { 1342 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot)); 1343 1344 /* Only allow permission changes for now */ 1345 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)), 1346 pud_val(new_pud))) 1347 return 0; 1348 1349 VM_BUG_ON(phys & ~PUD_MASK); 1350 set_pud(pudp, new_pud); 1351 return 1; 1352 } 1353 1354 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot) 1355 { 1356 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot)); 1357 1358 /* Only allow permission changes for now */ 1359 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)), 1360 pmd_val(new_pmd))) 1361 return 0; 1362 1363 VM_BUG_ON(phys & ~PMD_MASK); 1364 set_pmd(pmdp, new_pmd); 1365 return 1; 1366 } 1367 1368 int pud_clear_huge(pud_t *pudp) 1369 { 1370 if (!pud_sect(READ_ONCE(*pudp))) 1371 return 0; 1372 pud_clear(pudp); 1373 return 1; 1374 } 1375 1376 int pmd_clear_huge(pmd_t *pmdp) 1377 { 1378 if (!pmd_sect(READ_ONCE(*pmdp))) 1379 return 0; 1380 pmd_clear(pmdp); 1381 return 1; 1382 } 1383 1384 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr) 1385 { 1386 pte_t *table; 1387 pmd_t pmd; 1388 1389 pmd = READ_ONCE(*pmdp); 1390 1391 if (!pmd_table(pmd)) { 1392 VM_WARN_ON(1); 1393 return 1; 1394 } 1395 1396 table = pte_offset_kernel(pmdp, addr); 1397 pmd_clear(pmdp); 1398 __flush_tlb_kernel_pgtable(addr); 1399 pte_free_kernel(NULL, table); 1400 return 1; 1401 } 1402 1403 int pud_free_pmd_page(pud_t *pudp, unsigned long addr) 1404 { 1405 pmd_t *table; 1406 pmd_t *pmdp; 1407 pud_t pud; 1408 unsigned long next, end; 1409 1410 pud = READ_ONCE(*pudp); 1411 1412 if (!pud_table(pud)) { 1413 VM_WARN_ON(1); 1414 return 1; 1415 } 1416 1417 table = pmd_offset(pudp, addr); 1418 pmdp = table; 1419 next = addr; 1420 end = addr + PUD_SIZE; 1421 do { 1422 pmd_free_pte_page(pmdp, next); 1423 } while (pmdp++, next += PMD_SIZE, next != end); 1424 1425 pud_clear(pudp); 1426 __flush_tlb_kernel_pgtable(addr); 1427 pmd_free(NULL, table); 1428 return 1; 1429 } 1430 1431 int p4d_free_pud_page(p4d_t *p4d, unsigned long addr) 1432 { 1433 return 0; /* Don't attempt a block mapping */ 1434 } 1435 1436 #ifdef CONFIG_MEMORY_HOTPLUG 1437 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size) 1438 { 1439 unsigned long end = start + size; 1440 1441 WARN_ON(pgdir != init_mm.pgd); 1442 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END)); 1443 1444 unmap_hotplug_range(start, end, false, NULL); 1445 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END); 1446 } 1447 1448 struct range arch_get_mappable_range(void) 1449 { 1450 struct range mhp_range; 1451 u64 start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual)); 1452 u64 end_linear_pa = __pa(PAGE_END - 1); 1453 1454 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 1455 /* 1456 * Check for a wrap, it is possible because of randomized linear 1457 * mapping the start physical address is actually bigger than 1458 * the end physical address. In this case set start to zero 1459 * because [0, end_linear_pa] range must still be able to cover 1460 * all addressable physical addresses. 1461 */ 1462 if (start_linear_pa > end_linear_pa) 1463 start_linear_pa = 0; 1464 } 1465 1466 WARN_ON(start_linear_pa > end_linear_pa); 1467 1468 /* 1469 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)] 1470 * accommodating both its ends but excluding PAGE_END. Max physical 1471 * range which can be mapped inside this linear mapping range, must 1472 * also be derived from its end points. 1473 */ 1474 mhp_range.start = start_linear_pa; 1475 mhp_range.end = end_linear_pa; 1476 1477 return mhp_range; 1478 } 1479 1480 int arch_add_memory(int nid, u64 start, u64 size, 1481 struct mhp_params *params) 1482 { 1483 int ret, flags = 0; 1484 1485 VM_BUG_ON(!mhp_range_allowed(start, size, true)); 1486 1487 /* 1488 * KFENCE requires linear map to be mapped at page granularity, so that 1489 * it is possible to protect/unprotect single pages in the KFENCE pool. 1490 */ 1491 if (rodata_full || debug_pagealloc_enabled() || 1492 IS_ENABLED(CONFIG_KFENCE)) 1493 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1494 1495 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1496 size, params->pgprot, __pgd_pgtable_alloc, 1497 flags); 1498 1499 memblock_clear_nomap(start, size); 1500 1501 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, 1502 params); 1503 if (ret) 1504 __remove_pgd_mapping(swapper_pg_dir, 1505 __phys_to_virt(start), size); 1506 return ret; 1507 } 1508 1509 void arch_remove_memory(int nid, u64 start, u64 size, 1510 struct vmem_altmap *altmap) 1511 { 1512 unsigned long start_pfn = start >> PAGE_SHIFT; 1513 unsigned long nr_pages = size >> PAGE_SHIFT; 1514 1515 __remove_pages(start_pfn, nr_pages, altmap); 1516 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size); 1517 } 1518 1519 /* 1520 * This memory hotplug notifier helps prevent boot memory from being 1521 * inadvertently removed as it blocks pfn range offlining process in 1522 * __offline_pages(). Hence this prevents both offlining as well as 1523 * removal process for boot memory which is initially always online. 1524 * In future if and when boot memory could be removed, this notifier 1525 * should be dropped and free_hotplug_page_range() should handle any 1526 * reserved pages allocated during boot. 1527 */ 1528 static int prevent_bootmem_remove_notifier(struct notifier_block *nb, 1529 unsigned long action, void *data) 1530 { 1531 struct mem_section *ms; 1532 struct memory_notify *arg = data; 1533 unsigned long end_pfn = arg->start_pfn + arg->nr_pages; 1534 unsigned long pfn = arg->start_pfn; 1535 1536 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE)) 1537 return NOTIFY_OK; 1538 1539 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 1540 unsigned long start = PFN_PHYS(pfn); 1541 unsigned long end = start + (1UL << PA_SECTION_SHIFT); 1542 1543 ms = __pfn_to_section(pfn); 1544 if (!early_section(ms)) 1545 continue; 1546 1547 if (action == MEM_GOING_OFFLINE) { 1548 /* 1549 * Boot memory removal is not supported. Prevent 1550 * it via blocking any attempted offline request 1551 * for the boot memory and just report it. 1552 */ 1553 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end); 1554 return NOTIFY_BAD; 1555 } else if (action == MEM_OFFLINE) { 1556 /* 1557 * This should have never happened. Boot memory 1558 * offlining should have been prevented by this 1559 * very notifier. Probably some memory removal 1560 * procedure might have changed which would then 1561 * require further debug. 1562 */ 1563 pr_err("Boot memory [%lx %lx] offlined\n", start, end); 1564 1565 /* 1566 * Core memory hotplug does not process a return 1567 * code from the notifier for MEM_OFFLINE events. 1568 * The error condition has been reported. Return 1569 * from here as if ignored. 1570 */ 1571 return NOTIFY_DONE; 1572 } 1573 } 1574 return NOTIFY_OK; 1575 } 1576 1577 static struct notifier_block prevent_bootmem_remove_nb = { 1578 .notifier_call = prevent_bootmem_remove_notifier, 1579 }; 1580 1581 /* 1582 * This ensures that boot memory sections on the platform are online 1583 * from early boot. Memory sections could not be prevented from being 1584 * offlined, unless for some reason they are not online to begin with. 1585 * This helps validate the basic assumption on which the above memory 1586 * event notifier works to prevent boot memory section offlining and 1587 * its possible removal. 1588 */ 1589 static void validate_bootmem_online(void) 1590 { 1591 phys_addr_t start, end, addr; 1592 struct mem_section *ms; 1593 u64 i; 1594 1595 /* 1596 * Scanning across all memblock might be expensive 1597 * on some big memory systems. Hence enable this 1598 * validation only with DEBUG_VM. 1599 */ 1600 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 1601 return; 1602 1603 for_each_mem_range(i, &start, &end) { 1604 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) { 1605 ms = __pfn_to_section(PHYS_PFN(addr)); 1606 1607 /* 1608 * All memory ranges in the system at this point 1609 * should have been marked as early sections. 1610 */ 1611 WARN_ON(!early_section(ms)); 1612 1613 /* 1614 * Memory notifier mechanism here to prevent boot 1615 * memory offlining depends on the fact that each 1616 * early section memory on the system is initially 1617 * online. Otherwise a given memory section which 1618 * is already offline will be overlooked and can 1619 * be removed completely. Call out such sections. 1620 */ 1621 if (!online_section(ms)) 1622 pr_err("Boot memory [%llx %llx] is offline, can be removed\n", 1623 addr, addr + (1UL << PA_SECTION_SHIFT)); 1624 } 1625 } 1626 } 1627 1628 static int __init prevent_bootmem_remove_init(void) 1629 { 1630 int ret = 0; 1631 1632 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) 1633 return ret; 1634 1635 validate_bootmem_online(); 1636 ret = register_memory_notifier(&prevent_bootmem_remove_nb); 1637 if (ret) 1638 pr_err("%s: Notifier registration failed %d\n", __func__, ret); 1639 1640 return ret; 1641 } 1642 early_initcall(prevent_bootmem_remove_init); 1643 #endif 1644