1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Based on arch/arm/mm/mmu.c 4 * 5 * Copyright (C) 1995-2005 Russell King 6 * Copyright (C) 2012 ARM Ltd. 7 */ 8 9 #include <linux/cache.h> 10 #include <linux/export.h> 11 #include <linux/kernel.h> 12 #include <linux/errno.h> 13 #include <linux/init.h> 14 #include <linux/ioport.h> 15 #include <linux/kexec.h> 16 #include <linux/libfdt.h> 17 #include <linux/mman.h> 18 #include <linux/nodemask.h> 19 #include <linux/memblock.h> 20 #include <linux/memremap.h> 21 #include <linux/memory.h> 22 #include <linux/fs.h> 23 #include <linux/io.h> 24 #include <linux/mm.h> 25 #include <linux/vmalloc.h> 26 #include <linux/set_memory.h> 27 #include <linux/kfence.h> 28 #include <linux/pkeys.h> 29 #include <linux/mm_inline.h> 30 31 #include <asm/barrier.h> 32 #include <asm/cputype.h> 33 #include <asm/fixmap.h> 34 #include <asm/kasan.h> 35 #include <asm/kernel-pgtable.h> 36 #include <asm/sections.h> 37 #include <asm/setup.h> 38 #include <linux/sizes.h> 39 #include <asm/tlb.h> 40 #include <asm/mmu_context.h> 41 #include <asm/ptdump.h> 42 #include <asm/tlbflush.h> 43 #include <asm/pgalloc.h> 44 #include <asm/kfence.h> 45 46 #define NO_BLOCK_MAPPINGS BIT(0) 47 #define NO_CONT_MAPPINGS BIT(1) 48 #define NO_EXEC_MAPPINGS BIT(2) /* assumes FEAT_HPDS is not used */ 49 50 u64 kimage_voffset __ro_after_init; 51 EXPORT_SYMBOL(kimage_voffset); 52 53 u32 __boot_cpu_mode[] = { BOOT_CPU_MODE_EL2, BOOT_CPU_MODE_EL1 }; 54 55 static bool rodata_is_rw __ro_after_init = true; 56 57 /* 58 * The booting CPU updates the failed status @__early_cpu_boot_status, 59 * with MMU turned off. 60 */ 61 long __section(".mmuoff.data.write") __early_cpu_boot_status; 62 63 /* 64 * Empty_zero_page is a special page that is used for zero-initialized data 65 * and COW. 66 */ 67 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss; 68 EXPORT_SYMBOL(empty_zero_page); 69 70 static DEFINE_SPINLOCK(swapper_pgdir_lock); 71 static DEFINE_MUTEX(fixmap_lock); 72 73 void noinstr set_swapper_pgd(pgd_t *pgdp, pgd_t pgd) 74 { 75 pgd_t *fixmap_pgdp; 76 77 /* 78 * Don't bother with the fixmap if swapper_pg_dir is still mapped 79 * writable in the kernel mapping. 80 */ 81 if (rodata_is_rw) { 82 WRITE_ONCE(*pgdp, pgd); 83 dsb(ishst); 84 isb(); 85 return; 86 } 87 88 spin_lock(&swapper_pgdir_lock); 89 fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp)); 90 WRITE_ONCE(*fixmap_pgdp, pgd); 91 /* 92 * We need dsb(ishst) here to ensure the page-table-walker sees 93 * our new entry before set_p?d() returns. The fixmap's 94 * flush_tlb_kernel_range() via clear_fixmap() does this for us. 95 */ 96 pgd_clear_fixmap(); 97 spin_unlock(&swapper_pgdir_lock); 98 } 99 100 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn, 101 unsigned long size, pgprot_t vma_prot) 102 { 103 if (!pfn_is_map_memory(pfn)) 104 return pgprot_noncached(vma_prot); 105 else if (file->f_flags & O_SYNC) 106 return pgprot_writecombine(vma_prot); 107 return vma_prot; 108 } 109 EXPORT_SYMBOL(phys_mem_access_prot); 110 111 static phys_addr_t __init early_pgtable_alloc(enum pgtable_type pgtable_type) 112 { 113 phys_addr_t phys; 114 115 phys = memblock_phys_alloc_range(PAGE_SIZE, PAGE_SIZE, 0, 116 MEMBLOCK_ALLOC_NOLEAKTRACE); 117 if (!phys) 118 panic("Failed to allocate page table page\n"); 119 120 return phys; 121 } 122 123 bool pgattr_change_is_safe(pteval_t old, pteval_t new) 124 { 125 /* 126 * The following mapping attributes may be updated in live 127 * kernel mappings without the need for break-before-make. 128 */ 129 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG | 130 PTE_SWBITS_MASK; 131 132 /* creating or taking down mappings is always safe */ 133 if (!pte_valid(__pte(old)) || !pte_valid(__pte(new))) 134 return true; 135 136 /* A live entry's pfn should not change */ 137 if (pte_pfn(__pte(old)) != pte_pfn(__pte(new))) 138 return false; 139 140 /* live contiguous mappings may not be manipulated at all */ 141 if ((old | new) & PTE_CONT) 142 return false; 143 144 /* Transitioning from Non-Global to Global is unsafe */ 145 if (old & ~new & PTE_NG) 146 return false; 147 148 /* 149 * Changing the memory type between Normal and Normal-Tagged is safe 150 * since Tagged is considered a permission attribute from the 151 * mismatched attribute aliases perspective. 152 */ 153 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 154 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) && 155 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 156 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED))) 157 mask |= PTE_ATTRINDX_MASK; 158 159 return ((old ^ new) & ~mask) == 0; 160 } 161 162 static void init_clear_pgtable(void *table) 163 { 164 clear_page(table); 165 166 /* Ensure the zeroing is observed by page table walks. */ 167 dsb(ishst); 168 } 169 170 static void init_pte(pte_t *ptep, unsigned long addr, unsigned long end, 171 phys_addr_t phys, pgprot_t prot) 172 { 173 do { 174 pte_t old_pte = __ptep_get(ptep); 175 176 /* 177 * Required barriers to make this visible to the table walker 178 * are deferred to the end of alloc_init_cont_pte(). 179 */ 180 __set_pte_nosync(ptep, pfn_pte(__phys_to_pfn(phys), prot)); 181 182 /* 183 * After the PTE entry has been populated once, we 184 * only allow updates to the permission attributes. 185 */ 186 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte), 187 pte_val(__ptep_get(ptep)))); 188 189 phys += PAGE_SIZE; 190 } while (ptep++, addr += PAGE_SIZE, addr != end); 191 } 192 193 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr, 194 unsigned long end, phys_addr_t phys, 195 pgprot_t prot, 196 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 197 int flags) 198 { 199 unsigned long next; 200 pmd_t pmd = READ_ONCE(*pmdp); 201 pte_t *ptep; 202 203 BUG_ON(pmd_sect(pmd)); 204 if (pmd_none(pmd)) { 205 pmdval_t pmdval = PMD_TYPE_TABLE | PMD_TABLE_UXN | PMD_TABLE_AF; 206 phys_addr_t pte_phys; 207 208 if (flags & NO_EXEC_MAPPINGS) 209 pmdval |= PMD_TABLE_PXN; 210 BUG_ON(!pgtable_alloc); 211 pte_phys = pgtable_alloc(TABLE_PTE); 212 ptep = pte_set_fixmap(pte_phys); 213 init_clear_pgtable(ptep); 214 ptep += pte_index(addr); 215 __pmd_populate(pmdp, pte_phys, pmdval); 216 } else { 217 BUG_ON(pmd_bad(pmd)); 218 ptep = pte_set_fixmap_offset(pmdp, addr); 219 } 220 221 do { 222 pgprot_t __prot = prot; 223 224 next = pte_cont_addr_end(addr, end); 225 226 /* use a contiguous mapping if the range is suitably aligned */ 227 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) && 228 (flags & NO_CONT_MAPPINGS) == 0) 229 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 230 231 init_pte(ptep, addr, next, phys, __prot); 232 233 ptep += pte_index(next) - pte_index(addr); 234 phys += next - addr; 235 } while (addr = next, addr != end); 236 237 /* 238 * Note: barriers and maintenance necessary to clear the fixmap slot 239 * ensure that all previous pgtable writes are visible to the table 240 * walker. 241 */ 242 pte_clear_fixmap(); 243 } 244 245 static void init_pmd(pmd_t *pmdp, unsigned long addr, unsigned long end, 246 phys_addr_t phys, pgprot_t prot, 247 phys_addr_t (*pgtable_alloc)(enum pgtable_type), int flags) 248 { 249 unsigned long next; 250 251 do { 252 pmd_t old_pmd = READ_ONCE(*pmdp); 253 254 next = pmd_addr_end(addr, end); 255 256 /* try section mapping first */ 257 if (((addr | next | phys) & ~PMD_MASK) == 0 && 258 (flags & NO_BLOCK_MAPPINGS) == 0) { 259 pmd_set_huge(pmdp, phys, prot); 260 261 /* 262 * After the PMD entry has been populated once, we 263 * only allow updates to the permission attributes. 264 */ 265 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd), 266 READ_ONCE(pmd_val(*pmdp)))); 267 } else { 268 alloc_init_cont_pte(pmdp, addr, next, phys, prot, 269 pgtable_alloc, flags); 270 271 BUG_ON(pmd_val(old_pmd) != 0 && 272 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp))); 273 } 274 phys += next - addr; 275 } while (pmdp++, addr = next, addr != end); 276 } 277 278 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr, 279 unsigned long end, phys_addr_t phys, 280 pgprot_t prot, 281 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 282 int flags) 283 { 284 unsigned long next; 285 pud_t pud = READ_ONCE(*pudp); 286 pmd_t *pmdp; 287 288 /* 289 * Check for initial section mappings in the pgd/pud. 290 */ 291 BUG_ON(pud_sect(pud)); 292 if (pud_none(pud)) { 293 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF; 294 phys_addr_t pmd_phys; 295 296 if (flags & NO_EXEC_MAPPINGS) 297 pudval |= PUD_TABLE_PXN; 298 BUG_ON(!pgtable_alloc); 299 pmd_phys = pgtable_alloc(TABLE_PMD); 300 pmdp = pmd_set_fixmap(pmd_phys); 301 init_clear_pgtable(pmdp); 302 pmdp += pmd_index(addr); 303 __pud_populate(pudp, pmd_phys, pudval); 304 } else { 305 BUG_ON(pud_bad(pud)); 306 pmdp = pmd_set_fixmap_offset(pudp, addr); 307 } 308 309 do { 310 pgprot_t __prot = prot; 311 312 next = pmd_cont_addr_end(addr, end); 313 314 /* use a contiguous mapping if the range is suitably aligned */ 315 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) && 316 (flags & NO_CONT_MAPPINGS) == 0) 317 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 318 319 init_pmd(pmdp, addr, next, phys, __prot, pgtable_alloc, flags); 320 321 pmdp += pmd_index(next) - pmd_index(addr); 322 phys += next - addr; 323 } while (addr = next, addr != end); 324 325 pmd_clear_fixmap(); 326 } 327 328 static void alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end, 329 phys_addr_t phys, pgprot_t prot, 330 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 331 int flags) 332 { 333 unsigned long next; 334 p4d_t p4d = READ_ONCE(*p4dp); 335 pud_t *pudp; 336 337 if (p4d_none(p4d)) { 338 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN | P4D_TABLE_AF; 339 phys_addr_t pud_phys; 340 341 if (flags & NO_EXEC_MAPPINGS) 342 p4dval |= P4D_TABLE_PXN; 343 BUG_ON(!pgtable_alloc); 344 pud_phys = pgtable_alloc(TABLE_PUD); 345 pudp = pud_set_fixmap(pud_phys); 346 init_clear_pgtable(pudp); 347 pudp += pud_index(addr); 348 __p4d_populate(p4dp, pud_phys, p4dval); 349 } else { 350 BUG_ON(p4d_bad(p4d)); 351 pudp = pud_set_fixmap_offset(p4dp, addr); 352 } 353 354 do { 355 pud_t old_pud = READ_ONCE(*pudp); 356 357 next = pud_addr_end(addr, end); 358 359 /* 360 * For 4K granule only, attempt to put down a 1GB block 361 */ 362 if (pud_sect_supported() && 363 ((addr | next | phys) & ~PUD_MASK) == 0 && 364 (flags & NO_BLOCK_MAPPINGS) == 0) { 365 pud_set_huge(pudp, phys, prot); 366 367 /* 368 * After the PUD entry has been populated once, we 369 * only allow updates to the permission attributes. 370 */ 371 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud), 372 READ_ONCE(pud_val(*pudp)))); 373 } else { 374 alloc_init_cont_pmd(pudp, addr, next, phys, prot, 375 pgtable_alloc, flags); 376 377 BUG_ON(pud_val(old_pud) != 0 && 378 pud_val(old_pud) != READ_ONCE(pud_val(*pudp))); 379 } 380 phys += next - addr; 381 } while (pudp++, addr = next, addr != end); 382 383 pud_clear_fixmap(); 384 } 385 386 static void alloc_init_p4d(pgd_t *pgdp, unsigned long addr, unsigned long end, 387 phys_addr_t phys, pgprot_t prot, 388 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 389 int flags) 390 { 391 unsigned long next; 392 pgd_t pgd = READ_ONCE(*pgdp); 393 p4d_t *p4dp; 394 395 if (pgd_none(pgd)) { 396 pgdval_t pgdval = PGD_TYPE_TABLE | PGD_TABLE_UXN | PGD_TABLE_AF; 397 phys_addr_t p4d_phys; 398 399 if (flags & NO_EXEC_MAPPINGS) 400 pgdval |= PGD_TABLE_PXN; 401 BUG_ON(!pgtable_alloc); 402 p4d_phys = pgtable_alloc(TABLE_P4D); 403 p4dp = p4d_set_fixmap(p4d_phys); 404 init_clear_pgtable(p4dp); 405 p4dp += p4d_index(addr); 406 __pgd_populate(pgdp, p4d_phys, pgdval); 407 } else { 408 BUG_ON(pgd_bad(pgd)); 409 p4dp = p4d_set_fixmap_offset(pgdp, addr); 410 } 411 412 do { 413 p4d_t old_p4d = READ_ONCE(*p4dp); 414 415 next = p4d_addr_end(addr, end); 416 417 alloc_init_pud(p4dp, addr, next, phys, prot, 418 pgtable_alloc, flags); 419 420 BUG_ON(p4d_val(old_p4d) != 0 && 421 p4d_val(old_p4d) != READ_ONCE(p4d_val(*p4dp))); 422 423 phys += next - addr; 424 } while (p4dp++, addr = next, addr != end); 425 426 p4d_clear_fixmap(); 427 } 428 429 static void __create_pgd_mapping_locked(pgd_t *pgdir, phys_addr_t phys, 430 unsigned long virt, phys_addr_t size, 431 pgprot_t prot, 432 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 433 int flags) 434 { 435 unsigned long addr, end, next; 436 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt); 437 438 /* 439 * If the virtual and physical address don't have the same offset 440 * within a page, we cannot map the region as the caller expects. 441 */ 442 if (WARN_ON((phys ^ virt) & ~PAGE_MASK)) 443 return; 444 445 phys &= PAGE_MASK; 446 addr = virt & PAGE_MASK; 447 end = PAGE_ALIGN(virt + size); 448 449 do { 450 next = pgd_addr_end(addr, end); 451 alloc_init_p4d(pgdp, addr, next, phys, prot, pgtable_alloc, 452 flags); 453 phys += next - addr; 454 } while (pgdp++, addr = next, addr != end); 455 } 456 457 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, 458 unsigned long virt, phys_addr_t size, 459 pgprot_t prot, 460 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 461 int flags) 462 { 463 mutex_lock(&fixmap_lock); 464 __create_pgd_mapping_locked(pgdir, phys, virt, size, prot, 465 pgtable_alloc, flags); 466 mutex_unlock(&fixmap_lock); 467 } 468 469 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0 470 extern __alias(__create_pgd_mapping_locked) 471 void create_kpti_ng_temp_pgd(pgd_t *pgdir, phys_addr_t phys, unsigned long virt, 472 phys_addr_t size, pgprot_t prot, 473 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 474 int flags); 475 #endif 476 477 static phys_addr_t __pgd_pgtable_alloc(struct mm_struct *mm, 478 enum pgtable_type pgtable_type) 479 { 480 /* Page is zeroed by init_clear_pgtable() so don't duplicate effort. */ 481 struct ptdesc *ptdesc = pagetable_alloc(GFP_PGTABLE_KERNEL & ~__GFP_ZERO, 0); 482 phys_addr_t pa; 483 484 BUG_ON(!ptdesc); 485 pa = page_to_phys(ptdesc_page(ptdesc)); 486 487 switch (pgtable_type) { 488 case TABLE_PTE: 489 BUG_ON(!pagetable_pte_ctor(mm, ptdesc)); 490 break; 491 case TABLE_PMD: 492 BUG_ON(!pagetable_pmd_ctor(mm, ptdesc)); 493 break; 494 case TABLE_PUD: 495 pagetable_pud_ctor(ptdesc); 496 break; 497 case TABLE_P4D: 498 pagetable_p4d_ctor(ptdesc); 499 break; 500 } 501 502 return pa; 503 } 504 505 static phys_addr_t __maybe_unused 506 pgd_pgtable_alloc_init_mm(enum pgtable_type pgtable_type) 507 { 508 return __pgd_pgtable_alloc(&init_mm, pgtable_type); 509 } 510 511 static phys_addr_t 512 pgd_pgtable_alloc_special_mm(enum pgtable_type pgtable_type) 513 { 514 return __pgd_pgtable_alloc(NULL, pgtable_type); 515 } 516 517 /* 518 * This function can only be used to modify existing table entries, 519 * without allocating new levels of table. Note that this permits the 520 * creation of new section or page entries. 521 */ 522 void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt, 523 phys_addr_t size, pgprot_t prot) 524 { 525 if (virt < PAGE_OFFSET) { 526 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n", 527 &phys, virt); 528 return; 529 } 530 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 531 NO_CONT_MAPPINGS); 532 } 533 534 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys, 535 unsigned long virt, phys_addr_t size, 536 pgprot_t prot, bool page_mappings_only) 537 { 538 int flags = 0; 539 540 BUG_ON(mm == &init_mm); 541 542 if (page_mappings_only) 543 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 544 545 __create_pgd_mapping(mm->pgd, phys, virt, size, prot, 546 pgd_pgtable_alloc_special_mm, flags); 547 } 548 549 static void update_mapping_prot(phys_addr_t phys, unsigned long virt, 550 phys_addr_t size, pgprot_t prot) 551 { 552 if (virt < PAGE_OFFSET) { 553 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n", 554 &phys, virt); 555 return; 556 } 557 558 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 559 NO_CONT_MAPPINGS); 560 561 /* flush the TLBs after updating live kernel mappings */ 562 flush_tlb_kernel_range(virt, virt + size); 563 } 564 565 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start, 566 phys_addr_t end, pgprot_t prot, int flags) 567 { 568 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start, 569 prot, early_pgtable_alloc, flags); 570 } 571 572 void __init mark_linear_text_alias_ro(void) 573 { 574 /* 575 * Remove the write permissions from the linear alias of .text/.rodata 576 */ 577 update_mapping_prot(__pa_symbol(_stext), (unsigned long)lm_alias(_stext), 578 (unsigned long)__init_begin - (unsigned long)_stext, 579 PAGE_KERNEL_RO); 580 } 581 582 #ifdef CONFIG_KFENCE 583 584 bool __ro_after_init kfence_early_init = !!CONFIG_KFENCE_SAMPLE_INTERVAL; 585 586 /* early_param() will be parsed before map_mem() below. */ 587 static int __init parse_kfence_early_init(char *arg) 588 { 589 int val; 590 591 if (get_option(&arg, &val)) 592 kfence_early_init = !!val; 593 return 0; 594 } 595 early_param("kfence.sample_interval", parse_kfence_early_init); 596 597 static phys_addr_t __init arm64_kfence_alloc_pool(void) 598 { 599 phys_addr_t kfence_pool; 600 601 if (!kfence_early_init) 602 return 0; 603 604 kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE); 605 if (!kfence_pool) { 606 pr_err("failed to allocate kfence pool\n"); 607 kfence_early_init = false; 608 return 0; 609 } 610 611 /* Temporarily mark as NOMAP. */ 612 memblock_mark_nomap(kfence_pool, KFENCE_POOL_SIZE); 613 614 return kfence_pool; 615 } 616 617 static void __init arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) 618 { 619 if (!kfence_pool) 620 return; 621 622 /* KFENCE pool needs page-level mapping. */ 623 __map_memblock(pgdp, kfence_pool, kfence_pool + KFENCE_POOL_SIZE, 624 pgprot_tagged(PAGE_KERNEL), 625 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS); 626 memblock_clear_nomap(kfence_pool, KFENCE_POOL_SIZE); 627 __kfence_pool = phys_to_virt(kfence_pool); 628 } 629 #else /* CONFIG_KFENCE */ 630 631 static inline phys_addr_t arm64_kfence_alloc_pool(void) { return 0; } 632 static inline void arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) { } 633 634 #endif /* CONFIG_KFENCE */ 635 636 static void __init map_mem(pgd_t *pgdp) 637 { 638 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN); 639 phys_addr_t kernel_start = __pa_symbol(_stext); 640 phys_addr_t kernel_end = __pa_symbol(__init_begin); 641 phys_addr_t start, end; 642 phys_addr_t early_kfence_pool; 643 int flags = NO_EXEC_MAPPINGS; 644 u64 i; 645 646 /* 647 * Setting hierarchical PXNTable attributes on table entries covering 648 * the linear region is only possible if it is guaranteed that no table 649 * entries at any level are being shared between the linear region and 650 * the vmalloc region. Check whether this is true for the PGD level, in 651 * which case it is guaranteed to be true for all other levels as well. 652 * (Unless we are running with support for LPA2, in which case the 653 * entire reduced VA space is covered by a single pgd_t which will have 654 * been populated without the PXNTable attribute by the time we get here.) 655 */ 656 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end) && 657 pgd_index(_PAGE_OFFSET(VA_BITS_MIN)) != PTRS_PER_PGD - 1); 658 659 early_kfence_pool = arm64_kfence_alloc_pool(); 660 661 if (can_set_direct_map()) 662 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 663 664 /* 665 * Take care not to create a writable alias for the 666 * read-only text and rodata sections of the kernel image. 667 * So temporarily mark them as NOMAP to skip mappings in 668 * the following for-loop 669 */ 670 memblock_mark_nomap(kernel_start, kernel_end - kernel_start); 671 672 /* map all the memory banks */ 673 for_each_mem_range(i, &start, &end) { 674 if (start >= end) 675 break; 676 /* 677 * The linear map must allow allocation tags reading/writing 678 * if MTE is present. Otherwise, it has the same attributes as 679 * PAGE_KERNEL. 680 */ 681 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL), 682 flags); 683 } 684 685 /* 686 * Map the linear alias of the [_stext, __init_begin) interval 687 * as non-executable now, and remove the write permission in 688 * mark_linear_text_alias_ro() below (which will be called after 689 * alternative patching has completed). This makes the contents 690 * of the region accessible to subsystems such as hibernate, 691 * but protects it from inadvertent modification or execution. 692 * Note that contiguous mappings cannot be remapped in this way, 693 * so we should avoid them here. 694 */ 695 __map_memblock(pgdp, kernel_start, kernel_end, 696 PAGE_KERNEL, NO_CONT_MAPPINGS); 697 memblock_clear_nomap(kernel_start, kernel_end - kernel_start); 698 arm64_kfence_map_pool(early_kfence_pool, pgdp); 699 } 700 701 void mark_rodata_ro(void) 702 { 703 unsigned long section_size; 704 705 /* 706 * mark .rodata as read only. Use __init_begin rather than __end_rodata 707 * to cover NOTES and EXCEPTION_TABLE. 708 */ 709 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata; 710 WRITE_ONCE(rodata_is_rw, false); 711 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata, 712 section_size, PAGE_KERNEL_RO); 713 } 714 715 static void __init declare_vma(struct vm_struct *vma, 716 void *va_start, void *va_end, 717 unsigned long vm_flags) 718 { 719 phys_addr_t pa_start = __pa_symbol(va_start); 720 unsigned long size = va_end - va_start; 721 722 BUG_ON(!PAGE_ALIGNED(pa_start)); 723 BUG_ON(!PAGE_ALIGNED(size)); 724 725 if (!(vm_flags & VM_NO_GUARD)) 726 size += PAGE_SIZE; 727 728 vma->addr = va_start; 729 vma->phys_addr = pa_start; 730 vma->size = size; 731 vma->flags = VM_MAP | vm_flags; 732 vma->caller = __builtin_return_address(0); 733 734 vm_area_add_early(vma); 735 } 736 737 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0 738 static pgprot_t kernel_exec_prot(void) 739 { 740 return rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 741 } 742 743 static int __init map_entry_trampoline(void) 744 { 745 int i; 746 747 if (!arm64_kernel_unmapped_at_el0()) 748 return 0; 749 750 pgprot_t prot = kernel_exec_prot(); 751 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start); 752 753 /* The trampoline is always mapped and can therefore be global */ 754 pgprot_val(prot) &= ~PTE_NG; 755 756 /* Map only the text into the trampoline page table */ 757 memset(tramp_pg_dir, 0, PGD_SIZE); 758 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, 759 entry_tramp_text_size(), prot, 760 pgd_pgtable_alloc_init_mm, NO_BLOCK_MAPPINGS); 761 762 /* Map both the text and data into the kernel page table */ 763 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++) 764 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 765 pa_start + i * PAGE_SIZE, prot); 766 767 if (IS_ENABLED(CONFIG_RELOCATABLE)) 768 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 769 pa_start + i * PAGE_SIZE, PAGE_KERNEL_RO); 770 771 return 0; 772 } 773 core_initcall(map_entry_trampoline); 774 #endif 775 776 /* 777 * Declare the VMA areas for the kernel 778 */ 779 static void __init declare_kernel_vmas(void) 780 { 781 static struct vm_struct vmlinux_seg[KERNEL_SEGMENT_COUNT]; 782 783 declare_vma(&vmlinux_seg[0], _stext, _etext, VM_NO_GUARD); 784 declare_vma(&vmlinux_seg[1], __start_rodata, __inittext_begin, VM_NO_GUARD); 785 declare_vma(&vmlinux_seg[2], __inittext_begin, __inittext_end, VM_NO_GUARD); 786 declare_vma(&vmlinux_seg[3], __initdata_begin, __initdata_end, VM_NO_GUARD); 787 declare_vma(&vmlinux_seg[4], _data, _end, 0); 788 } 789 790 void __pi_map_range(u64 *pgd, u64 start, u64 end, u64 pa, pgprot_t prot, 791 int level, pte_t *tbl, bool may_use_cont, u64 va_offset); 792 793 static u8 idmap_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init, 794 kpti_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init; 795 796 static void __init create_idmap(void) 797 { 798 u64 start = __pa_symbol(__idmap_text_start); 799 u64 end = __pa_symbol(__idmap_text_end); 800 u64 ptep = __pa_symbol(idmap_ptes); 801 802 __pi_map_range(&ptep, start, end, start, PAGE_KERNEL_ROX, 803 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 804 __phys_to_virt(ptep) - ptep); 805 806 if (IS_ENABLED(CONFIG_UNMAP_KERNEL_AT_EL0) && !arm64_use_ng_mappings) { 807 extern u32 __idmap_kpti_flag; 808 u64 pa = __pa_symbol(&__idmap_kpti_flag); 809 810 /* 811 * The KPTI G-to-nG conversion code needs a read-write mapping 812 * of its synchronization flag in the ID map. 813 */ 814 ptep = __pa_symbol(kpti_ptes); 815 __pi_map_range(&ptep, pa, pa + sizeof(u32), pa, PAGE_KERNEL, 816 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 817 __phys_to_virt(ptep) - ptep); 818 } 819 } 820 821 void __init paging_init(void) 822 { 823 map_mem(swapper_pg_dir); 824 825 memblock_allow_resize(); 826 827 create_idmap(); 828 declare_kernel_vmas(); 829 } 830 831 #ifdef CONFIG_MEMORY_HOTPLUG 832 static void free_hotplug_page_range(struct page *page, size_t size, 833 struct vmem_altmap *altmap) 834 { 835 if (altmap) { 836 vmem_altmap_free(altmap, size >> PAGE_SHIFT); 837 } else { 838 WARN_ON(PageReserved(page)); 839 free_pages((unsigned long)page_address(page), get_order(size)); 840 } 841 } 842 843 static void free_hotplug_pgtable_page(struct page *page) 844 { 845 free_hotplug_page_range(page, PAGE_SIZE, NULL); 846 } 847 848 static bool pgtable_range_aligned(unsigned long start, unsigned long end, 849 unsigned long floor, unsigned long ceiling, 850 unsigned long mask) 851 { 852 start &= mask; 853 if (start < floor) 854 return false; 855 856 if (ceiling) { 857 ceiling &= mask; 858 if (!ceiling) 859 return false; 860 } 861 862 if (end - 1 > ceiling - 1) 863 return false; 864 return true; 865 } 866 867 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr, 868 unsigned long end, bool free_mapped, 869 struct vmem_altmap *altmap) 870 { 871 pte_t *ptep, pte; 872 873 do { 874 ptep = pte_offset_kernel(pmdp, addr); 875 pte = __ptep_get(ptep); 876 if (pte_none(pte)) 877 continue; 878 879 WARN_ON(!pte_present(pte)); 880 __pte_clear(&init_mm, addr, ptep); 881 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 882 if (free_mapped) 883 free_hotplug_page_range(pte_page(pte), 884 PAGE_SIZE, altmap); 885 } while (addr += PAGE_SIZE, addr < end); 886 } 887 888 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr, 889 unsigned long end, bool free_mapped, 890 struct vmem_altmap *altmap) 891 { 892 unsigned long next; 893 pmd_t *pmdp, pmd; 894 895 do { 896 next = pmd_addr_end(addr, end); 897 pmdp = pmd_offset(pudp, addr); 898 pmd = READ_ONCE(*pmdp); 899 if (pmd_none(pmd)) 900 continue; 901 902 WARN_ON(!pmd_present(pmd)); 903 if (pmd_sect(pmd)) { 904 pmd_clear(pmdp); 905 906 /* 907 * One TLBI should be sufficient here as the PMD_SIZE 908 * range is mapped with a single block entry. 909 */ 910 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 911 if (free_mapped) 912 free_hotplug_page_range(pmd_page(pmd), 913 PMD_SIZE, altmap); 914 continue; 915 } 916 WARN_ON(!pmd_table(pmd)); 917 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap); 918 } while (addr = next, addr < end); 919 } 920 921 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr, 922 unsigned long end, bool free_mapped, 923 struct vmem_altmap *altmap) 924 { 925 unsigned long next; 926 pud_t *pudp, pud; 927 928 do { 929 next = pud_addr_end(addr, end); 930 pudp = pud_offset(p4dp, addr); 931 pud = READ_ONCE(*pudp); 932 if (pud_none(pud)) 933 continue; 934 935 WARN_ON(!pud_present(pud)); 936 if (pud_sect(pud)) { 937 pud_clear(pudp); 938 939 /* 940 * One TLBI should be sufficient here as the PUD_SIZE 941 * range is mapped with a single block entry. 942 */ 943 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 944 if (free_mapped) 945 free_hotplug_page_range(pud_page(pud), 946 PUD_SIZE, altmap); 947 continue; 948 } 949 WARN_ON(!pud_table(pud)); 950 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap); 951 } while (addr = next, addr < end); 952 } 953 954 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr, 955 unsigned long end, bool free_mapped, 956 struct vmem_altmap *altmap) 957 { 958 unsigned long next; 959 p4d_t *p4dp, p4d; 960 961 do { 962 next = p4d_addr_end(addr, end); 963 p4dp = p4d_offset(pgdp, addr); 964 p4d = READ_ONCE(*p4dp); 965 if (p4d_none(p4d)) 966 continue; 967 968 WARN_ON(!p4d_present(p4d)); 969 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap); 970 } while (addr = next, addr < end); 971 } 972 973 static void unmap_hotplug_range(unsigned long addr, unsigned long end, 974 bool free_mapped, struct vmem_altmap *altmap) 975 { 976 unsigned long next; 977 pgd_t *pgdp, pgd; 978 979 /* 980 * altmap can only be used as vmemmap mapping backing memory. 981 * In case the backing memory itself is not being freed, then 982 * altmap is irrelevant. Warn about this inconsistency when 983 * encountered. 984 */ 985 WARN_ON(!free_mapped && altmap); 986 987 do { 988 next = pgd_addr_end(addr, end); 989 pgdp = pgd_offset_k(addr); 990 pgd = READ_ONCE(*pgdp); 991 if (pgd_none(pgd)) 992 continue; 993 994 WARN_ON(!pgd_present(pgd)); 995 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap); 996 } while (addr = next, addr < end); 997 } 998 999 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr, 1000 unsigned long end, unsigned long floor, 1001 unsigned long ceiling) 1002 { 1003 pte_t *ptep, pte; 1004 unsigned long i, start = addr; 1005 1006 do { 1007 ptep = pte_offset_kernel(pmdp, addr); 1008 pte = __ptep_get(ptep); 1009 1010 /* 1011 * This is just a sanity check here which verifies that 1012 * pte clearing has been done by earlier unmap loops. 1013 */ 1014 WARN_ON(!pte_none(pte)); 1015 } while (addr += PAGE_SIZE, addr < end); 1016 1017 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK)) 1018 return; 1019 1020 /* 1021 * Check whether we can free the pte page if the rest of the 1022 * entries are empty. Overlap with other regions have been 1023 * handled by the floor/ceiling check. 1024 */ 1025 ptep = pte_offset_kernel(pmdp, 0UL); 1026 for (i = 0; i < PTRS_PER_PTE; i++) { 1027 if (!pte_none(__ptep_get(&ptep[i]))) 1028 return; 1029 } 1030 1031 pmd_clear(pmdp); 1032 __flush_tlb_kernel_pgtable(start); 1033 free_hotplug_pgtable_page(virt_to_page(ptep)); 1034 } 1035 1036 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr, 1037 unsigned long end, unsigned long floor, 1038 unsigned long ceiling) 1039 { 1040 pmd_t *pmdp, pmd; 1041 unsigned long i, next, start = addr; 1042 1043 do { 1044 next = pmd_addr_end(addr, end); 1045 pmdp = pmd_offset(pudp, addr); 1046 pmd = READ_ONCE(*pmdp); 1047 if (pmd_none(pmd)) 1048 continue; 1049 1050 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd)); 1051 free_empty_pte_table(pmdp, addr, next, floor, ceiling); 1052 } while (addr = next, addr < end); 1053 1054 if (CONFIG_PGTABLE_LEVELS <= 2) 1055 return; 1056 1057 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK)) 1058 return; 1059 1060 /* 1061 * Check whether we can free the pmd page if the rest of the 1062 * entries are empty. Overlap with other regions have been 1063 * handled by the floor/ceiling check. 1064 */ 1065 pmdp = pmd_offset(pudp, 0UL); 1066 for (i = 0; i < PTRS_PER_PMD; i++) { 1067 if (!pmd_none(READ_ONCE(pmdp[i]))) 1068 return; 1069 } 1070 1071 pud_clear(pudp); 1072 __flush_tlb_kernel_pgtable(start); 1073 free_hotplug_pgtable_page(virt_to_page(pmdp)); 1074 } 1075 1076 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr, 1077 unsigned long end, unsigned long floor, 1078 unsigned long ceiling) 1079 { 1080 pud_t *pudp, pud; 1081 unsigned long i, next, start = addr; 1082 1083 do { 1084 next = pud_addr_end(addr, end); 1085 pudp = pud_offset(p4dp, addr); 1086 pud = READ_ONCE(*pudp); 1087 if (pud_none(pud)) 1088 continue; 1089 1090 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud)); 1091 free_empty_pmd_table(pudp, addr, next, floor, ceiling); 1092 } while (addr = next, addr < end); 1093 1094 if (!pgtable_l4_enabled()) 1095 return; 1096 1097 if (!pgtable_range_aligned(start, end, floor, ceiling, P4D_MASK)) 1098 return; 1099 1100 /* 1101 * Check whether we can free the pud page if the rest of the 1102 * entries are empty. Overlap with other regions have been 1103 * handled by the floor/ceiling check. 1104 */ 1105 pudp = pud_offset(p4dp, 0UL); 1106 for (i = 0; i < PTRS_PER_PUD; i++) { 1107 if (!pud_none(READ_ONCE(pudp[i]))) 1108 return; 1109 } 1110 1111 p4d_clear(p4dp); 1112 __flush_tlb_kernel_pgtable(start); 1113 free_hotplug_pgtable_page(virt_to_page(pudp)); 1114 } 1115 1116 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr, 1117 unsigned long end, unsigned long floor, 1118 unsigned long ceiling) 1119 { 1120 p4d_t *p4dp, p4d; 1121 unsigned long i, next, start = addr; 1122 1123 do { 1124 next = p4d_addr_end(addr, end); 1125 p4dp = p4d_offset(pgdp, addr); 1126 p4d = READ_ONCE(*p4dp); 1127 if (p4d_none(p4d)) 1128 continue; 1129 1130 WARN_ON(!p4d_present(p4d)); 1131 free_empty_pud_table(p4dp, addr, next, floor, ceiling); 1132 } while (addr = next, addr < end); 1133 1134 if (!pgtable_l5_enabled()) 1135 return; 1136 1137 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK)) 1138 return; 1139 1140 /* 1141 * Check whether we can free the p4d page if the rest of the 1142 * entries are empty. Overlap with other regions have been 1143 * handled by the floor/ceiling check. 1144 */ 1145 p4dp = p4d_offset(pgdp, 0UL); 1146 for (i = 0; i < PTRS_PER_P4D; i++) { 1147 if (!p4d_none(READ_ONCE(p4dp[i]))) 1148 return; 1149 } 1150 1151 pgd_clear(pgdp); 1152 __flush_tlb_kernel_pgtable(start); 1153 free_hotplug_pgtable_page(virt_to_page(p4dp)); 1154 } 1155 1156 static void free_empty_tables(unsigned long addr, unsigned long end, 1157 unsigned long floor, unsigned long ceiling) 1158 { 1159 unsigned long next; 1160 pgd_t *pgdp, pgd; 1161 1162 do { 1163 next = pgd_addr_end(addr, end); 1164 pgdp = pgd_offset_k(addr); 1165 pgd = READ_ONCE(*pgdp); 1166 if (pgd_none(pgd)) 1167 continue; 1168 1169 WARN_ON(!pgd_present(pgd)); 1170 free_empty_p4d_table(pgdp, addr, next, floor, ceiling); 1171 } while (addr = next, addr < end); 1172 } 1173 #endif 1174 1175 void __meminit vmemmap_set_pmd(pmd_t *pmdp, void *p, int node, 1176 unsigned long addr, unsigned long next) 1177 { 1178 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL)); 1179 } 1180 1181 int __meminit vmemmap_check_pmd(pmd_t *pmdp, int node, 1182 unsigned long addr, unsigned long next) 1183 { 1184 vmemmap_verify((pte_t *)pmdp, node, addr, next); 1185 1186 return pmd_sect(READ_ONCE(*pmdp)); 1187 } 1188 1189 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1190 struct vmem_altmap *altmap) 1191 { 1192 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1193 /* [start, end] should be within one section */ 1194 WARN_ON_ONCE(end - start > PAGES_PER_SECTION * sizeof(struct page)); 1195 1196 if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || 1197 (end - start < PAGES_PER_SECTION * sizeof(struct page))) 1198 return vmemmap_populate_basepages(start, end, node, altmap); 1199 else 1200 return vmemmap_populate_hugepages(start, end, node, altmap); 1201 } 1202 1203 #ifdef CONFIG_MEMORY_HOTPLUG 1204 void vmemmap_free(unsigned long start, unsigned long end, 1205 struct vmem_altmap *altmap) 1206 { 1207 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1208 1209 unmap_hotplug_range(start, end, true, altmap); 1210 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END); 1211 } 1212 #endif /* CONFIG_MEMORY_HOTPLUG */ 1213 1214 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot) 1215 { 1216 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot)); 1217 1218 /* Only allow permission changes for now */ 1219 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)), 1220 pud_val(new_pud))) 1221 return 0; 1222 1223 VM_BUG_ON(phys & ~PUD_MASK); 1224 set_pud(pudp, new_pud); 1225 return 1; 1226 } 1227 1228 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot) 1229 { 1230 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot)); 1231 1232 /* Only allow permission changes for now */ 1233 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)), 1234 pmd_val(new_pmd))) 1235 return 0; 1236 1237 VM_BUG_ON(phys & ~PMD_MASK); 1238 set_pmd(pmdp, new_pmd); 1239 return 1; 1240 } 1241 1242 #ifndef __PAGETABLE_P4D_FOLDED 1243 void p4d_clear_huge(p4d_t *p4dp) 1244 { 1245 } 1246 #endif 1247 1248 int pud_clear_huge(pud_t *pudp) 1249 { 1250 if (!pud_sect(READ_ONCE(*pudp))) 1251 return 0; 1252 pud_clear(pudp); 1253 return 1; 1254 } 1255 1256 int pmd_clear_huge(pmd_t *pmdp) 1257 { 1258 if (!pmd_sect(READ_ONCE(*pmdp))) 1259 return 0; 1260 pmd_clear(pmdp); 1261 return 1; 1262 } 1263 1264 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr) 1265 { 1266 pte_t *table; 1267 pmd_t pmd; 1268 1269 pmd = READ_ONCE(*pmdp); 1270 1271 if (!pmd_table(pmd)) { 1272 VM_WARN_ON(1); 1273 return 1; 1274 } 1275 1276 table = pte_offset_kernel(pmdp, addr); 1277 pmd_clear(pmdp); 1278 __flush_tlb_kernel_pgtable(addr); 1279 pte_free_kernel(NULL, table); 1280 return 1; 1281 } 1282 1283 int pud_free_pmd_page(pud_t *pudp, unsigned long addr) 1284 { 1285 pmd_t *table; 1286 pmd_t *pmdp; 1287 pud_t pud; 1288 unsigned long next, end; 1289 1290 pud = READ_ONCE(*pudp); 1291 1292 if (!pud_table(pud)) { 1293 VM_WARN_ON(1); 1294 return 1; 1295 } 1296 1297 table = pmd_offset(pudp, addr); 1298 pmdp = table; 1299 next = addr; 1300 end = addr + PUD_SIZE; 1301 do { 1302 if (pmd_present(pmdp_get(pmdp))) 1303 pmd_free_pte_page(pmdp, next); 1304 } while (pmdp++, next += PMD_SIZE, next != end); 1305 1306 pud_clear(pudp); 1307 __flush_tlb_kernel_pgtable(addr); 1308 pmd_free(NULL, table); 1309 return 1; 1310 } 1311 1312 #ifdef CONFIG_MEMORY_HOTPLUG 1313 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size) 1314 { 1315 unsigned long end = start + size; 1316 1317 WARN_ON(pgdir != init_mm.pgd); 1318 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END)); 1319 1320 unmap_hotplug_range(start, end, false, NULL); 1321 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END); 1322 } 1323 1324 struct range arch_get_mappable_range(void) 1325 { 1326 struct range mhp_range; 1327 u64 start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual)); 1328 u64 end_linear_pa = __pa(PAGE_END - 1); 1329 1330 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 1331 /* 1332 * Check for a wrap, it is possible because of randomized linear 1333 * mapping the start physical address is actually bigger than 1334 * the end physical address. In this case set start to zero 1335 * because [0, end_linear_pa] range must still be able to cover 1336 * all addressable physical addresses. 1337 */ 1338 if (start_linear_pa > end_linear_pa) 1339 start_linear_pa = 0; 1340 } 1341 1342 WARN_ON(start_linear_pa > end_linear_pa); 1343 1344 /* 1345 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)] 1346 * accommodating both its ends but excluding PAGE_END. Max physical 1347 * range which can be mapped inside this linear mapping range, must 1348 * also be derived from its end points. 1349 */ 1350 mhp_range.start = start_linear_pa; 1351 mhp_range.end = end_linear_pa; 1352 1353 return mhp_range; 1354 } 1355 1356 int arch_add_memory(int nid, u64 start, u64 size, 1357 struct mhp_params *params) 1358 { 1359 int ret, flags = NO_EXEC_MAPPINGS; 1360 1361 VM_BUG_ON(!mhp_range_allowed(start, size, true)); 1362 1363 if (can_set_direct_map()) 1364 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1365 1366 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1367 size, params->pgprot, pgd_pgtable_alloc_init_mm, 1368 flags); 1369 1370 memblock_clear_nomap(start, size); 1371 1372 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, 1373 params); 1374 if (ret) 1375 __remove_pgd_mapping(swapper_pg_dir, 1376 __phys_to_virt(start), size); 1377 else { 1378 /* Address of hotplugged memory can be smaller */ 1379 max_pfn = max(max_pfn, PFN_UP(start + size)); 1380 max_low_pfn = max_pfn; 1381 } 1382 1383 return ret; 1384 } 1385 1386 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap) 1387 { 1388 unsigned long start_pfn = start >> PAGE_SHIFT; 1389 unsigned long nr_pages = size >> PAGE_SHIFT; 1390 1391 __remove_pages(start_pfn, nr_pages, altmap); 1392 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size); 1393 } 1394 1395 /* 1396 * This memory hotplug notifier helps prevent boot memory from being 1397 * inadvertently removed as it blocks pfn range offlining process in 1398 * __offline_pages(). Hence this prevents both offlining as well as 1399 * removal process for boot memory which is initially always online. 1400 * In future if and when boot memory could be removed, this notifier 1401 * should be dropped and free_hotplug_page_range() should handle any 1402 * reserved pages allocated during boot. 1403 */ 1404 static int prevent_bootmem_remove_notifier(struct notifier_block *nb, 1405 unsigned long action, void *data) 1406 { 1407 struct mem_section *ms; 1408 struct memory_notify *arg = data; 1409 unsigned long end_pfn = arg->start_pfn + arg->nr_pages; 1410 unsigned long pfn = arg->start_pfn; 1411 1412 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE)) 1413 return NOTIFY_OK; 1414 1415 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 1416 unsigned long start = PFN_PHYS(pfn); 1417 unsigned long end = start + (1UL << PA_SECTION_SHIFT); 1418 1419 ms = __pfn_to_section(pfn); 1420 if (!early_section(ms)) 1421 continue; 1422 1423 if (action == MEM_GOING_OFFLINE) { 1424 /* 1425 * Boot memory removal is not supported. Prevent 1426 * it via blocking any attempted offline request 1427 * for the boot memory and just report it. 1428 */ 1429 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end); 1430 return NOTIFY_BAD; 1431 } else if (action == MEM_OFFLINE) { 1432 /* 1433 * This should have never happened. Boot memory 1434 * offlining should have been prevented by this 1435 * very notifier. Probably some memory removal 1436 * procedure might have changed which would then 1437 * require further debug. 1438 */ 1439 pr_err("Boot memory [%lx %lx] offlined\n", start, end); 1440 1441 /* 1442 * Core memory hotplug does not process a return 1443 * code from the notifier for MEM_OFFLINE events. 1444 * The error condition has been reported. Return 1445 * from here as if ignored. 1446 */ 1447 return NOTIFY_DONE; 1448 } 1449 } 1450 return NOTIFY_OK; 1451 } 1452 1453 static struct notifier_block prevent_bootmem_remove_nb = { 1454 .notifier_call = prevent_bootmem_remove_notifier, 1455 }; 1456 1457 /* 1458 * This ensures that boot memory sections on the platform are online 1459 * from early boot. Memory sections could not be prevented from being 1460 * offlined, unless for some reason they are not online to begin with. 1461 * This helps validate the basic assumption on which the above memory 1462 * event notifier works to prevent boot memory section offlining and 1463 * its possible removal. 1464 */ 1465 static void validate_bootmem_online(void) 1466 { 1467 phys_addr_t start, end, addr; 1468 struct mem_section *ms; 1469 u64 i; 1470 1471 /* 1472 * Scanning across all memblock might be expensive 1473 * on some big memory systems. Hence enable this 1474 * validation only with DEBUG_VM. 1475 */ 1476 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 1477 return; 1478 1479 for_each_mem_range(i, &start, &end) { 1480 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) { 1481 ms = __pfn_to_section(PHYS_PFN(addr)); 1482 1483 /* 1484 * All memory ranges in the system at this point 1485 * should have been marked as early sections. 1486 */ 1487 WARN_ON(!early_section(ms)); 1488 1489 /* 1490 * Memory notifier mechanism here to prevent boot 1491 * memory offlining depends on the fact that each 1492 * early section memory on the system is initially 1493 * online. Otherwise a given memory section which 1494 * is already offline will be overlooked and can 1495 * be removed completely. Call out such sections. 1496 */ 1497 if (!online_section(ms)) 1498 pr_err("Boot memory [%llx %llx] is offline, can be removed\n", 1499 addr, addr + (1UL << PA_SECTION_SHIFT)); 1500 } 1501 } 1502 } 1503 1504 static int __init prevent_bootmem_remove_init(void) 1505 { 1506 int ret = 0; 1507 1508 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) 1509 return ret; 1510 1511 validate_bootmem_online(); 1512 ret = register_memory_notifier(&prevent_bootmem_remove_nb); 1513 if (ret) 1514 pr_err("%s: Notifier registration failed %d\n", __func__, ret); 1515 1516 return ret; 1517 } 1518 early_initcall(prevent_bootmem_remove_init); 1519 #endif 1520 1521 pte_t modify_prot_start_ptes(struct vm_area_struct *vma, unsigned long addr, 1522 pte_t *ptep, unsigned int nr) 1523 { 1524 pte_t pte = get_and_clear_ptes(vma->vm_mm, addr, ptep, nr); 1525 1526 if (alternative_has_cap_unlikely(ARM64_WORKAROUND_2645198)) { 1527 /* 1528 * Break-before-make (BBM) is required for all user space mappings 1529 * when the permission changes from executable to non-executable 1530 * in cases where cpu is affected with errata #2645198. 1531 */ 1532 if (pte_accessible(vma->vm_mm, pte) && pte_user_exec(pte)) 1533 __flush_tlb_range(vma, addr, nr * PAGE_SIZE, 1534 PAGE_SIZE, true, 3); 1535 } 1536 1537 return pte; 1538 } 1539 1540 pte_t ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep) 1541 { 1542 return modify_prot_start_ptes(vma, addr, ptep, 1); 1543 } 1544 1545 void modify_prot_commit_ptes(struct vm_area_struct *vma, unsigned long addr, 1546 pte_t *ptep, pte_t old_pte, pte_t pte, 1547 unsigned int nr) 1548 { 1549 set_ptes(vma->vm_mm, addr, ptep, pte, nr); 1550 } 1551 1552 void ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, 1553 pte_t old_pte, pte_t pte) 1554 { 1555 modify_prot_commit_ptes(vma, addr, ptep, old_pte, pte, 1); 1556 } 1557 1558 /* 1559 * Atomically replaces the active TTBR1_EL1 PGD with a new VA-compatible PGD, 1560 * avoiding the possibility of conflicting TLB entries being allocated. 1561 */ 1562 void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp) 1563 { 1564 typedef void (ttbr_replace_func)(phys_addr_t); 1565 extern ttbr_replace_func idmap_cpu_replace_ttbr1; 1566 ttbr_replace_func *replace_phys; 1567 unsigned long daif; 1568 1569 /* phys_to_ttbr() zeros lower 2 bits of ttbr with 52-bit PA */ 1570 phys_addr_t ttbr1 = phys_to_ttbr(virt_to_phys(pgdp)); 1571 1572 if (cnp) 1573 ttbr1 |= TTBR_CNP_BIT; 1574 1575 replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1); 1576 1577 cpu_install_idmap(); 1578 1579 /* 1580 * We really don't want to take *any* exceptions while TTBR1 is 1581 * in the process of being replaced so mask everything. 1582 */ 1583 daif = local_daif_save(); 1584 replace_phys(ttbr1); 1585 local_daif_restore(daif); 1586 1587 cpu_uninstall_idmap(); 1588 } 1589 1590 #ifdef CONFIG_ARCH_HAS_PKEYS 1591 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey, unsigned long init_val) 1592 { 1593 u64 new_por; 1594 u64 old_por; 1595 1596 if (!system_supports_poe()) 1597 return -ENOSPC; 1598 1599 /* 1600 * This code should only be called with valid 'pkey' 1601 * values originating from in-kernel users. Complain 1602 * if a bad value is observed. 1603 */ 1604 if (WARN_ON_ONCE(pkey >= arch_max_pkey())) 1605 return -EINVAL; 1606 1607 /* Set the bits we need in POR: */ 1608 new_por = POE_RWX; 1609 if (init_val & PKEY_DISABLE_WRITE) 1610 new_por &= ~POE_W; 1611 if (init_val & PKEY_DISABLE_ACCESS) 1612 new_por &= ~POE_RW; 1613 if (init_val & PKEY_DISABLE_READ) 1614 new_por &= ~POE_R; 1615 if (init_val & PKEY_DISABLE_EXECUTE) 1616 new_por &= ~POE_X; 1617 1618 /* Shift the bits in to the correct place in POR for pkey: */ 1619 new_por = POR_ELx_PERM_PREP(pkey, new_por); 1620 1621 /* Get old POR and mask off any old bits in place: */ 1622 old_por = read_sysreg_s(SYS_POR_EL0); 1623 old_por &= ~(POE_MASK << POR_ELx_PERM_SHIFT(pkey)); 1624 1625 /* Write old part along with new part: */ 1626 write_sysreg_s(old_por | new_por, SYS_POR_EL0); 1627 1628 return 0; 1629 } 1630 #endif 1631