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 #include <linux/pagewalk.h> 31 #include <linux/stop_machine.h> 32 33 #include <asm/barrier.h> 34 #include <asm/cputype.h> 35 #include <asm/fixmap.h> 36 #include <asm/kasan.h> 37 #include <asm/kernel-pgtable.h> 38 #include <asm/sections.h> 39 #include <asm/setup.h> 40 #include <linux/sizes.h> 41 #include <asm/tlb.h> 42 #include <asm/mmu_context.h> 43 #include <asm/ptdump.h> 44 #include <asm/tlbflush.h> 45 #include <asm/pgalloc.h> 46 #include <asm/kfence.h> 47 48 #define NO_BLOCK_MAPPINGS BIT(0) 49 #define NO_CONT_MAPPINGS BIT(1) 50 #define NO_EXEC_MAPPINGS BIT(2) /* assumes FEAT_HPDS is not used */ 51 52 DEFINE_STATIC_KEY_FALSE(arm64_ptdump_lock_key); 53 54 u64 kimage_voffset __ro_after_init; 55 EXPORT_SYMBOL(kimage_voffset); 56 57 u32 __boot_cpu_mode[] = { BOOT_CPU_MODE_EL2, BOOT_CPU_MODE_EL1 }; 58 59 static bool rodata_is_rw __ro_after_init = true; 60 61 /* 62 * The booting CPU updates the failed status @__early_cpu_boot_status, 63 * with MMU turned off. 64 */ 65 long __section(".mmuoff.data.write") __early_cpu_boot_status; 66 67 static DEFINE_SPINLOCK(swapper_pgdir_lock); 68 static DEFINE_MUTEX(fixmap_lock); 69 70 void noinstr set_swapper_pgd(pgd_t *pgdp, pgd_t pgd) 71 { 72 pgd_t *fixmap_pgdp; 73 74 /* 75 * Don't bother with the fixmap if swapper_pg_dir is still mapped 76 * writable in the kernel mapping. 77 */ 78 if (rodata_is_rw) { 79 WRITE_ONCE(*pgdp, pgd); 80 dsb(ishst); 81 isb(); 82 return; 83 } 84 85 spin_lock(&swapper_pgdir_lock); 86 fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp)); 87 WRITE_ONCE(*fixmap_pgdp, pgd); 88 /* 89 * We need dsb(ishst) here to ensure the page-table-walker sees 90 * our new entry before set_p?d() returns. The fixmap's 91 * flush_tlb_kernel_range() via clear_fixmap() does this for us. 92 */ 93 pgd_clear_fixmap(); 94 spin_unlock(&swapper_pgdir_lock); 95 } 96 97 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn, 98 unsigned long size, pgprot_t vma_prot) 99 { 100 if (!pfn_is_map_memory(pfn)) 101 return pgprot_noncached(vma_prot); 102 else if (file->f_flags & O_SYNC) 103 return pgprot_writecombine(vma_prot); 104 return vma_prot; 105 } 106 EXPORT_SYMBOL(phys_mem_access_prot); 107 108 static phys_addr_t __init early_pgtable_alloc(enum pgtable_level pgtable_level) 109 { 110 phys_addr_t phys; 111 112 phys = memblock_phys_alloc_range(PAGE_SIZE, PAGE_SIZE, 0, 113 MEMBLOCK_ALLOC_NOLEAKTRACE); 114 if (!phys) 115 panic("Failed to allocate page table page\n"); 116 117 return phys; 118 } 119 120 bool pgattr_change_is_safe(pteval_t old, pteval_t new) 121 { 122 /* 123 * The following mapping attributes may be updated in live 124 * kernel mappings without the need for break-before-make. 125 */ 126 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG | 127 PTE_SWBITS_MASK; 128 129 /* creating or taking down mappings is always safe */ 130 if (!pte_valid(__pte(old)) || !pte_valid(__pte(new))) 131 return true; 132 133 /* A live entry's pfn should not change */ 134 if (pte_pfn(__pte(old)) != pte_pfn(__pte(new))) 135 return false; 136 137 /* Transitioning from Non-Global to Global is unsafe */ 138 if (old & ~new & PTE_NG) 139 return false; 140 141 /* 142 * Changing the memory type between Normal and Normal-Tagged is safe 143 * since Tagged is considered a permission attribute from the 144 * mismatched attribute aliases perspective. 145 */ 146 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 147 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) && 148 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 149 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED))) 150 mask |= PTE_ATTRINDX_MASK; 151 152 return ((old ^ new) & ~mask) == 0; 153 } 154 155 static void init_clear_pgtable(void *table) 156 { 157 clear_page(table); 158 159 /* Ensure the zeroing is observed by page table walks. */ 160 dsb(ishst); 161 } 162 163 static void init_pte(pte_t *ptep, unsigned long addr, unsigned long end, 164 phys_addr_t phys, pgprot_t prot) 165 { 166 do { 167 pte_t old_pte = __ptep_get(ptep); 168 169 /* 170 * Required barriers to make this visible to the table walker 171 * are deferred to the end of alloc_init_cont_pte(). 172 */ 173 __set_pte_nosync(ptep, pfn_pte(__phys_to_pfn(phys), prot)); 174 175 /* 176 * After the PTE entry has been populated once, we 177 * only allow updates to the permission attributes. 178 */ 179 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte), 180 pte_val(__ptep_get(ptep)))); 181 182 phys += PAGE_SIZE; 183 } while (ptep++, addr += PAGE_SIZE, addr != end); 184 } 185 186 static bool pte_range_has_valid_noncont(pte_t *ptep) 187 { 188 for (int i = 0; i < CONT_PTES; i++) { 189 pte_t pte = __ptep_get(&ptep[i]); 190 191 if (pte_valid(pte) && !pte_cont(pte)) 192 return true; 193 } 194 return false; 195 } 196 197 static int alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr, 198 unsigned long end, phys_addr_t phys, 199 pgprot_t prot, 200 phys_addr_t (*pgtable_alloc)(enum pgtable_level), 201 int flags) 202 { 203 unsigned long next; 204 pmd_t pmd = READ_ONCE(*pmdp); 205 pte_t *ptep; 206 207 BUG_ON(pmd_leaf(pmd)); 208 if (pmd_none(pmd)) { 209 pmdval_t pmdval = PMD_TYPE_TABLE | PMD_TABLE_UXN | PMD_TABLE_AF; 210 phys_addr_t pte_phys; 211 212 if (flags & NO_EXEC_MAPPINGS) 213 pmdval |= PMD_TABLE_PXN; 214 BUG_ON(!pgtable_alloc); 215 pte_phys = pgtable_alloc(PGTABLE_LEVEL_PTE); 216 if (pte_phys == INVALID_PHYS_ADDR) 217 return -ENOMEM; 218 ptep = pte_set_fixmap(pte_phys); 219 init_clear_pgtable(ptep); 220 ptep += pte_index(addr); 221 __pmd_populate(pmdp, pte_phys, pmdval); 222 } else { 223 BUG_ON(pmd_bad(pmd)); 224 ptep = pte_set_fixmap_offset(pmdp, addr); 225 } 226 227 do { 228 pgprot_t __prot = prot; 229 230 next = pte_cont_addr_end(addr, end); 231 232 /* use a contiguous mapping if the range is suitably aligned */ 233 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) && 234 (flags & NO_CONT_MAPPINGS) == 0 && 235 !pte_range_has_valid_noncont(ptep)) 236 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 237 238 init_pte(ptep, addr, next, phys, __prot); 239 240 ptep += pte_index(next) - pte_index(addr); 241 phys += next - addr; 242 } while (addr = next, addr != end); 243 244 /* 245 * Note: barriers and maintenance necessary to clear the fixmap slot 246 * ensure that all previous pgtable writes are visible to the table 247 * walker. 248 */ 249 pte_clear_fixmap(); 250 251 return 0; 252 } 253 254 static int init_pmd(pmd_t *pmdp, unsigned long addr, unsigned long end, 255 phys_addr_t phys, pgprot_t prot, 256 phys_addr_t (*pgtable_alloc)(enum pgtable_level), int flags) 257 { 258 unsigned long next; 259 260 do { 261 pmd_t old_pmd = READ_ONCE(*pmdp); 262 263 next = pmd_addr_end(addr, end); 264 265 /* try section mapping first */ 266 if (((addr | next | phys) & ~PMD_MASK) == 0 && 267 (flags & NO_BLOCK_MAPPINGS) == 0 && 268 !pmd_table(old_pmd)) { 269 WARN_ON(!pmd_set_huge(pmdp, phys, prot)); 270 271 /* 272 * After the PMD entry has been populated once, we 273 * only allow updates to the permission attributes. 274 */ 275 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd), 276 READ_ONCE(pmd_val(*pmdp)))); 277 } else { 278 int ret; 279 280 ret = alloc_init_cont_pte(pmdp, addr, next, phys, prot, 281 pgtable_alloc, flags); 282 if (ret) 283 return ret; 284 285 VM_WARN_ON_ONCE(pmd_val(old_pmd) != 0 && 286 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp))); 287 } 288 phys += next - addr; 289 } while (pmdp++, addr = next, addr != end); 290 291 return 0; 292 } 293 294 static bool pmd_range_has_valid_noncont(pmd_t *pmdp) 295 { 296 for (int i = 0; i < CONT_PMDS; i++) { 297 pte_t pte = pmd_pte(READ_ONCE(pmdp[i])); 298 299 if (pte_valid(pte) && !pte_cont(pte)) 300 return true; 301 } 302 return false; 303 } 304 305 static int alloc_init_cont_pmd(pud_t *pudp, unsigned long addr, 306 unsigned long end, phys_addr_t phys, 307 pgprot_t prot, 308 phys_addr_t (*pgtable_alloc)(enum pgtable_level), 309 int flags) 310 { 311 int ret; 312 unsigned long next; 313 pud_t pud = READ_ONCE(*pudp); 314 pmd_t *pmdp; 315 316 /* 317 * Check for initial section mappings in the pgd/pud. 318 */ 319 BUG_ON(pud_leaf(pud)); 320 if (pud_none(pud)) { 321 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF; 322 phys_addr_t pmd_phys; 323 324 if (flags & NO_EXEC_MAPPINGS) 325 pudval |= PUD_TABLE_PXN; 326 BUG_ON(!pgtable_alloc); 327 pmd_phys = pgtable_alloc(PGTABLE_LEVEL_PMD); 328 if (pmd_phys == INVALID_PHYS_ADDR) 329 return -ENOMEM; 330 pmdp = pmd_set_fixmap(pmd_phys); 331 init_clear_pgtable(pmdp); 332 pmdp += pmd_index(addr); 333 __pud_populate(pudp, pmd_phys, pudval); 334 } else { 335 BUG_ON(pud_bad(pud)); 336 pmdp = pmd_set_fixmap_offset(pudp, addr); 337 } 338 339 do { 340 pgprot_t __prot = prot; 341 342 next = pmd_cont_addr_end(addr, end); 343 344 /* use a contiguous mapping if the range is suitably aligned */ 345 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) && 346 (flags & NO_CONT_MAPPINGS) == 0 && 347 !pmd_range_has_valid_noncont(pmdp)) 348 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 349 350 ret = init_pmd(pmdp, addr, next, phys, __prot, pgtable_alloc, flags); 351 if (ret) 352 goto out; 353 354 pmdp += pmd_index(next) - pmd_index(addr); 355 phys += next - addr; 356 } while (addr = next, addr != end); 357 358 out: 359 pmd_clear_fixmap(); 360 361 return ret; 362 } 363 364 static int alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end, 365 phys_addr_t phys, pgprot_t prot, 366 phys_addr_t (*pgtable_alloc)(enum pgtable_level), 367 int flags) 368 { 369 int ret = 0; 370 unsigned long next; 371 p4d_t p4d = READ_ONCE(*p4dp); 372 pud_t *pudp; 373 374 if (p4d_none(p4d)) { 375 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN | P4D_TABLE_AF; 376 phys_addr_t pud_phys; 377 378 if (flags & NO_EXEC_MAPPINGS) 379 p4dval |= P4D_TABLE_PXN; 380 BUG_ON(!pgtable_alloc); 381 pud_phys = pgtable_alloc(PGTABLE_LEVEL_PUD); 382 if (pud_phys == INVALID_PHYS_ADDR) 383 return -ENOMEM; 384 pudp = pud_set_fixmap(pud_phys); 385 init_clear_pgtable(pudp); 386 pudp += pud_index(addr); 387 __p4d_populate(p4dp, pud_phys, p4dval); 388 } else { 389 BUG_ON(p4d_bad(p4d)); 390 pudp = pud_set_fixmap_offset(p4dp, addr); 391 } 392 393 do { 394 pud_t old_pud = READ_ONCE(*pudp); 395 396 next = pud_addr_end(addr, end); 397 398 /* 399 * For 4K granule only, attempt to put down a 1GB block 400 */ 401 if (pud_sect_supported() && 402 ((addr | next | phys) & ~PUD_MASK) == 0 && 403 (flags & NO_BLOCK_MAPPINGS) == 0 && 404 !pud_table(old_pud)) { 405 WARN_ON(!pud_set_huge(pudp, phys, prot)); 406 407 /* 408 * After the PUD entry has been populated once, we 409 * only allow updates to the permission attributes. 410 */ 411 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud), 412 READ_ONCE(pud_val(*pudp)))); 413 } else { 414 ret = alloc_init_cont_pmd(pudp, addr, next, phys, prot, 415 pgtable_alloc, flags); 416 if (ret) 417 goto out; 418 419 VM_WARN_ON_ONCE(pud_val(old_pud) != 0 && 420 pud_val(old_pud) != READ_ONCE(pud_val(*pudp))); 421 } 422 phys += next - addr; 423 } while (pudp++, addr = next, addr != end); 424 425 out: 426 pud_clear_fixmap(); 427 428 return ret; 429 } 430 431 static int alloc_init_p4d(pgd_t *pgdp, unsigned long addr, unsigned long end, 432 phys_addr_t phys, pgprot_t prot, 433 phys_addr_t (*pgtable_alloc)(enum pgtable_level), 434 int flags) 435 { 436 int ret; 437 unsigned long next; 438 pgd_t pgd = READ_ONCE(*pgdp); 439 p4d_t *p4dp; 440 441 if (pgd_none(pgd)) { 442 pgdval_t pgdval = PGD_TYPE_TABLE | PGD_TABLE_UXN | PGD_TABLE_AF; 443 phys_addr_t p4d_phys; 444 445 if (flags & NO_EXEC_MAPPINGS) 446 pgdval |= PGD_TABLE_PXN; 447 BUG_ON(!pgtable_alloc); 448 p4d_phys = pgtable_alloc(PGTABLE_LEVEL_P4D); 449 if (p4d_phys == INVALID_PHYS_ADDR) 450 return -ENOMEM; 451 p4dp = p4d_set_fixmap(p4d_phys); 452 init_clear_pgtable(p4dp); 453 p4dp += p4d_index(addr); 454 __pgd_populate(pgdp, p4d_phys, pgdval); 455 } else { 456 BUG_ON(pgd_bad(pgd)); 457 p4dp = p4d_set_fixmap_offset(pgdp, addr); 458 } 459 460 do { 461 p4d_t old_p4d = READ_ONCE(*p4dp); 462 463 next = p4d_addr_end(addr, end); 464 465 ret = alloc_init_pud(p4dp, addr, next, phys, prot, 466 pgtable_alloc, flags); 467 if (ret) 468 goto out; 469 470 VM_WARN_ON_ONCE(p4d_val(old_p4d) != 0 && 471 p4d_val(old_p4d) != READ_ONCE(p4d_val(*p4dp))); 472 473 phys += next - addr; 474 } while (p4dp++, addr = next, addr != end); 475 476 out: 477 p4d_clear_fixmap(); 478 479 return ret; 480 } 481 482 static int __create_pgd_mapping_locked(pgd_t *pgdir, phys_addr_t phys, 483 unsigned long virt, phys_addr_t size, 484 pgprot_t prot, 485 phys_addr_t (*pgtable_alloc)(enum pgtable_level), 486 int flags) 487 { 488 int ret; 489 unsigned long addr, end, next; 490 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt); 491 492 /* 493 * If the virtual and physical address don't have the same offset 494 * within a page, we cannot map the region as the caller expects. 495 */ 496 if (WARN_ON((phys ^ virt) & ~PAGE_MASK)) 497 return -EINVAL; 498 499 phys &= PAGE_MASK; 500 addr = virt & PAGE_MASK; 501 end = PAGE_ALIGN(virt + size); 502 503 do { 504 next = pgd_addr_end(addr, end); 505 ret = alloc_init_p4d(pgdp, addr, next, phys, prot, pgtable_alloc, 506 flags); 507 if (ret) 508 return ret; 509 phys += next - addr; 510 } while (pgdp++, addr = next, addr != end); 511 512 return 0; 513 } 514 515 static int __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, 516 unsigned long virt, phys_addr_t size, 517 pgprot_t prot, 518 phys_addr_t (*pgtable_alloc)(enum pgtable_level), 519 int flags) 520 { 521 int ret; 522 523 mutex_lock(&fixmap_lock); 524 ret = __create_pgd_mapping_locked(pgdir, phys, virt, size, prot, 525 pgtable_alloc, flags); 526 mutex_unlock(&fixmap_lock); 527 528 return ret; 529 } 530 531 static void early_create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, 532 unsigned long virt, phys_addr_t size, 533 pgprot_t prot, 534 phys_addr_t (*pgtable_alloc)(enum pgtable_level), 535 int flags) 536 { 537 int ret; 538 539 ret = __create_pgd_mapping(pgdir, phys, virt, size, prot, pgtable_alloc, 540 flags); 541 if (ret) 542 panic("Failed to create page tables\n"); 543 } 544 545 static phys_addr_t __pgd_pgtable_alloc(struct mm_struct *mm, gfp_t gfp, 546 enum pgtable_level pgtable_level) 547 { 548 /* Page is zeroed by init_clear_pgtable() so don't duplicate effort. */ 549 struct ptdesc *ptdesc = pagetable_alloc(gfp & ~__GFP_ZERO, 0); 550 phys_addr_t pa; 551 552 if (!ptdesc) 553 return INVALID_PHYS_ADDR; 554 555 pa = page_to_phys(ptdesc_page(ptdesc)); 556 557 switch (pgtable_level) { 558 case PGTABLE_LEVEL_PTE: 559 BUG_ON(!pagetable_pte_ctor(mm, ptdesc)); 560 break; 561 case PGTABLE_LEVEL_PMD: 562 BUG_ON(!pagetable_pmd_ctor(mm, ptdesc)); 563 break; 564 case PGTABLE_LEVEL_PUD: 565 pagetable_pud_ctor(ptdesc); 566 break; 567 case PGTABLE_LEVEL_P4D: 568 pagetable_p4d_ctor(ptdesc); 569 break; 570 case PGTABLE_LEVEL_PGD: 571 VM_WARN_ON(1); 572 break; 573 } 574 575 return pa; 576 } 577 578 static phys_addr_t 579 pgd_pgtable_alloc_init_mm_gfp(enum pgtable_level pgtable_level, gfp_t gfp) 580 { 581 return __pgd_pgtable_alloc(&init_mm, gfp, pgtable_level); 582 } 583 584 static phys_addr_t __maybe_unused 585 pgd_pgtable_alloc_init_mm(enum pgtable_level pgtable_level) 586 { 587 return pgd_pgtable_alloc_init_mm_gfp(pgtable_level, GFP_PGTABLE_KERNEL); 588 } 589 590 static phys_addr_t 591 pgd_pgtable_alloc_special_mm(enum pgtable_level pgtable_level) 592 { 593 return __pgd_pgtable_alloc(NULL, GFP_PGTABLE_KERNEL, pgtable_level); 594 } 595 596 static void split_contpte(pte_t *ptep) 597 { 598 int i; 599 600 ptep = PTR_ALIGN_DOWN(ptep, sizeof(*ptep) * CONT_PTES); 601 for (i = 0; i < CONT_PTES; i++, ptep++) 602 __set_pte(ptep, pte_mknoncont(__ptep_get(ptep))); 603 } 604 605 static int split_pmd(pmd_t *pmdp, pmd_t pmd, gfp_t gfp, bool to_cont) 606 { 607 pmdval_t tableprot = PMD_TYPE_TABLE | PMD_TABLE_UXN | PMD_TABLE_AF; 608 unsigned long pfn = pmd_pfn(pmd); 609 pgprot_t prot = pmd_pgprot(pmd); 610 phys_addr_t pte_phys; 611 pte_t *ptep; 612 int i; 613 614 pte_phys = pgd_pgtable_alloc_init_mm_gfp(PGTABLE_LEVEL_PTE, gfp); 615 if (pte_phys == INVALID_PHYS_ADDR) 616 return -ENOMEM; 617 ptep = (pte_t *)phys_to_virt(pte_phys); 618 619 if (pgprot_val(prot) & PMD_SECT_PXN) 620 tableprot |= PMD_TABLE_PXN; 621 622 prot = __pgprot((pgprot_val(prot) & ~PTE_TYPE_MASK) | PTE_TYPE_PAGE); 623 if (!pmd_valid(pmd)) 624 prot = pte_pgprot(pte_mkinvalid(pfn_pte(0, prot))); 625 prot = __pgprot(pgprot_val(prot) & ~PTE_CONT); 626 if (to_cont) 627 prot = __pgprot(pgprot_val(prot) | PTE_CONT); 628 629 for (i = 0; i < PTRS_PER_PTE; i++, ptep++, pfn++) 630 __set_pte(ptep, pfn_pte(pfn, prot)); 631 632 /* 633 * Ensure the pte entries are visible to the table walker by the time 634 * the pmd entry that points to the ptes is visible. 635 */ 636 dsb(ishst); 637 __pmd_populate(pmdp, pte_phys, tableprot); 638 639 return 0; 640 } 641 642 static void split_contpmd(pmd_t *pmdp) 643 { 644 int i; 645 646 pmdp = PTR_ALIGN_DOWN(pmdp, sizeof(*pmdp) * CONT_PMDS); 647 for (i = 0; i < CONT_PMDS; i++, pmdp++) 648 set_pmd(pmdp, pmd_mknoncont(pmdp_get(pmdp))); 649 } 650 651 static int split_pud(pud_t *pudp, pud_t pud, gfp_t gfp, bool to_cont) 652 { 653 pudval_t tableprot = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF; 654 unsigned int step = PMD_SIZE >> PAGE_SHIFT; 655 unsigned long pfn = pud_pfn(pud); 656 pgprot_t prot = pud_pgprot(pud); 657 phys_addr_t pmd_phys; 658 pmd_t *pmdp; 659 int i; 660 661 pmd_phys = pgd_pgtable_alloc_init_mm_gfp(PGTABLE_LEVEL_PMD, gfp); 662 if (pmd_phys == INVALID_PHYS_ADDR) 663 return -ENOMEM; 664 pmdp = (pmd_t *)phys_to_virt(pmd_phys); 665 666 if (pgprot_val(prot) & PMD_SECT_PXN) 667 tableprot |= PUD_TABLE_PXN; 668 669 prot = __pgprot((pgprot_val(prot) & ~PMD_TYPE_MASK) | PMD_TYPE_SECT); 670 if (!pud_valid(pud)) 671 prot = pmd_pgprot(pmd_mkinvalid(pfn_pmd(0, prot))); 672 prot = __pgprot(pgprot_val(prot) & ~PTE_CONT); 673 if (to_cont) 674 prot = __pgprot(pgprot_val(prot) | PTE_CONT); 675 676 for (i = 0; i < PTRS_PER_PMD; i++, pmdp++, pfn += step) 677 set_pmd(pmdp, pfn_pmd(pfn, prot)); 678 679 /* 680 * Ensure the pmd entries are visible to the table walker by the time 681 * the pud entry that points to the pmds is visible. 682 */ 683 dsb(ishst); 684 __pud_populate(pudp, pmd_phys, tableprot); 685 686 return 0; 687 } 688 689 static int split_kernel_leaf_mapping_locked(unsigned long addr) 690 { 691 pgd_t *pgdp, pgd; 692 p4d_t *p4dp, p4d; 693 pud_t *pudp, pud; 694 pmd_t *pmdp, pmd; 695 pte_t *ptep, pte; 696 int ret = 0; 697 698 /* 699 * PGD: If addr is PGD aligned then addr already describes a leaf 700 * boundary. If not present then there is nothing to split. 701 */ 702 if (ALIGN_DOWN(addr, PGDIR_SIZE) == addr) 703 goto out; 704 pgdp = pgd_offset_k(addr); 705 pgd = pgdp_get(pgdp); 706 if (!pgd_present(pgd)) 707 goto out; 708 709 /* 710 * P4D: If addr is P4D aligned then addr already describes a leaf 711 * boundary. If not present then there is nothing to split. 712 */ 713 if (ALIGN_DOWN(addr, P4D_SIZE) == addr) 714 goto out; 715 p4dp = p4d_offset(pgdp, addr); 716 p4d = p4dp_get(p4dp); 717 if (!p4d_present(p4d)) 718 goto out; 719 720 /* 721 * PUD: If addr is PUD aligned then addr already describes a leaf 722 * boundary. If not present then there is nothing to split. Otherwise, 723 * if we have a pud leaf, split to contpmd. 724 */ 725 if (ALIGN_DOWN(addr, PUD_SIZE) == addr) 726 goto out; 727 pudp = pud_offset(p4dp, addr); 728 pud = pudp_get(pudp); 729 if (!pud_present(pud)) 730 goto out; 731 if (pud_leaf(pud)) { 732 ret = split_pud(pudp, pud, GFP_PGTABLE_KERNEL, true); 733 if (ret) 734 goto out; 735 } 736 737 /* 738 * CONTPMD: If addr is CONTPMD aligned then addr already describes a 739 * leaf boundary. If not present then there is nothing to split. 740 * Otherwise, if we have a contpmd leaf, split to pmd. 741 */ 742 if (ALIGN_DOWN(addr, CONT_PMD_SIZE) == addr) 743 goto out; 744 pmdp = pmd_offset(pudp, addr); 745 pmd = pmdp_get(pmdp); 746 if (!pmd_present(pmd)) 747 goto out; 748 if (pmd_leaf(pmd)) { 749 if (pmd_cont(pmd)) 750 split_contpmd(pmdp); 751 /* 752 * PMD: If addr is PMD aligned then addr already describes a 753 * leaf boundary. Otherwise, split to contpte. 754 */ 755 if (ALIGN_DOWN(addr, PMD_SIZE) == addr) 756 goto out; 757 ret = split_pmd(pmdp, pmd, GFP_PGTABLE_KERNEL, true); 758 if (ret) 759 goto out; 760 } 761 762 /* 763 * CONTPTE: If addr is CONTPTE aligned then addr already describes a 764 * leaf boundary. If not present then there is nothing to split. 765 * Otherwise, if we have a contpte leaf, split to pte. 766 */ 767 if (ALIGN_DOWN(addr, CONT_PTE_SIZE) == addr) 768 goto out; 769 ptep = pte_offset_kernel(pmdp, addr); 770 pte = __ptep_get(ptep); 771 if (!pte_present(pte)) 772 goto out; 773 if (pte_cont(pte)) 774 split_contpte(ptep); 775 776 out: 777 return ret; 778 } 779 780 static inline bool force_pte_mapping(void) 781 { 782 const bool bbml2 = system_capabilities_finalized() ? 783 system_supports_bbml2_noabort() : cpu_supports_bbml2_noabort(); 784 785 if (debug_pagealloc_enabled()) 786 return true; 787 if (bbml2) 788 return false; 789 return rodata_full || arm64_kfence_can_set_direct_map() || is_realm_world(); 790 } 791 792 static DEFINE_MUTEX(pgtable_split_lock); 793 static bool linear_map_requires_bbml2; 794 795 int split_kernel_leaf_mapping(unsigned long start, unsigned long end) 796 { 797 int ret; 798 799 /* 800 * If the region is within a pte-mapped area, there is no need to try to 801 * split. Additionally, CONFIG_DEBUG_PAGEALLOC and CONFIG_KFENCE may 802 * change permissions from atomic context so for those cases (which are 803 * always pte-mapped), we must not go any further because taking the 804 * mutex below may sleep. Do not call force_pte_mapping() here because 805 * it could return a confusing result if called from a secondary cpu 806 * prior to finalizing caps. Instead, linear_map_requires_bbml2 gives us 807 * what we need. 808 */ 809 if (!linear_map_requires_bbml2 || is_kfence_address((void *)start)) 810 return 0; 811 812 if (!system_supports_bbml2_noabort()) { 813 /* 814 * !BBML2_NOABORT systems should not be trying to change 815 * permissions on anything that is not pte-mapped in the first 816 * place. Just return early and let the permission change code 817 * raise a warning if not already pte-mapped. 818 */ 819 if (system_capabilities_finalized()) 820 return 0; 821 822 /* 823 * Boot-time: split_kernel_leaf_mapping_locked() allocates from 824 * page allocator. Can't split until it's available. 825 */ 826 if (WARN_ON(!page_alloc_available)) 827 return -EBUSY; 828 829 /* 830 * Boot-time: Started secondary cpus but don't know if they 831 * support BBML2_NOABORT yet. Can't allow splitting in this 832 * window in case they don't. 833 */ 834 if (WARN_ON(num_online_cpus() > 1)) 835 return -EBUSY; 836 } 837 838 /* 839 * Ensure start and end are at least page-aligned since this is the 840 * finest granularity we can split to. 841 */ 842 if (start != PAGE_ALIGN(start) || end != PAGE_ALIGN(end)) 843 return -EINVAL; 844 845 mutex_lock(&pgtable_split_lock); 846 lazy_mmu_mode_enable(); 847 848 /* 849 * The split_kernel_leaf_mapping_locked() may sleep, it is not a 850 * problem for ARM64 since ARM64's lazy MMU implementation allows 851 * sleeping. 852 * 853 * Optimize for the common case of splitting out a single page from a 854 * larger mapping. Here we can just split on the "least aligned" of 855 * start and end and this will guarantee that there must also be a split 856 * on the more aligned address since the both addresses must be in the 857 * same contpte block and it must have been split to ptes. 858 */ 859 if (end - start == PAGE_SIZE) { 860 start = __ffs(start) < __ffs(end) ? start : end; 861 ret = split_kernel_leaf_mapping_locked(start); 862 } else { 863 ret = split_kernel_leaf_mapping_locked(start); 864 if (!ret) 865 ret = split_kernel_leaf_mapping_locked(end); 866 } 867 868 lazy_mmu_mode_disable(); 869 mutex_unlock(&pgtable_split_lock); 870 return ret; 871 } 872 873 static int split_to_ptes_pud_entry(pud_t *pudp, unsigned long addr, 874 unsigned long next, struct mm_walk *walk) 875 { 876 gfp_t gfp = *(gfp_t *)walk->private; 877 pud_t pud = pudp_get(pudp); 878 int ret = 0; 879 880 if (pud_leaf(pud)) 881 ret = split_pud(pudp, pud, gfp, false); 882 883 return ret; 884 } 885 886 static int split_to_ptes_pmd_entry(pmd_t *pmdp, unsigned long addr, 887 unsigned long next, struct mm_walk *walk) 888 { 889 gfp_t gfp = *(gfp_t *)walk->private; 890 pmd_t pmd = pmdp_get(pmdp); 891 int ret = 0; 892 893 if (pmd_leaf(pmd)) { 894 if (pmd_cont(pmd)) 895 split_contpmd(pmdp); 896 ret = split_pmd(pmdp, pmd, gfp, false); 897 898 /* 899 * We have split the pmd directly to ptes so there is no need to 900 * visit each pte to check if they are contpte. 901 */ 902 walk->action = ACTION_CONTINUE; 903 } 904 905 return ret; 906 } 907 908 static int split_to_ptes_pte_entry(pte_t *ptep, unsigned long addr, 909 unsigned long next, struct mm_walk *walk) 910 { 911 pte_t pte = __ptep_get(ptep); 912 913 if (pte_cont(pte)) 914 split_contpte(ptep); 915 916 return 0; 917 } 918 919 static const struct mm_walk_ops split_to_ptes_ops = { 920 .pud_entry = split_to_ptes_pud_entry, 921 .pmd_entry = split_to_ptes_pmd_entry, 922 .pte_entry = split_to_ptes_pte_entry, 923 }; 924 925 static int range_split_to_ptes(unsigned long start, unsigned long end, gfp_t gfp) 926 { 927 int ret; 928 929 lazy_mmu_mode_enable(); 930 ret = walk_kernel_page_table_range_lockless(start, end, 931 &split_to_ptes_ops, NULL, &gfp); 932 lazy_mmu_mode_disable(); 933 934 return ret; 935 } 936 937 u32 idmap_kpti_bbml2_flag; 938 939 static void __init init_idmap_kpti_bbml2_flag(void) 940 { 941 WRITE_ONCE(idmap_kpti_bbml2_flag, 1); 942 /* Must be visible to other CPUs before stop_machine() is called. */ 943 smp_mb(); 944 } 945 946 static int __init linear_map_split_to_ptes(void *__unused) 947 { 948 /* 949 * Repainting the linear map must be done by CPU0 (the boot CPU) because 950 * that's the only CPU that we know supports BBML2. The other CPUs will 951 * be held in a waiting area with the idmap active. 952 */ 953 if (!smp_processor_id()) { 954 unsigned long lstart = _PAGE_OFFSET(vabits_actual); 955 unsigned long lend = PAGE_END; 956 unsigned long kstart = (unsigned long)lm_alias(_stext); 957 unsigned long kend = (unsigned long)lm_alias(__init_begin); 958 int ret; 959 960 /* 961 * Wait for all secondary CPUs to be put into the waiting area. 962 */ 963 smp_cond_load_acquire(&idmap_kpti_bbml2_flag, VAL == num_online_cpus()); 964 965 /* 966 * Walk all of the linear map [lstart, lend), except the kernel 967 * linear map alias [kstart, kend), and split all mappings to 968 * PTE. The kernel alias remains static throughout runtime so 969 * can continue to be safely mapped with large mappings. 970 */ 971 ret = range_split_to_ptes(lstart, kstart, GFP_ATOMIC); 972 if (!ret) 973 ret = range_split_to_ptes(kend, lend, GFP_ATOMIC); 974 if (ret) 975 panic("Failed to split linear map\n"); 976 flush_tlb_kernel_range(lstart, lend); 977 978 /* 979 * Relies on dsb in flush_tlb_kernel_range() to avoid reordering 980 * before any page table split operations. 981 */ 982 WRITE_ONCE(idmap_kpti_bbml2_flag, 0); 983 } else { 984 typedef void (wait_split_fn)(void); 985 extern wait_split_fn wait_linear_map_split_to_ptes; 986 wait_split_fn *wait_fn; 987 988 wait_fn = (void *)__pa_symbol(wait_linear_map_split_to_ptes); 989 990 /* 991 * At least one secondary CPU doesn't support BBML2 so cannot 992 * tolerate the size of the live mappings changing. So have the 993 * secondary CPUs wait for the boot CPU to make the changes 994 * with the idmap active and init_mm inactive. 995 */ 996 cpu_install_idmap(); 997 wait_fn(); 998 cpu_uninstall_idmap(); 999 } 1000 1001 return 0; 1002 } 1003 1004 void __init linear_map_maybe_split_to_ptes(void) 1005 { 1006 if (linear_map_requires_bbml2 && !system_supports_bbml2_noabort()) { 1007 init_idmap_kpti_bbml2_flag(); 1008 stop_machine(linear_map_split_to_ptes, NULL, cpu_online_mask); 1009 } 1010 } 1011 1012 /* 1013 * This function can only be used to modify existing table entries, 1014 * without allocating new levels of table. Note that this permits the 1015 * creation of new section or page entries. 1016 */ 1017 void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt, 1018 phys_addr_t size, pgprot_t prot) 1019 { 1020 if (virt < PAGE_OFFSET) { 1021 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n", 1022 &phys, virt); 1023 return; 1024 } 1025 early_create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 0); 1026 } 1027 1028 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys, 1029 unsigned long virt, phys_addr_t size, 1030 pgprot_t prot, bool page_mappings_only) 1031 { 1032 int flags = 0; 1033 1034 BUG_ON(mm == &init_mm); 1035 1036 if (page_mappings_only) 1037 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1038 1039 early_create_pgd_mapping(mm->pgd, phys, virt, size, prot, 1040 pgd_pgtable_alloc_special_mm, flags); 1041 } 1042 1043 static void update_mapping_prot(phys_addr_t phys, unsigned long virt, 1044 phys_addr_t size, pgprot_t prot) 1045 { 1046 if (virt < PAGE_OFFSET) { 1047 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n", 1048 &phys, virt); 1049 return; 1050 } 1051 1052 early_create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 0); 1053 1054 /* flush the TLBs after updating live kernel mappings */ 1055 flush_tlb_kernel_range(virt, virt + size); 1056 } 1057 1058 static void __init __map_memblock(phys_addr_t start, phys_addr_t end, 1059 pgprot_t prot, int flags) 1060 { 1061 early_create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1062 end - start, prot, early_pgtable_alloc, flags); 1063 } 1064 1065 void __init mark_linear_text_alias_ro(void) 1066 { 1067 /* 1068 * Remove the write permissions from the linear alias of .text/.rodata 1069 */ 1070 update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text), 1071 (unsigned long)__init_begin - (unsigned long)_text, 1072 PAGE_KERNEL_RO); 1073 } 1074 1075 #ifdef CONFIG_KFENCE 1076 1077 bool __ro_after_init kfence_early_init = !!CONFIG_KFENCE_SAMPLE_INTERVAL; 1078 1079 /* early_param() will be parsed before map_mem() below. */ 1080 static int __init parse_kfence_early_init(char *arg) 1081 { 1082 int val; 1083 1084 if (get_option(&arg, &val)) 1085 kfence_early_init = !!val; 1086 return 0; 1087 } 1088 early_param("kfence.sample_interval", parse_kfence_early_init); 1089 1090 static void __init arm64_kfence_map_pool(void) 1091 { 1092 phys_addr_t kfence_pool; 1093 1094 if (!kfence_early_init) 1095 return; 1096 1097 kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE); 1098 if (!kfence_pool) { 1099 pr_err("failed to allocate kfence pool\n"); 1100 kfence_early_init = false; 1101 return; 1102 } 1103 1104 /* KFENCE pool needs page-level mapping. */ 1105 __map_memblock(kfence_pool, kfence_pool + KFENCE_POOL_SIZE, 1106 pgprot_tagged(PAGE_KERNEL), 1107 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS | NO_EXEC_MAPPINGS); 1108 __kfence_pool = phys_to_virt(kfence_pool); 1109 } 1110 1111 bool arch_kfence_init_pool(void) 1112 { 1113 unsigned long start = (unsigned long)__kfence_pool; 1114 unsigned long end = start + KFENCE_POOL_SIZE; 1115 int ret; 1116 1117 /* Exit early if we know the linear map is already pte-mapped. */ 1118 if (force_pte_mapping()) 1119 return true; 1120 1121 /* Kfence pool is already pte-mapped for the early init case. */ 1122 if (kfence_early_init) 1123 return true; 1124 1125 mutex_lock(&pgtable_split_lock); 1126 ret = range_split_to_ptes(start, end, GFP_PGTABLE_KERNEL); 1127 mutex_unlock(&pgtable_split_lock); 1128 1129 /* 1130 * Since the system supports bbml2_noabort, tlb invalidation is not 1131 * required here; the pgtable mappings have been split to pte but larger 1132 * entries may safely linger in the TLB. 1133 */ 1134 1135 return !ret; 1136 } 1137 #else /* CONFIG_KFENCE */ 1138 1139 static inline void arm64_kfence_map_pool(void) { } 1140 1141 #endif /* CONFIG_KFENCE */ 1142 1143 static void __init map_mem(void) 1144 { 1145 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN); 1146 phys_addr_t kernel_start = __pa_symbol(_text); 1147 phys_addr_t init_begin = __pa_symbol(__init_begin); 1148 phys_addr_t init_end = __pa_symbol(__init_end); 1149 phys_addr_t kernel_end = __pa_symbol(__bss_stop); 1150 phys_addr_t start, end; 1151 int flags = NO_EXEC_MAPPINGS; 1152 u64 i; 1153 1154 /* 1155 * Setting hierarchical PXNTable attributes on table entries covering 1156 * the linear region is only possible if it is guaranteed that no table 1157 * entries at any level are being shared between the linear region and 1158 * the vmalloc region. Check whether this is true for the PGD level, in 1159 * which case it is guaranteed to be true for all other levels as well. 1160 * (Unless we are running with support for LPA2, in which case the 1161 * entire reduced VA space is covered by a single pgd_t which will have 1162 * been populated without the PXNTable attribute by the time we get here.) 1163 */ 1164 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end) && 1165 pgd_index(_PAGE_OFFSET(VA_BITS_MIN)) != PTRS_PER_PGD - 1); 1166 1167 arm64_kfence_map_pool(); 1168 1169 linear_map_requires_bbml2 = !force_pte_mapping() && can_set_direct_map(); 1170 1171 if (force_pte_mapping()) 1172 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1173 1174 /* 1175 * Map the linear alias of the [_text, __init_begin) interval first 1176 * so that its write permissions can be removed later without the need 1177 * to split any block mappings created by the loop below. 1178 * 1179 * Write permissions are needed for alternatives patching, and will be 1180 * removed later by mark_linear_text_alias_ro() above. This makes the 1181 * contents of the region accessible to subsystems such as hibernate, 1182 * but protects it from inadvertent modification or execution. 1183 */ 1184 __map_memblock(kernel_start, init_begin, pgprot_tagged(PAGE_KERNEL), 1185 flags); 1186 1187 /* Map the kernel data/bss so it can be remapped later */ 1188 __map_memblock(init_end, kernel_end, pgprot_tagged(PAGE_KERNEL), 1189 flags); 1190 1191 /* map all the memory banks */ 1192 for_each_mem_range(i, &start, &end) { 1193 /* 1194 * The linear map must allow allocation tags reading/writing 1195 * if MTE is present. Otherwise, it has the same attributes as 1196 * PAGE_KERNEL. 1197 */ 1198 __map_memblock(start, end, pgprot_tagged(PAGE_KERNEL), 1199 flags); 1200 } 1201 } 1202 1203 void mark_rodata_ro(void) 1204 { 1205 unsigned long section_size; 1206 1207 /* 1208 * mark .rodata as read only. Use __init_begin rather than __end_rodata 1209 * to cover NOTES and EXCEPTION_TABLE. 1210 */ 1211 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata; 1212 WRITE_ONCE(rodata_is_rw, false); 1213 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata, 1214 section_size, PAGE_KERNEL_RO); 1215 /* mark the range between _text and _stext as read only. */ 1216 update_mapping_prot(__pa_symbol(_text), (unsigned long)_text, 1217 (unsigned long)_stext - (unsigned long)_text, 1218 PAGE_KERNEL_RO); 1219 1220 /* Map the kernel data/bss read-only in the linear map */ 1221 update_mapping_prot(__pa_symbol(__init_end), 1222 (unsigned long)lm_alias(__init_end), 1223 (unsigned long)__bss_stop - (unsigned long)__init_end, 1224 PAGE_KERNEL_RO); 1225 } 1226 1227 static void __init declare_vma(struct vm_struct *vma, 1228 void *va_start, void *va_end, 1229 unsigned long vm_flags) 1230 { 1231 phys_addr_t pa_start = __pa_symbol(va_start); 1232 unsigned long size = va_end - va_start; 1233 1234 BUG_ON(!PAGE_ALIGNED(pa_start)); 1235 BUG_ON(!PAGE_ALIGNED(size)); 1236 1237 if (!(vm_flags & VM_NO_GUARD)) 1238 size += PAGE_SIZE; 1239 1240 vma->addr = va_start; 1241 vma->phys_addr = pa_start; 1242 vma->size = size; 1243 vma->flags = VM_MAP | vm_flags; 1244 vma->caller = __builtin_return_address(0); 1245 1246 vm_area_add_early(vma); 1247 } 1248 1249 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0 1250 #define KPTI_NG_TEMP_VA (-(1UL << PMD_SHIFT)) 1251 1252 static phys_addr_t kpti_ng_temp_alloc __initdata; 1253 1254 static phys_addr_t __init kpti_ng_pgd_alloc(enum pgtable_level pgtable_level) 1255 { 1256 kpti_ng_temp_alloc -= PAGE_SIZE; 1257 return kpti_ng_temp_alloc; 1258 } 1259 1260 static int __init __kpti_install_ng_mappings(void *__unused) 1261 { 1262 typedef void (kpti_remap_fn)(int, int, phys_addr_t, unsigned long); 1263 extern kpti_remap_fn idmap_kpti_install_ng_mappings; 1264 kpti_remap_fn *remap_fn; 1265 1266 int cpu = smp_processor_id(); 1267 int levels = CONFIG_PGTABLE_LEVELS; 1268 int order = order_base_2(levels); 1269 u64 kpti_ng_temp_pgd_pa = 0; 1270 pgd_t *kpti_ng_temp_pgd; 1271 u64 alloc = 0; 1272 1273 if (levels == 5 && !pgtable_l5_enabled()) 1274 levels = 4; 1275 else if (levels == 4 && !pgtable_l4_enabled()) 1276 levels = 3; 1277 1278 remap_fn = (void *)__pa_symbol(idmap_kpti_install_ng_mappings); 1279 1280 if (!cpu) { 1281 int ret; 1282 1283 alloc = __get_free_pages(GFP_ATOMIC | __GFP_ZERO, order); 1284 kpti_ng_temp_pgd = (pgd_t *)(alloc + (levels - 1) * PAGE_SIZE); 1285 kpti_ng_temp_alloc = kpti_ng_temp_pgd_pa = __pa(kpti_ng_temp_pgd); 1286 1287 // 1288 // Create a minimal page table hierarchy that permits us to map 1289 // the swapper page tables temporarily as we traverse them. 1290 // 1291 // The physical pages are laid out as follows: 1292 // 1293 // +--------+-/-------+-/------ +-/------ +-\\\--------+ 1294 // : PTE[] : | PMD[] : | PUD[] : | P4D[] : ||| PGD[] : 1295 // +--------+-\-------+-\------ +-\------ +-///--------+ 1296 // ^ 1297 // The first page is mapped into this hierarchy at a PMD_SHIFT 1298 // aligned virtual address, so that we can manipulate the PTE 1299 // level entries while the mapping is active. The first entry 1300 // covers the PTE[] page itself, the remaining entries are free 1301 // to be used as a ad-hoc fixmap. 1302 // 1303 ret = __create_pgd_mapping_locked(kpti_ng_temp_pgd, __pa(alloc), 1304 KPTI_NG_TEMP_VA, PAGE_SIZE, PAGE_KERNEL, 1305 kpti_ng_pgd_alloc, 0); 1306 if (ret) 1307 panic("Failed to create page tables\n"); 1308 } 1309 1310 cpu_install_idmap(); 1311 remap_fn(cpu, num_online_cpus(), kpti_ng_temp_pgd_pa, KPTI_NG_TEMP_VA); 1312 cpu_uninstall_idmap(); 1313 1314 if (!cpu) { 1315 free_pages(alloc, order); 1316 arm64_use_ng_mappings = true; 1317 } 1318 1319 return 0; 1320 } 1321 1322 void __init kpti_install_ng_mappings(void) 1323 { 1324 /* Check whether KPTI is going to be used */ 1325 if (!arm64_kernel_unmapped_at_el0()) 1326 return; 1327 1328 /* 1329 * We don't need to rewrite the page-tables if either we've done 1330 * it already or we have KASLR enabled and therefore have not 1331 * created any global mappings at all. 1332 */ 1333 if (arm64_use_ng_mappings) 1334 return; 1335 1336 init_idmap_kpti_bbml2_flag(); 1337 stop_machine(__kpti_install_ng_mappings, NULL, cpu_online_mask); 1338 } 1339 1340 static pgprot_t __init kernel_exec_prot(void) 1341 { 1342 return rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 1343 } 1344 1345 static int __init map_entry_trampoline(void) 1346 { 1347 int i; 1348 1349 if (!arm64_kernel_unmapped_at_el0()) 1350 return 0; 1351 1352 pgprot_t prot = kernel_exec_prot(); 1353 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start); 1354 1355 /* The trampoline is always mapped and can therefore be global */ 1356 pgprot_val(prot) &= ~PTE_NG; 1357 1358 /* Map only the text into the trampoline page table */ 1359 memset(tramp_pg_dir, 0, PGD_SIZE); 1360 early_create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, 1361 entry_tramp_text_size(), prot, 1362 pgd_pgtable_alloc_init_mm, NO_BLOCK_MAPPINGS); 1363 1364 /* Map both the text and data into the kernel page table */ 1365 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++) 1366 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 1367 pa_start + i * PAGE_SIZE, prot); 1368 1369 if (IS_ENABLED(CONFIG_RELOCATABLE)) 1370 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 1371 pa_start + i * PAGE_SIZE, PAGE_KERNEL_RO); 1372 1373 return 0; 1374 } 1375 core_initcall(map_entry_trampoline); 1376 #endif 1377 1378 /* 1379 * Declare the VMA areas for the kernel 1380 */ 1381 static void __init declare_kernel_vmas(void) 1382 { 1383 static struct vm_struct vmlinux_seg[KERNEL_SEGMENT_COUNT]; 1384 1385 declare_vma(&vmlinux_seg[0], _text, _etext, VM_NO_GUARD); 1386 declare_vma(&vmlinux_seg[1], __start_rodata, __inittext_begin, VM_NO_GUARD); 1387 declare_vma(&vmlinux_seg[2], __inittext_begin, __inittext_end, VM_NO_GUARD); 1388 declare_vma(&vmlinux_seg[3], __initdata_begin, __initdata_end, VM_NO_GUARD); 1389 declare_vma(&vmlinux_seg[4], _data, _end, 0); 1390 } 1391 1392 void __pi_map_range(phys_addr_t *pte, u64 start, u64 end, phys_addr_t pa, 1393 pgprot_t prot, int level, pte_t *tbl, bool may_use_cont, 1394 u64 va_offset); 1395 1396 static u8 idmap_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init, 1397 kpti_bbml2_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init; 1398 1399 static void __init create_idmap(void) 1400 { 1401 phys_addr_t start = __pa_symbol(__idmap_text_start); 1402 phys_addr_t end = __pa_symbol(__idmap_text_end); 1403 phys_addr_t ptep = __pa_symbol(idmap_ptes); 1404 1405 __pi_map_range(&ptep, start, end, start, PAGE_KERNEL_ROX, 1406 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 1407 __phys_to_virt(ptep) - ptep); 1408 1409 if (linear_map_requires_bbml2 || 1410 (IS_ENABLED(CONFIG_UNMAP_KERNEL_AT_EL0) && !arm64_use_ng_mappings)) { 1411 phys_addr_t pa = __pa_symbol(&idmap_kpti_bbml2_flag); 1412 1413 /* 1414 * The KPTI G-to-nG conversion code needs a read-write mapping 1415 * of its synchronization flag in the ID map. This is also used 1416 * when splitting the linear map to ptes if a secondary CPU 1417 * doesn't support bbml2. 1418 */ 1419 ptep = __pa_symbol(kpti_bbml2_ptes); 1420 __pi_map_range(&ptep, pa, pa + sizeof(u32), pa, PAGE_KERNEL, 1421 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 1422 __phys_to_virt(ptep) - ptep); 1423 } 1424 } 1425 1426 void __init paging_init(void) 1427 { 1428 map_mem(); 1429 1430 memblock_allow_resize(); 1431 1432 create_idmap(); 1433 declare_kernel_vmas(); 1434 } 1435 1436 #ifdef CONFIG_MEMORY_HOTPLUG 1437 static void free_hotplug_page_range(struct page *page, size_t size, 1438 struct vmem_altmap *altmap) 1439 { 1440 if (altmap) { 1441 vmem_altmap_free(altmap, size >> PAGE_SHIFT); 1442 } else { 1443 WARN_ON(PageReserved(page)); 1444 __free_pages(page, get_order(size)); 1445 } 1446 } 1447 1448 static void free_hotplug_pgtable_page(struct page *page) 1449 { 1450 pagetable_dtor(page_ptdesc(page)); 1451 free_hotplug_page_range(page, PAGE_SIZE, NULL); 1452 } 1453 1454 static bool pgtable_range_aligned(unsigned long start, unsigned long end, 1455 unsigned long floor, unsigned long ceiling, 1456 unsigned long mask) 1457 { 1458 start &= mask; 1459 if (start < floor) 1460 return false; 1461 1462 if (ceiling) { 1463 ceiling &= mask; 1464 if (!ceiling) 1465 return false; 1466 } 1467 1468 if (end - 1 > ceiling - 1) 1469 return false; 1470 return true; 1471 } 1472 1473 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr, 1474 unsigned long end, bool free_mapped, 1475 struct vmem_altmap *altmap) 1476 { 1477 pte_t *ptep, pte; 1478 1479 do { 1480 ptep = pte_offset_kernel(pmdp, addr); 1481 pte = __ptep_get(ptep); 1482 if (pte_none(pte)) 1483 continue; 1484 1485 WARN_ON(!pte_present(pte)); 1486 __pte_clear(&init_mm, addr, ptep); 1487 if (free_mapped) { 1488 /* CONT blocks are not supported in the vmemmap */ 1489 WARN_ON(pte_cont(pte)); 1490 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1491 free_hotplug_page_range(pte_page(pte), 1492 PAGE_SIZE, altmap); 1493 } 1494 /* unmap_hotplug_range() flushes TLB for !free_mapped */ 1495 } while (addr += PAGE_SIZE, addr < end); 1496 } 1497 1498 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr, 1499 unsigned long end, bool free_mapped, 1500 struct vmem_altmap *altmap) 1501 { 1502 unsigned long next; 1503 pmd_t *pmdp, pmd; 1504 1505 do { 1506 next = pmd_addr_end(addr, end); 1507 pmdp = pmd_offset(pudp, addr); 1508 pmd = READ_ONCE(*pmdp); 1509 if (pmd_none(pmd)) 1510 continue; 1511 1512 WARN_ON(!pmd_present(pmd)); 1513 if (pmd_leaf(pmd)) { 1514 pmd_clear(pmdp); 1515 if (free_mapped) { 1516 /* CONT blocks are not supported in the vmemmap */ 1517 WARN_ON(pmd_cont(pmd)); 1518 flush_tlb_kernel_range(addr, addr + PMD_SIZE); 1519 free_hotplug_page_range(pmd_page(pmd), 1520 PMD_SIZE, altmap); 1521 } 1522 /* unmap_hotplug_range() flushes TLB for !free_mapped */ 1523 continue; 1524 } 1525 WARN_ON(!pmd_table(pmd)); 1526 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap); 1527 } while (addr = next, addr < end); 1528 } 1529 1530 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr, 1531 unsigned long end, bool free_mapped, 1532 struct vmem_altmap *altmap) 1533 { 1534 unsigned long next; 1535 pud_t *pudp, pud; 1536 1537 do { 1538 next = pud_addr_end(addr, end); 1539 pudp = pud_offset(p4dp, addr); 1540 pud = READ_ONCE(*pudp); 1541 if (pud_none(pud)) 1542 continue; 1543 1544 WARN_ON(!pud_present(pud)); 1545 if (pud_leaf(pud)) { 1546 pud_clear(pudp); 1547 if (free_mapped) { 1548 flush_tlb_kernel_range(addr, addr + PUD_SIZE); 1549 free_hotplug_page_range(pud_page(pud), 1550 PUD_SIZE, altmap); 1551 } 1552 /* unmap_hotplug_range() flushes TLB for !free_mapped */ 1553 continue; 1554 } 1555 WARN_ON(!pud_table(pud)); 1556 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap); 1557 } while (addr = next, addr < end); 1558 } 1559 1560 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr, 1561 unsigned long end, bool free_mapped, 1562 struct vmem_altmap *altmap) 1563 { 1564 unsigned long next; 1565 p4d_t *p4dp, p4d; 1566 1567 do { 1568 next = p4d_addr_end(addr, end); 1569 p4dp = p4d_offset(pgdp, addr); 1570 p4d = READ_ONCE(*p4dp); 1571 if (p4d_none(p4d)) 1572 continue; 1573 1574 WARN_ON(!p4d_present(p4d)); 1575 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap); 1576 } while (addr = next, addr < end); 1577 } 1578 1579 static void unmap_hotplug_range(unsigned long addr, unsigned long end, 1580 bool free_mapped, struct vmem_altmap *altmap) 1581 { 1582 unsigned long start = addr; 1583 unsigned long next; 1584 pgd_t *pgdp, pgd; 1585 1586 /* 1587 * altmap can only be used as vmemmap mapping backing memory. 1588 * In case the backing memory itself is not being freed, then 1589 * altmap is irrelevant. Warn about this inconsistency when 1590 * encountered. 1591 */ 1592 WARN_ON(!free_mapped && altmap); 1593 1594 do { 1595 next = pgd_addr_end(addr, end); 1596 pgdp = pgd_offset_k(addr); 1597 pgd = READ_ONCE(*pgdp); 1598 if (pgd_none(pgd)) 1599 continue; 1600 1601 WARN_ON(!pgd_present(pgd)); 1602 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap); 1603 } while (addr = next, addr < end); 1604 1605 if (!free_mapped) 1606 flush_tlb_kernel_range(start, end); 1607 } 1608 1609 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr, 1610 unsigned long end, unsigned long floor, 1611 unsigned long ceiling) 1612 { 1613 pte_t *ptep, pte; 1614 unsigned long i, start = addr; 1615 1616 do { 1617 ptep = pte_offset_kernel(pmdp, addr); 1618 pte = __ptep_get(ptep); 1619 1620 /* 1621 * This is just a sanity check here which verifies that 1622 * pte clearing has been done by earlier unmap loops. 1623 */ 1624 WARN_ON(!pte_none(pte)); 1625 } while (addr += PAGE_SIZE, addr < end); 1626 1627 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK)) 1628 return; 1629 1630 /* 1631 * Check whether we can free the pte page if the rest of the 1632 * entries are empty. Overlap with other regions have been 1633 * handled by the floor/ceiling check. 1634 */ 1635 ptep = pte_offset_kernel(pmdp, 0UL); 1636 for (i = 0; i < PTRS_PER_PTE; i++) { 1637 if (!pte_none(__ptep_get(&ptep[i]))) 1638 return; 1639 } 1640 1641 pmd_clear(pmdp); 1642 __flush_tlb_kernel_pgtable(start); 1643 free_hotplug_pgtable_page(virt_to_page(ptep)); 1644 } 1645 1646 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr, 1647 unsigned long end, unsigned long floor, 1648 unsigned long ceiling) 1649 { 1650 pmd_t *pmdp, pmd; 1651 unsigned long i, next, start = addr; 1652 1653 do { 1654 next = pmd_addr_end(addr, end); 1655 pmdp = pmd_offset(pudp, addr); 1656 pmd = READ_ONCE(*pmdp); 1657 if (pmd_none(pmd)) 1658 continue; 1659 1660 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd)); 1661 free_empty_pte_table(pmdp, addr, next, floor, ceiling); 1662 } while (addr = next, addr < end); 1663 1664 if (CONFIG_PGTABLE_LEVELS <= 2) 1665 return; 1666 1667 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK)) 1668 return; 1669 1670 /* 1671 * Check whether we can free the pmd page if the rest of the 1672 * entries are empty. Overlap with other regions have been 1673 * handled by the floor/ceiling check. 1674 */ 1675 pmdp = pmd_offset(pudp, 0UL); 1676 for (i = 0; i < PTRS_PER_PMD; i++) { 1677 if (!pmd_none(READ_ONCE(pmdp[i]))) 1678 return; 1679 } 1680 1681 pud_clear(pudp); 1682 __flush_tlb_kernel_pgtable(start); 1683 free_hotplug_pgtable_page(virt_to_page(pmdp)); 1684 } 1685 1686 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr, 1687 unsigned long end, unsigned long floor, 1688 unsigned long ceiling) 1689 { 1690 pud_t *pudp, pud; 1691 unsigned long i, next, start = addr; 1692 1693 do { 1694 next = pud_addr_end(addr, end); 1695 pudp = pud_offset(p4dp, addr); 1696 pud = READ_ONCE(*pudp); 1697 if (pud_none(pud)) 1698 continue; 1699 1700 WARN_ON(!pud_present(pud) || !pud_table(pud)); 1701 free_empty_pmd_table(pudp, addr, next, floor, ceiling); 1702 } while (addr = next, addr < end); 1703 1704 if (!pgtable_l4_enabled()) 1705 return; 1706 1707 if (!pgtable_range_aligned(start, end, floor, ceiling, P4D_MASK)) 1708 return; 1709 1710 /* 1711 * Check whether we can free the pud page if the rest of the 1712 * entries are empty. Overlap with other regions have been 1713 * handled by the floor/ceiling check. 1714 */ 1715 pudp = pud_offset(p4dp, 0UL); 1716 for (i = 0; i < PTRS_PER_PUD; i++) { 1717 if (!pud_none(READ_ONCE(pudp[i]))) 1718 return; 1719 } 1720 1721 p4d_clear(p4dp); 1722 __flush_tlb_kernel_pgtable(start); 1723 free_hotplug_pgtable_page(virt_to_page(pudp)); 1724 } 1725 1726 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr, 1727 unsigned long end, unsigned long floor, 1728 unsigned long ceiling) 1729 { 1730 p4d_t *p4dp, p4d; 1731 unsigned long i, next, start = addr; 1732 1733 do { 1734 next = p4d_addr_end(addr, end); 1735 p4dp = p4d_offset(pgdp, addr); 1736 p4d = READ_ONCE(*p4dp); 1737 if (p4d_none(p4d)) 1738 continue; 1739 1740 WARN_ON(!p4d_present(p4d)); 1741 free_empty_pud_table(p4dp, addr, next, floor, ceiling); 1742 } while (addr = next, addr < end); 1743 1744 if (!pgtable_l5_enabled()) 1745 return; 1746 1747 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK)) 1748 return; 1749 1750 /* 1751 * Check whether we can free the p4d page if the rest of the 1752 * entries are empty. Overlap with other regions have been 1753 * handled by the floor/ceiling check. 1754 */ 1755 p4dp = p4d_offset(pgdp, 0UL); 1756 for (i = 0; i < PTRS_PER_P4D; i++) { 1757 if (!p4d_none(READ_ONCE(p4dp[i]))) 1758 return; 1759 } 1760 1761 pgd_clear(pgdp); 1762 __flush_tlb_kernel_pgtable(start); 1763 free_hotplug_pgtable_page(virt_to_page(p4dp)); 1764 } 1765 1766 static void free_empty_tables(unsigned long addr, unsigned long end, 1767 unsigned long floor, unsigned long ceiling) 1768 { 1769 unsigned long next; 1770 pgd_t *pgdp, pgd; 1771 1772 do { 1773 next = pgd_addr_end(addr, end); 1774 pgdp = pgd_offset_k(addr); 1775 pgd = READ_ONCE(*pgdp); 1776 if (pgd_none(pgd)) 1777 continue; 1778 1779 WARN_ON(!pgd_present(pgd)); 1780 free_empty_p4d_table(pgdp, addr, next, floor, ceiling); 1781 } while (addr = next, addr < end); 1782 } 1783 #endif 1784 1785 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1786 struct vmem_altmap *altmap) 1787 { 1788 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1789 /* [start, end] should be within one section */ 1790 WARN_ON_ONCE(end - start > PAGES_PER_SECTION * sizeof(struct page)); 1791 1792 if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || 1793 (end - start < PAGES_PER_SECTION * sizeof(struct page))) 1794 return vmemmap_populate_basepages(start, end, node, altmap); 1795 else 1796 return vmemmap_populate_hugepages(start, end, node, altmap); 1797 } 1798 1799 #ifdef CONFIG_MEMORY_HOTPLUG 1800 void vmemmap_free(unsigned long start, unsigned long end, 1801 struct vmem_altmap *altmap) 1802 { 1803 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1804 1805 unmap_hotplug_range(start, end, true, altmap); 1806 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END); 1807 } 1808 #endif /* CONFIG_MEMORY_HOTPLUG */ 1809 1810 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot) 1811 { 1812 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot)); 1813 1814 /* Only allow permission changes for now */ 1815 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)), 1816 pud_val(new_pud))) 1817 return 0; 1818 1819 VM_BUG_ON(phys & ~PUD_MASK); 1820 set_pud(pudp, new_pud); 1821 return 1; 1822 } 1823 1824 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot) 1825 { 1826 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot)); 1827 1828 /* Only allow permission changes for now */ 1829 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)), 1830 pmd_val(new_pmd))) 1831 return 0; 1832 1833 VM_BUG_ON(phys & ~PMD_MASK); 1834 set_pmd(pmdp, new_pmd); 1835 return 1; 1836 } 1837 1838 #ifndef __PAGETABLE_P4D_FOLDED 1839 void p4d_clear_huge(p4d_t *p4dp) 1840 { 1841 } 1842 #endif 1843 1844 int pud_clear_huge(pud_t *pudp) 1845 { 1846 if (!pud_leaf(READ_ONCE(*pudp))) 1847 return 0; 1848 pud_clear(pudp); 1849 return 1; 1850 } 1851 1852 int pmd_clear_huge(pmd_t *pmdp) 1853 { 1854 if (!pmd_leaf(READ_ONCE(*pmdp))) 1855 return 0; 1856 pmd_clear(pmdp); 1857 return 1; 1858 } 1859 1860 static int __pmd_free_pte_page(pmd_t *pmdp, unsigned long addr, 1861 bool acquire_mmap_lock) 1862 { 1863 pte_t *table; 1864 pmd_t pmd; 1865 1866 pmd = READ_ONCE(*pmdp); 1867 1868 if (!pmd_table(pmd)) { 1869 VM_WARN_ON(1); 1870 return 1; 1871 } 1872 1873 /* See comment in pud_free_pmd_page for static key logic */ 1874 table = pte_offset_kernel(pmdp, addr); 1875 pmd_clear(pmdp); 1876 __flush_tlb_kernel_pgtable(addr); 1877 if (static_branch_unlikely(&arm64_ptdump_lock_key) && acquire_mmap_lock) { 1878 mmap_read_lock(&init_mm); 1879 mmap_read_unlock(&init_mm); 1880 } 1881 1882 pte_free_kernel(NULL, table); 1883 return 1; 1884 } 1885 1886 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr) 1887 { 1888 /* If ptdump is walking the pagetables, acquire init_mm.mmap_lock */ 1889 return __pmd_free_pte_page(pmdp, addr, /* acquire_mmap_lock = */ true); 1890 } 1891 1892 int pud_free_pmd_page(pud_t *pudp, unsigned long addr) 1893 { 1894 pmd_t *table; 1895 pmd_t *pmdp; 1896 pud_t pud; 1897 unsigned long next, end; 1898 1899 pud = READ_ONCE(*pudp); 1900 1901 if (!pud_table(pud)) { 1902 VM_WARN_ON(1); 1903 return 1; 1904 } 1905 1906 table = pmd_offset(pudp, addr); 1907 1908 /* 1909 * Our objective is to prevent ptdump from reading a PMD table which has 1910 * been freed. In this race, if pud_free_pmd_page observes the key on 1911 * (which got flipped by ptdump) then the mmap lock sequence here will, 1912 * as a result of the mmap write lock/unlock sequence in ptdump, give 1913 * us the correct synchronization. If not, this means that ptdump has 1914 * yet not started walking the pagetables - the sequence of barriers 1915 * issued by __flush_tlb_kernel_pgtable() guarantees that ptdump will 1916 * observe an empty PUD. 1917 */ 1918 pud_clear(pudp); 1919 __flush_tlb_kernel_pgtable(addr); 1920 if (static_branch_unlikely(&arm64_ptdump_lock_key)) { 1921 mmap_read_lock(&init_mm); 1922 mmap_read_unlock(&init_mm); 1923 } 1924 1925 pmdp = table; 1926 next = addr; 1927 end = addr + PUD_SIZE; 1928 do { 1929 if (pmd_present(pmdp_get(pmdp))) 1930 /* 1931 * PMD has been isolated, so ptdump won't see it. No 1932 * need to acquire init_mm.mmap_lock. 1933 */ 1934 __pmd_free_pte_page(pmdp, next, /* acquire_mmap_lock = */ false); 1935 } while (pmdp++, next += PMD_SIZE, next != end); 1936 1937 pmd_free(NULL, table); 1938 return 1; 1939 } 1940 1941 #ifdef CONFIG_MEMORY_HOTPLUG 1942 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size) 1943 { 1944 unsigned long end = start + size; 1945 1946 WARN_ON(pgdir != init_mm.pgd); 1947 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END)); 1948 1949 unmap_hotplug_range(start, end, false, NULL); 1950 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END); 1951 } 1952 1953 struct range arch_get_mappable_range(void) 1954 { 1955 struct range mhp_range; 1956 phys_addr_t start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual)); 1957 phys_addr_t end_linear_pa = __pa(PAGE_END - 1); 1958 1959 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 1960 /* 1961 * Check for a wrap, it is possible because of randomized linear 1962 * mapping the start physical address is actually bigger than 1963 * the end physical address. In this case set start to zero 1964 * because [0, end_linear_pa] range must still be able to cover 1965 * all addressable physical addresses. 1966 */ 1967 if (start_linear_pa > end_linear_pa) 1968 start_linear_pa = 0; 1969 } 1970 1971 WARN_ON(start_linear_pa > end_linear_pa); 1972 1973 /* 1974 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)] 1975 * accommodating both its ends but excluding PAGE_END. Max physical 1976 * range which can be mapped inside this linear mapping range, must 1977 * also be derived from its end points. 1978 */ 1979 mhp_range.start = start_linear_pa; 1980 mhp_range.end = end_linear_pa; 1981 1982 return mhp_range; 1983 } 1984 1985 int arch_add_memory(int nid, u64 start, u64 size, 1986 struct mhp_params *params) 1987 { 1988 int ret, flags = NO_EXEC_MAPPINGS; 1989 1990 VM_BUG_ON(!mhp_range_allowed(start, size, true)); 1991 1992 if (force_pte_mapping()) 1993 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1994 1995 ret = __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1996 size, params->pgprot, pgd_pgtable_alloc_init_mm, 1997 flags); 1998 if (ret) 1999 goto err; 2000 2001 memblock_clear_nomap(start, size); 2002 2003 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, 2004 params); 2005 if (ret) 2006 goto err; 2007 2008 /* Address of hotplugged memory can be smaller */ 2009 max_pfn = max(max_pfn, PFN_UP(start + size)); 2010 max_low_pfn = max_pfn; 2011 2012 return 0; 2013 2014 err: 2015 __remove_pgd_mapping(swapper_pg_dir, 2016 __phys_to_virt(start), size); 2017 return ret; 2018 } 2019 2020 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap, 2021 struct dev_pagemap *pgmap) 2022 { 2023 unsigned long start_pfn = start >> PAGE_SHIFT; 2024 unsigned long nr_pages = size >> PAGE_SHIFT; 2025 2026 __remove_pages(start_pfn, nr_pages, altmap, pgmap); 2027 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size); 2028 } 2029 2030 2031 static bool addr_splits_kernel_leaf(unsigned long addr) 2032 { 2033 pgd_t *pgdp, pgd; 2034 p4d_t *p4dp, p4d; 2035 pud_t *pudp, pud; 2036 pmd_t *pmdp, pmd; 2037 pte_t *ptep, pte; 2038 2039 /* 2040 * If the given address points at a the start address of 2041 * a possible leaf, we certainly won't split. Otherwise, 2042 * check if we would actually split a leaf by traversing 2043 * the page tables further. 2044 */ 2045 if (IS_ALIGNED(addr, PGDIR_SIZE)) 2046 return false; 2047 2048 pgdp = pgd_offset_k(addr); 2049 pgd = pgdp_get(pgdp); 2050 if (!pgd_present(pgd)) 2051 return false; 2052 2053 if (IS_ALIGNED(addr, P4D_SIZE)) 2054 return false; 2055 2056 p4dp = p4d_offset(pgdp, addr); 2057 p4d = p4dp_get(p4dp); 2058 if (!p4d_present(p4d)) 2059 return false; 2060 2061 if (IS_ALIGNED(addr, PUD_SIZE)) 2062 return false; 2063 2064 pudp = pud_offset(p4dp, addr); 2065 pud = pudp_get(pudp); 2066 if (!pud_present(pud)) 2067 return false; 2068 2069 if (pud_leaf(pud)) 2070 return true; 2071 2072 if (IS_ALIGNED(addr, CONT_PMD_SIZE)) 2073 return false; 2074 2075 pmdp = pmd_offset(pudp, addr); 2076 pmd = pmdp_get(pmdp); 2077 if (!pmd_present(pmd)) 2078 return false; 2079 2080 if (pmd_cont(pmd)) 2081 return true; 2082 2083 if (IS_ALIGNED(addr, PMD_SIZE)) 2084 return false; 2085 2086 if (pmd_leaf(pmd)) 2087 return true; 2088 2089 if (IS_ALIGNED(addr, CONT_PTE_SIZE)) 2090 return false; 2091 2092 ptep = pte_offset_kernel(pmdp, addr); 2093 pte = __ptep_get(ptep); 2094 if (!pte_present(pte)) 2095 return false; 2096 2097 if (pte_cont(pte)) 2098 return true; 2099 2100 return !IS_ALIGNED(addr, PAGE_SIZE); 2101 } 2102 2103 static bool can_unmap_without_split(unsigned long pfn, unsigned long nr_pages) 2104 { 2105 unsigned long phys_start, phys_end, start, end; 2106 2107 phys_start = PFN_PHYS(pfn); 2108 phys_end = phys_start + nr_pages * PAGE_SIZE; 2109 2110 /* PFN range's linear map edges are leaf entry aligned */ 2111 start = __phys_to_virt(phys_start); 2112 end = __phys_to_virt(phys_end); 2113 if (addr_splits_kernel_leaf(start) || addr_splits_kernel_leaf(end)) { 2114 pr_warn("[%lx %lx] splits a leaf entry in linear map\n", 2115 phys_start, phys_end); 2116 return false; 2117 } 2118 2119 /* PFN range's vmemmap edges are leaf entry aligned */ 2120 BUILD_BUG_ON(!IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP)); 2121 start = (unsigned long)pfn_to_page(pfn); 2122 end = (unsigned long)pfn_to_page(pfn + nr_pages); 2123 if (addr_splits_kernel_leaf(start) || addr_splits_kernel_leaf(end)) { 2124 pr_warn("[%lx %lx] splits a leaf entry in vmemmap\n", 2125 phys_start, phys_end); 2126 return false; 2127 } 2128 return true; 2129 } 2130 2131 /* 2132 * This memory hotplug notifier helps prevent boot memory from being 2133 * inadvertently removed as it blocks pfn range offlining process in 2134 * __offline_pages(). Hence this prevents both offlining as well as 2135 * removal process for boot memory which is initially always online. 2136 * In future if and when boot memory could be removed, this notifier 2137 * should be dropped and free_hotplug_page_range() should handle any 2138 * reserved pages allocated during boot. 2139 * 2140 * This also blocks any memory remove that would have caused a split 2141 * in leaf entry in kernel linear or vmemmap mapping. 2142 */ 2143 static int prevent_memory_remove_notifier(struct notifier_block *nb, 2144 unsigned long action, void *data) 2145 { 2146 struct mem_section *ms; 2147 struct memory_notify *arg = data; 2148 unsigned long end_pfn = arg->start_pfn + arg->nr_pages; 2149 unsigned long pfn = arg->start_pfn; 2150 2151 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE)) 2152 return NOTIFY_OK; 2153 2154 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 2155 unsigned long start = PFN_PHYS(pfn); 2156 unsigned long end = start + (1UL << PA_SECTION_SHIFT); 2157 2158 ms = __pfn_to_section(pfn); 2159 if (!early_section(ms)) 2160 continue; 2161 2162 if (action == MEM_GOING_OFFLINE) { 2163 /* 2164 * Boot memory removal is not supported. Prevent 2165 * it via blocking any attempted offline request 2166 * for the boot memory and just report it. 2167 */ 2168 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end); 2169 return NOTIFY_BAD; 2170 } else if (action == MEM_OFFLINE) { 2171 /* 2172 * This should have never happened. Boot memory 2173 * offlining should have been prevented by this 2174 * very notifier. Probably some memory removal 2175 * procedure might have changed which would then 2176 * require further debug. 2177 */ 2178 pr_err("Boot memory [%lx %lx] offlined\n", start, end); 2179 2180 /* 2181 * Core memory hotplug does not process a return 2182 * code from the notifier for MEM_OFFLINE events. 2183 * The error condition has been reported. Return 2184 * from here as if ignored. 2185 */ 2186 return NOTIFY_DONE; 2187 } 2188 } 2189 2190 if (!can_unmap_without_split(pfn, arg->nr_pages)) 2191 return NOTIFY_BAD; 2192 2193 return NOTIFY_OK; 2194 } 2195 2196 static struct notifier_block prevent_memory_remove_nb = { 2197 .notifier_call = prevent_memory_remove_notifier, 2198 }; 2199 2200 /* 2201 * This ensures that boot memory sections on the platform are online 2202 * from early boot. Memory sections could not be prevented from being 2203 * offlined, unless for some reason they are not online to begin with. 2204 * This helps validate the basic assumption on which the above memory 2205 * event notifier works to prevent boot memory section offlining and 2206 * its possible removal. 2207 */ 2208 static void validate_bootmem_online(void) 2209 { 2210 phys_addr_t start, end, addr; 2211 struct mem_section *ms; 2212 u64 i; 2213 2214 /* 2215 * Scanning across all memblock might be expensive 2216 * on some big memory systems. Hence enable this 2217 * validation only with DEBUG_VM. 2218 */ 2219 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 2220 return; 2221 2222 for_each_mem_range(i, &start, &end) { 2223 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) { 2224 ms = __pfn_to_section(PHYS_PFN(addr)); 2225 2226 /* 2227 * All memory ranges in the system at this point 2228 * should have been marked as early sections. 2229 */ 2230 WARN_ON(!early_section(ms)); 2231 2232 /* 2233 * Memory notifier mechanism here to prevent boot 2234 * memory offlining depends on the fact that each 2235 * early section memory on the system is initially 2236 * online. Otherwise a given memory section which 2237 * is already offline will be overlooked and can 2238 * be removed completely. Call out such sections. 2239 */ 2240 if (!online_section(ms)) 2241 pr_err("Boot memory [%llx %llx] is offline, can be removed\n", 2242 addr, addr + (1UL << PA_SECTION_SHIFT)); 2243 } 2244 } 2245 } 2246 2247 static int __init prevent_memory_remove_init(void) 2248 { 2249 int ret = 0; 2250 2251 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) 2252 return ret; 2253 2254 validate_bootmem_online(); 2255 ret = register_memory_notifier(&prevent_memory_remove_nb); 2256 if (ret) 2257 pr_err("%s: Notifier registration failed %d\n", __func__, ret); 2258 2259 return ret; 2260 } 2261 early_initcall(prevent_memory_remove_init); 2262 #endif 2263 2264 pte_t modify_prot_start_ptes(struct vm_area_struct *vma, unsigned long addr, 2265 pte_t *ptep, unsigned int nr) 2266 { 2267 pte_t pte = get_and_clear_ptes(vma->vm_mm, addr, ptep, nr); 2268 2269 if (alternative_has_cap_unlikely(ARM64_WORKAROUND_2645198)) { 2270 /* 2271 * Break-before-make (BBM) is required for all user space mappings 2272 * when the permission changes from executable to non-executable 2273 * in cases where cpu is affected with errata #2645198. 2274 */ 2275 if (pte_accessible(vma->vm_mm, pte) && pte_user_exec(pte)) 2276 __flush_tlb_range(vma, addr, nr * PAGE_SIZE, 2277 PAGE_SIZE, 3, TLBF_NOWALKCACHE); 2278 } 2279 2280 return pte; 2281 } 2282 2283 pte_t ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep) 2284 { 2285 return modify_prot_start_ptes(vma, addr, ptep, 1); 2286 } 2287 2288 void modify_prot_commit_ptes(struct vm_area_struct *vma, unsigned long addr, 2289 pte_t *ptep, pte_t old_pte, pte_t pte, 2290 unsigned int nr) 2291 { 2292 set_ptes(vma->vm_mm, addr, ptep, pte, nr); 2293 } 2294 2295 void ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, 2296 pte_t old_pte, pte_t pte) 2297 { 2298 modify_prot_commit_ptes(vma, addr, ptep, old_pte, pte, 1); 2299 } 2300 2301 /* 2302 * Atomically replaces the active TTBR1_EL1 PGD with a new VA-compatible PGD, 2303 * avoiding the possibility of conflicting TLB entries being allocated. 2304 */ 2305 void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp) 2306 { 2307 typedef void (ttbr_replace_func)(phys_addr_t); 2308 extern ttbr_replace_func idmap_cpu_replace_ttbr1; 2309 ttbr_replace_func *replace_phys; 2310 unsigned long daif; 2311 2312 /* phys_to_ttbr() zeros lower 2 bits of ttbr with 52-bit PA */ 2313 phys_addr_t ttbr1 = phys_to_ttbr(virt_to_phys(pgdp)); 2314 2315 if (cnp) 2316 ttbr1 |= TTBRx_EL1_CnP; 2317 2318 replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1); 2319 2320 cpu_install_idmap(); 2321 2322 /* 2323 * We really don't want to take *any* exceptions while TTBR1 is 2324 * in the process of being replaced so mask everything. 2325 */ 2326 daif = local_daif_save(); 2327 replace_phys(ttbr1); 2328 local_daif_restore(daif); 2329 2330 cpu_uninstall_idmap(); 2331 } 2332 2333 #ifdef CONFIG_ARCH_HAS_PKEYS 2334 int arch_set_user_pkey_access(int pkey, unsigned long init_val) 2335 { 2336 u64 new_por; 2337 u64 old_por; 2338 2339 if (!system_supports_poe()) 2340 return -ENOSPC; 2341 2342 /* 2343 * This code should only be called with valid 'pkey' 2344 * values originating from in-kernel users. Complain 2345 * if a bad value is observed. 2346 */ 2347 if (WARN_ON_ONCE(pkey >= arch_max_pkey())) 2348 return -EINVAL; 2349 2350 /* Set the bits we need in POR: */ 2351 new_por = POE_RWX; 2352 if (init_val & PKEY_DISABLE_WRITE) 2353 new_por &= ~POE_W; 2354 if (init_val & PKEY_DISABLE_ACCESS) 2355 new_por &= ~POE_RW; 2356 if (init_val & PKEY_DISABLE_READ) 2357 new_por &= ~POE_R; 2358 if (init_val & PKEY_DISABLE_EXECUTE) 2359 new_por &= ~POE_X; 2360 2361 /* Shift the bits in to the correct place in POR for pkey: */ 2362 new_por = POR_ELx_PERM_PREP(pkey, new_por); 2363 2364 /* Get old POR and mask off any old bits in place: */ 2365 old_por = read_sysreg_s(SYS_POR_EL0); 2366 old_por &= ~(POE_MASK << POR_ELx_PERM_SHIFT(pkey)); 2367 2368 /* Write old part along with new part: */ 2369 write_sysreg_s(old_por | new_por, SYS_POR_EL0); 2370 2371 return 0; 2372 } 2373 #endif 2374