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 /* 68 * Empty_zero_page is a special page that is used for zero-initialized data 69 * and COW. 70 */ 71 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss; 72 EXPORT_SYMBOL(empty_zero_page); 73 74 static DEFINE_SPINLOCK(swapper_pgdir_lock); 75 static DEFINE_MUTEX(fixmap_lock); 76 77 void noinstr set_swapper_pgd(pgd_t *pgdp, pgd_t pgd) 78 { 79 pgd_t *fixmap_pgdp; 80 81 /* 82 * Don't bother with the fixmap if swapper_pg_dir is still mapped 83 * writable in the kernel mapping. 84 */ 85 if (rodata_is_rw) { 86 WRITE_ONCE(*pgdp, pgd); 87 dsb(ishst); 88 isb(); 89 return; 90 } 91 92 spin_lock(&swapper_pgdir_lock); 93 fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp)); 94 WRITE_ONCE(*fixmap_pgdp, pgd); 95 /* 96 * We need dsb(ishst) here to ensure the page-table-walker sees 97 * our new entry before set_p?d() returns. The fixmap's 98 * flush_tlb_kernel_range() via clear_fixmap() does this for us. 99 */ 100 pgd_clear_fixmap(); 101 spin_unlock(&swapper_pgdir_lock); 102 } 103 104 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn, 105 unsigned long size, pgprot_t vma_prot) 106 { 107 if (!pfn_is_map_memory(pfn)) 108 return pgprot_noncached(vma_prot); 109 else if (file->f_flags & O_SYNC) 110 return pgprot_writecombine(vma_prot); 111 return vma_prot; 112 } 113 EXPORT_SYMBOL(phys_mem_access_prot); 114 115 static phys_addr_t __init early_pgtable_alloc(enum pgtable_type pgtable_type) 116 { 117 phys_addr_t phys; 118 119 phys = memblock_phys_alloc_range(PAGE_SIZE, PAGE_SIZE, 0, 120 MEMBLOCK_ALLOC_NOLEAKTRACE); 121 if (!phys) 122 panic("Failed to allocate page table page\n"); 123 124 return phys; 125 } 126 127 bool pgattr_change_is_safe(pteval_t old, pteval_t new) 128 { 129 /* 130 * The following mapping attributes may be updated in live 131 * kernel mappings without the need for break-before-make. 132 */ 133 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG | 134 PTE_SWBITS_MASK; 135 136 /* creating or taking down mappings is always safe */ 137 if (!pte_valid(__pte(old)) || !pte_valid(__pte(new))) 138 return true; 139 140 /* A live entry's pfn should not change */ 141 if (pte_pfn(__pte(old)) != pte_pfn(__pte(new))) 142 return false; 143 144 /* live contiguous mappings may not be manipulated at all */ 145 if ((old | new) & PTE_CONT) 146 return false; 147 148 /* Transitioning from Non-Global to Global is unsafe */ 149 if (old & ~new & PTE_NG) 150 return false; 151 152 /* 153 * Changing the memory type between Normal and Normal-Tagged is safe 154 * since Tagged is considered a permission attribute from the 155 * mismatched attribute aliases perspective. 156 */ 157 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 158 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) && 159 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) || 160 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED))) 161 mask |= PTE_ATTRINDX_MASK; 162 163 return ((old ^ new) & ~mask) == 0; 164 } 165 166 static void init_clear_pgtable(void *table) 167 { 168 clear_page(table); 169 170 /* Ensure the zeroing is observed by page table walks. */ 171 dsb(ishst); 172 } 173 174 static void init_pte(pte_t *ptep, unsigned long addr, unsigned long end, 175 phys_addr_t phys, pgprot_t prot) 176 { 177 do { 178 pte_t old_pte = __ptep_get(ptep); 179 180 /* 181 * Required barriers to make this visible to the table walker 182 * are deferred to the end of alloc_init_cont_pte(). 183 */ 184 __set_pte_nosync(ptep, pfn_pte(__phys_to_pfn(phys), prot)); 185 186 /* 187 * After the PTE entry has been populated once, we 188 * only allow updates to the permission attributes. 189 */ 190 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte), 191 pte_val(__ptep_get(ptep)))); 192 193 phys += PAGE_SIZE; 194 } while (ptep++, addr += PAGE_SIZE, addr != end); 195 } 196 197 static void 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_type), 201 int flags) 202 { 203 unsigned long next; 204 pmd_t pmd = READ_ONCE(*pmdp); 205 pte_t *ptep; 206 207 BUG_ON(pmd_sect(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(TABLE_PTE); 216 ptep = pte_set_fixmap(pte_phys); 217 init_clear_pgtable(ptep); 218 ptep += pte_index(addr); 219 __pmd_populate(pmdp, pte_phys, pmdval); 220 } else { 221 BUG_ON(pmd_bad(pmd)); 222 ptep = pte_set_fixmap_offset(pmdp, addr); 223 } 224 225 do { 226 pgprot_t __prot = prot; 227 228 next = pte_cont_addr_end(addr, end); 229 230 /* use a contiguous mapping if the range is suitably aligned */ 231 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) && 232 (flags & NO_CONT_MAPPINGS) == 0) 233 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 234 235 init_pte(ptep, addr, next, phys, __prot); 236 237 ptep += pte_index(next) - pte_index(addr); 238 phys += next - addr; 239 } while (addr = next, addr != end); 240 241 /* 242 * Note: barriers and maintenance necessary to clear the fixmap slot 243 * ensure that all previous pgtable writes are visible to the table 244 * walker. 245 */ 246 pte_clear_fixmap(); 247 } 248 249 static void init_pmd(pmd_t *pmdp, unsigned long addr, unsigned long end, 250 phys_addr_t phys, pgprot_t prot, 251 phys_addr_t (*pgtable_alloc)(enum pgtable_type), int flags) 252 { 253 unsigned long next; 254 255 do { 256 pmd_t old_pmd = READ_ONCE(*pmdp); 257 258 next = pmd_addr_end(addr, end); 259 260 /* try section mapping first */ 261 if (((addr | next | phys) & ~PMD_MASK) == 0 && 262 (flags & NO_BLOCK_MAPPINGS) == 0) { 263 pmd_set_huge(pmdp, phys, prot); 264 265 /* 266 * After the PMD entry has been populated once, we 267 * only allow updates to the permission attributes. 268 */ 269 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd), 270 READ_ONCE(pmd_val(*pmdp)))); 271 } else { 272 alloc_init_cont_pte(pmdp, addr, next, phys, prot, 273 pgtable_alloc, flags); 274 275 BUG_ON(pmd_val(old_pmd) != 0 && 276 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp))); 277 } 278 phys += next - addr; 279 } while (pmdp++, addr = next, addr != end); 280 } 281 282 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr, 283 unsigned long end, phys_addr_t phys, 284 pgprot_t prot, 285 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 286 int flags) 287 { 288 unsigned long next; 289 pud_t pud = READ_ONCE(*pudp); 290 pmd_t *pmdp; 291 292 /* 293 * Check for initial section mappings in the pgd/pud. 294 */ 295 BUG_ON(pud_sect(pud)); 296 if (pud_none(pud)) { 297 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF; 298 phys_addr_t pmd_phys; 299 300 if (flags & NO_EXEC_MAPPINGS) 301 pudval |= PUD_TABLE_PXN; 302 BUG_ON(!pgtable_alloc); 303 pmd_phys = pgtable_alloc(TABLE_PMD); 304 pmdp = pmd_set_fixmap(pmd_phys); 305 init_clear_pgtable(pmdp); 306 pmdp += pmd_index(addr); 307 __pud_populate(pudp, pmd_phys, pudval); 308 } else { 309 BUG_ON(pud_bad(pud)); 310 pmdp = pmd_set_fixmap_offset(pudp, addr); 311 } 312 313 do { 314 pgprot_t __prot = prot; 315 316 next = pmd_cont_addr_end(addr, end); 317 318 /* use a contiguous mapping if the range is suitably aligned */ 319 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) && 320 (flags & NO_CONT_MAPPINGS) == 0) 321 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 322 323 init_pmd(pmdp, addr, next, phys, __prot, pgtable_alloc, flags); 324 325 pmdp += pmd_index(next) - pmd_index(addr); 326 phys += next - addr; 327 } while (addr = next, addr != end); 328 329 pmd_clear_fixmap(); 330 } 331 332 static void alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end, 333 phys_addr_t phys, pgprot_t prot, 334 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 335 int flags) 336 { 337 unsigned long next; 338 p4d_t p4d = READ_ONCE(*p4dp); 339 pud_t *pudp; 340 341 if (p4d_none(p4d)) { 342 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN | P4D_TABLE_AF; 343 phys_addr_t pud_phys; 344 345 if (flags & NO_EXEC_MAPPINGS) 346 p4dval |= P4D_TABLE_PXN; 347 BUG_ON(!pgtable_alloc); 348 pud_phys = pgtable_alloc(TABLE_PUD); 349 pudp = pud_set_fixmap(pud_phys); 350 init_clear_pgtable(pudp); 351 pudp += pud_index(addr); 352 __p4d_populate(p4dp, pud_phys, p4dval); 353 } else { 354 BUG_ON(p4d_bad(p4d)); 355 pudp = pud_set_fixmap_offset(p4dp, addr); 356 } 357 358 do { 359 pud_t old_pud = READ_ONCE(*pudp); 360 361 next = pud_addr_end(addr, end); 362 363 /* 364 * For 4K granule only, attempt to put down a 1GB block 365 */ 366 if (pud_sect_supported() && 367 ((addr | next | phys) & ~PUD_MASK) == 0 && 368 (flags & NO_BLOCK_MAPPINGS) == 0) { 369 pud_set_huge(pudp, phys, prot); 370 371 /* 372 * After the PUD entry has been populated once, we 373 * only allow updates to the permission attributes. 374 */ 375 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud), 376 READ_ONCE(pud_val(*pudp)))); 377 } else { 378 alloc_init_cont_pmd(pudp, addr, next, phys, prot, 379 pgtable_alloc, flags); 380 381 BUG_ON(pud_val(old_pud) != 0 && 382 pud_val(old_pud) != READ_ONCE(pud_val(*pudp))); 383 } 384 phys += next - addr; 385 } while (pudp++, addr = next, addr != end); 386 387 pud_clear_fixmap(); 388 } 389 390 static void alloc_init_p4d(pgd_t *pgdp, unsigned long addr, unsigned long end, 391 phys_addr_t phys, pgprot_t prot, 392 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 393 int flags) 394 { 395 unsigned long next; 396 pgd_t pgd = READ_ONCE(*pgdp); 397 p4d_t *p4dp; 398 399 if (pgd_none(pgd)) { 400 pgdval_t pgdval = PGD_TYPE_TABLE | PGD_TABLE_UXN | PGD_TABLE_AF; 401 phys_addr_t p4d_phys; 402 403 if (flags & NO_EXEC_MAPPINGS) 404 pgdval |= PGD_TABLE_PXN; 405 BUG_ON(!pgtable_alloc); 406 p4d_phys = pgtable_alloc(TABLE_P4D); 407 p4dp = p4d_set_fixmap(p4d_phys); 408 init_clear_pgtable(p4dp); 409 p4dp += p4d_index(addr); 410 __pgd_populate(pgdp, p4d_phys, pgdval); 411 } else { 412 BUG_ON(pgd_bad(pgd)); 413 p4dp = p4d_set_fixmap_offset(pgdp, addr); 414 } 415 416 do { 417 p4d_t old_p4d = READ_ONCE(*p4dp); 418 419 next = p4d_addr_end(addr, end); 420 421 alloc_init_pud(p4dp, addr, next, phys, prot, 422 pgtable_alloc, flags); 423 424 BUG_ON(p4d_val(old_p4d) != 0 && 425 p4d_val(old_p4d) != READ_ONCE(p4d_val(*p4dp))); 426 427 phys += next - addr; 428 } while (p4dp++, addr = next, addr != end); 429 430 p4d_clear_fixmap(); 431 } 432 433 static void __create_pgd_mapping_locked(pgd_t *pgdir, phys_addr_t phys, 434 unsigned long virt, phys_addr_t size, 435 pgprot_t prot, 436 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 437 int flags) 438 { 439 unsigned long addr, end, next; 440 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt); 441 442 /* 443 * If the virtual and physical address don't have the same offset 444 * within a page, we cannot map the region as the caller expects. 445 */ 446 if (WARN_ON((phys ^ virt) & ~PAGE_MASK)) 447 return; 448 449 phys &= PAGE_MASK; 450 addr = virt & PAGE_MASK; 451 end = PAGE_ALIGN(virt + size); 452 453 do { 454 next = pgd_addr_end(addr, end); 455 alloc_init_p4d(pgdp, addr, next, phys, prot, pgtable_alloc, 456 flags); 457 phys += next - addr; 458 } while (pgdp++, addr = next, addr != end); 459 } 460 461 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, 462 unsigned long virt, phys_addr_t size, 463 pgprot_t prot, 464 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 465 int flags) 466 { 467 mutex_lock(&fixmap_lock); 468 __create_pgd_mapping_locked(pgdir, phys, virt, size, prot, 469 pgtable_alloc, flags); 470 mutex_unlock(&fixmap_lock); 471 } 472 473 #define INVALID_PHYS_ADDR (-1ULL) 474 475 static phys_addr_t __pgd_pgtable_alloc(struct mm_struct *mm, gfp_t gfp, 476 enum pgtable_type pgtable_type) 477 { 478 /* Page is zeroed by init_clear_pgtable() so don't duplicate effort. */ 479 struct ptdesc *ptdesc = pagetable_alloc(gfp & ~__GFP_ZERO, 0); 480 phys_addr_t pa; 481 482 if (!ptdesc) 483 return INVALID_PHYS_ADDR; 484 485 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 506 try_pgd_pgtable_alloc_init_mm(enum pgtable_type pgtable_type, gfp_t gfp) 507 { 508 return __pgd_pgtable_alloc(&init_mm, gfp, pgtable_type); 509 } 510 511 static phys_addr_t __maybe_unused 512 pgd_pgtable_alloc_init_mm(enum pgtable_type pgtable_type) 513 { 514 phys_addr_t pa; 515 516 pa = __pgd_pgtable_alloc(&init_mm, GFP_PGTABLE_KERNEL, pgtable_type); 517 BUG_ON(pa == INVALID_PHYS_ADDR); 518 return pa; 519 } 520 521 static phys_addr_t 522 pgd_pgtable_alloc_special_mm(enum pgtable_type pgtable_type) 523 { 524 phys_addr_t pa; 525 526 pa = __pgd_pgtable_alloc(NULL, GFP_PGTABLE_KERNEL, pgtable_type); 527 BUG_ON(pa == INVALID_PHYS_ADDR); 528 return pa; 529 } 530 531 static void split_contpte(pte_t *ptep) 532 { 533 int i; 534 535 ptep = PTR_ALIGN_DOWN(ptep, sizeof(*ptep) * CONT_PTES); 536 for (i = 0; i < CONT_PTES; i++, ptep++) 537 __set_pte(ptep, pte_mknoncont(__ptep_get(ptep))); 538 } 539 540 static int split_pmd(pmd_t *pmdp, pmd_t pmd, gfp_t gfp, bool to_cont) 541 { 542 pmdval_t tableprot = PMD_TYPE_TABLE | PMD_TABLE_UXN | PMD_TABLE_AF; 543 unsigned long pfn = pmd_pfn(pmd); 544 pgprot_t prot = pmd_pgprot(pmd); 545 phys_addr_t pte_phys; 546 pte_t *ptep; 547 int i; 548 549 pte_phys = try_pgd_pgtable_alloc_init_mm(TABLE_PTE, gfp); 550 if (pte_phys == INVALID_PHYS_ADDR) 551 return -ENOMEM; 552 ptep = (pte_t *)phys_to_virt(pte_phys); 553 554 if (pgprot_val(prot) & PMD_SECT_PXN) 555 tableprot |= PMD_TABLE_PXN; 556 557 prot = __pgprot((pgprot_val(prot) & ~PTE_TYPE_MASK) | PTE_TYPE_PAGE); 558 prot = __pgprot(pgprot_val(prot) & ~PTE_CONT); 559 if (to_cont) 560 prot = __pgprot(pgprot_val(prot) | PTE_CONT); 561 562 for (i = 0; i < PTRS_PER_PTE; i++, ptep++, pfn++) 563 __set_pte(ptep, pfn_pte(pfn, prot)); 564 565 /* 566 * Ensure the pte entries are visible to the table walker by the time 567 * the pmd entry that points to the ptes is visible. 568 */ 569 dsb(ishst); 570 __pmd_populate(pmdp, pte_phys, tableprot); 571 572 return 0; 573 } 574 575 static void split_contpmd(pmd_t *pmdp) 576 { 577 int i; 578 579 pmdp = PTR_ALIGN_DOWN(pmdp, sizeof(*pmdp) * CONT_PMDS); 580 for (i = 0; i < CONT_PMDS; i++, pmdp++) 581 set_pmd(pmdp, pmd_mknoncont(pmdp_get(pmdp))); 582 } 583 584 static int split_pud(pud_t *pudp, pud_t pud, gfp_t gfp, bool to_cont) 585 { 586 pudval_t tableprot = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF; 587 unsigned int step = PMD_SIZE >> PAGE_SHIFT; 588 unsigned long pfn = pud_pfn(pud); 589 pgprot_t prot = pud_pgprot(pud); 590 phys_addr_t pmd_phys; 591 pmd_t *pmdp; 592 int i; 593 594 pmd_phys = try_pgd_pgtable_alloc_init_mm(TABLE_PMD, gfp); 595 if (pmd_phys == INVALID_PHYS_ADDR) 596 return -ENOMEM; 597 pmdp = (pmd_t *)phys_to_virt(pmd_phys); 598 599 if (pgprot_val(prot) & PMD_SECT_PXN) 600 tableprot |= PUD_TABLE_PXN; 601 602 prot = __pgprot((pgprot_val(prot) & ~PMD_TYPE_MASK) | PMD_TYPE_SECT); 603 prot = __pgprot(pgprot_val(prot) & ~PTE_CONT); 604 if (to_cont) 605 prot = __pgprot(pgprot_val(prot) | PTE_CONT); 606 607 for (i = 0; i < PTRS_PER_PMD; i++, pmdp++, pfn += step) 608 set_pmd(pmdp, pfn_pmd(pfn, prot)); 609 610 /* 611 * Ensure the pmd entries are visible to the table walker by the time 612 * the pud entry that points to the pmds is visible. 613 */ 614 dsb(ishst); 615 __pud_populate(pudp, pmd_phys, tableprot); 616 617 return 0; 618 } 619 620 static int split_kernel_leaf_mapping_locked(unsigned long addr) 621 { 622 pgd_t *pgdp, pgd; 623 p4d_t *p4dp, p4d; 624 pud_t *pudp, pud; 625 pmd_t *pmdp, pmd; 626 pte_t *ptep, pte; 627 int ret = 0; 628 629 /* 630 * PGD: If addr is PGD aligned then addr already describes a leaf 631 * boundary. If not present then there is nothing to split. 632 */ 633 if (ALIGN_DOWN(addr, PGDIR_SIZE) == addr) 634 goto out; 635 pgdp = pgd_offset_k(addr); 636 pgd = pgdp_get(pgdp); 637 if (!pgd_present(pgd)) 638 goto out; 639 640 /* 641 * P4D: If addr is P4D aligned then addr already describes a leaf 642 * boundary. If not present then there is nothing to split. 643 */ 644 if (ALIGN_DOWN(addr, P4D_SIZE) == addr) 645 goto out; 646 p4dp = p4d_offset(pgdp, addr); 647 p4d = p4dp_get(p4dp); 648 if (!p4d_present(p4d)) 649 goto out; 650 651 /* 652 * PUD: If addr is PUD aligned then addr already describes a leaf 653 * boundary. If not present then there is nothing to split. Otherwise, 654 * if we have a pud leaf, split to contpmd. 655 */ 656 if (ALIGN_DOWN(addr, PUD_SIZE) == addr) 657 goto out; 658 pudp = pud_offset(p4dp, addr); 659 pud = pudp_get(pudp); 660 if (!pud_present(pud)) 661 goto out; 662 if (pud_leaf(pud)) { 663 ret = split_pud(pudp, pud, GFP_PGTABLE_KERNEL, true); 664 if (ret) 665 goto out; 666 } 667 668 /* 669 * CONTPMD: If addr is CONTPMD aligned then addr already describes a 670 * leaf boundary. If not present then there is nothing to split. 671 * Otherwise, if we have a contpmd leaf, split to pmd. 672 */ 673 if (ALIGN_DOWN(addr, CONT_PMD_SIZE) == addr) 674 goto out; 675 pmdp = pmd_offset(pudp, addr); 676 pmd = pmdp_get(pmdp); 677 if (!pmd_present(pmd)) 678 goto out; 679 if (pmd_leaf(pmd)) { 680 if (pmd_cont(pmd)) 681 split_contpmd(pmdp); 682 /* 683 * PMD: If addr is PMD aligned then addr already describes a 684 * leaf boundary. Otherwise, split to contpte. 685 */ 686 if (ALIGN_DOWN(addr, PMD_SIZE) == addr) 687 goto out; 688 ret = split_pmd(pmdp, pmd, GFP_PGTABLE_KERNEL, true); 689 if (ret) 690 goto out; 691 } 692 693 /* 694 * CONTPTE: If addr is CONTPTE aligned then addr already describes a 695 * leaf boundary. If not present then there is nothing to split. 696 * Otherwise, if we have a contpte leaf, split to pte. 697 */ 698 if (ALIGN_DOWN(addr, CONT_PTE_SIZE) == addr) 699 goto out; 700 ptep = pte_offset_kernel(pmdp, addr); 701 pte = __ptep_get(ptep); 702 if (!pte_present(pte)) 703 goto out; 704 if (pte_cont(pte)) 705 split_contpte(ptep); 706 707 out: 708 return ret; 709 } 710 711 static inline bool force_pte_mapping(void) 712 { 713 const bool bbml2 = system_capabilities_finalized() ? 714 system_supports_bbml2_noabort() : cpu_supports_bbml2_noabort(); 715 716 if (debug_pagealloc_enabled()) 717 return true; 718 if (bbml2) 719 return false; 720 return rodata_full || arm64_kfence_can_set_direct_map() || is_realm_world(); 721 } 722 723 static inline bool split_leaf_mapping_possible(void) 724 { 725 /* 726 * !BBML2_NOABORT systems should never run into scenarios where we would 727 * have to split. So exit early and let calling code detect it and raise 728 * a warning. 729 */ 730 if (!system_supports_bbml2_noabort()) 731 return false; 732 return !force_pte_mapping(); 733 } 734 735 static DEFINE_MUTEX(pgtable_split_lock); 736 737 int split_kernel_leaf_mapping(unsigned long start, unsigned long end) 738 { 739 int ret; 740 741 /* 742 * Exit early if the region is within a pte-mapped area or if we can't 743 * split. For the latter case, the permission change code will raise a 744 * warning if not already pte-mapped. 745 */ 746 if (!split_leaf_mapping_possible() || is_kfence_address((void *)start)) 747 return 0; 748 749 /* 750 * Ensure start and end are at least page-aligned since this is the 751 * finest granularity we can split to. 752 */ 753 if (start != PAGE_ALIGN(start) || end != PAGE_ALIGN(end)) 754 return -EINVAL; 755 756 mutex_lock(&pgtable_split_lock); 757 arch_enter_lazy_mmu_mode(); 758 759 /* 760 * The split_kernel_leaf_mapping_locked() may sleep, it is not a 761 * problem for ARM64 since ARM64's lazy MMU implementation allows 762 * sleeping. 763 * 764 * Optimize for the common case of splitting out a single page from a 765 * larger mapping. Here we can just split on the "least aligned" of 766 * start and end and this will guarantee that there must also be a split 767 * on the more aligned address since the both addresses must be in the 768 * same contpte block and it must have been split to ptes. 769 */ 770 if (end - start == PAGE_SIZE) { 771 start = __ffs(start) < __ffs(end) ? start : end; 772 ret = split_kernel_leaf_mapping_locked(start); 773 } else { 774 ret = split_kernel_leaf_mapping_locked(start); 775 if (!ret) 776 ret = split_kernel_leaf_mapping_locked(end); 777 } 778 779 arch_leave_lazy_mmu_mode(); 780 mutex_unlock(&pgtable_split_lock); 781 return ret; 782 } 783 784 static int split_to_ptes_pud_entry(pud_t *pudp, unsigned long addr, 785 unsigned long next, struct mm_walk *walk) 786 { 787 gfp_t gfp = *(gfp_t *)walk->private; 788 pud_t pud = pudp_get(pudp); 789 int ret = 0; 790 791 if (pud_leaf(pud)) 792 ret = split_pud(pudp, pud, gfp, false); 793 794 return ret; 795 } 796 797 static int split_to_ptes_pmd_entry(pmd_t *pmdp, unsigned long addr, 798 unsigned long next, struct mm_walk *walk) 799 { 800 gfp_t gfp = *(gfp_t *)walk->private; 801 pmd_t pmd = pmdp_get(pmdp); 802 int ret = 0; 803 804 if (pmd_leaf(pmd)) { 805 if (pmd_cont(pmd)) 806 split_contpmd(pmdp); 807 ret = split_pmd(pmdp, pmd, gfp, false); 808 809 /* 810 * We have split the pmd directly to ptes so there is no need to 811 * visit each pte to check if they are contpte. 812 */ 813 walk->action = ACTION_CONTINUE; 814 } 815 816 return ret; 817 } 818 819 static int split_to_ptes_pte_entry(pte_t *ptep, unsigned long addr, 820 unsigned long next, struct mm_walk *walk) 821 { 822 pte_t pte = __ptep_get(ptep); 823 824 if (pte_cont(pte)) 825 split_contpte(ptep); 826 827 return 0; 828 } 829 830 static const struct mm_walk_ops split_to_ptes_ops = { 831 .pud_entry = split_to_ptes_pud_entry, 832 .pmd_entry = split_to_ptes_pmd_entry, 833 .pte_entry = split_to_ptes_pte_entry, 834 }; 835 836 static int range_split_to_ptes(unsigned long start, unsigned long end, gfp_t gfp) 837 { 838 int ret; 839 840 arch_enter_lazy_mmu_mode(); 841 ret = walk_kernel_page_table_range_lockless(start, end, 842 &split_to_ptes_ops, NULL, &gfp); 843 arch_leave_lazy_mmu_mode(); 844 845 return ret; 846 } 847 848 static bool linear_map_requires_bbml2 __initdata; 849 850 u32 idmap_kpti_bbml2_flag; 851 852 static void __init init_idmap_kpti_bbml2_flag(void) 853 { 854 WRITE_ONCE(idmap_kpti_bbml2_flag, 1); 855 /* Must be visible to other CPUs before stop_machine() is called. */ 856 smp_mb(); 857 } 858 859 static int __init linear_map_split_to_ptes(void *__unused) 860 { 861 /* 862 * Repainting the linear map must be done by CPU0 (the boot CPU) because 863 * that's the only CPU that we know supports BBML2. The other CPUs will 864 * be held in a waiting area with the idmap active. 865 */ 866 if (!smp_processor_id()) { 867 unsigned long lstart = _PAGE_OFFSET(vabits_actual); 868 unsigned long lend = PAGE_END; 869 unsigned long kstart = (unsigned long)lm_alias(_stext); 870 unsigned long kend = (unsigned long)lm_alias(__init_begin); 871 int ret; 872 873 /* 874 * Wait for all secondary CPUs to be put into the waiting area. 875 */ 876 smp_cond_load_acquire(&idmap_kpti_bbml2_flag, VAL == num_online_cpus()); 877 878 /* 879 * Walk all of the linear map [lstart, lend), except the kernel 880 * linear map alias [kstart, kend), and split all mappings to 881 * PTE. The kernel alias remains static throughout runtime so 882 * can continue to be safely mapped with large mappings. 883 */ 884 ret = range_split_to_ptes(lstart, kstart, GFP_ATOMIC); 885 if (!ret) 886 ret = range_split_to_ptes(kend, lend, GFP_ATOMIC); 887 if (ret) 888 panic("Failed to split linear map\n"); 889 flush_tlb_kernel_range(lstart, lend); 890 891 /* 892 * Relies on dsb in flush_tlb_kernel_range() to avoid reordering 893 * before any page table split operations. 894 */ 895 WRITE_ONCE(idmap_kpti_bbml2_flag, 0); 896 } else { 897 typedef void (wait_split_fn)(void); 898 extern wait_split_fn wait_linear_map_split_to_ptes; 899 wait_split_fn *wait_fn; 900 901 wait_fn = (void *)__pa_symbol(wait_linear_map_split_to_ptes); 902 903 /* 904 * At least one secondary CPU doesn't support BBML2 so cannot 905 * tolerate the size of the live mappings changing. So have the 906 * secondary CPUs wait for the boot CPU to make the changes 907 * with the idmap active and init_mm inactive. 908 */ 909 cpu_install_idmap(); 910 wait_fn(); 911 cpu_uninstall_idmap(); 912 } 913 914 return 0; 915 } 916 917 void __init linear_map_maybe_split_to_ptes(void) 918 { 919 if (linear_map_requires_bbml2 && !system_supports_bbml2_noabort()) { 920 init_idmap_kpti_bbml2_flag(); 921 stop_machine(linear_map_split_to_ptes, NULL, cpu_online_mask); 922 } 923 } 924 925 /* 926 * This function can only be used to modify existing table entries, 927 * without allocating new levels of table. Note that this permits the 928 * creation of new section or page entries. 929 */ 930 void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt, 931 phys_addr_t size, pgprot_t prot) 932 { 933 if (virt < PAGE_OFFSET) { 934 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n", 935 &phys, virt); 936 return; 937 } 938 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 939 NO_CONT_MAPPINGS); 940 } 941 942 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys, 943 unsigned long virt, phys_addr_t size, 944 pgprot_t prot, bool page_mappings_only) 945 { 946 int flags = 0; 947 948 BUG_ON(mm == &init_mm); 949 950 if (page_mappings_only) 951 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 952 953 __create_pgd_mapping(mm->pgd, phys, virt, size, prot, 954 pgd_pgtable_alloc_special_mm, flags); 955 } 956 957 static void update_mapping_prot(phys_addr_t phys, unsigned long virt, 958 phys_addr_t size, pgprot_t prot) 959 { 960 if (virt < PAGE_OFFSET) { 961 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n", 962 &phys, virt); 963 return; 964 } 965 966 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 967 NO_CONT_MAPPINGS); 968 969 /* flush the TLBs after updating live kernel mappings */ 970 flush_tlb_kernel_range(virt, virt + size); 971 } 972 973 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start, 974 phys_addr_t end, pgprot_t prot, int flags) 975 { 976 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start, 977 prot, early_pgtable_alloc, flags); 978 } 979 980 void __init mark_linear_text_alias_ro(void) 981 { 982 /* 983 * Remove the write permissions from the linear alias of .text/.rodata 984 */ 985 update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text), 986 (unsigned long)__init_begin - (unsigned long)_text, 987 PAGE_KERNEL_RO); 988 } 989 990 #ifdef CONFIG_KFENCE 991 992 bool __ro_after_init kfence_early_init = !!CONFIG_KFENCE_SAMPLE_INTERVAL; 993 994 /* early_param() will be parsed before map_mem() below. */ 995 static int __init parse_kfence_early_init(char *arg) 996 { 997 int val; 998 999 if (get_option(&arg, &val)) 1000 kfence_early_init = !!val; 1001 return 0; 1002 } 1003 early_param("kfence.sample_interval", parse_kfence_early_init); 1004 1005 static phys_addr_t __init arm64_kfence_alloc_pool(void) 1006 { 1007 phys_addr_t kfence_pool; 1008 1009 if (!kfence_early_init) 1010 return 0; 1011 1012 kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE); 1013 if (!kfence_pool) { 1014 pr_err("failed to allocate kfence pool\n"); 1015 kfence_early_init = false; 1016 return 0; 1017 } 1018 1019 /* Temporarily mark as NOMAP. */ 1020 memblock_mark_nomap(kfence_pool, KFENCE_POOL_SIZE); 1021 1022 return kfence_pool; 1023 } 1024 1025 static void __init arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) 1026 { 1027 if (!kfence_pool) 1028 return; 1029 1030 /* KFENCE pool needs page-level mapping. */ 1031 __map_memblock(pgdp, kfence_pool, kfence_pool + KFENCE_POOL_SIZE, 1032 pgprot_tagged(PAGE_KERNEL), 1033 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS); 1034 memblock_clear_nomap(kfence_pool, KFENCE_POOL_SIZE); 1035 __kfence_pool = phys_to_virt(kfence_pool); 1036 } 1037 1038 bool arch_kfence_init_pool(void) 1039 { 1040 unsigned long start = (unsigned long)__kfence_pool; 1041 unsigned long end = start + KFENCE_POOL_SIZE; 1042 int ret; 1043 1044 /* Exit early if we know the linear map is already pte-mapped. */ 1045 if (!split_leaf_mapping_possible()) 1046 return true; 1047 1048 /* Kfence pool is already pte-mapped for the early init case. */ 1049 if (kfence_early_init) 1050 return true; 1051 1052 mutex_lock(&pgtable_split_lock); 1053 ret = range_split_to_ptes(start, end, GFP_PGTABLE_KERNEL); 1054 mutex_unlock(&pgtable_split_lock); 1055 1056 /* 1057 * Since the system supports bbml2_noabort, tlb invalidation is not 1058 * required here; the pgtable mappings have been split to pte but larger 1059 * entries may safely linger in the TLB. 1060 */ 1061 1062 return !ret; 1063 } 1064 #else /* CONFIG_KFENCE */ 1065 1066 static inline phys_addr_t arm64_kfence_alloc_pool(void) { return 0; } 1067 static inline void arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) { } 1068 1069 #endif /* CONFIG_KFENCE */ 1070 1071 static void __init map_mem(pgd_t *pgdp) 1072 { 1073 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN); 1074 phys_addr_t kernel_start = __pa_symbol(_text); 1075 phys_addr_t kernel_end = __pa_symbol(__init_begin); 1076 phys_addr_t start, end; 1077 phys_addr_t early_kfence_pool; 1078 int flags = NO_EXEC_MAPPINGS; 1079 u64 i; 1080 1081 /* 1082 * Setting hierarchical PXNTable attributes on table entries covering 1083 * the linear region is only possible if it is guaranteed that no table 1084 * entries at any level are being shared between the linear region and 1085 * the vmalloc region. Check whether this is true for the PGD level, in 1086 * which case it is guaranteed to be true for all other levels as well. 1087 * (Unless we are running with support for LPA2, in which case the 1088 * entire reduced VA space is covered by a single pgd_t which will have 1089 * been populated without the PXNTable attribute by the time we get here.) 1090 */ 1091 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end) && 1092 pgd_index(_PAGE_OFFSET(VA_BITS_MIN)) != PTRS_PER_PGD - 1); 1093 1094 early_kfence_pool = arm64_kfence_alloc_pool(); 1095 1096 linear_map_requires_bbml2 = !force_pte_mapping() && can_set_direct_map(); 1097 1098 if (force_pte_mapping()) 1099 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1100 1101 /* 1102 * Take care not to create a writable alias for the 1103 * read-only text and rodata sections of the kernel image. 1104 * So temporarily mark them as NOMAP to skip mappings in 1105 * the following for-loop 1106 */ 1107 memblock_mark_nomap(kernel_start, kernel_end - kernel_start); 1108 1109 /* map all the memory banks */ 1110 for_each_mem_range(i, &start, &end) { 1111 if (start >= end) 1112 break; 1113 /* 1114 * The linear map must allow allocation tags reading/writing 1115 * if MTE is present. Otherwise, it has the same attributes as 1116 * PAGE_KERNEL. 1117 */ 1118 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL), 1119 flags); 1120 } 1121 1122 /* 1123 * Map the linear alias of the [_text, __init_begin) interval 1124 * as non-executable now, and remove the write permission in 1125 * mark_linear_text_alias_ro() below (which will be called after 1126 * alternative patching has completed). This makes the contents 1127 * of the region accessible to subsystems such as hibernate, 1128 * but protects it from inadvertent modification or execution. 1129 * Note that contiguous mappings cannot be remapped in this way, 1130 * so we should avoid them here. 1131 */ 1132 __map_memblock(pgdp, kernel_start, kernel_end, 1133 PAGE_KERNEL, NO_CONT_MAPPINGS); 1134 memblock_clear_nomap(kernel_start, kernel_end - kernel_start); 1135 arm64_kfence_map_pool(early_kfence_pool, pgdp); 1136 } 1137 1138 void mark_rodata_ro(void) 1139 { 1140 unsigned long section_size; 1141 1142 /* 1143 * mark .rodata as read only. Use __init_begin rather than __end_rodata 1144 * to cover NOTES and EXCEPTION_TABLE. 1145 */ 1146 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata; 1147 WRITE_ONCE(rodata_is_rw, false); 1148 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata, 1149 section_size, PAGE_KERNEL_RO); 1150 /* mark the range between _text and _stext as read only. */ 1151 update_mapping_prot(__pa_symbol(_text), (unsigned long)_text, 1152 (unsigned long)_stext - (unsigned long)_text, 1153 PAGE_KERNEL_RO); 1154 } 1155 1156 static void __init declare_vma(struct vm_struct *vma, 1157 void *va_start, void *va_end, 1158 unsigned long vm_flags) 1159 { 1160 phys_addr_t pa_start = __pa_symbol(va_start); 1161 unsigned long size = va_end - va_start; 1162 1163 BUG_ON(!PAGE_ALIGNED(pa_start)); 1164 BUG_ON(!PAGE_ALIGNED(size)); 1165 1166 if (!(vm_flags & VM_NO_GUARD)) 1167 size += PAGE_SIZE; 1168 1169 vma->addr = va_start; 1170 vma->phys_addr = pa_start; 1171 vma->size = size; 1172 vma->flags = VM_MAP | vm_flags; 1173 vma->caller = __builtin_return_address(0); 1174 1175 vm_area_add_early(vma); 1176 } 1177 1178 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0 1179 #define KPTI_NG_TEMP_VA (-(1UL << PMD_SHIFT)) 1180 1181 static phys_addr_t kpti_ng_temp_alloc __initdata; 1182 1183 static phys_addr_t __init kpti_ng_pgd_alloc(enum pgtable_type type) 1184 { 1185 kpti_ng_temp_alloc -= PAGE_SIZE; 1186 return kpti_ng_temp_alloc; 1187 } 1188 1189 static int __init __kpti_install_ng_mappings(void *__unused) 1190 { 1191 typedef void (kpti_remap_fn)(int, int, phys_addr_t, unsigned long); 1192 extern kpti_remap_fn idmap_kpti_install_ng_mappings; 1193 kpti_remap_fn *remap_fn; 1194 1195 int cpu = smp_processor_id(); 1196 int levels = CONFIG_PGTABLE_LEVELS; 1197 int order = order_base_2(levels); 1198 u64 kpti_ng_temp_pgd_pa = 0; 1199 pgd_t *kpti_ng_temp_pgd; 1200 u64 alloc = 0; 1201 1202 if (levels == 5 && !pgtable_l5_enabled()) 1203 levels = 4; 1204 else if (levels == 4 && !pgtable_l4_enabled()) 1205 levels = 3; 1206 1207 remap_fn = (void *)__pa_symbol(idmap_kpti_install_ng_mappings); 1208 1209 if (!cpu) { 1210 alloc = __get_free_pages(GFP_ATOMIC | __GFP_ZERO, order); 1211 kpti_ng_temp_pgd = (pgd_t *)(alloc + (levels - 1) * PAGE_SIZE); 1212 kpti_ng_temp_alloc = kpti_ng_temp_pgd_pa = __pa(kpti_ng_temp_pgd); 1213 1214 // 1215 // Create a minimal page table hierarchy that permits us to map 1216 // the swapper page tables temporarily as we traverse them. 1217 // 1218 // The physical pages are laid out as follows: 1219 // 1220 // +--------+-/-------+-/------ +-/------ +-\\\--------+ 1221 // : PTE[] : | PMD[] : | PUD[] : | P4D[] : ||| PGD[] : 1222 // +--------+-\-------+-\------ +-\------ +-///--------+ 1223 // ^ 1224 // The first page is mapped into this hierarchy at a PMD_SHIFT 1225 // aligned virtual address, so that we can manipulate the PTE 1226 // level entries while the mapping is active. The first entry 1227 // covers the PTE[] page itself, the remaining entries are free 1228 // to be used as a ad-hoc fixmap. 1229 // 1230 __create_pgd_mapping_locked(kpti_ng_temp_pgd, __pa(alloc), 1231 KPTI_NG_TEMP_VA, PAGE_SIZE, PAGE_KERNEL, 1232 kpti_ng_pgd_alloc, 0); 1233 } 1234 1235 cpu_install_idmap(); 1236 remap_fn(cpu, num_online_cpus(), kpti_ng_temp_pgd_pa, KPTI_NG_TEMP_VA); 1237 cpu_uninstall_idmap(); 1238 1239 if (!cpu) { 1240 free_pages(alloc, order); 1241 arm64_use_ng_mappings = true; 1242 } 1243 1244 return 0; 1245 } 1246 1247 void __init kpti_install_ng_mappings(void) 1248 { 1249 /* Check whether KPTI is going to be used */ 1250 if (!arm64_kernel_unmapped_at_el0()) 1251 return; 1252 1253 /* 1254 * We don't need to rewrite the page-tables if either we've done 1255 * it already or we have KASLR enabled and therefore have not 1256 * created any global mappings at all. 1257 */ 1258 if (arm64_use_ng_mappings) 1259 return; 1260 1261 init_idmap_kpti_bbml2_flag(); 1262 stop_machine(__kpti_install_ng_mappings, NULL, cpu_online_mask); 1263 } 1264 1265 static pgprot_t __init kernel_exec_prot(void) 1266 { 1267 return rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 1268 } 1269 1270 static int __init map_entry_trampoline(void) 1271 { 1272 int i; 1273 1274 if (!arm64_kernel_unmapped_at_el0()) 1275 return 0; 1276 1277 pgprot_t prot = kernel_exec_prot(); 1278 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start); 1279 1280 /* The trampoline is always mapped and can therefore be global */ 1281 pgprot_val(prot) &= ~PTE_NG; 1282 1283 /* Map only the text into the trampoline page table */ 1284 memset(tramp_pg_dir, 0, PGD_SIZE); 1285 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, 1286 entry_tramp_text_size(), prot, 1287 pgd_pgtable_alloc_init_mm, NO_BLOCK_MAPPINGS); 1288 1289 /* Map both the text and data into the kernel page table */ 1290 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++) 1291 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 1292 pa_start + i * PAGE_SIZE, prot); 1293 1294 if (IS_ENABLED(CONFIG_RELOCATABLE)) 1295 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 1296 pa_start + i * PAGE_SIZE, PAGE_KERNEL_RO); 1297 1298 return 0; 1299 } 1300 core_initcall(map_entry_trampoline); 1301 #endif 1302 1303 /* 1304 * Declare the VMA areas for the kernel 1305 */ 1306 static void __init declare_kernel_vmas(void) 1307 { 1308 static struct vm_struct vmlinux_seg[KERNEL_SEGMENT_COUNT]; 1309 1310 declare_vma(&vmlinux_seg[0], _text, _etext, VM_NO_GUARD); 1311 declare_vma(&vmlinux_seg[1], __start_rodata, __inittext_begin, VM_NO_GUARD); 1312 declare_vma(&vmlinux_seg[2], __inittext_begin, __inittext_end, VM_NO_GUARD); 1313 declare_vma(&vmlinux_seg[3], __initdata_begin, __initdata_end, VM_NO_GUARD); 1314 declare_vma(&vmlinux_seg[4], _data, _end, 0); 1315 } 1316 1317 void __pi_map_range(phys_addr_t *pte, u64 start, u64 end, phys_addr_t pa, 1318 pgprot_t prot, int level, pte_t *tbl, bool may_use_cont, 1319 u64 va_offset); 1320 1321 static u8 idmap_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init, 1322 kpti_bbml2_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init; 1323 1324 static void __init create_idmap(void) 1325 { 1326 phys_addr_t start = __pa_symbol(__idmap_text_start); 1327 phys_addr_t end = __pa_symbol(__idmap_text_end); 1328 phys_addr_t ptep = __pa_symbol(idmap_ptes); 1329 1330 __pi_map_range(&ptep, start, end, start, PAGE_KERNEL_ROX, 1331 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 1332 __phys_to_virt(ptep) - ptep); 1333 1334 if (linear_map_requires_bbml2 || 1335 (IS_ENABLED(CONFIG_UNMAP_KERNEL_AT_EL0) && !arm64_use_ng_mappings)) { 1336 phys_addr_t pa = __pa_symbol(&idmap_kpti_bbml2_flag); 1337 1338 /* 1339 * The KPTI G-to-nG conversion code needs a read-write mapping 1340 * of its synchronization flag in the ID map. This is also used 1341 * when splitting the linear map to ptes if a secondary CPU 1342 * doesn't support bbml2. 1343 */ 1344 ptep = __pa_symbol(kpti_bbml2_ptes); 1345 __pi_map_range(&ptep, pa, pa + sizeof(u32), pa, PAGE_KERNEL, 1346 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 1347 __phys_to_virt(ptep) - ptep); 1348 } 1349 } 1350 1351 void __init paging_init(void) 1352 { 1353 map_mem(swapper_pg_dir); 1354 1355 memblock_allow_resize(); 1356 1357 create_idmap(); 1358 declare_kernel_vmas(); 1359 } 1360 1361 #ifdef CONFIG_MEMORY_HOTPLUG 1362 static void free_hotplug_page_range(struct page *page, size_t size, 1363 struct vmem_altmap *altmap) 1364 { 1365 if (altmap) { 1366 vmem_altmap_free(altmap, size >> PAGE_SHIFT); 1367 } else { 1368 WARN_ON(PageReserved(page)); 1369 __free_pages(page, get_order(size)); 1370 } 1371 } 1372 1373 static void free_hotplug_pgtable_page(struct page *page) 1374 { 1375 free_hotplug_page_range(page, PAGE_SIZE, NULL); 1376 } 1377 1378 static bool pgtable_range_aligned(unsigned long start, unsigned long end, 1379 unsigned long floor, unsigned long ceiling, 1380 unsigned long mask) 1381 { 1382 start &= mask; 1383 if (start < floor) 1384 return false; 1385 1386 if (ceiling) { 1387 ceiling &= mask; 1388 if (!ceiling) 1389 return false; 1390 } 1391 1392 if (end - 1 > ceiling - 1) 1393 return false; 1394 return true; 1395 } 1396 1397 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr, 1398 unsigned long end, bool free_mapped, 1399 struct vmem_altmap *altmap) 1400 { 1401 pte_t *ptep, pte; 1402 1403 do { 1404 ptep = pte_offset_kernel(pmdp, addr); 1405 pte = __ptep_get(ptep); 1406 if (pte_none(pte)) 1407 continue; 1408 1409 WARN_ON(!pte_present(pte)); 1410 __pte_clear(&init_mm, addr, ptep); 1411 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1412 if (free_mapped) 1413 free_hotplug_page_range(pte_page(pte), 1414 PAGE_SIZE, altmap); 1415 } while (addr += PAGE_SIZE, addr < end); 1416 } 1417 1418 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr, 1419 unsigned long end, bool free_mapped, 1420 struct vmem_altmap *altmap) 1421 { 1422 unsigned long next; 1423 pmd_t *pmdp, pmd; 1424 1425 do { 1426 next = pmd_addr_end(addr, end); 1427 pmdp = pmd_offset(pudp, addr); 1428 pmd = READ_ONCE(*pmdp); 1429 if (pmd_none(pmd)) 1430 continue; 1431 1432 WARN_ON(!pmd_present(pmd)); 1433 if (pmd_sect(pmd)) { 1434 pmd_clear(pmdp); 1435 1436 /* 1437 * One TLBI should be sufficient here as the PMD_SIZE 1438 * range is mapped with a single block entry. 1439 */ 1440 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1441 if (free_mapped) 1442 free_hotplug_page_range(pmd_page(pmd), 1443 PMD_SIZE, altmap); 1444 continue; 1445 } 1446 WARN_ON(!pmd_table(pmd)); 1447 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap); 1448 } while (addr = next, addr < end); 1449 } 1450 1451 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr, 1452 unsigned long end, bool free_mapped, 1453 struct vmem_altmap *altmap) 1454 { 1455 unsigned long next; 1456 pud_t *pudp, pud; 1457 1458 do { 1459 next = pud_addr_end(addr, end); 1460 pudp = pud_offset(p4dp, addr); 1461 pud = READ_ONCE(*pudp); 1462 if (pud_none(pud)) 1463 continue; 1464 1465 WARN_ON(!pud_present(pud)); 1466 if (pud_sect(pud)) { 1467 pud_clear(pudp); 1468 1469 /* 1470 * One TLBI should be sufficient here as the PUD_SIZE 1471 * range is mapped with a single block entry. 1472 */ 1473 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1474 if (free_mapped) 1475 free_hotplug_page_range(pud_page(pud), 1476 PUD_SIZE, altmap); 1477 continue; 1478 } 1479 WARN_ON(!pud_table(pud)); 1480 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap); 1481 } while (addr = next, addr < end); 1482 } 1483 1484 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr, 1485 unsigned long end, bool free_mapped, 1486 struct vmem_altmap *altmap) 1487 { 1488 unsigned long next; 1489 p4d_t *p4dp, p4d; 1490 1491 do { 1492 next = p4d_addr_end(addr, end); 1493 p4dp = p4d_offset(pgdp, addr); 1494 p4d = READ_ONCE(*p4dp); 1495 if (p4d_none(p4d)) 1496 continue; 1497 1498 WARN_ON(!p4d_present(p4d)); 1499 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap); 1500 } while (addr = next, addr < end); 1501 } 1502 1503 static void unmap_hotplug_range(unsigned long addr, unsigned long end, 1504 bool free_mapped, struct vmem_altmap *altmap) 1505 { 1506 unsigned long next; 1507 pgd_t *pgdp, pgd; 1508 1509 /* 1510 * altmap can only be used as vmemmap mapping backing memory. 1511 * In case the backing memory itself is not being freed, then 1512 * altmap is irrelevant. Warn about this inconsistency when 1513 * encountered. 1514 */ 1515 WARN_ON(!free_mapped && altmap); 1516 1517 do { 1518 next = pgd_addr_end(addr, end); 1519 pgdp = pgd_offset_k(addr); 1520 pgd = READ_ONCE(*pgdp); 1521 if (pgd_none(pgd)) 1522 continue; 1523 1524 WARN_ON(!pgd_present(pgd)); 1525 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap); 1526 } while (addr = next, addr < end); 1527 } 1528 1529 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr, 1530 unsigned long end, unsigned long floor, 1531 unsigned long ceiling) 1532 { 1533 pte_t *ptep, pte; 1534 unsigned long i, start = addr; 1535 1536 do { 1537 ptep = pte_offset_kernel(pmdp, addr); 1538 pte = __ptep_get(ptep); 1539 1540 /* 1541 * This is just a sanity check here which verifies that 1542 * pte clearing has been done by earlier unmap loops. 1543 */ 1544 WARN_ON(!pte_none(pte)); 1545 } while (addr += PAGE_SIZE, addr < end); 1546 1547 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK)) 1548 return; 1549 1550 /* 1551 * Check whether we can free the pte page if the rest of the 1552 * entries are empty. Overlap with other regions have been 1553 * handled by the floor/ceiling check. 1554 */ 1555 ptep = pte_offset_kernel(pmdp, 0UL); 1556 for (i = 0; i < PTRS_PER_PTE; i++) { 1557 if (!pte_none(__ptep_get(&ptep[i]))) 1558 return; 1559 } 1560 1561 pmd_clear(pmdp); 1562 __flush_tlb_kernel_pgtable(start); 1563 free_hotplug_pgtable_page(virt_to_page(ptep)); 1564 } 1565 1566 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr, 1567 unsigned long end, unsigned long floor, 1568 unsigned long ceiling) 1569 { 1570 pmd_t *pmdp, pmd; 1571 unsigned long i, next, start = addr; 1572 1573 do { 1574 next = pmd_addr_end(addr, end); 1575 pmdp = pmd_offset(pudp, addr); 1576 pmd = READ_ONCE(*pmdp); 1577 if (pmd_none(pmd)) 1578 continue; 1579 1580 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd)); 1581 free_empty_pte_table(pmdp, addr, next, floor, ceiling); 1582 } while (addr = next, addr < end); 1583 1584 if (CONFIG_PGTABLE_LEVELS <= 2) 1585 return; 1586 1587 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK)) 1588 return; 1589 1590 /* 1591 * Check whether we can free the pmd page if the rest of the 1592 * entries are empty. Overlap with other regions have been 1593 * handled by the floor/ceiling check. 1594 */ 1595 pmdp = pmd_offset(pudp, 0UL); 1596 for (i = 0; i < PTRS_PER_PMD; i++) { 1597 if (!pmd_none(READ_ONCE(pmdp[i]))) 1598 return; 1599 } 1600 1601 pud_clear(pudp); 1602 __flush_tlb_kernel_pgtable(start); 1603 free_hotplug_pgtable_page(virt_to_page(pmdp)); 1604 } 1605 1606 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr, 1607 unsigned long end, unsigned long floor, 1608 unsigned long ceiling) 1609 { 1610 pud_t *pudp, pud; 1611 unsigned long i, next, start = addr; 1612 1613 do { 1614 next = pud_addr_end(addr, end); 1615 pudp = pud_offset(p4dp, addr); 1616 pud = READ_ONCE(*pudp); 1617 if (pud_none(pud)) 1618 continue; 1619 1620 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud)); 1621 free_empty_pmd_table(pudp, addr, next, floor, ceiling); 1622 } while (addr = next, addr < end); 1623 1624 if (!pgtable_l4_enabled()) 1625 return; 1626 1627 if (!pgtable_range_aligned(start, end, floor, ceiling, P4D_MASK)) 1628 return; 1629 1630 /* 1631 * Check whether we can free the pud 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 pudp = pud_offset(p4dp, 0UL); 1636 for (i = 0; i < PTRS_PER_PUD; i++) { 1637 if (!pud_none(READ_ONCE(pudp[i]))) 1638 return; 1639 } 1640 1641 p4d_clear(p4dp); 1642 __flush_tlb_kernel_pgtable(start); 1643 free_hotplug_pgtable_page(virt_to_page(pudp)); 1644 } 1645 1646 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr, 1647 unsigned long end, unsigned long floor, 1648 unsigned long ceiling) 1649 { 1650 p4d_t *p4dp, p4d; 1651 unsigned long i, next, start = addr; 1652 1653 do { 1654 next = p4d_addr_end(addr, end); 1655 p4dp = p4d_offset(pgdp, addr); 1656 p4d = READ_ONCE(*p4dp); 1657 if (p4d_none(p4d)) 1658 continue; 1659 1660 WARN_ON(!p4d_present(p4d)); 1661 free_empty_pud_table(p4dp, addr, next, floor, ceiling); 1662 } while (addr = next, addr < end); 1663 1664 if (!pgtable_l5_enabled()) 1665 return; 1666 1667 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK)) 1668 return; 1669 1670 /* 1671 * Check whether we can free the p4d 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 p4dp = p4d_offset(pgdp, 0UL); 1676 for (i = 0; i < PTRS_PER_P4D; i++) { 1677 if (!p4d_none(READ_ONCE(p4dp[i]))) 1678 return; 1679 } 1680 1681 pgd_clear(pgdp); 1682 __flush_tlb_kernel_pgtable(start); 1683 free_hotplug_pgtable_page(virt_to_page(p4dp)); 1684 } 1685 1686 static void free_empty_tables(unsigned long addr, unsigned long end, 1687 unsigned long floor, unsigned long ceiling) 1688 { 1689 unsigned long next; 1690 pgd_t *pgdp, pgd; 1691 1692 do { 1693 next = pgd_addr_end(addr, end); 1694 pgdp = pgd_offset_k(addr); 1695 pgd = READ_ONCE(*pgdp); 1696 if (pgd_none(pgd)) 1697 continue; 1698 1699 WARN_ON(!pgd_present(pgd)); 1700 free_empty_p4d_table(pgdp, addr, next, floor, ceiling); 1701 } while (addr = next, addr < end); 1702 } 1703 #endif 1704 1705 void __meminit vmemmap_set_pmd(pmd_t *pmdp, void *p, int node, 1706 unsigned long addr, unsigned long next) 1707 { 1708 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL)); 1709 } 1710 1711 int __meminit vmemmap_check_pmd(pmd_t *pmdp, int node, 1712 unsigned long addr, unsigned long next) 1713 { 1714 vmemmap_verify((pte_t *)pmdp, node, addr, next); 1715 1716 return pmd_sect(READ_ONCE(*pmdp)); 1717 } 1718 1719 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1720 struct vmem_altmap *altmap) 1721 { 1722 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1723 /* [start, end] should be within one section */ 1724 WARN_ON_ONCE(end - start > PAGES_PER_SECTION * sizeof(struct page)); 1725 1726 if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || 1727 (end - start < PAGES_PER_SECTION * sizeof(struct page))) 1728 return vmemmap_populate_basepages(start, end, node, altmap); 1729 else 1730 return vmemmap_populate_hugepages(start, end, node, altmap); 1731 } 1732 1733 #ifdef CONFIG_MEMORY_HOTPLUG 1734 void vmemmap_free(unsigned long start, unsigned long end, 1735 struct vmem_altmap *altmap) 1736 { 1737 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1738 1739 unmap_hotplug_range(start, end, true, altmap); 1740 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END); 1741 } 1742 #endif /* CONFIG_MEMORY_HOTPLUG */ 1743 1744 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot) 1745 { 1746 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot)); 1747 1748 /* Only allow permission changes for now */ 1749 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)), 1750 pud_val(new_pud))) 1751 return 0; 1752 1753 VM_BUG_ON(phys & ~PUD_MASK); 1754 set_pud(pudp, new_pud); 1755 return 1; 1756 } 1757 1758 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot) 1759 { 1760 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot)); 1761 1762 /* Only allow permission changes for now */ 1763 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)), 1764 pmd_val(new_pmd))) 1765 return 0; 1766 1767 VM_BUG_ON(phys & ~PMD_MASK); 1768 set_pmd(pmdp, new_pmd); 1769 return 1; 1770 } 1771 1772 #ifndef __PAGETABLE_P4D_FOLDED 1773 void p4d_clear_huge(p4d_t *p4dp) 1774 { 1775 } 1776 #endif 1777 1778 int pud_clear_huge(pud_t *pudp) 1779 { 1780 if (!pud_sect(READ_ONCE(*pudp))) 1781 return 0; 1782 pud_clear(pudp); 1783 return 1; 1784 } 1785 1786 int pmd_clear_huge(pmd_t *pmdp) 1787 { 1788 if (!pmd_sect(READ_ONCE(*pmdp))) 1789 return 0; 1790 pmd_clear(pmdp); 1791 return 1; 1792 } 1793 1794 static int __pmd_free_pte_page(pmd_t *pmdp, unsigned long addr, 1795 bool acquire_mmap_lock) 1796 { 1797 pte_t *table; 1798 pmd_t pmd; 1799 1800 pmd = READ_ONCE(*pmdp); 1801 1802 if (!pmd_table(pmd)) { 1803 VM_WARN_ON(1); 1804 return 1; 1805 } 1806 1807 /* See comment in pud_free_pmd_page for static key logic */ 1808 table = pte_offset_kernel(pmdp, addr); 1809 pmd_clear(pmdp); 1810 __flush_tlb_kernel_pgtable(addr); 1811 if (static_branch_unlikely(&arm64_ptdump_lock_key) && acquire_mmap_lock) { 1812 mmap_read_lock(&init_mm); 1813 mmap_read_unlock(&init_mm); 1814 } 1815 1816 pte_free_kernel(NULL, table); 1817 return 1; 1818 } 1819 1820 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr) 1821 { 1822 /* If ptdump is walking the pagetables, acquire init_mm.mmap_lock */ 1823 return __pmd_free_pte_page(pmdp, addr, /* acquire_mmap_lock = */ true); 1824 } 1825 1826 int pud_free_pmd_page(pud_t *pudp, unsigned long addr) 1827 { 1828 pmd_t *table; 1829 pmd_t *pmdp; 1830 pud_t pud; 1831 unsigned long next, end; 1832 1833 pud = READ_ONCE(*pudp); 1834 1835 if (!pud_table(pud)) { 1836 VM_WARN_ON(1); 1837 return 1; 1838 } 1839 1840 table = pmd_offset(pudp, addr); 1841 1842 /* 1843 * Our objective is to prevent ptdump from reading a PMD table which has 1844 * been freed. In this race, if pud_free_pmd_page observes the key on 1845 * (which got flipped by ptdump) then the mmap lock sequence here will, 1846 * as a result of the mmap write lock/unlock sequence in ptdump, give 1847 * us the correct synchronization. If not, this means that ptdump has 1848 * yet not started walking the pagetables - the sequence of barriers 1849 * issued by __flush_tlb_kernel_pgtable() guarantees that ptdump will 1850 * observe an empty PUD. 1851 */ 1852 pud_clear(pudp); 1853 __flush_tlb_kernel_pgtable(addr); 1854 if (static_branch_unlikely(&arm64_ptdump_lock_key)) { 1855 mmap_read_lock(&init_mm); 1856 mmap_read_unlock(&init_mm); 1857 } 1858 1859 pmdp = table; 1860 next = addr; 1861 end = addr + PUD_SIZE; 1862 do { 1863 if (pmd_present(pmdp_get(pmdp))) 1864 /* 1865 * PMD has been isolated, so ptdump won't see it. No 1866 * need to acquire init_mm.mmap_lock. 1867 */ 1868 __pmd_free_pte_page(pmdp, next, /* acquire_mmap_lock = */ false); 1869 } while (pmdp++, next += PMD_SIZE, next != end); 1870 1871 pmd_free(NULL, table); 1872 return 1; 1873 } 1874 1875 #ifdef CONFIG_MEMORY_HOTPLUG 1876 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size) 1877 { 1878 unsigned long end = start + size; 1879 1880 WARN_ON(pgdir != init_mm.pgd); 1881 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END)); 1882 1883 unmap_hotplug_range(start, end, false, NULL); 1884 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END); 1885 } 1886 1887 struct range arch_get_mappable_range(void) 1888 { 1889 struct range mhp_range; 1890 phys_addr_t start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual)); 1891 phys_addr_t end_linear_pa = __pa(PAGE_END - 1); 1892 1893 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 1894 /* 1895 * Check for a wrap, it is possible because of randomized linear 1896 * mapping the start physical address is actually bigger than 1897 * the end physical address. In this case set start to zero 1898 * because [0, end_linear_pa] range must still be able to cover 1899 * all addressable physical addresses. 1900 */ 1901 if (start_linear_pa > end_linear_pa) 1902 start_linear_pa = 0; 1903 } 1904 1905 WARN_ON(start_linear_pa > end_linear_pa); 1906 1907 /* 1908 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)] 1909 * accommodating both its ends but excluding PAGE_END. Max physical 1910 * range which can be mapped inside this linear mapping range, must 1911 * also be derived from its end points. 1912 */ 1913 mhp_range.start = start_linear_pa; 1914 mhp_range.end = end_linear_pa; 1915 1916 return mhp_range; 1917 } 1918 1919 int arch_add_memory(int nid, u64 start, u64 size, 1920 struct mhp_params *params) 1921 { 1922 int ret, flags = NO_EXEC_MAPPINGS; 1923 1924 VM_BUG_ON(!mhp_range_allowed(start, size, true)); 1925 1926 if (force_pte_mapping()) 1927 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1928 1929 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1930 size, params->pgprot, pgd_pgtable_alloc_init_mm, 1931 flags); 1932 1933 memblock_clear_nomap(start, size); 1934 1935 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, 1936 params); 1937 if (ret) 1938 __remove_pgd_mapping(swapper_pg_dir, 1939 __phys_to_virt(start), size); 1940 else { 1941 /* Address of hotplugged memory can be smaller */ 1942 max_pfn = max(max_pfn, PFN_UP(start + size)); 1943 max_low_pfn = max_pfn; 1944 } 1945 1946 return ret; 1947 } 1948 1949 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap) 1950 { 1951 unsigned long start_pfn = start >> PAGE_SHIFT; 1952 unsigned long nr_pages = size >> PAGE_SHIFT; 1953 1954 __remove_pages(start_pfn, nr_pages, altmap); 1955 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size); 1956 } 1957 1958 /* 1959 * This memory hotplug notifier helps prevent boot memory from being 1960 * inadvertently removed as it blocks pfn range offlining process in 1961 * __offline_pages(). Hence this prevents both offlining as well as 1962 * removal process for boot memory which is initially always online. 1963 * In future if and when boot memory could be removed, this notifier 1964 * should be dropped and free_hotplug_page_range() should handle any 1965 * reserved pages allocated during boot. 1966 */ 1967 static int prevent_bootmem_remove_notifier(struct notifier_block *nb, 1968 unsigned long action, void *data) 1969 { 1970 struct mem_section *ms; 1971 struct memory_notify *arg = data; 1972 unsigned long end_pfn = arg->start_pfn + arg->nr_pages; 1973 unsigned long pfn = arg->start_pfn; 1974 1975 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE)) 1976 return NOTIFY_OK; 1977 1978 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 1979 unsigned long start = PFN_PHYS(pfn); 1980 unsigned long end = start + (1UL << PA_SECTION_SHIFT); 1981 1982 ms = __pfn_to_section(pfn); 1983 if (!early_section(ms)) 1984 continue; 1985 1986 if (action == MEM_GOING_OFFLINE) { 1987 /* 1988 * Boot memory removal is not supported. Prevent 1989 * it via blocking any attempted offline request 1990 * for the boot memory and just report it. 1991 */ 1992 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end); 1993 return NOTIFY_BAD; 1994 } else if (action == MEM_OFFLINE) { 1995 /* 1996 * This should have never happened. Boot memory 1997 * offlining should have been prevented by this 1998 * very notifier. Probably some memory removal 1999 * procedure might have changed which would then 2000 * require further debug. 2001 */ 2002 pr_err("Boot memory [%lx %lx] offlined\n", start, end); 2003 2004 /* 2005 * Core memory hotplug does not process a return 2006 * code from the notifier for MEM_OFFLINE events. 2007 * The error condition has been reported. Return 2008 * from here as if ignored. 2009 */ 2010 return NOTIFY_DONE; 2011 } 2012 } 2013 return NOTIFY_OK; 2014 } 2015 2016 static struct notifier_block prevent_bootmem_remove_nb = { 2017 .notifier_call = prevent_bootmem_remove_notifier, 2018 }; 2019 2020 /* 2021 * This ensures that boot memory sections on the platform are online 2022 * from early boot. Memory sections could not be prevented from being 2023 * offlined, unless for some reason they are not online to begin with. 2024 * This helps validate the basic assumption on which the above memory 2025 * event notifier works to prevent boot memory section offlining and 2026 * its possible removal. 2027 */ 2028 static void validate_bootmem_online(void) 2029 { 2030 phys_addr_t start, end, addr; 2031 struct mem_section *ms; 2032 u64 i; 2033 2034 /* 2035 * Scanning across all memblock might be expensive 2036 * on some big memory systems. Hence enable this 2037 * validation only with DEBUG_VM. 2038 */ 2039 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 2040 return; 2041 2042 for_each_mem_range(i, &start, &end) { 2043 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) { 2044 ms = __pfn_to_section(PHYS_PFN(addr)); 2045 2046 /* 2047 * All memory ranges in the system at this point 2048 * should have been marked as early sections. 2049 */ 2050 WARN_ON(!early_section(ms)); 2051 2052 /* 2053 * Memory notifier mechanism here to prevent boot 2054 * memory offlining depends on the fact that each 2055 * early section memory on the system is initially 2056 * online. Otherwise a given memory section which 2057 * is already offline will be overlooked and can 2058 * be removed completely. Call out such sections. 2059 */ 2060 if (!online_section(ms)) 2061 pr_err("Boot memory [%llx %llx] is offline, can be removed\n", 2062 addr, addr + (1UL << PA_SECTION_SHIFT)); 2063 } 2064 } 2065 } 2066 2067 static int __init prevent_bootmem_remove_init(void) 2068 { 2069 int ret = 0; 2070 2071 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) 2072 return ret; 2073 2074 validate_bootmem_online(); 2075 ret = register_memory_notifier(&prevent_bootmem_remove_nb); 2076 if (ret) 2077 pr_err("%s: Notifier registration failed %d\n", __func__, ret); 2078 2079 return ret; 2080 } 2081 early_initcall(prevent_bootmem_remove_init); 2082 #endif 2083 2084 pte_t modify_prot_start_ptes(struct vm_area_struct *vma, unsigned long addr, 2085 pte_t *ptep, unsigned int nr) 2086 { 2087 pte_t pte = get_and_clear_ptes(vma->vm_mm, addr, ptep, nr); 2088 2089 if (alternative_has_cap_unlikely(ARM64_WORKAROUND_2645198)) { 2090 /* 2091 * Break-before-make (BBM) is required for all user space mappings 2092 * when the permission changes from executable to non-executable 2093 * in cases where cpu is affected with errata #2645198. 2094 */ 2095 if (pte_accessible(vma->vm_mm, pte) && pte_user_exec(pte)) 2096 __flush_tlb_range(vma, addr, nr * PAGE_SIZE, 2097 PAGE_SIZE, true, 3); 2098 } 2099 2100 return pte; 2101 } 2102 2103 pte_t ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep) 2104 { 2105 return modify_prot_start_ptes(vma, addr, ptep, 1); 2106 } 2107 2108 void modify_prot_commit_ptes(struct vm_area_struct *vma, unsigned long addr, 2109 pte_t *ptep, pte_t old_pte, pte_t pte, 2110 unsigned int nr) 2111 { 2112 set_ptes(vma->vm_mm, addr, ptep, pte, nr); 2113 } 2114 2115 void ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, 2116 pte_t old_pte, pte_t pte) 2117 { 2118 modify_prot_commit_ptes(vma, addr, ptep, old_pte, pte, 1); 2119 } 2120 2121 /* 2122 * Atomically replaces the active TTBR1_EL1 PGD with a new VA-compatible PGD, 2123 * avoiding the possibility of conflicting TLB entries being allocated. 2124 */ 2125 void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp) 2126 { 2127 typedef void (ttbr_replace_func)(phys_addr_t); 2128 extern ttbr_replace_func idmap_cpu_replace_ttbr1; 2129 ttbr_replace_func *replace_phys; 2130 unsigned long daif; 2131 2132 /* phys_to_ttbr() zeros lower 2 bits of ttbr with 52-bit PA */ 2133 phys_addr_t ttbr1 = phys_to_ttbr(virt_to_phys(pgdp)); 2134 2135 if (cnp) 2136 ttbr1 |= TTBR_CNP_BIT; 2137 2138 replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1); 2139 2140 cpu_install_idmap(); 2141 2142 /* 2143 * We really don't want to take *any* exceptions while TTBR1 is 2144 * in the process of being replaced so mask everything. 2145 */ 2146 daif = local_daif_save(); 2147 replace_phys(ttbr1); 2148 local_daif_restore(daif); 2149 2150 cpu_uninstall_idmap(); 2151 } 2152 2153 #ifdef CONFIG_ARCH_HAS_PKEYS 2154 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey, unsigned long init_val) 2155 { 2156 u64 new_por; 2157 u64 old_por; 2158 2159 if (!system_supports_poe()) 2160 return -ENOSPC; 2161 2162 /* 2163 * This code should only be called with valid 'pkey' 2164 * values originating from in-kernel users. Complain 2165 * if a bad value is observed. 2166 */ 2167 if (WARN_ON_ONCE(pkey >= arch_max_pkey())) 2168 return -EINVAL; 2169 2170 /* Set the bits we need in POR: */ 2171 new_por = POE_RWX; 2172 if (init_val & PKEY_DISABLE_WRITE) 2173 new_por &= ~POE_W; 2174 if (init_val & PKEY_DISABLE_ACCESS) 2175 new_por &= ~POE_RW; 2176 if (init_val & PKEY_DISABLE_READ) 2177 new_por &= ~POE_R; 2178 if (init_val & PKEY_DISABLE_EXECUTE) 2179 new_por &= ~POE_X; 2180 2181 /* Shift the bits in to the correct place in POR for pkey: */ 2182 new_por = POR_ELx_PERM_PREP(pkey, new_por); 2183 2184 /* Get old POR and mask off any old bits in place: */ 2185 old_por = read_sysreg_s(SYS_POR_EL0); 2186 old_por &= ~(POE_MASK << POR_ELx_PERM_SHIFT(pkey)); 2187 2188 /* Write old part along with new part: */ 2189 write_sysreg_s(old_por | new_por, SYS_POR_EL0); 2190 2191 return 0; 2192 } 2193 #endif 2194