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 DEFINE_MUTEX(pgtable_split_lock); 712 713 int split_kernel_leaf_mapping(unsigned long start, unsigned long end) 714 { 715 int ret; 716 717 /* 718 * !BBML2_NOABORT systems should not be trying to change permissions on 719 * anything that is not pte-mapped in the first place. Just return early 720 * and let the permission change code raise a warning if not already 721 * pte-mapped. 722 */ 723 if (!system_supports_bbml2_noabort()) 724 return 0; 725 726 /* 727 * Ensure start and end are at least page-aligned since this is the 728 * finest granularity we can split to. 729 */ 730 if (start != PAGE_ALIGN(start) || end != PAGE_ALIGN(end)) 731 return -EINVAL; 732 733 mutex_lock(&pgtable_split_lock); 734 arch_enter_lazy_mmu_mode(); 735 736 /* 737 * The split_kernel_leaf_mapping_locked() may sleep, it is not a 738 * problem for ARM64 since ARM64's lazy MMU implementation allows 739 * sleeping. 740 * 741 * Optimize for the common case of splitting out a single page from a 742 * larger mapping. Here we can just split on the "least aligned" of 743 * start and end and this will guarantee that there must also be a split 744 * on the more aligned address since the both addresses must be in the 745 * same contpte block and it must have been split to ptes. 746 */ 747 if (end - start == PAGE_SIZE) { 748 start = __ffs(start) < __ffs(end) ? start : end; 749 ret = split_kernel_leaf_mapping_locked(start); 750 } else { 751 ret = split_kernel_leaf_mapping_locked(start); 752 if (!ret) 753 ret = split_kernel_leaf_mapping_locked(end); 754 } 755 756 arch_leave_lazy_mmu_mode(); 757 mutex_unlock(&pgtable_split_lock); 758 return ret; 759 } 760 761 static int __init split_to_ptes_pud_entry(pud_t *pudp, unsigned long addr, 762 unsigned long next, 763 struct mm_walk *walk) 764 { 765 pud_t pud = pudp_get(pudp); 766 int ret = 0; 767 768 if (pud_leaf(pud)) 769 ret = split_pud(pudp, pud, GFP_ATOMIC, false); 770 771 return ret; 772 } 773 774 static int __init split_to_ptes_pmd_entry(pmd_t *pmdp, unsigned long addr, 775 unsigned long next, 776 struct mm_walk *walk) 777 { 778 pmd_t pmd = pmdp_get(pmdp); 779 int ret = 0; 780 781 if (pmd_leaf(pmd)) { 782 if (pmd_cont(pmd)) 783 split_contpmd(pmdp); 784 ret = split_pmd(pmdp, pmd, GFP_ATOMIC, false); 785 786 /* 787 * We have split the pmd directly to ptes so there is no need to 788 * visit each pte to check if they are contpte. 789 */ 790 walk->action = ACTION_CONTINUE; 791 } 792 793 return ret; 794 } 795 796 static int __init split_to_ptes_pte_entry(pte_t *ptep, unsigned long addr, 797 unsigned long next, 798 struct mm_walk *walk) 799 { 800 pte_t pte = __ptep_get(ptep); 801 802 if (pte_cont(pte)) 803 split_contpte(ptep); 804 805 return 0; 806 } 807 808 static const struct mm_walk_ops split_to_ptes_ops __initconst = { 809 .pud_entry = split_to_ptes_pud_entry, 810 .pmd_entry = split_to_ptes_pmd_entry, 811 .pte_entry = split_to_ptes_pte_entry, 812 }; 813 814 static bool linear_map_requires_bbml2 __initdata; 815 816 u32 idmap_kpti_bbml2_flag; 817 818 static void __init init_idmap_kpti_bbml2_flag(void) 819 { 820 WRITE_ONCE(idmap_kpti_bbml2_flag, 1); 821 /* Must be visible to other CPUs before stop_machine() is called. */ 822 smp_mb(); 823 } 824 825 static int __init linear_map_split_to_ptes(void *__unused) 826 { 827 /* 828 * Repainting the linear map must be done by CPU0 (the boot CPU) because 829 * that's the only CPU that we know supports BBML2. The other CPUs will 830 * be held in a waiting area with the idmap active. 831 */ 832 if (!smp_processor_id()) { 833 unsigned long lstart = _PAGE_OFFSET(vabits_actual); 834 unsigned long lend = PAGE_END; 835 unsigned long kstart = (unsigned long)lm_alias(_stext); 836 unsigned long kend = (unsigned long)lm_alias(__init_begin); 837 int ret; 838 839 /* 840 * Wait for all secondary CPUs to be put into the waiting area. 841 */ 842 smp_cond_load_acquire(&idmap_kpti_bbml2_flag, VAL == num_online_cpus()); 843 844 /* 845 * Walk all of the linear map [lstart, lend), except the kernel 846 * linear map alias [kstart, kend), and split all mappings to 847 * PTE. The kernel alias remains static throughout runtime so 848 * can continue to be safely mapped with large mappings. 849 */ 850 ret = walk_kernel_page_table_range_lockless(lstart, kstart, 851 &split_to_ptes_ops, NULL, NULL); 852 if (!ret) 853 ret = walk_kernel_page_table_range_lockless(kend, lend, 854 &split_to_ptes_ops, NULL, NULL); 855 if (ret) 856 panic("Failed to split linear map\n"); 857 flush_tlb_kernel_range(lstart, lend); 858 859 /* 860 * Relies on dsb in flush_tlb_kernel_range() to avoid reordering 861 * before any page table split operations. 862 */ 863 WRITE_ONCE(idmap_kpti_bbml2_flag, 0); 864 } else { 865 typedef void (wait_split_fn)(void); 866 extern wait_split_fn wait_linear_map_split_to_ptes; 867 wait_split_fn *wait_fn; 868 869 wait_fn = (void *)__pa_symbol(wait_linear_map_split_to_ptes); 870 871 /* 872 * At least one secondary CPU doesn't support BBML2 so cannot 873 * tolerate the size of the live mappings changing. So have the 874 * secondary CPUs wait for the boot CPU to make the changes 875 * with the idmap active and init_mm inactive. 876 */ 877 cpu_install_idmap(); 878 wait_fn(); 879 cpu_uninstall_idmap(); 880 } 881 882 return 0; 883 } 884 885 void __init linear_map_maybe_split_to_ptes(void) 886 { 887 if (linear_map_requires_bbml2 && !system_supports_bbml2_noabort()) { 888 init_idmap_kpti_bbml2_flag(); 889 stop_machine(linear_map_split_to_ptes, NULL, cpu_online_mask); 890 } 891 } 892 893 /* 894 * This function can only be used to modify existing table entries, 895 * without allocating new levels of table. Note that this permits the 896 * creation of new section or page entries. 897 */ 898 void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt, 899 phys_addr_t size, pgprot_t prot) 900 { 901 if (virt < PAGE_OFFSET) { 902 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n", 903 &phys, virt); 904 return; 905 } 906 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 907 NO_CONT_MAPPINGS); 908 } 909 910 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys, 911 unsigned long virt, phys_addr_t size, 912 pgprot_t prot, bool page_mappings_only) 913 { 914 int flags = 0; 915 916 BUG_ON(mm == &init_mm); 917 918 if (page_mappings_only) 919 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 920 921 __create_pgd_mapping(mm->pgd, phys, virt, size, prot, 922 pgd_pgtable_alloc_special_mm, flags); 923 } 924 925 static void update_mapping_prot(phys_addr_t phys, unsigned long virt, 926 phys_addr_t size, pgprot_t prot) 927 { 928 if (virt < PAGE_OFFSET) { 929 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n", 930 &phys, virt); 931 return; 932 } 933 934 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 935 NO_CONT_MAPPINGS); 936 937 /* flush the TLBs after updating live kernel mappings */ 938 flush_tlb_kernel_range(virt, virt + size); 939 } 940 941 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start, 942 phys_addr_t end, pgprot_t prot, int flags) 943 { 944 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start, 945 prot, early_pgtable_alloc, flags); 946 } 947 948 void __init mark_linear_text_alias_ro(void) 949 { 950 /* 951 * Remove the write permissions from the linear alias of .text/.rodata 952 */ 953 update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text), 954 (unsigned long)__init_begin - (unsigned long)_text, 955 PAGE_KERNEL_RO); 956 } 957 958 #ifdef CONFIG_KFENCE 959 960 bool __ro_after_init kfence_early_init = !!CONFIG_KFENCE_SAMPLE_INTERVAL; 961 962 /* early_param() will be parsed before map_mem() below. */ 963 static int __init parse_kfence_early_init(char *arg) 964 { 965 int val; 966 967 if (get_option(&arg, &val)) 968 kfence_early_init = !!val; 969 return 0; 970 } 971 early_param("kfence.sample_interval", parse_kfence_early_init); 972 973 static phys_addr_t __init arm64_kfence_alloc_pool(void) 974 { 975 phys_addr_t kfence_pool; 976 977 if (!kfence_early_init) 978 return 0; 979 980 kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE); 981 if (!kfence_pool) { 982 pr_err("failed to allocate kfence pool\n"); 983 kfence_early_init = false; 984 return 0; 985 } 986 987 /* Temporarily mark as NOMAP. */ 988 memblock_mark_nomap(kfence_pool, KFENCE_POOL_SIZE); 989 990 return kfence_pool; 991 } 992 993 static void __init arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) 994 { 995 if (!kfence_pool) 996 return; 997 998 /* KFENCE pool needs page-level mapping. */ 999 __map_memblock(pgdp, kfence_pool, kfence_pool + KFENCE_POOL_SIZE, 1000 pgprot_tagged(PAGE_KERNEL), 1001 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS); 1002 memblock_clear_nomap(kfence_pool, KFENCE_POOL_SIZE); 1003 __kfence_pool = phys_to_virt(kfence_pool); 1004 } 1005 #else /* CONFIG_KFENCE */ 1006 1007 static inline phys_addr_t arm64_kfence_alloc_pool(void) { return 0; } 1008 static inline void arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) { } 1009 1010 #endif /* CONFIG_KFENCE */ 1011 1012 static inline bool force_pte_mapping(void) 1013 { 1014 bool bbml2 = system_capabilities_finalized() ? 1015 system_supports_bbml2_noabort() : cpu_supports_bbml2_noabort(); 1016 1017 return (!bbml2 && (rodata_full || arm64_kfence_can_set_direct_map() || 1018 is_realm_world())) || 1019 debug_pagealloc_enabled(); 1020 } 1021 1022 static void __init map_mem(pgd_t *pgdp) 1023 { 1024 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN); 1025 phys_addr_t kernel_start = __pa_symbol(_text); 1026 phys_addr_t kernel_end = __pa_symbol(__init_begin); 1027 phys_addr_t start, end; 1028 phys_addr_t early_kfence_pool; 1029 int flags = NO_EXEC_MAPPINGS; 1030 u64 i; 1031 1032 /* 1033 * Setting hierarchical PXNTable attributes on table entries covering 1034 * the linear region is only possible if it is guaranteed that no table 1035 * entries at any level are being shared between the linear region and 1036 * the vmalloc region. Check whether this is true for the PGD level, in 1037 * which case it is guaranteed to be true for all other levels as well. 1038 * (Unless we are running with support for LPA2, in which case the 1039 * entire reduced VA space is covered by a single pgd_t which will have 1040 * been populated without the PXNTable attribute by the time we get here.) 1041 */ 1042 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end) && 1043 pgd_index(_PAGE_OFFSET(VA_BITS_MIN)) != PTRS_PER_PGD - 1); 1044 1045 early_kfence_pool = arm64_kfence_alloc_pool(); 1046 1047 linear_map_requires_bbml2 = !force_pte_mapping() && can_set_direct_map(); 1048 1049 if (force_pte_mapping()) 1050 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1051 1052 /* 1053 * Take care not to create a writable alias for the 1054 * read-only text and rodata sections of the kernel image. 1055 * So temporarily mark them as NOMAP to skip mappings in 1056 * the following for-loop 1057 */ 1058 memblock_mark_nomap(kernel_start, kernel_end - kernel_start); 1059 1060 /* map all the memory banks */ 1061 for_each_mem_range(i, &start, &end) { 1062 if (start >= end) 1063 break; 1064 /* 1065 * The linear map must allow allocation tags reading/writing 1066 * if MTE is present. Otherwise, it has the same attributes as 1067 * PAGE_KERNEL. 1068 */ 1069 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL), 1070 flags); 1071 } 1072 1073 /* 1074 * Map the linear alias of the [_text, __init_begin) interval 1075 * as non-executable now, and remove the write permission in 1076 * mark_linear_text_alias_ro() below (which will be called after 1077 * alternative patching has completed). This makes the contents 1078 * of the region accessible to subsystems such as hibernate, 1079 * but protects it from inadvertent modification or execution. 1080 * Note that contiguous mappings cannot be remapped in this way, 1081 * so we should avoid them here. 1082 */ 1083 __map_memblock(pgdp, kernel_start, kernel_end, 1084 PAGE_KERNEL, NO_CONT_MAPPINGS); 1085 memblock_clear_nomap(kernel_start, kernel_end - kernel_start); 1086 arm64_kfence_map_pool(early_kfence_pool, pgdp); 1087 } 1088 1089 void mark_rodata_ro(void) 1090 { 1091 unsigned long section_size; 1092 1093 /* 1094 * mark .rodata as read only. Use __init_begin rather than __end_rodata 1095 * to cover NOTES and EXCEPTION_TABLE. 1096 */ 1097 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata; 1098 WRITE_ONCE(rodata_is_rw, false); 1099 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata, 1100 section_size, PAGE_KERNEL_RO); 1101 /* mark the range between _text and _stext as read only. */ 1102 update_mapping_prot(__pa_symbol(_text), (unsigned long)_text, 1103 (unsigned long)_stext - (unsigned long)_text, 1104 PAGE_KERNEL_RO); 1105 } 1106 1107 static void __init declare_vma(struct vm_struct *vma, 1108 void *va_start, void *va_end, 1109 unsigned long vm_flags) 1110 { 1111 phys_addr_t pa_start = __pa_symbol(va_start); 1112 unsigned long size = va_end - va_start; 1113 1114 BUG_ON(!PAGE_ALIGNED(pa_start)); 1115 BUG_ON(!PAGE_ALIGNED(size)); 1116 1117 if (!(vm_flags & VM_NO_GUARD)) 1118 size += PAGE_SIZE; 1119 1120 vma->addr = va_start; 1121 vma->phys_addr = pa_start; 1122 vma->size = size; 1123 vma->flags = VM_MAP | vm_flags; 1124 vma->caller = __builtin_return_address(0); 1125 1126 vm_area_add_early(vma); 1127 } 1128 1129 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0 1130 #define KPTI_NG_TEMP_VA (-(1UL << PMD_SHIFT)) 1131 1132 static phys_addr_t kpti_ng_temp_alloc __initdata; 1133 1134 static phys_addr_t __init kpti_ng_pgd_alloc(enum pgtable_type type) 1135 { 1136 kpti_ng_temp_alloc -= PAGE_SIZE; 1137 return kpti_ng_temp_alloc; 1138 } 1139 1140 static int __init __kpti_install_ng_mappings(void *__unused) 1141 { 1142 typedef void (kpti_remap_fn)(int, int, phys_addr_t, unsigned long); 1143 extern kpti_remap_fn idmap_kpti_install_ng_mappings; 1144 kpti_remap_fn *remap_fn; 1145 1146 int cpu = smp_processor_id(); 1147 int levels = CONFIG_PGTABLE_LEVELS; 1148 int order = order_base_2(levels); 1149 u64 kpti_ng_temp_pgd_pa = 0; 1150 pgd_t *kpti_ng_temp_pgd; 1151 u64 alloc = 0; 1152 1153 if (levels == 5 && !pgtable_l5_enabled()) 1154 levels = 4; 1155 else if (levels == 4 && !pgtable_l4_enabled()) 1156 levels = 3; 1157 1158 remap_fn = (void *)__pa_symbol(idmap_kpti_install_ng_mappings); 1159 1160 if (!cpu) { 1161 alloc = __get_free_pages(GFP_ATOMIC | __GFP_ZERO, order); 1162 kpti_ng_temp_pgd = (pgd_t *)(alloc + (levels - 1) * PAGE_SIZE); 1163 kpti_ng_temp_alloc = kpti_ng_temp_pgd_pa = __pa(kpti_ng_temp_pgd); 1164 1165 // 1166 // Create a minimal page table hierarchy that permits us to map 1167 // the swapper page tables temporarily as we traverse them. 1168 // 1169 // The physical pages are laid out as follows: 1170 // 1171 // +--------+-/-------+-/------ +-/------ +-\\\--------+ 1172 // : PTE[] : | PMD[] : | PUD[] : | P4D[] : ||| PGD[] : 1173 // +--------+-\-------+-\------ +-\------ +-///--------+ 1174 // ^ 1175 // The first page is mapped into this hierarchy at a PMD_SHIFT 1176 // aligned virtual address, so that we can manipulate the PTE 1177 // level entries while the mapping is active. The first entry 1178 // covers the PTE[] page itself, the remaining entries are free 1179 // to be used as a ad-hoc fixmap. 1180 // 1181 __create_pgd_mapping_locked(kpti_ng_temp_pgd, __pa(alloc), 1182 KPTI_NG_TEMP_VA, PAGE_SIZE, PAGE_KERNEL, 1183 kpti_ng_pgd_alloc, 0); 1184 } 1185 1186 cpu_install_idmap(); 1187 remap_fn(cpu, num_online_cpus(), kpti_ng_temp_pgd_pa, KPTI_NG_TEMP_VA); 1188 cpu_uninstall_idmap(); 1189 1190 if (!cpu) { 1191 free_pages(alloc, order); 1192 arm64_use_ng_mappings = true; 1193 } 1194 1195 return 0; 1196 } 1197 1198 void __init kpti_install_ng_mappings(void) 1199 { 1200 /* Check whether KPTI is going to be used */ 1201 if (!arm64_kernel_unmapped_at_el0()) 1202 return; 1203 1204 /* 1205 * We don't need to rewrite the page-tables if either we've done 1206 * it already or we have KASLR enabled and therefore have not 1207 * created any global mappings at all. 1208 */ 1209 if (arm64_use_ng_mappings) 1210 return; 1211 1212 init_idmap_kpti_bbml2_flag(); 1213 stop_machine(__kpti_install_ng_mappings, NULL, cpu_online_mask); 1214 } 1215 1216 static pgprot_t __init kernel_exec_prot(void) 1217 { 1218 return rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 1219 } 1220 1221 static int __init map_entry_trampoline(void) 1222 { 1223 int i; 1224 1225 if (!arm64_kernel_unmapped_at_el0()) 1226 return 0; 1227 1228 pgprot_t prot = kernel_exec_prot(); 1229 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start); 1230 1231 /* The trampoline is always mapped and can therefore be global */ 1232 pgprot_val(prot) &= ~PTE_NG; 1233 1234 /* Map only the text into the trampoline page table */ 1235 memset(tramp_pg_dir, 0, PGD_SIZE); 1236 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, 1237 entry_tramp_text_size(), prot, 1238 pgd_pgtable_alloc_init_mm, NO_BLOCK_MAPPINGS); 1239 1240 /* Map both the text and data into the kernel page table */ 1241 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++) 1242 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 1243 pa_start + i * PAGE_SIZE, prot); 1244 1245 if (IS_ENABLED(CONFIG_RELOCATABLE)) 1246 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 1247 pa_start + i * PAGE_SIZE, PAGE_KERNEL_RO); 1248 1249 return 0; 1250 } 1251 core_initcall(map_entry_trampoline); 1252 #endif 1253 1254 /* 1255 * Declare the VMA areas for the kernel 1256 */ 1257 static void __init declare_kernel_vmas(void) 1258 { 1259 static struct vm_struct vmlinux_seg[KERNEL_SEGMENT_COUNT]; 1260 1261 declare_vma(&vmlinux_seg[0], _text, _etext, VM_NO_GUARD); 1262 declare_vma(&vmlinux_seg[1], __start_rodata, __inittext_begin, VM_NO_GUARD); 1263 declare_vma(&vmlinux_seg[2], __inittext_begin, __inittext_end, VM_NO_GUARD); 1264 declare_vma(&vmlinux_seg[3], __initdata_begin, __initdata_end, VM_NO_GUARD); 1265 declare_vma(&vmlinux_seg[4], _data, _end, 0); 1266 } 1267 1268 void __pi_map_range(phys_addr_t *pte, u64 start, u64 end, phys_addr_t pa, 1269 pgprot_t prot, int level, pte_t *tbl, bool may_use_cont, 1270 u64 va_offset); 1271 1272 static u8 idmap_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init, 1273 kpti_bbml2_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init; 1274 1275 static void __init create_idmap(void) 1276 { 1277 phys_addr_t start = __pa_symbol(__idmap_text_start); 1278 phys_addr_t end = __pa_symbol(__idmap_text_end); 1279 phys_addr_t ptep = __pa_symbol(idmap_ptes); 1280 1281 __pi_map_range(&ptep, start, end, start, PAGE_KERNEL_ROX, 1282 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 1283 __phys_to_virt(ptep) - ptep); 1284 1285 if (linear_map_requires_bbml2 || 1286 (IS_ENABLED(CONFIG_UNMAP_KERNEL_AT_EL0) && !arm64_use_ng_mappings)) { 1287 phys_addr_t pa = __pa_symbol(&idmap_kpti_bbml2_flag); 1288 1289 /* 1290 * The KPTI G-to-nG conversion code needs a read-write mapping 1291 * of its synchronization flag in the ID map. This is also used 1292 * when splitting the linear map to ptes if a secondary CPU 1293 * doesn't support bbml2. 1294 */ 1295 ptep = __pa_symbol(kpti_bbml2_ptes); 1296 __pi_map_range(&ptep, pa, pa + sizeof(u32), pa, PAGE_KERNEL, 1297 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 1298 __phys_to_virt(ptep) - ptep); 1299 } 1300 } 1301 1302 void __init paging_init(void) 1303 { 1304 map_mem(swapper_pg_dir); 1305 1306 memblock_allow_resize(); 1307 1308 create_idmap(); 1309 declare_kernel_vmas(); 1310 } 1311 1312 #ifdef CONFIG_MEMORY_HOTPLUG 1313 static void free_hotplug_page_range(struct page *page, size_t size, 1314 struct vmem_altmap *altmap) 1315 { 1316 if (altmap) { 1317 vmem_altmap_free(altmap, size >> PAGE_SHIFT); 1318 } else { 1319 WARN_ON(PageReserved(page)); 1320 __free_pages(page, get_order(size)); 1321 } 1322 } 1323 1324 static void free_hotplug_pgtable_page(struct page *page) 1325 { 1326 free_hotplug_page_range(page, PAGE_SIZE, NULL); 1327 } 1328 1329 static bool pgtable_range_aligned(unsigned long start, unsigned long end, 1330 unsigned long floor, unsigned long ceiling, 1331 unsigned long mask) 1332 { 1333 start &= mask; 1334 if (start < floor) 1335 return false; 1336 1337 if (ceiling) { 1338 ceiling &= mask; 1339 if (!ceiling) 1340 return false; 1341 } 1342 1343 if (end - 1 > ceiling - 1) 1344 return false; 1345 return true; 1346 } 1347 1348 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr, 1349 unsigned long end, bool free_mapped, 1350 struct vmem_altmap *altmap) 1351 { 1352 pte_t *ptep, pte; 1353 1354 do { 1355 ptep = pte_offset_kernel(pmdp, addr); 1356 pte = __ptep_get(ptep); 1357 if (pte_none(pte)) 1358 continue; 1359 1360 WARN_ON(!pte_present(pte)); 1361 __pte_clear(&init_mm, addr, ptep); 1362 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1363 if (free_mapped) 1364 free_hotplug_page_range(pte_page(pte), 1365 PAGE_SIZE, altmap); 1366 } while (addr += PAGE_SIZE, addr < end); 1367 } 1368 1369 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr, 1370 unsigned long end, bool free_mapped, 1371 struct vmem_altmap *altmap) 1372 { 1373 unsigned long next; 1374 pmd_t *pmdp, pmd; 1375 1376 do { 1377 next = pmd_addr_end(addr, end); 1378 pmdp = pmd_offset(pudp, addr); 1379 pmd = READ_ONCE(*pmdp); 1380 if (pmd_none(pmd)) 1381 continue; 1382 1383 WARN_ON(!pmd_present(pmd)); 1384 if (pmd_sect(pmd)) { 1385 pmd_clear(pmdp); 1386 1387 /* 1388 * One TLBI should be sufficient here as the PMD_SIZE 1389 * range is mapped with a single block entry. 1390 */ 1391 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1392 if (free_mapped) 1393 free_hotplug_page_range(pmd_page(pmd), 1394 PMD_SIZE, altmap); 1395 continue; 1396 } 1397 WARN_ON(!pmd_table(pmd)); 1398 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap); 1399 } while (addr = next, addr < end); 1400 } 1401 1402 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr, 1403 unsigned long end, bool free_mapped, 1404 struct vmem_altmap *altmap) 1405 { 1406 unsigned long next; 1407 pud_t *pudp, pud; 1408 1409 do { 1410 next = pud_addr_end(addr, end); 1411 pudp = pud_offset(p4dp, addr); 1412 pud = READ_ONCE(*pudp); 1413 if (pud_none(pud)) 1414 continue; 1415 1416 WARN_ON(!pud_present(pud)); 1417 if (pud_sect(pud)) { 1418 pud_clear(pudp); 1419 1420 /* 1421 * One TLBI should be sufficient here as the PUD_SIZE 1422 * range is mapped with a single block entry. 1423 */ 1424 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1425 if (free_mapped) 1426 free_hotplug_page_range(pud_page(pud), 1427 PUD_SIZE, altmap); 1428 continue; 1429 } 1430 WARN_ON(!pud_table(pud)); 1431 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap); 1432 } while (addr = next, addr < end); 1433 } 1434 1435 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr, 1436 unsigned long end, bool free_mapped, 1437 struct vmem_altmap *altmap) 1438 { 1439 unsigned long next; 1440 p4d_t *p4dp, p4d; 1441 1442 do { 1443 next = p4d_addr_end(addr, end); 1444 p4dp = p4d_offset(pgdp, addr); 1445 p4d = READ_ONCE(*p4dp); 1446 if (p4d_none(p4d)) 1447 continue; 1448 1449 WARN_ON(!p4d_present(p4d)); 1450 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap); 1451 } while (addr = next, addr < end); 1452 } 1453 1454 static void unmap_hotplug_range(unsigned long addr, unsigned long end, 1455 bool free_mapped, struct vmem_altmap *altmap) 1456 { 1457 unsigned long next; 1458 pgd_t *pgdp, pgd; 1459 1460 /* 1461 * altmap can only be used as vmemmap mapping backing memory. 1462 * In case the backing memory itself is not being freed, then 1463 * altmap is irrelevant. Warn about this inconsistency when 1464 * encountered. 1465 */ 1466 WARN_ON(!free_mapped && altmap); 1467 1468 do { 1469 next = pgd_addr_end(addr, end); 1470 pgdp = pgd_offset_k(addr); 1471 pgd = READ_ONCE(*pgdp); 1472 if (pgd_none(pgd)) 1473 continue; 1474 1475 WARN_ON(!pgd_present(pgd)); 1476 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap); 1477 } while (addr = next, addr < end); 1478 } 1479 1480 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr, 1481 unsigned long end, unsigned long floor, 1482 unsigned long ceiling) 1483 { 1484 pte_t *ptep, pte; 1485 unsigned long i, start = addr; 1486 1487 do { 1488 ptep = pte_offset_kernel(pmdp, addr); 1489 pte = __ptep_get(ptep); 1490 1491 /* 1492 * This is just a sanity check here which verifies that 1493 * pte clearing has been done by earlier unmap loops. 1494 */ 1495 WARN_ON(!pte_none(pte)); 1496 } while (addr += PAGE_SIZE, addr < end); 1497 1498 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK)) 1499 return; 1500 1501 /* 1502 * Check whether we can free the pte page if the rest of the 1503 * entries are empty. Overlap with other regions have been 1504 * handled by the floor/ceiling check. 1505 */ 1506 ptep = pte_offset_kernel(pmdp, 0UL); 1507 for (i = 0; i < PTRS_PER_PTE; i++) { 1508 if (!pte_none(__ptep_get(&ptep[i]))) 1509 return; 1510 } 1511 1512 pmd_clear(pmdp); 1513 __flush_tlb_kernel_pgtable(start); 1514 free_hotplug_pgtable_page(virt_to_page(ptep)); 1515 } 1516 1517 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr, 1518 unsigned long end, unsigned long floor, 1519 unsigned long ceiling) 1520 { 1521 pmd_t *pmdp, pmd; 1522 unsigned long i, next, start = addr; 1523 1524 do { 1525 next = pmd_addr_end(addr, end); 1526 pmdp = pmd_offset(pudp, addr); 1527 pmd = READ_ONCE(*pmdp); 1528 if (pmd_none(pmd)) 1529 continue; 1530 1531 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd)); 1532 free_empty_pte_table(pmdp, addr, next, floor, ceiling); 1533 } while (addr = next, addr < end); 1534 1535 if (CONFIG_PGTABLE_LEVELS <= 2) 1536 return; 1537 1538 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK)) 1539 return; 1540 1541 /* 1542 * Check whether we can free the pmd page if the rest of the 1543 * entries are empty. Overlap with other regions have been 1544 * handled by the floor/ceiling check. 1545 */ 1546 pmdp = pmd_offset(pudp, 0UL); 1547 for (i = 0; i < PTRS_PER_PMD; i++) { 1548 if (!pmd_none(READ_ONCE(pmdp[i]))) 1549 return; 1550 } 1551 1552 pud_clear(pudp); 1553 __flush_tlb_kernel_pgtable(start); 1554 free_hotplug_pgtable_page(virt_to_page(pmdp)); 1555 } 1556 1557 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr, 1558 unsigned long end, unsigned long floor, 1559 unsigned long ceiling) 1560 { 1561 pud_t *pudp, pud; 1562 unsigned long i, next, start = addr; 1563 1564 do { 1565 next = pud_addr_end(addr, end); 1566 pudp = pud_offset(p4dp, addr); 1567 pud = READ_ONCE(*pudp); 1568 if (pud_none(pud)) 1569 continue; 1570 1571 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud)); 1572 free_empty_pmd_table(pudp, addr, next, floor, ceiling); 1573 } while (addr = next, addr < end); 1574 1575 if (!pgtable_l4_enabled()) 1576 return; 1577 1578 if (!pgtable_range_aligned(start, end, floor, ceiling, P4D_MASK)) 1579 return; 1580 1581 /* 1582 * Check whether we can free the pud page if the rest of the 1583 * entries are empty. Overlap with other regions have been 1584 * handled by the floor/ceiling check. 1585 */ 1586 pudp = pud_offset(p4dp, 0UL); 1587 for (i = 0; i < PTRS_PER_PUD; i++) { 1588 if (!pud_none(READ_ONCE(pudp[i]))) 1589 return; 1590 } 1591 1592 p4d_clear(p4dp); 1593 __flush_tlb_kernel_pgtable(start); 1594 free_hotplug_pgtable_page(virt_to_page(pudp)); 1595 } 1596 1597 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr, 1598 unsigned long end, unsigned long floor, 1599 unsigned long ceiling) 1600 { 1601 p4d_t *p4dp, p4d; 1602 unsigned long i, next, start = addr; 1603 1604 do { 1605 next = p4d_addr_end(addr, end); 1606 p4dp = p4d_offset(pgdp, addr); 1607 p4d = READ_ONCE(*p4dp); 1608 if (p4d_none(p4d)) 1609 continue; 1610 1611 WARN_ON(!p4d_present(p4d)); 1612 free_empty_pud_table(p4dp, addr, next, floor, ceiling); 1613 } while (addr = next, addr < end); 1614 1615 if (!pgtable_l5_enabled()) 1616 return; 1617 1618 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK)) 1619 return; 1620 1621 /* 1622 * Check whether we can free the p4d page if the rest of the 1623 * entries are empty. Overlap with other regions have been 1624 * handled by the floor/ceiling check. 1625 */ 1626 p4dp = p4d_offset(pgdp, 0UL); 1627 for (i = 0; i < PTRS_PER_P4D; i++) { 1628 if (!p4d_none(READ_ONCE(p4dp[i]))) 1629 return; 1630 } 1631 1632 pgd_clear(pgdp); 1633 __flush_tlb_kernel_pgtable(start); 1634 free_hotplug_pgtable_page(virt_to_page(p4dp)); 1635 } 1636 1637 static void free_empty_tables(unsigned long addr, unsigned long end, 1638 unsigned long floor, unsigned long ceiling) 1639 { 1640 unsigned long next; 1641 pgd_t *pgdp, pgd; 1642 1643 do { 1644 next = pgd_addr_end(addr, end); 1645 pgdp = pgd_offset_k(addr); 1646 pgd = READ_ONCE(*pgdp); 1647 if (pgd_none(pgd)) 1648 continue; 1649 1650 WARN_ON(!pgd_present(pgd)); 1651 free_empty_p4d_table(pgdp, addr, next, floor, ceiling); 1652 } while (addr = next, addr < end); 1653 } 1654 #endif 1655 1656 void __meminit vmemmap_set_pmd(pmd_t *pmdp, void *p, int node, 1657 unsigned long addr, unsigned long next) 1658 { 1659 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL)); 1660 } 1661 1662 int __meminit vmemmap_check_pmd(pmd_t *pmdp, int node, 1663 unsigned long addr, unsigned long next) 1664 { 1665 vmemmap_verify((pte_t *)pmdp, node, addr, next); 1666 1667 return pmd_sect(READ_ONCE(*pmdp)); 1668 } 1669 1670 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1671 struct vmem_altmap *altmap) 1672 { 1673 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1674 /* [start, end] should be within one section */ 1675 WARN_ON_ONCE(end - start > PAGES_PER_SECTION * sizeof(struct page)); 1676 1677 if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || 1678 (end - start < PAGES_PER_SECTION * sizeof(struct page))) 1679 return vmemmap_populate_basepages(start, end, node, altmap); 1680 else 1681 return vmemmap_populate_hugepages(start, end, node, altmap); 1682 } 1683 1684 #ifdef CONFIG_MEMORY_HOTPLUG 1685 void vmemmap_free(unsigned long start, unsigned long end, 1686 struct vmem_altmap *altmap) 1687 { 1688 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1689 1690 unmap_hotplug_range(start, end, true, altmap); 1691 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END); 1692 } 1693 #endif /* CONFIG_MEMORY_HOTPLUG */ 1694 1695 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot) 1696 { 1697 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot)); 1698 1699 /* Only allow permission changes for now */ 1700 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)), 1701 pud_val(new_pud))) 1702 return 0; 1703 1704 VM_BUG_ON(phys & ~PUD_MASK); 1705 set_pud(pudp, new_pud); 1706 return 1; 1707 } 1708 1709 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot) 1710 { 1711 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot)); 1712 1713 /* Only allow permission changes for now */ 1714 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)), 1715 pmd_val(new_pmd))) 1716 return 0; 1717 1718 VM_BUG_ON(phys & ~PMD_MASK); 1719 set_pmd(pmdp, new_pmd); 1720 return 1; 1721 } 1722 1723 #ifndef __PAGETABLE_P4D_FOLDED 1724 void p4d_clear_huge(p4d_t *p4dp) 1725 { 1726 } 1727 #endif 1728 1729 int pud_clear_huge(pud_t *pudp) 1730 { 1731 if (!pud_sect(READ_ONCE(*pudp))) 1732 return 0; 1733 pud_clear(pudp); 1734 return 1; 1735 } 1736 1737 int pmd_clear_huge(pmd_t *pmdp) 1738 { 1739 if (!pmd_sect(READ_ONCE(*pmdp))) 1740 return 0; 1741 pmd_clear(pmdp); 1742 return 1; 1743 } 1744 1745 static int __pmd_free_pte_page(pmd_t *pmdp, unsigned long addr, 1746 bool acquire_mmap_lock) 1747 { 1748 pte_t *table; 1749 pmd_t pmd; 1750 1751 pmd = READ_ONCE(*pmdp); 1752 1753 if (!pmd_table(pmd)) { 1754 VM_WARN_ON(1); 1755 return 1; 1756 } 1757 1758 /* See comment in pud_free_pmd_page for static key logic */ 1759 table = pte_offset_kernel(pmdp, addr); 1760 pmd_clear(pmdp); 1761 __flush_tlb_kernel_pgtable(addr); 1762 if (static_branch_unlikely(&arm64_ptdump_lock_key) && acquire_mmap_lock) { 1763 mmap_read_lock(&init_mm); 1764 mmap_read_unlock(&init_mm); 1765 } 1766 1767 pte_free_kernel(NULL, table); 1768 return 1; 1769 } 1770 1771 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr) 1772 { 1773 /* If ptdump is walking the pagetables, acquire init_mm.mmap_lock */ 1774 return __pmd_free_pte_page(pmdp, addr, /* acquire_mmap_lock = */ true); 1775 } 1776 1777 int pud_free_pmd_page(pud_t *pudp, unsigned long addr) 1778 { 1779 pmd_t *table; 1780 pmd_t *pmdp; 1781 pud_t pud; 1782 unsigned long next, end; 1783 1784 pud = READ_ONCE(*pudp); 1785 1786 if (!pud_table(pud)) { 1787 VM_WARN_ON(1); 1788 return 1; 1789 } 1790 1791 table = pmd_offset(pudp, addr); 1792 1793 /* 1794 * Our objective is to prevent ptdump from reading a PMD table which has 1795 * been freed. In this race, if pud_free_pmd_page observes the key on 1796 * (which got flipped by ptdump) then the mmap lock sequence here will, 1797 * as a result of the mmap write lock/unlock sequence in ptdump, give 1798 * us the correct synchronization. If not, this means that ptdump has 1799 * yet not started walking the pagetables - the sequence of barriers 1800 * issued by __flush_tlb_kernel_pgtable() guarantees that ptdump will 1801 * observe an empty PUD. 1802 */ 1803 pud_clear(pudp); 1804 __flush_tlb_kernel_pgtable(addr); 1805 if (static_branch_unlikely(&arm64_ptdump_lock_key)) { 1806 mmap_read_lock(&init_mm); 1807 mmap_read_unlock(&init_mm); 1808 } 1809 1810 pmdp = table; 1811 next = addr; 1812 end = addr + PUD_SIZE; 1813 do { 1814 if (pmd_present(pmdp_get(pmdp))) 1815 /* 1816 * PMD has been isolated, so ptdump won't see it. No 1817 * need to acquire init_mm.mmap_lock. 1818 */ 1819 __pmd_free_pte_page(pmdp, next, /* acquire_mmap_lock = */ false); 1820 } while (pmdp++, next += PMD_SIZE, next != end); 1821 1822 pmd_free(NULL, table); 1823 return 1; 1824 } 1825 1826 #ifdef CONFIG_MEMORY_HOTPLUG 1827 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size) 1828 { 1829 unsigned long end = start + size; 1830 1831 WARN_ON(pgdir != init_mm.pgd); 1832 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END)); 1833 1834 unmap_hotplug_range(start, end, false, NULL); 1835 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END); 1836 } 1837 1838 struct range arch_get_mappable_range(void) 1839 { 1840 struct range mhp_range; 1841 phys_addr_t start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual)); 1842 phys_addr_t end_linear_pa = __pa(PAGE_END - 1); 1843 1844 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 1845 /* 1846 * Check for a wrap, it is possible because of randomized linear 1847 * mapping the start physical address is actually bigger than 1848 * the end physical address. In this case set start to zero 1849 * because [0, end_linear_pa] range must still be able to cover 1850 * all addressable physical addresses. 1851 */ 1852 if (start_linear_pa > end_linear_pa) 1853 start_linear_pa = 0; 1854 } 1855 1856 WARN_ON(start_linear_pa > end_linear_pa); 1857 1858 /* 1859 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)] 1860 * accommodating both its ends but excluding PAGE_END. Max physical 1861 * range which can be mapped inside this linear mapping range, must 1862 * also be derived from its end points. 1863 */ 1864 mhp_range.start = start_linear_pa; 1865 mhp_range.end = end_linear_pa; 1866 1867 return mhp_range; 1868 } 1869 1870 int arch_add_memory(int nid, u64 start, u64 size, 1871 struct mhp_params *params) 1872 { 1873 int ret, flags = NO_EXEC_MAPPINGS; 1874 1875 VM_BUG_ON(!mhp_range_allowed(start, size, true)); 1876 1877 if (force_pte_mapping()) 1878 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1879 1880 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1881 size, params->pgprot, pgd_pgtable_alloc_init_mm, 1882 flags); 1883 1884 memblock_clear_nomap(start, size); 1885 1886 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, 1887 params); 1888 if (ret) 1889 __remove_pgd_mapping(swapper_pg_dir, 1890 __phys_to_virt(start), size); 1891 else { 1892 /* Address of hotplugged memory can be smaller */ 1893 max_pfn = max(max_pfn, PFN_UP(start + size)); 1894 max_low_pfn = max_pfn; 1895 } 1896 1897 return ret; 1898 } 1899 1900 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap) 1901 { 1902 unsigned long start_pfn = start >> PAGE_SHIFT; 1903 unsigned long nr_pages = size >> PAGE_SHIFT; 1904 1905 __remove_pages(start_pfn, nr_pages, altmap); 1906 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size); 1907 } 1908 1909 /* 1910 * This memory hotplug notifier helps prevent boot memory from being 1911 * inadvertently removed as it blocks pfn range offlining process in 1912 * __offline_pages(). Hence this prevents both offlining as well as 1913 * removal process for boot memory which is initially always online. 1914 * In future if and when boot memory could be removed, this notifier 1915 * should be dropped and free_hotplug_page_range() should handle any 1916 * reserved pages allocated during boot. 1917 */ 1918 static int prevent_bootmem_remove_notifier(struct notifier_block *nb, 1919 unsigned long action, void *data) 1920 { 1921 struct mem_section *ms; 1922 struct memory_notify *arg = data; 1923 unsigned long end_pfn = arg->start_pfn + arg->nr_pages; 1924 unsigned long pfn = arg->start_pfn; 1925 1926 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE)) 1927 return NOTIFY_OK; 1928 1929 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 1930 unsigned long start = PFN_PHYS(pfn); 1931 unsigned long end = start + (1UL << PA_SECTION_SHIFT); 1932 1933 ms = __pfn_to_section(pfn); 1934 if (!early_section(ms)) 1935 continue; 1936 1937 if (action == MEM_GOING_OFFLINE) { 1938 /* 1939 * Boot memory removal is not supported. Prevent 1940 * it via blocking any attempted offline request 1941 * for the boot memory and just report it. 1942 */ 1943 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end); 1944 return NOTIFY_BAD; 1945 } else if (action == MEM_OFFLINE) { 1946 /* 1947 * This should have never happened. Boot memory 1948 * offlining should have been prevented by this 1949 * very notifier. Probably some memory removal 1950 * procedure might have changed which would then 1951 * require further debug. 1952 */ 1953 pr_err("Boot memory [%lx %lx] offlined\n", start, end); 1954 1955 /* 1956 * Core memory hotplug does not process a return 1957 * code from the notifier for MEM_OFFLINE events. 1958 * The error condition has been reported. Return 1959 * from here as if ignored. 1960 */ 1961 return NOTIFY_DONE; 1962 } 1963 } 1964 return NOTIFY_OK; 1965 } 1966 1967 static struct notifier_block prevent_bootmem_remove_nb = { 1968 .notifier_call = prevent_bootmem_remove_notifier, 1969 }; 1970 1971 /* 1972 * This ensures that boot memory sections on the platform are online 1973 * from early boot. Memory sections could not be prevented from being 1974 * offlined, unless for some reason they are not online to begin with. 1975 * This helps validate the basic assumption on which the above memory 1976 * event notifier works to prevent boot memory section offlining and 1977 * its possible removal. 1978 */ 1979 static void validate_bootmem_online(void) 1980 { 1981 phys_addr_t start, end, addr; 1982 struct mem_section *ms; 1983 u64 i; 1984 1985 /* 1986 * Scanning across all memblock might be expensive 1987 * on some big memory systems. Hence enable this 1988 * validation only with DEBUG_VM. 1989 */ 1990 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 1991 return; 1992 1993 for_each_mem_range(i, &start, &end) { 1994 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) { 1995 ms = __pfn_to_section(PHYS_PFN(addr)); 1996 1997 /* 1998 * All memory ranges in the system at this point 1999 * should have been marked as early sections. 2000 */ 2001 WARN_ON(!early_section(ms)); 2002 2003 /* 2004 * Memory notifier mechanism here to prevent boot 2005 * memory offlining depends on the fact that each 2006 * early section memory on the system is initially 2007 * online. Otherwise a given memory section which 2008 * is already offline will be overlooked and can 2009 * be removed completely. Call out such sections. 2010 */ 2011 if (!online_section(ms)) 2012 pr_err("Boot memory [%llx %llx] is offline, can be removed\n", 2013 addr, addr + (1UL << PA_SECTION_SHIFT)); 2014 } 2015 } 2016 } 2017 2018 static int __init prevent_bootmem_remove_init(void) 2019 { 2020 int ret = 0; 2021 2022 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) 2023 return ret; 2024 2025 validate_bootmem_online(); 2026 ret = register_memory_notifier(&prevent_bootmem_remove_nb); 2027 if (ret) 2028 pr_err("%s: Notifier registration failed %d\n", __func__, ret); 2029 2030 return ret; 2031 } 2032 early_initcall(prevent_bootmem_remove_init); 2033 #endif 2034 2035 pte_t modify_prot_start_ptes(struct vm_area_struct *vma, unsigned long addr, 2036 pte_t *ptep, unsigned int nr) 2037 { 2038 pte_t pte = get_and_clear_ptes(vma->vm_mm, addr, ptep, nr); 2039 2040 if (alternative_has_cap_unlikely(ARM64_WORKAROUND_2645198)) { 2041 /* 2042 * Break-before-make (BBM) is required for all user space mappings 2043 * when the permission changes from executable to non-executable 2044 * in cases where cpu is affected with errata #2645198. 2045 */ 2046 if (pte_accessible(vma->vm_mm, pte) && pte_user_exec(pte)) 2047 __flush_tlb_range(vma, addr, nr * PAGE_SIZE, 2048 PAGE_SIZE, true, 3); 2049 } 2050 2051 return pte; 2052 } 2053 2054 pte_t ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep) 2055 { 2056 return modify_prot_start_ptes(vma, addr, ptep, 1); 2057 } 2058 2059 void modify_prot_commit_ptes(struct vm_area_struct *vma, unsigned long addr, 2060 pte_t *ptep, pte_t old_pte, pte_t pte, 2061 unsigned int nr) 2062 { 2063 set_ptes(vma->vm_mm, addr, ptep, pte, nr); 2064 } 2065 2066 void ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, 2067 pte_t old_pte, pte_t pte) 2068 { 2069 modify_prot_commit_ptes(vma, addr, ptep, old_pte, pte, 1); 2070 } 2071 2072 /* 2073 * Atomically replaces the active TTBR1_EL1 PGD with a new VA-compatible PGD, 2074 * avoiding the possibility of conflicting TLB entries being allocated. 2075 */ 2076 void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp) 2077 { 2078 typedef void (ttbr_replace_func)(phys_addr_t); 2079 extern ttbr_replace_func idmap_cpu_replace_ttbr1; 2080 ttbr_replace_func *replace_phys; 2081 unsigned long daif; 2082 2083 /* phys_to_ttbr() zeros lower 2 bits of ttbr with 52-bit PA */ 2084 phys_addr_t ttbr1 = phys_to_ttbr(virt_to_phys(pgdp)); 2085 2086 if (cnp) 2087 ttbr1 |= TTBR_CNP_BIT; 2088 2089 replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1); 2090 2091 cpu_install_idmap(); 2092 2093 /* 2094 * We really don't want to take *any* exceptions while TTBR1 is 2095 * in the process of being replaced so mask everything. 2096 */ 2097 daif = local_daif_save(); 2098 replace_phys(ttbr1); 2099 local_daif_restore(daif); 2100 2101 cpu_uninstall_idmap(); 2102 } 2103 2104 #ifdef CONFIG_ARCH_HAS_PKEYS 2105 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey, unsigned long init_val) 2106 { 2107 u64 new_por; 2108 u64 old_por; 2109 2110 if (!system_supports_poe()) 2111 return -ENOSPC; 2112 2113 /* 2114 * This code should only be called with valid 'pkey' 2115 * values originating from in-kernel users. Complain 2116 * if a bad value is observed. 2117 */ 2118 if (WARN_ON_ONCE(pkey >= arch_max_pkey())) 2119 return -EINVAL; 2120 2121 /* Set the bits we need in POR: */ 2122 new_por = POE_RWX; 2123 if (init_val & PKEY_DISABLE_WRITE) 2124 new_por &= ~POE_W; 2125 if (init_val & PKEY_DISABLE_ACCESS) 2126 new_por &= ~POE_RW; 2127 if (init_val & PKEY_DISABLE_READ) 2128 new_por &= ~POE_R; 2129 if (init_val & PKEY_DISABLE_EXECUTE) 2130 new_por &= ~POE_X; 2131 2132 /* Shift the bits in to the correct place in POR for pkey: */ 2133 new_por = POR_ELx_PERM_PREP(pkey, new_por); 2134 2135 /* Get old POR and mask off any old bits in place: */ 2136 old_por = read_sysreg_s(SYS_POR_EL0); 2137 old_por &= ~(POE_MASK << POR_ELx_PERM_SHIFT(pkey)); 2138 2139 /* Write old part along with new part: */ 2140 write_sysreg_s(old_por | new_por, SYS_POR_EL0); 2141 2142 return 0; 2143 } 2144 #endif 2145