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 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0 474 extern __alias(__create_pgd_mapping_locked) 475 void create_kpti_ng_temp_pgd(pgd_t *pgdir, phys_addr_t phys, unsigned long virt, 476 phys_addr_t size, pgprot_t prot, 477 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 478 int flags); 479 #endif 480 481 #define INVALID_PHYS_ADDR (-1ULL) 482 483 static phys_addr_t __pgd_pgtable_alloc(struct mm_struct *mm, gfp_t gfp, 484 enum pgtable_type pgtable_type) 485 { 486 /* Page is zeroed by init_clear_pgtable() so don't duplicate effort. */ 487 struct ptdesc *ptdesc = pagetable_alloc(gfp & ~__GFP_ZERO, 0); 488 phys_addr_t pa; 489 490 if (!ptdesc) 491 return INVALID_PHYS_ADDR; 492 493 pa = page_to_phys(ptdesc_page(ptdesc)); 494 495 switch (pgtable_type) { 496 case TABLE_PTE: 497 BUG_ON(!pagetable_pte_ctor(mm, ptdesc)); 498 break; 499 case TABLE_PMD: 500 BUG_ON(!pagetable_pmd_ctor(mm, ptdesc)); 501 break; 502 case TABLE_PUD: 503 pagetable_pud_ctor(ptdesc); 504 break; 505 case TABLE_P4D: 506 pagetable_p4d_ctor(ptdesc); 507 break; 508 } 509 510 return pa; 511 } 512 513 static phys_addr_t 514 try_pgd_pgtable_alloc_init_mm(enum pgtable_type pgtable_type, gfp_t gfp) 515 { 516 return __pgd_pgtable_alloc(&init_mm, gfp, pgtable_type); 517 } 518 519 static phys_addr_t __maybe_unused 520 pgd_pgtable_alloc_init_mm(enum pgtable_type pgtable_type) 521 { 522 phys_addr_t pa; 523 524 pa = __pgd_pgtable_alloc(&init_mm, GFP_PGTABLE_KERNEL, pgtable_type); 525 BUG_ON(pa == INVALID_PHYS_ADDR); 526 return pa; 527 } 528 529 static phys_addr_t 530 pgd_pgtable_alloc_special_mm(enum pgtable_type pgtable_type) 531 { 532 phys_addr_t pa; 533 534 pa = __pgd_pgtable_alloc(NULL, GFP_PGTABLE_KERNEL, pgtable_type); 535 BUG_ON(pa == INVALID_PHYS_ADDR); 536 return pa; 537 } 538 539 static void split_contpte(pte_t *ptep) 540 { 541 int i; 542 543 ptep = PTR_ALIGN_DOWN(ptep, sizeof(*ptep) * CONT_PTES); 544 for (i = 0; i < CONT_PTES; i++, ptep++) 545 __set_pte(ptep, pte_mknoncont(__ptep_get(ptep))); 546 } 547 548 static int split_pmd(pmd_t *pmdp, pmd_t pmd, gfp_t gfp, bool to_cont) 549 { 550 pmdval_t tableprot = PMD_TYPE_TABLE | PMD_TABLE_UXN | PMD_TABLE_AF; 551 unsigned long pfn = pmd_pfn(pmd); 552 pgprot_t prot = pmd_pgprot(pmd); 553 phys_addr_t pte_phys; 554 pte_t *ptep; 555 int i; 556 557 pte_phys = try_pgd_pgtable_alloc_init_mm(TABLE_PTE, gfp); 558 if (pte_phys == INVALID_PHYS_ADDR) 559 return -ENOMEM; 560 ptep = (pte_t *)phys_to_virt(pte_phys); 561 562 if (pgprot_val(prot) & PMD_SECT_PXN) 563 tableprot |= PMD_TABLE_PXN; 564 565 prot = __pgprot((pgprot_val(prot) & ~PTE_TYPE_MASK) | PTE_TYPE_PAGE); 566 prot = __pgprot(pgprot_val(prot) & ~PTE_CONT); 567 if (to_cont) 568 prot = __pgprot(pgprot_val(prot) | PTE_CONT); 569 570 for (i = 0; i < PTRS_PER_PTE; i++, ptep++, pfn++) 571 __set_pte(ptep, pfn_pte(pfn, prot)); 572 573 /* 574 * Ensure the pte entries are visible to the table walker by the time 575 * the pmd entry that points to the ptes is visible. 576 */ 577 dsb(ishst); 578 __pmd_populate(pmdp, pte_phys, tableprot); 579 580 return 0; 581 } 582 583 static void split_contpmd(pmd_t *pmdp) 584 { 585 int i; 586 587 pmdp = PTR_ALIGN_DOWN(pmdp, sizeof(*pmdp) * CONT_PMDS); 588 for (i = 0; i < CONT_PMDS; i++, pmdp++) 589 set_pmd(pmdp, pmd_mknoncont(pmdp_get(pmdp))); 590 } 591 592 static int split_pud(pud_t *pudp, pud_t pud, gfp_t gfp, bool to_cont) 593 { 594 pudval_t tableprot = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF; 595 unsigned int step = PMD_SIZE >> PAGE_SHIFT; 596 unsigned long pfn = pud_pfn(pud); 597 pgprot_t prot = pud_pgprot(pud); 598 phys_addr_t pmd_phys; 599 pmd_t *pmdp; 600 int i; 601 602 pmd_phys = try_pgd_pgtable_alloc_init_mm(TABLE_PMD, gfp); 603 if (pmd_phys == INVALID_PHYS_ADDR) 604 return -ENOMEM; 605 pmdp = (pmd_t *)phys_to_virt(pmd_phys); 606 607 if (pgprot_val(prot) & PMD_SECT_PXN) 608 tableprot |= PUD_TABLE_PXN; 609 610 prot = __pgprot((pgprot_val(prot) & ~PMD_TYPE_MASK) | PMD_TYPE_SECT); 611 prot = __pgprot(pgprot_val(prot) & ~PTE_CONT); 612 if (to_cont) 613 prot = __pgprot(pgprot_val(prot) | PTE_CONT); 614 615 for (i = 0; i < PTRS_PER_PMD; i++, pmdp++, pfn += step) 616 set_pmd(pmdp, pfn_pmd(pfn, prot)); 617 618 /* 619 * Ensure the pmd entries are visible to the table walker by the time 620 * the pud entry that points to the pmds is visible. 621 */ 622 dsb(ishst); 623 __pud_populate(pudp, pmd_phys, tableprot); 624 625 return 0; 626 } 627 628 static int split_kernel_leaf_mapping_locked(unsigned long addr) 629 { 630 pgd_t *pgdp, pgd; 631 p4d_t *p4dp, p4d; 632 pud_t *pudp, pud; 633 pmd_t *pmdp, pmd; 634 pte_t *ptep, pte; 635 int ret = 0; 636 637 /* 638 * PGD: If addr is PGD aligned then addr already describes a leaf 639 * boundary. If not present then there is nothing to split. 640 */ 641 if (ALIGN_DOWN(addr, PGDIR_SIZE) == addr) 642 goto out; 643 pgdp = pgd_offset_k(addr); 644 pgd = pgdp_get(pgdp); 645 if (!pgd_present(pgd)) 646 goto out; 647 648 /* 649 * P4D: If addr is P4D aligned then addr already describes a leaf 650 * boundary. If not present then there is nothing to split. 651 */ 652 if (ALIGN_DOWN(addr, P4D_SIZE) == addr) 653 goto out; 654 p4dp = p4d_offset(pgdp, addr); 655 p4d = p4dp_get(p4dp); 656 if (!p4d_present(p4d)) 657 goto out; 658 659 /* 660 * PUD: If addr is PUD aligned then addr already describes a leaf 661 * boundary. If not present then there is nothing to split. Otherwise, 662 * if we have a pud leaf, split to contpmd. 663 */ 664 if (ALIGN_DOWN(addr, PUD_SIZE) == addr) 665 goto out; 666 pudp = pud_offset(p4dp, addr); 667 pud = pudp_get(pudp); 668 if (!pud_present(pud)) 669 goto out; 670 if (pud_leaf(pud)) { 671 ret = split_pud(pudp, pud, GFP_PGTABLE_KERNEL, true); 672 if (ret) 673 goto out; 674 } 675 676 /* 677 * CONTPMD: If addr is CONTPMD aligned then addr already describes a 678 * leaf boundary. If not present then there is nothing to split. 679 * Otherwise, if we have a contpmd leaf, split to pmd. 680 */ 681 if (ALIGN_DOWN(addr, CONT_PMD_SIZE) == addr) 682 goto out; 683 pmdp = pmd_offset(pudp, addr); 684 pmd = pmdp_get(pmdp); 685 if (!pmd_present(pmd)) 686 goto out; 687 if (pmd_leaf(pmd)) { 688 if (pmd_cont(pmd)) 689 split_contpmd(pmdp); 690 /* 691 * PMD: If addr is PMD aligned then addr already describes a 692 * leaf boundary. Otherwise, split to contpte. 693 */ 694 if (ALIGN_DOWN(addr, PMD_SIZE) == addr) 695 goto out; 696 ret = split_pmd(pmdp, pmd, GFP_PGTABLE_KERNEL, true); 697 if (ret) 698 goto out; 699 } 700 701 /* 702 * CONTPTE: If addr is CONTPTE aligned then addr already describes a 703 * leaf boundary. If not present then there is nothing to split. 704 * Otherwise, if we have a contpte leaf, split to pte. 705 */ 706 if (ALIGN_DOWN(addr, CONT_PTE_SIZE) == addr) 707 goto out; 708 ptep = pte_offset_kernel(pmdp, addr); 709 pte = __ptep_get(ptep); 710 if (!pte_present(pte)) 711 goto out; 712 if (pte_cont(pte)) 713 split_contpte(ptep); 714 715 out: 716 return ret; 717 } 718 719 static DEFINE_MUTEX(pgtable_split_lock); 720 721 int split_kernel_leaf_mapping(unsigned long start, unsigned long end) 722 { 723 int ret; 724 725 /* 726 * !BBML2_NOABORT systems should not be trying to change permissions on 727 * anything that is not pte-mapped in the first place. Just return early 728 * and let the permission change code raise a warning if not already 729 * pte-mapped. 730 */ 731 if (!system_supports_bbml2_noabort()) 732 return 0; 733 734 /* 735 * Ensure start and end are at least page-aligned since this is the 736 * finest granularity we can split to. 737 */ 738 if (start != PAGE_ALIGN(start) || end != PAGE_ALIGN(end)) 739 return -EINVAL; 740 741 mutex_lock(&pgtable_split_lock); 742 arch_enter_lazy_mmu_mode(); 743 744 /* 745 * The split_kernel_leaf_mapping_locked() may sleep, it is not a 746 * problem for ARM64 since ARM64's lazy MMU implementation allows 747 * sleeping. 748 * 749 * Optimize for the common case of splitting out a single page from a 750 * larger mapping. Here we can just split on the "least aligned" of 751 * start and end and this will guarantee that there must also be a split 752 * on the more aligned address since the both addresses must be in the 753 * same contpte block and it must have been split to ptes. 754 */ 755 if (end - start == PAGE_SIZE) { 756 start = __ffs(start) < __ffs(end) ? start : end; 757 ret = split_kernel_leaf_mapping_locked(start); 758 } else { 759 ret = split_kernel_leaf_mapping_locked(start); 760 if (!ret) 761 ret = split_kernel_leaf_mapping_locked(end); 762 } 763 764 arch_leave_lazy_mmu_mode(); 765 mutex_unlock(&pgtable_split_lock); 766 return ret; 767 } 768 769 static int __init split_to_ptes_pud_entry(pud_t *pudp, unsigned long addr, 770 unsigned long next, 771 struct mm_walk *walk) 772 { 773 pud_t pud = pudp_get(pudp); 774 int ret = 0; 775 776 if (pud_leaf(pud)) 777 ret = split_pud(pudp, pud, GFP_ATOMIC, false); 778 779 return ret; 780 } 781 782 static int __init split_to_ptes_pmd_entry(pmd_t *pmdp, unsigned long addr, 783 unsigned long next, 784 struct mm_walk *walk) 785 { 786 pmd_t pmd = pmdp_get(pmdp); 787 int ret = 0; 788 789 if (pmd_leaf(pmd)) { 790 if (pmd_cont(pmd)) 791 split_contpmd(pmdp); 792 ret = split_pmd(pmdp, pmd, GFP_ATOMIC, false); 793 794 /* 795 * We have split the pmd directly to ptes so there is no need to 796 * visit each pte to check if they are contpte. 797 */ 798 walk->action = ACTION_CONTINUE; 799 } 800 801 return ret; 802 } 803 804 static int __init split_to_ptes_pte_entry(pte_t *ptep, unsigned long addr, 805 unsigned long next, 806 struct mm_walk *walk) 807 { 808 pte_t pte = __ptep_get(ptep); 809 810 if (pte_cont(pte)) 811 split_contpte(ptep); 812 813 return 0; 814 } 815 816 static const struct mm_walk_ops split_to_ptes_ops __initconst = { 817 .pud_entry = split_to_ptes_pud_entry, 818 .pmd_entry = split_to_ptes_pmd_entry, 819 .pte_entry = split_to_ptes_pte_entry, 820 }; 821 822 static bool linear_map_requires_bbml2 __initdata; 823 824 u32 idmap_kpti_bbml2_flag; 825 826 void __init init_idmap_kpti_bbml2_flag(void) 827 { 828 WRITE_ONCE(idmap_kpti_bbml2_flag, 1); 829 /* Must be visible to other CPUs before stop_machine() is called. */ 830 smp_mb(); 831 } 832 833 static int __init linear_map_split_to_ptes(void *__unused) 834 { 835 /* 836 * Repainting the linear map must be done by CPU0 (the boot CPU) because 837 * that's the only CPU that we know supports BBML2. The other CPUs will 838 * be held in a waiting area with the idmap active. 839 */ 840 if (!smp_processor_id()) { 841 unsigned long lstart = _PAGE_OFFSET(vabits_actual); 842 unsigned long lend = PAGE_END; 843 unsigned long kstart = (unsigned long)lm_alias(_stext); 844 unsigned long kend = (unsigned long)lm_alias(__init_begin); 845 int ret; 846 847 /* 848 * Wait for all secondary CPUs to be put into the waiting area. 849 */ 850 smp_cond_load_acquire(&idmap_kpti_bbml2_flag, VAL == num_online_cpus()); 851 852 /* 853 * Walk all of the linear map [lstart, lend), except the kernel 854 * linear map alias [kstart, kend), and split all mappings to 855 * PTE. The kernel alias remains static throughout runtime so 856 * can continue to be safely mapped with large mappings. 857 */ 858 ret = walk_kernel_page_table_range_lockless(lstart, kstart, 859 &split_to_ptes_ops, NULL, NULL); 860 if (!ret) 861 ret = walk_kernel_page_table_range_lockless(kend, lend, 862 &split_to_ptes_ops, NULL, NULL); 863 if (ret) 864 panic("Failed to split linear map\n"); 865 flush_tlb_kernel_range(lstart, lend); 866 867 /* 868 * Relies on dsb in flush_tlb_kernel_range() to avoid reordering 869 * before any page table split operations. 870 */ 871 WRITE_ONCE(idmap_kpti_bbml2_flag, 0); 872 } else { 873 typedef void (wait_split_fn)(void); 874 extern wait_split_fn wait_linear_map_split_to_ptes; 875 wait_split_fn *wait_fn; 876 877 wait_fn = (void *)__pa_symbol(wait_linear_map_split_to_ptes); 878 879 /* 880 * At least one secondary CPU doesn't support BBML2 so cannot 881 * tolerate the size of the live mappings changing. So have the 882 * secondary CPUs wait for the boot CPU to make the changes 883 * with the idmap active and init_mm inactive. 884 */ 885 cpu_install_idmap(); 886 wait_fn(); 887 cpu_uninstall_idmap(); 888 } 889 890 return 0; 891 } 892 893 void __init linear_map_maybe_split_to_ptes(void) 894 { 895 if (linear_map_requires_bbml2 && !system_supports_bbml2_noabort()) { 896 init_idmap_kpti_bbml2_flag(); 897 stop_machine(linear_map_split_to_ptes, NULL, cpu_online_mask); 898 } 899 } 900 901 /* 902 * This function can only be used to modify existing table entries, 903 * without allocating new levels of table. Note that this permits the 904 * creation of new section or page entries. 905 */ 906 void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt, 907 phys_addr_t size, pgprot_t prot) 908 { 909 if (virt < PAGE_OFFSET) { 910 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n", 911 &phys, virt); 912 return; 913 } 914 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 915 NO_CONT_MAPPINGS); 916 } 917 918 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys, 919 unsigned long virt, phys_addr_t size, 920 pgprot_t prot, bool page_mappings_only) 921 { 922 int flags = 0; 923 924 BUG_ON(mm == &init_mm); 925 926 if (page_mappings_only) 927 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 928 929 __create_pgd_mapping(mm->pgd, phys, virt, size, prot, 930 pgd_pgtable_alloc_special_mm, flags); 931 } 932 933 static void update_mapping_prot(phys_addr_t phys, unsigned long virt, 934 phys_addr_t size, pgprot_t prot) 935 { 936 if (virt < PAGE_OFFSET) { 937 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n", 938 &phys, virt); 939 return; 940 } 941 942 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 943 NO_CONT_MAPPINGS); 944 945 /* flush the TLBs after updating live kernel mappings */ 946 flush_tlb_kernel_range(virt, virt + size); 947 } 948 949 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start, 950 phys_addr_t end, pgprot_t prot, int flags) 951 { 952 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start, 953 prot, early_pgtable_alloc, flags); 954 } 955 956 void __init mark_linear_text_alias_ro(void) 957 { 958 /* 959 * Remove the write permissions from the linear alias of .text/.rodata 960 */ 961 update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text), 962 (unsigned long)__init_begin - (unsigned long)_text, 963 PAGE_KERNEL_RO); 964 } 965 966 #ifdef CONFIG_KFENCE 967 968 bool __ro_after_init kfence_early_init = !!CONFIG_KFENCE_SAMPLE_INTERVAL; 969 970 /* early_param() will be parsed before map_mem() below. */ 971 static int __init parse_kfence_early_init(char *arg) 972 { 973 int val; 974 975 if (get_option(&arg, &val)) 976 kfence_early_init = !!val; 977 return 0; 978 } 979 early_param("kfence.sample_interval", parse_kfence_early_init); 980 981 static phys_addr_t __init arm64_kfence_alloc_pool(void) 982 { 983 phys_addr_t kfence_pool; 984 985 if (!kfence_early_init) 986 return 0; 987 988 kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE); 989 if (!kfence_pool) { 990 pr_err("failed to allocate kfence pool\n"); 991 kfence_early_init = false; 992 return 0; 993 } 994 995 /* Temporarily mark as NOMAP. */ 996 memblock_mark_nomap(kfence_pool, KFENCE_POOL_SIZE); 997 998 return kfence_pool; 999 } 1000 1001 static void __init arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) 1002 { 1003 if (!kfence_pool) 1004 return; 1005 1006 /* KFENCE pool needs page-level mapping. */ 1007 __map_memblock(pgdp, kfence_pool, kfence_pool + KFENCE_POOL_SIZE, 1008 pgprot_tagged(PAGE_KERNEL), 1009 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS); 1010 memblock_clear_nomap(kfence_pool, KFENCE_POOL_SIZE); 1011 __kfence_pool = phys_to_virt(kfence_pool); 1012 } 1013 #else /* CONFIG_KFENCE */ 1014 1015 static inline phys_addr_t arm64_kfence_alloc_pool(void) { return 0; } 1016 static inline void arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) { } 1017 1018 #endif /* CONFIG_KFENCE */ 1019 1020 static inline bool force_pte_mapping(void) 1021 { 1022 bool bbml2 = system_capabilities_finalized() ? 1023 system_supports_bbml2_noabort() : cpu_supports_bbml2_noabort(); 1024 1025 return (!bbml2 && (rodata_full || arm64_kfence_can_set_direct_map() || 1026 is_realm_world())) || 1027 debug_pagealloc_enabled(); 1028 } 1029 1030 static void __init map_mem(pgd_t *pgdp) 1031 { 1032 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN); 1033 phys_addr_t kernel_start = __pa_symbol(_text); 1034 phys_addr_t kernel_end = __pa_symbol(__init_begin); 1035 phys_addr_t start, end; 1036 phys_addr_t early_kfence_pool; 1037 int flags = NO_EXEC_MAPPINGS; 1038 u64 i; 1039 1040 /* 1041 * Setting hierarchical PXNTable attributes on table entries covering 1042 * the linear region is only possible if it is guaranteed that no table 1043 * entries at any level are being shared between the linear region and 1044 * the vmalloc region. Check whether this is true for the PGD level, in 1045 * which case it is guaranteed to be true for all other levels as well. 1046 * (Unless we are running with support for LPA2, in which case the 1047 * entire reduced VA space is covered by a single pgd_t which will have 1048 * been populated without the PXNTable attribute by the time we get here.) 1049 */ 1050 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end) && 1051 pgd_index(_PAGE_OFFSET(VA_BITS_MIN)) != PTRS_PER_PGD - 1); 1052 1053 early_kfence_pool = arm64_kfence_alloc_pool(); 1054 1055 linear_map_requires_bbml2 = !force_pte_mapping() && can_set_direct_map(); 1056 1057 if (force_pte_mapping()) 1058 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1059 1060 /* 1061 * Take care not to create a writable alias for the 1062 * read-only text and rodata sections of the kernel image. 1063 * So temporarily mark them as NOMAP to skip mappings in 1064 * the following for-loop 1065 */ 1066 memblock_mark_nomap(kernel_start, kernel_end - kernel_start); 1067 1068 /* map all the memory banks */ 1069 for_each_mem_range(i, &start, &end) { 1070 if (start >= end) 1071 break; 1072 /* 1073 * The linear map must allow allocation tags reading/writing 1074 * if MTE is present. Otherwise, it has the same attributes as 1075 * PAGE_KERNEL. 1076 */ 1077 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL), 1078 flags); 1079 } 1080 1081 /* 1082 * Map the linear alias of the [_text, __init_begin) interval 1083 * as non-executable now, and remove the write permission in 1084 * mark_linear_text_alias_ro() below (which will be called after 1085 * alternative patching has completed). This makes the contents 1086 * of the region accessible to subsystems such as hibernate, 1087 * but protects it from inadvertent modification or execution. 1088 * Note that contiguous mappings cannot be remapped in this way, 1089 * so we should avoid them here. 1090 */ 1091 __map_memblock(pgdp, kernel_start, kernel_end, 1092 PAGE_KERNEL, NO_CONT_MAPPINGS); 1093 memblock_clear_nomap(kernel_start, kernel_end - kernel_start); 1094 arm64_kfence_map_pool(early_kfence_pool, pgdp); 1095 } 1096 1097 void mark_rodata_ro(void) 1098 { 1099 unsigned long section_size; 1100 1101 /* 1102 * mark .rodata as read only. Use __init_begin rather than __end_rodata 1103 * to cover NOTES and EXCEPTION_TABLE. 1104 */ 1105 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata; 1106 WRITE_ONCE(rodata_is_rw, false); 1107 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata, 1108 section_size, PAGE_KERNEL_RO); 1109 /* mark the range between _text and _stext as read only. */ 1110 update_mapping_prot(__pa_symbol(_text), (unsigned long)_text, 1111 (unsigned long)_stext - (unsigned long)_text, 1112 PAGE_KERNEL_RO); 1113 } 1114 1115 static void __init declare_vma(struct vm_struct *vma, 1116 void *va_start, void *va_end, 1117 unsigned long vm_flags) 1118 { 1119 phys_addr_t pa_start = __pa_symbol(va_start); 1120 unsigned long size = va_end - va_start; 1121 1122 BUG_ON(!PAGE_ALIGNED(pa_start)); 1123 BUG_ON(!PAGE_ALIGNED(size)); 1124 1125 if (!(vm_flags & VM_NO_GUARD)) 1126 size += PAGE_SIZE; 1127 1128 vma->addr = va_start; 1129 vma->phys_addr = pa_start; 1130 vma->size = size; 1131 vma->flags = VM_MAP | vm_flags; 1132 vma->caller = __builtin_return_address(0); 1133 1134 vm_area_add_early(vma); 1135 } 1136 1137 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0 1138 static pgprot_t kernel_exec_prot(void) 1139 { 1140 return rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 1141 } 1142 1143 static int __init map_entry_trampoline(void) 1144 { 1145 int i; 1146 1147 if (!arm64_kernel_unmapped_at_el0()) 1148 return 0; 1149 1150 pgprot_t prot = kernel_exec_prot(); 1151 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start); 1152 1153 /* The trampoline is always mapped and can therefore be global */ 1154 pgprot_val(prot) &= ~PTE_NG; 1155 1156 /* Map only the text into the trampoline page table */ 1157 memset(tramp_pg_dir, 0, PGD_SIZE); 1158 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, 1159 entry_tramp_text_size(), prot, 1160 pgd_pgtable_alloc_init_mm, NO_BLOCK_MAPPINGS); 1161 1162 /* Map both the text and data into the kernel page table */ 1163 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++) 1164 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 1165 pa_start + i * PAGE_SIZE, prot); 1166 1167 if (IS_ENABLED(CONFIG_RELOCATABLE)) 1168 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 1169 pa_start + i * PAGE_SIZE, PAGE_KERNEL_RO); 1170 1171 return 0; 1172 } 1173 core_initcall(map_entry_trampoline); 1174 #endif 1175 1176 /* 1177 * Declare the VMA areas for the kernel 1178 */ 1179 static void __init declare_kernel_vmas(void) 1180 { 1181 static struct vm_struct vmlinux_seg[KERNEL_SEGMENT_COUNT]; 1182 1183 declare_vma(&vmlinux_seg[0], _text, _etext, VM_NO_GUARD); 1184 declare_vma(&vmlinux_seg[1], __start_rodata, __inittext_begin, VM_NO_GUARD); 1185 declare_vma(&vmlinux_seg[2], __inittext_begin, __inittext_end, VM_NO_GUARD); 1186 declare_vma(&vmlinux_seg[3], __initdata_begin, __initdata_end, VM_NO_GUARD); 1187 declare_vma(&vmlinux_seg[4], _data, _end, 0); 1188 } 1189 1190 void __pi_map_range(phys_addr_t *pte, u64 start, u64 end, phys_addr_t pa, 1191 pgprot_t prot, int level, pte_t *tbl, bool may_use_cont, 1192 u64 va_offset); 1193 1194 static u8 idmap_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init, 1195 kpti_bbml2_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init; 1196 1197 static void __init create_idmap(void) 1198 { 1199 phys_addr_t start = __pa_symbol(__idmap_text_start); 1200 phys_addr_t end = __pa_symbol(__idmap_text_end); 1201 phys_addr_t ptep = __pa_symbol(idmap_ptes); 1202 1203 __pi_map_range(&ptep, start, end, start, PAGE_KERNEL_ROX, 1204 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 1205 __phys_to_virt(ptep) - ptep); 1206 1207 if (linear_map_requires_bbml2 || 1208 (IS_ENABLED(CONFIG_UNMAP_KERNEL_AT_EL0) && !arm64_use_ng_mappings)) { 1209 phys_addr_t pa = __pa_symbol(&idmap_kpti_bbml2_flag); 1210 1211 /* 1212 * The KPTI G-to-nG conversion code needs a read-write mapping 1213 * of its synchronization flag in the ID map. This is also used 1214 * when splitting the linear map to ptes if a secondary CPU 1215 * doesn't support bbml2. 1216 */ 1217 ptep = __pa_symbol(kpti_bbml2_ptes); 1218 __pi_map_range(&ptep, pa, pa + sizeof(u32), pa, PAGE_KERNEL, 1219 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 1220 __phys_to_virt(ptep) - ptep); 1221 } 1222 } 1223 1224 void __init paging_init(void) 1225 { 1226 map_mem(swapper_pg_dir); 1227 1228 memblock_allow_resize(); 1229 1230 create_idmap(); 1231 declare_kernel_vmas(); 1232 } 1233 1234 #ifdef CONFIG_MEMORY_HOTPLUG 1235 static void free_hotplug_page_range(struct page *page, size_t size, 1236 struct vmem_altmap *altmap) 1237 { 1238 if (altmap) { 1239 vmem_altmap_free(altmap, size >> PAGE_SHIFT); 1240 } else { 1241 WARN_ON(PageReserved(page)); 1242 __free_pages(page, get_order(size)); 1243 } 1244 } 1245 1246 static void free_hotplug_pgtable_page(struct page *page) 1247 { 1248 free_hotplug_page_range(page, PAGE_SIZE, NULL); 1249 } 1250 1251 static bool pgtable_range_aligned(unsigned long start, unsigned long end, 1252 unsigned long floor, unsigned long ceiling, 1253 unsigned long mask) 1254 { 1255 start &= mask; 1256 if (start < floor) 1257 return false; 1258 1259 if (ceiling) { 1260 ceiling &= mask; 1261 if (!ceiling) 1262 return false; 1263 } 1264 1265 if (end - 1 > ceiling - 1) 1266 return false; 1267 return true; 1268 } 1269 1270 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr, 1271 unsigned long end, bool free_mapped, 1272 struct vmem_altmap *altmap) 1273 { 1274 pte_t *ptep, pte; 1275 1276 do { 1277 ptep = pte_offset_kernel(pmdp, addr); 1278 pte = __ptep_get(ptep); 1279 if (pte_none(pte)) 1280 continue; 1281 1282 WARN_ON(!pte_present(pte)); 1283 __pte_clear(&init_mm, addr, ptep); 1284 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1285 if (free_mapped) 1286 free_hotplug_page_range(pte_page(pte), 1287 PAGE_SIZE, altmap); 1288 } while (addr += PAGE_SIZE, addr < end); 1289 } 1290 1291 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr, 1292 unsigned long end, bool free_mapped, 1293 struct vmem_altmap *altmap) 1294 { 1295 unsigned long next; 1296 pmd_t *pmdp, pmd; 1297 1298 do { 1299 next = pmd_addr_end(addr, end); 1300 pmdp = pmd_offset(pudp, addr); 1301 pmd = READ_ONCE(*pmdp); 1302 if (pmd_none(pmd)) 1303 continue; 1304 1305 WARN_ON(!pmd_present(pmd)); 1306 if (pmd_sect(pmd)) { 1307 pmd_clear(pmdp); 1308 1309 /* 1310 * One TLBI should be sufficient here as the PMD_SIZE 1311 * range is mapped with a single block entry. 1312 */ 1313 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1314 if (free_mapped) 1315 free_hotplug_page_range(pmd_page(pmd), 1316 PMD_SIZE, altmap); 1317 continue; 1318 } 1319 WARN_ON(!pmd_table(pmd)); 1320 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap); 1321 } while (addr = next, addr < end); 1322 } 1323 1324 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr, 1325 unsigned long end, bool free_mapped, 1326 struct vmem_altmap *altmap) 1327 { 1328 unsigned long next; 1329 pud_t *pudp, pud; 1330 1331 do { 1332 next = pud_addr_end(addr, end); 1333 pudp = pud_offset(p4dp, addr); 1334 pud = READ_ONCE(*pudp); 1335 if (pud_none(pud)) 1336 continue; 1337 1338 WARN_ON(!pud_present(pud)); 1339 if (pud_sect(pud)) { 1340 pud_clear(pudp); 1341 1342 /* 1343 * One TLBI should be sufficient here as the PUD_SIZE 1344 * range is mapped with a single block entry. 1345 */ 1346 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1347 if (free_mapped) 1348 free_hotplug_page_range(pud_page(pud), 1349 PUD_SIZE, altmap); 1350 continue; 1351 } 1352 WARN_ON(!pud_table(pud)); 1353 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap); 1354 } while (addr = next, addr < end); 1355 } 1356 1357 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr, 1358 unsigned long end, bool free_mapped, 1359 struct vmem_altmap *altmap) 1360 { 1361 unsigned long next; 1362 p4d_t *p4dp, p4d; 1363 1364 do { 1365 next = p4d_addr_end(addr, end); 1366 p4dp = p4d_offset(pgdp, addr); 1367 p4d = READ_ONCE(*p4dp); 1368 if (p4d_none(p4d)) 1369 continue; 1370 1371 WARN_ON(!p4d_present(p4d)); 1372 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap); 1373 } while (addr = next, addr < end); 1374 } 1375 1376 static void unmap_hotplug_range(unsigned long addr, unsigned long end, 1377 bool free_mapped, struct vmem_altmap *altmap) 1378 { 1379 unsigned long next; 1380 pgd_t *pgdp, pgd; 1381 1382 /* 1383 * altmap can only be used as vmemmap mapping backing memory. 1384 * In case the backing memory itself is not being freed, then 1385 * altmap is irrelevant. Warn about this inconsistency when 1386 * encountered. 1387 */ 1388 WARN_ON(!free_mapped && altmap); 1389 1390 do { 1391 next = pgd_addr_end(addr, end); 1392 pgdp = pgd_offset_k(addr); 1393 pgd = READ_ONCE(*pgdp); 1394 if (pgd_none(pgd)) 1395 continue; 1396 1397 WARN_ON(!pgd_present(pgd)); 1398 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap); 1399 } while (addr = next, addr < end); 1400 } 1401 1402 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr, 1403 unsigned long end, unsigned long floor, 1404 unsigned long ceiling) 1405 { 1406 pte_t *ptep, pte; 1407 unsigned long i, start = addr; 1408 1409 do { 1410 ptep = pte_offset_kernel(pmdp, addr); 1411 pte = __ptep_get(ptep); 1412 1413 /* 1414 * This is just a sanity check here which verifies that 1415 * pte clearing has been done by earlier unmap loops. 1416 */ 1417 WARN_ON(!pte_none(pte)); 1418 } while (addr += PAGE_SIZE, addr < end); 1419 1420 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK)) 1421 return; 1422 1423 /* 1424 * Check whether we can free the pte page if the rest of the 1425 * entries are empty. Overlap with other regions have been 1426 * handled by the floor/ceiling check. 1427 */ 1428 ptep = pte_offset_kernel(pmdp, 0UL); 1429 for (i = 0; i < PTRS_PER_PTE; i++) { 1430 if (!pte_none(__ptep_get(&ptep[i]))) 1431 return; 1432 } 1433 1434 pmd_clear(pmdp); 1435 __flush_tlb_kernel_pgtable(start); 1436 free_hotplug_pgtable_page(virt_to_page(ptep)); 1437 } 1438 1439 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr, 1440 unsigned long end, unsigned long floor, 1441 unsigned long ceiling) 1442 { 1443 pmd_t *pmdp, pmd; 1444 unsigned long i, next, start = addr; 1445 1446 do { 1447 next = pmd_addr_end(addr, end); 1448 pmdp = pmd_offset(pudp, addr); 1449 pmd = READ_ONCE(*pmdp); 1450 if (pmd_none(pmd)) 1451 continue; 1452 1453 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd)); 1454 free_empty_pte_table(pmdp, addr, next, floor, ceiling); 1455 } while (addr = next, addr < end); 1456 1457 if (CONFIG_PGTABLE_LEVELS <= 2) 1458 return; 1459 1460 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK)) 1461 return; 1462 1463 /* 1464 * Check whether we can free the pmd page if the rest of the 1465 * entries are empty. Overlap with other regions have been 1466 * handled by the floor/ceiling check. 1467 */ 1468 pmdp = pmd_offset(pudp, 0UL); 1469 for (i = 0; i < PTRS_PER_PMD; i++) { 1470 if (!pmd_none(READ_ONCE(pmdp[i]))) 1471 return; 1472 } 1473 1474 pud_clear(pudp); 1475 __flush_tlb_kernel_pgtable(start); 1476 free_hotplug_pgtable_page(virt_to_page(pmdp)); 1477 } 1478 1479 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr, 1480 unsigned long end, unsigned long floor, 1481 unsigned long ceiling) 1482 { 1483 pud_t *pudp, pud; 1484 unsigned long i, next, start = addr; 1485 1486 do { 1487 next = pud_addr_end(addr, end); 1488 pudp = pud_offset(p4dp, addr); 1489 pud = READ_ONCE(*pudp); 1490 if (pud_none(pud)) 1491 continue; 1492 1493 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud)); 1494 free_empty_pmd_table(pudp, addr, next, floor, ceiling); 1495 } while (addr = next, addr < end); 1496 1497 if (!pgtable_l4_enabled()) 1498 return; 1499 1500 if (!pgtable_range_aligned(start, end, floor, ceiling, P4D_MASK)) 1501 return; 1502 1503 /* 1504 * Check whether we can free the pud page if the rest of the 1505 * entries are empty. Overlap with other regions have been 1506 * handled by the floor/ceiling check. 1507 */ 1508 pudp = pud_offset(p4dp, 0UL); 1509 for (i = 0; i < PTRS_PER_PUD; i++) { 1510 if (!pud_none(READ_ONCE(pudp[i]))) 1511 return; 1512 } 1513 1514 p4d_clear(p4dp); 1515 __flush_tlb_kernel_pgtable(start); 1516 free_hotplug_pgtable_page(virt_to_page(pudp)); 1517 } 1518 1519 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr, 1520 unsigned long end, unsigned long floor, 1521 unsigned long ceiling) 1522 { 1523 p4d_t *p4dp, p4d; 1524 unsigned long i, next, start = addr; 1525 1526 do { 1527 next = p4d_addr_end(addr, end); 1528 p4dp = p4d_offset(pgdp, addr); 1529 p4d = READ_ONCE(*p4dp); 1530 if (p4d_none(p4d)) 1531 continue; 1532 1533 WARN_ON(!p4d_present(p4d)); 1534 free_empty_pud_table(p4dp, addr, next, floor, ceiling); 1535 } while (addr = next, addr < end); 1536 1537 if (!pgtable_l5_enabled()) 1538 return; 1539 1540 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK)) 1541 return; 1542 1543 /* 1544 * Check whether we can free the p4d page if the rest of the 1545 * entries are empty. Overlap with other regions have been 1546 * handled by the floor/ceiling check. 1547 */ 1548 p4dp = p4d_offset(pgdp, 0UL); 1549 for (i = 0; i < PTRS_PER_P4D; i++) { 1550 if (!p4d_none(READ_ONCE(p4dp[i]))) 1551 return; 1552 } 1553 1554 pgd_clear(pgdp); 1555 __flush_tlb_kernel_pgtable(start); 1556 free_hotplug_pgtable_page(virt_to_page(p4dp)); 1557 } 1558 1559 static void free_empty_tables(unsigned long addr, unsigned long end, 1560 unsigned long floor, unsigned long ceiling) 1561 { 1562 unsigned long next; 1563 pgd_t *pgdp, pgd; 1564 1565 do { 1566 next = pgd_addr_end(addr, end); 1567 pgdp = pgd_offset_k(addr); 1568 pgd = READ_ONCE(*pgdp); 1569 if (pgd_none(pgd)) 1570 continue; 1571 1572 WARN_ON(!pgd_present(pgd)); 1573 free_empty_p4d_table(pgdp, addr, next, floor, ceiling); 1574 } while (addr = next, addr < end); 1575 } 1576 #endif 1577 1578 void __meminit vmemmap_set_pmd(pmd_t *pmdp, void *p, int node, 1579 unsigned long addr, unsigned long next) 1580 { 1581 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL)); 1582 } 1583 1584 int __meminit vmemmap_check_pmd(pmd_t *pmdp, int node, 1585 unsigned long addr, unsigned long next) 1586 { 1587 vmemmap_verify((pte_t *)pmdp, node, addr, next); 1588 1589 return pmd_sect(READ_ONCE(*pmdp)); 1590 } 1591 1592 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1593 struct vmem_altmap *altmap) 1594 { 1595 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1596 /* [start, end] should be within one section */ 1597 WARN_ON_ONCE(end - start > PAGES_PER_SECTION * sizeof(struct page)); 1598 1599 if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || 1600 (end - start < PAGES_PER_SECTION * sizeof(struct page))) 1601 return vmemmap_populate_basepages(start, end, node, altmap); 1602 else 1603 return vmemmap_populate_hugepages(start, end, node, altmap); 1604 } 1605 1606 #ifdef CONFIG_MEMORY_HOTPLUG 1607 void vmemmap_free(unsigned long start, unsigned long end, 1608 struct vmem_altmap *altmap) 1609 { 1610 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1611 1612 unmap_hotplug_range(start, end, true, altmap); 1613 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END); 1614 } 1615 #endif /* CONFIG_MEMORY_HOTPLUG */ 1616 1617 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot) 1618 { 1619 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot)); 1620 1621 /* Only allow permission changes for now */ 1622 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)), 1623 pud_val(new_pud))) 1624 return 0; 1625 1626 VM_BUG_ON(phys & ~PUD_MASK); 1627 set_pud(pudp, new_pud); 1628 return 1; 1629 } 1630 1631 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot) 1632 { 1633 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot)); 1634 1635 /* Only allow permission changes for now */ 1636 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)), 1637 pmd_val(new_pmd))) 1638 return 0; 1639 1640 VM_BUG_ON(phys & ~PMD_MASK); 1641 set_pmd(pmdp, new_pmd); 1642 return 1; 1643 } 1644 1645 #ifndef __PAGETABLE_P4D_FOLDED 1646 void p4d_clear_huge(p4d_t *p4dp) 1647 { 1648 } 1649 #endif 1650 1651 int pud_clear_huge(pud_t *pudp) 1652 { 1653 if (!pud_sect(READ_ONCE(*pudp))) 1654 return 0; 1655 pud_clear(pudp); 1656 return 1; 1657 } 1658 1659 int pmd_clear_huge(pmd_t *pmdp) 1660 { 1661 if (!pmd_sect(READ_ONCE(*pmdp))) 1662 return 0; 1663 pmd_clear(pmdp); 1664 return 1; 1665 } 1666 1667 static int __pmd_free_pte_page(pmd_t *pmdp, unsigned long addr, 1668 bool acquire_mmap_lock) 1669 { 1670 pte_t *table; 1671 pmd_t pmd; 1672 1673 pmd = READ_ONCE(*pmdp); 1674 1675 if (!pmd_table(pmd)) { 1676 VM_WARN_ON(1); 1677 return 1; 1678 } 1679 1680 /* See comment in pud_free_pmd_page for static key logic */ 1681 table = pte_offset_kernel(pmdp, addr); 1682 pmd_clear(pmdp); 1683 __flush_tlb_kernel_pgtable(addr); 1684 if (static_branch_unlikely(&arm64_ptdump_lock_key) && acquire_mmap_lock) { 1685 mmap_read_lock(&init_mm); 1686 mmap_read_unlock(&init_mm); 1687 } 1688 1689 pte_free_kernel(NULL, table); 1690 return 1; 1691 } 1692 1693 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr) 1694 { 1695 /* If ptdump is walking the pagetables, acquire init_mm.mmap_lock */ 1696 return __pmd_free_pte_page(pmdp, addr, /* acquire_mmap_lock = */ true); 1697 } 1698 1699 int pud_free_pmd_page(pud_t *pudp, unsigned long addr) 1700 { 1701 pmd_t *table; 1702 pmd_t *pmdp; 1703 pud_t pud; 1704 unsigned long next, end; 1705 1706 pud = READ_ONCE(*pudp); 1707 1708 if (!pud_table(pud)) { 1709 VM_WARN_ON(1); 1710 return 1; 1711 } 1712 1713 table = pmd_offset(pudp, addr); 1714 1715 /* 1716 * Our objective is to prevent ptdump from reading a PMD table which has 1717 * been freed. In this race, if pud_free_pmd_page observes the key on 1718 * (which got flipped by ptdump) then the mmap lock sequence here will, 1719 * as a result of the mmap write lock/unlock sequence in ptdump, give 1720 * us the correct synchronization. If not, this means that ptdump has 1721 * yet not started walking the pagetables - the sequence of barriers 1722 * issued by __flush_tlb_kernel_pgtable() guarantees that ptdump will 1723 * observe an empty PUD. 1724 */ 1725 pud_clear(pudp); 1726 __flush_tlb_kernel_pgtable(addr); 1727 if (static_branch_unlikely(&arm64_ptdump_lock_key)) { 1728 mmap_read_lock(&init_mm); 1729 mmap_read_unlock(&init_mm); 1730 } 1731 1732 pmdp = table; 1733 next = addr; 1734 end = addr + PUD_SIZE; 1735 do { 1736 if (pmd_present(pmdp_get(pmdp))) 1737 /* 1738 * PMD has been isolated, so ptdump won't see it. No 1739 * need to acquire init_mm.mmap_lock. 1740 */ 1741 __pmd_free_pte_page(pmdp, next, /* acquire_mmap_lock = */ false); 1742 } while (pmdp++, next += PMD_SIZE, next != end); 1743 1744 pmd_free(NULL, table); 1745 return 1; 1746 } 1747 1748 #ifdef CONFIG_MEMORY_HOTPLUG 1749 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size) 1750 { 1751 unsigned long end = start + size; 1752 1753 WARN_ON(pgdir != init_mm.pgd); 1754 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END)); 1755 1756 unmap_hotplug_range(start, end, false, NULL); 1757 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END); 1758 } 1759 1760 struct range arch_get_mappable_range(void) 1761 { 1762 struct range mhp_range; 1763 phys_addr_t start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual)); 1764 phys_addr_t end_linear_pa = __pa(PAGE_END - 1); 1765 1766 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 1767 /* 1768 * Check for a wrap, it is possible because of randomized linear 1769 * mapping the start physical address is actually bigger than 1770 * the end physical address. In this case set start to zero 1771 * because [0, end_linear_pa] range must still be able to cover 1772 * all addressable physical addresses. 1773 */ 1774 if (start_linear_pa > end_linear_pa) 1775 start_linear_pa = 0; 1776 } 1777 1778 WARN_ON(start_linear_pa > end_linear_pa); 1779 1780 /* 1781 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)] 1782 * accommodating both its ends but excluding PAGE_END. Max physical 1783 * range which can be mapped inside this linear mapping range, must 1784 * also be derived from its end points. 1785 */ 1786 mhp_range.start = start_linear_pa; 1787 mhp_range.end = end_linear_pa; 1788 1789 return mhp_range; 1790 } 1791 1792 int arch_add_memory(int nid, u64 start, u64 size, 1793 struct mhp_params *params) 1794 { 1795 int ret, flags = NO_EXEC_MAPPINGS; 1796 1797 VM_BUG_ON(!mhp_range_allowed(start, size, true)); 1798 1799 if (force_pte_mapping()) 1800 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1801 1802 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1803 size, params->pgprot, pgd_pgtable_alloc_init_mm, 1804 flags); 1805 1806 memblock_clear_nomap(start, size); 1807 1808 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, 1809 params); 1810 if (ret) 1811 __remove_pgd_mapping(swapper_pg_dir, 1812 __phys_to_virt(start), size); 1813 else { 1814 /* Address of hotplugged memory can be smaller */ 1815 max_pfn = max(max_pfn, PFN_UP(start + size)); 1816 max_low_pfn = max_pfn; 1817 } 1818 1819 return ret; 1820 } 1821 1822 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap) 1823 { 1824 unsigned long start_pfn = start >> PAGE_SHIFT; 1825 unsigned long nr_pages = size >> PAGE_SHIFT; 1826 1827 __remove_pages(start_pfn, nr_pages, altmap); 1828 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size); 1829 } 1830 1831 /* 1832 * This memory hotplug notifier helps prevent boot memory from being 1833 * inadvertently removed as it blocks pfn range offlining process in 1834 * __offline_pages(). Hence this prevents both offlining as well as 1835 * removal process for boot memory which is initially always online. 1836 * In future if and when boot memory could be removed, this notifier 1837 * should be dropped and free_hotplug_page_range() should handle any 1838 * reserved pages allocated during boot. 1839 */ 1840 static int prevent_bootmem_remove_notifier(struct notifier_block *nb, 1841 unsigned long action, void *data) 1842 { 1843 struct mem_section *ms; 1844 struct memory_notify *arg = data; 1845 unsigned long end_pfn = arg->start_pfn + arg->nr_pages; 1846 unsigned long pfn = arg->start_pfn; 1847 1848 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE)) 1849 return NOTIFY_OK; 1850 1851 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 1852 unsigned long start = PFN_PHYS(pfn); 1853 unsigned long end = start + (1UL << PA_SECTION_SHIFT); 1854 1855 ms = __pfn_to_section(pfn); 1856 if (!early_section(ms)) 1857 continue; 1858 1859 if (action == MEM_GOING_OFFLINE) { 1860 /* 1861 * Boot memory removal is not supported. Prevent 1862 * it via blocking any attempted offline request 1863 * for the boot memory and just report it. 1864 */ 1865 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end); 1866 return NOTIFY_BAD; 1867 } else if (action == MEM_OFFLINE) { 1868 /* 1869 * This should have never happened. Boot memory 1870 * offlining should have been prevented by this 1871 * very notifier. Probably some memory removal 1872 * procedure might have changed which would then 1873 * require further debug. 1874 */ 1875 pr_err("Boot memory [%lx %lx] offlined\n", start, end); 1876 1877 /* 1878 * Core memory hotplug does not process a return 1879 * code from the notifier for MEM_OFFLINE events. 1880 * The error condition has been reported. Return 1881 * from here as if ignored. 1882 */ 1883 return NOTIFY_DONE; 1884 } 1885 } 1886 return NOTIFY_OK; 1887 } 1888 1889 static struct notifier_block prevent_bootmem_remove_nb = { 1890 .notifier_call = prevent_bootmem_remove_notifier, 1891 }; 1892 1893 /* 1894 * This ensures that boot memory sections on the platform are online 1895 * from early boot. Memory sections could not be prevented from being 1896 * offlined, unless for some reason they are not online to begin with. 1897 * This helps validate the basic assumption on which the above memory 1898 * event notifier works to prevent boot memory section offlining and 1899 * its possible removal. 1900 */ 1901 static void validate_bootmem_online(void) 1902 { 1903 phys_addr_t start, end, addr; 1904 struct mem_section *ms; 1905 u64 i; 1906 1907 /* 1908 * Scanning across all memblock might be expensive 1909 * on some big memory systems. Hence enable this 1910 * validation only with DEBUG_VM. 1911 */ 1912 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 1913 return; 1914 1915 for_each_mem_range(i, &start, &end) { 1916 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) { 1917 ms = __pfn_to_section(PHYS_PFN(addr)); 1918 1919 /* 1920 * All memory ranges in the system at this point 1921 * should have been marked as early sections. 1922 */ 1923 WARN_ON(!early_section(ms)); 1924 1925 /* 1926 * Memory notifier mechanism here to prevent boot 1927 * memory offlining depends on the fact that each 1928 * early section memory on the system is initially 1929 * online. Otherwise a given memory section which 1930 * is already offline will be overlooked and can 1931 * be removed completely. Call out such sections. 1932 */ 1933 if (!online_section(ms)) 1934 pr_err("Boot memory [%llx %llx] is offline, can be removed\n", 1935 addr, addr + (1UL << PA_SECTION_SHIFT)); 1936 } 1937 } 1938 } 1939 1940 static int __init prevent_bootmem_remove_init(void) 1941 { 1942 int ret = 0; 1943 1944 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) 1945 return ret; 1946 1947 validate_bootmem_online(); 1948 ret = register_memory_notifier(&prevent_bootmem_remove_nb); 1949 if (ret) 1950 pr_err("%s: Notifier registration failed %d\n", __func__, ret); 1951 1952 return ret; 1953 } 1954 early_initcall(prevent_bootmem_remove_init); 1955 #endif 1956 1957 pte_t modify_prot_start_ptes(struct vm_area_struct *vma, unsigned long addr, 1958 pte_t *ptep, unsigned int nr) 1959 { 1960 pte_t pte = get_and_clear_ptes(vma->vm_mm, addr, ptep, nr); 1961 1962 if (alternative_has_cap_unlikely(ARM64_WORKAROUND_2645198)) { 1963 /* 1964 * Break-before-make (BBM) is required for all user space mappings 1965 * when the permission changes from executable to non-executable 1966 * in cases where cpu is affected with errata #2645198. 1967 */ 1968 if (pte_accessible(vma->vm_mm, pte) && pte_user_exec(pte)) 1969 __flush_tlb_range(vma, addr, nr * PAGE_SIZE, 1970 PAGE_SIZE, true, 3); 1971 } 1972 1973 return pte; 1974 } 1975 1976 pte_t ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep) 1977 { 1978 return modify_prot_start_ptes(vma, addr, ptep, 1); 1979 } 1980 1981 void modify_prot_commit_ptes(struct vm_area_struct *vma, unsigned long addr, 1982 pte_t *ptep, pte_t old_pte, pte_t pte, 1983 unsigned int nr) 1984 { 1985 set_ptes(vma->vm_mm, addr, ptep, pte, nr); 1986 } 1987 1988 void ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, 1989 pte_t old_pte, pte_t pte) 1990 { 1991 modify_prot_commit_ptes(vma, addr, ptep, old_pte, pte, 1); 1992 } 1993 1994 /* 1995 * Atomically replaces the active TTBR1_EL1 PGD with a new VA-compatible PGD, 1996 * avoiding the possibility of conflicting TLB entries being allocated. 1997 */ 1998 void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp) 1999 { 2000 typedef void (ttbr_replace_func)(phys_addr_t); 2001 extern ttbr_replace_func idmap_cpu_replace_ttbr1; 2002 ttbr_replace_func *replace_phys; 2003 unsigned long daif; 2004 2005 /* phys_to_ttbr() zeros lower 2 bits of ttbr with 52-bit PA */ 2006 phys_addr_t ttbr1 = phys_to_ttbr(virt_to_phys(pgdp)); 2007 2008 if (cnp) 2009 ttbr1 |= TTBR_CNP_BIT; 2010 2011 replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1); 2012 2013 cpu_install_idmap(); 2014 2015 /* 2016 * We really don't want to take *any* exceptions while TTBR1 is 2017 * in the process of being replaced so mask everything. 2018 */ 2019 daif = local_daif_save(); 2020 replace_phys(ttbr1); 2021 local_daif_restore(daif); 2022 2023 cpu_uninstall_idmap(); 2024 } 2025 2026 #ifdef CONFIG_ARCH_HAS_PKEYS 2027 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey, unsigned long init_val) 2028 { 2029 u64 new_por; 2030 u64 old_por; 2031 2032 if (!system_supports_poe()) 2033 return -ENOSPC; 2034 2035 /* 2036 * This code should only be called with valid 'pkey' 2037 * values originating from in-kernel users. Complain 2038 * if a bad value is observed. 2039 */ 2040 if (WARN_ON_ONCE(pkey >= arch_max_pkey())) 2041 return -EINVAL; 2042 2043 /* Set the bits we need in POR: */ 2044 new_por = POE_RWX; 2045 if (init_val & PKEY_DISABLE_WRITE) 2046 new_por &= ~POE_W; 2047 if (init_val & PKEY_DISABLE_ACCESS) 2048 new_por &= ~POE_RW; 2049 if (init_val & PKEY_DISABLE_READ) 2050 new_por &= ~POE_R; 2051 if (init_val & PKEY_DISABLE_EXECUTE) 2052 new_por &= ~POE_X; 2053 2054 /* Shift the bits in to the correct place in POR for pkey: */ 2055 new_por = POR_ELx_PERM_PREP(pkey, new_por); 2056 2057 /* Get old POR and mask off any old bits in place: */ 2058 old_por = read_sysreg_s(SYS_POR_EL0); 2059 old_por &= ~(POE_MASK << POR_ELx_PERM_SHIFT(pkey)); 2060 2061 /* Write old part along with new part: */ 2062 write_sysreg_s(old_por | new_por, SYS_POR_EL0); 2063 2064 return 0; 2065 } 2066 #endif 2067