FILE
mmu.c
File preview with syntax highlighting.
linux-6.19.14/arch/arm64/mm/mmu.c
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 52DEFINE_STATIC_KEY_FALSE(arm64_ptdump_lock_key); 53 54u64 kimage_voffset __ro_after_init; 55EXPORT_SYMBOL(kimage_voffset); 56 57u32 __boot_cpu_mode[] = { BOOT_CPU_MODE_EL2, BOOT_CPU_MODE_EL1 }; 58 59static 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 */ 65long __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 */ 71unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss; 72EXPORT_SYMBOL(empty_zero_page); 73 74static DEFINE_SPINLOCK(swapper_pgdir_lock); 75static DEFINE_MUTEX(fixmap_lock); 76 77void 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 104pgprot_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} 113EXPORT_SYMBOL(phys_mem_access_prot); 114 115static 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 127bool 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 166static 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 174static 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 197static int alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr, 198 unsigned long end, phys_addr_t phys, 199 pgprot_t prot, 200 phys_addr_t (*pgtable_alloc)(enum pgtable_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 if (pte_phys == INVALID_PHYS_ADDR) 217 return -ENOMEM; 218 ptep = pte_set_fixmap(pte_phys); 219 init_clear_pgtable(ptep); 220 ptep += pte_index(addr); 221 __pmd_populate(pmdp, pte_phys, pmdval); 222 } else { 223 BUG_ON(pmd_bad(pmd)); 224 ptep = pte_set_fixmap_offset(pmdp, addr); 225 } 226 227 do { 228 pgprot_t __prot = prot; 229 230 next = pte_cont_addr_end(addr, end); 231 232 /* use a contiguous mapping if the range is suitably aligned */ 233 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) && 234 (flags & NO_CONT_MAPPINGS) == 0) 235 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 236 237 init_pte(ptep, addr, next, phys, __prot); 238 239 ptep += pte_index(next) - pte_index(addr); 240 phys += next - addr; 241 } while (addr = next, addr != end); 242 243 /* 244 * Note: barriers and maintenance necessary to clear the fixmap slot 245 * ensure that all previous pgtable writes are visible to the table 246 * walker. 247 */ 248 pte_clear_fixmap(); 249 250 return 0; 251} 252 253static int init_pmd(pmd_t *pmdp, unsigned long addr, unsigned long end, 254 phys_addr_t phys, pgprot_t prot, 255 phys_addr_t (*pgtable_alloc)(enum pgtable_type), int flags) 256{ 257 unsigned long next; 258 259 do { 260 pmd_t old_pmd = READ_ONCE(*pmdp); 261 262 next = pmd_addr_end(addr, end); 263 264 /* try section mapping first */ 265 if (((addr | next | phys) & ~PMD_MASK) == 0 && 266 (flags & NO_BLOCK_MAPPINGS) == 0) { 267 pmd_set_huge(pmdp, phys, prot); 268 269 /* 270 * After the PMD entry has been populated once, we 271 * only allow updates to the permission attributes. 272 */ 273 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd), 274 READ_ONCE(pmd_val(*pmdp)))); 275 } else { 276 int ret; 277 278 ret = alloc_init_cont_pte(pmdp, addr, next, phys, prot, 279 pgtable_alloc, flags); 280 if (ret) 281 return ret; 282 283 BUG_ON(pmd_val(old_pmd) != 0 && 284 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp))); 285 } 286 phys += next - addr; 287 } while (pmdp++, addr = next, addr != end); 288 289 return 0; 290} 291 292static int alloc_init_cont_pmd(pud_t *pudp, unsigned long addr, 293 unsigned long end, phys_addr_t phys, 294 pgprot_t prot, 295 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 296 int flags) 297{ 298 int ret; 299 unsigned long next; 300 pud_t pud = READ_ONCE(*pudp); 301 pmd_t *pmdp; 302 303 /* 304 * Check for initial section mappings in the pgd/pud. 305 */ 306 BUG_ON(pud_sect(pud)); 307 if (pud_none(pud)) { 308 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF; 309 phys_addr_t pmd_phys; 310 311 if (flags & NO_EXEC_MAPPINGS) 312 pudval |= PUD_TABLE_PXN; 313 BUG_ON(!pgtable_alloc); 314 pmd_phys = pgtable_alloc(TABLE_PMD); 315 if (pmd_phys == INVALID_PHYS_ADDR) 316 return -ENOMEM; 317 pmdp = pmd_set_fixmap(pmd_phys); 318 init_clear_pgtable(pmdp); 319 pmdp += pmd_index(addr); 320 __pud_populate(pudp, pmd_phys, pudval); 321 } else { 322 BUG_ON(pud_bad(pud)); 323 pmdp = pmd_set_fixmap_offset(pudp, addr); 324 } 325 326 do { 327 pgprot_t __prot = prot; 328 329 next = pmd_cont_addr_end(addr, end); 330 331 /* use a contiguous mapping if the range is suitably aligned */ 332 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) && 333 (flags & NO_CONT_MAPPINGS) == 0) 334 __prot = __pgprot(pgprot_val(prot) | PTE_CONT); 335 336 ret = init_pmd(pmdp, addr, next, phys, __prot, pgtable_alloc, flags); 337 if (ret) 338 goto out; 339 340 pmdp += pmd_index(next) - pmd_index(addr); 341 phys += next - addr; 342 } while (addr = next, addr != end); 343 344out: 345 pmd_clear_fixmap(); 346 347 return ret; 348} 349 350static int alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end, 351 phys_addr_t phys, pgprot_t prot, 352 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 353 int flags) 354{ 355 int ret = 0; 356 unsigned long next; 357 p4d_t p4d = READ_ONCE(*p4dp); 358 pud_t *pudp; 359 360 if (p4d_none(p4d)) { 361 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN | P4D_TABLE_AF; 362 phys_addr_t pud_phys; 363 364 if (flags & NO_EXEC_MAPPINGS) 365 p4dval |= P4D_TABLE_PXN; 366 BUG_ON(!pgtable_alloc); 367 pud_phys = pgtable_alloc(TABLE_PUD); 368 if (pud_phys == INVALID_PHYS_ADDR) 369 return -ENOMEM; 370 pudp = pud_set_fixmap(pud_phys); 371 init_clear_pgtable(pudp); 372 pudp += pud_index(addr); 373 __p4d_populate(p4dp, pud_phys, p4dval); 374 } else { 375 BUG_ON(p4d_bad(p4d)); 376 pudp = pud_set_fixmap_offset(p4dp, addr); 377 } 378 379 do { 380 pud_t old_pud = READ_ONCE(*pudp); 381 382 next = pud_addr_end(addr, end); 383 384 /* 385 * For 4K granule only, attempt to put down a 1GB block 386 */ 387 if (pud_sect_supported() && 388 ((addr | next | phys) & ~PUD_MASK) == 0 && 389 (flags & NO_BLOCK_MAPPINGS) == 0) { 390 pud_set_huge(pudp, phys, prot); 391 392 /* 393 * After the PUD entry has been populated once, we 394 * only allow updates to the permission attributes. 395 */ 396 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud), 397 READ_ONCE(pud_val(*pudp)))); 398 } else { 399 ret = alloc_init_cont_pmd(pudp, addr, next, phys, prot, 400 pgtable_alloc, flags); 401 if (ret) 402 goto out; 403 404 BUG_ON(pud_val(old_pud) != 0 && 405 pud_val(old_pud) != READ_ONCE(pud_val(*pudp))); 406 } 407 phys += next - addr; 408 } while (pudp++, addr = next, addr != end); 409 410out: 411 pud_clear_fixmap(); 412 413 return ret; 414} 415 416static int alloc_init_p4d(pgd_t *pgdp, unsigned long addr, unsigned long end, 417 phys_addr_t phys, pgprot_t prot, 418 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 419 int flags) 420{ 421 int ret; 422 unsigned long next; 423 pgd_t pgd = READ_ONCE(*pgdp); 424 p4d_t *p4dp; 425 426 if (pgd_none(pgd)) { 427 pgdval_t pgdval = PGD_TYPE_TABLE | PGD_TABLE_UXN | PGD_TABLE_AF; 428 phys_addr_t p4d_phys; 429 430 if (flags & NO_EXEC_MAPPINGS) 431 pgdval |= PGD_TABLE_PXN; 432 BUG_ON(!pgtable_alloc); 433 p4d_phys = pgtable_alloc(TABLE_P4D); 434 if (p4d_phys == INVALID_PHYS_ADDR) 435 return -ENOMEM; 436 p4dp = p4d_set_fixmap(p4d_phys); 437 init_clear_pgtable(p4dp); 438 p4dp += p4d_index(addr); 439 __pgd_populate(pgdp, p4d_phys, pgdval); 440 } else { 441 BUG_ON(pgd_bad(pgd)); 442 p4dp = p4d_set_fixmap_offset(pgdp, addr); 443 } 444 445 do { 446 p4d_t old_p4d = READ_ONCE(*p4dp); 447 448 next = p4d_addr_end(addr, end); 449 450 ret = alloc_init_pud(p4dp, addr, next, phys, prot, 451 pgtable_alloc, flags); 452 if (ret) 453 goto out; 454 455 BUG_ON(p4d_val(old_p4d) != 0 && 456 p4d_val(old_p4d) != READ_ONCE(p4d_val(*p4dp))); 457 458 phys += next - addr; 459 } while (p4dp++, addr = next, addr != end); 460 461out: 462 p4d_clear_fixmap(); 463 464 return ret; 465} 466 467static int __create_pgd_mapping_locked(pgd_t *pgdir, phys_addr_t phys, 468 unsigned long virt, phys_addr_t size, 469 pgprot_t prot, 470 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 471 int flags) 472{ 473 int ret; 474 unsigned long addr, end, next; 475 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt); 476 477 /* 478 * If the virtual and physical address don't have the same offset 479 * within a page, we cannot map the region as the caller expects. 480 */ 481 if (WARN_ON((phys ^ virt) & ~PAGE_MASK)) 482 return -EINVAL; 483 484 phys &= PAGE_MASK; 485 addr = virt & PAGE_MASK; 486 end = PAGE_ALIGN(virt + size); 487 488 do { 489 next = pgd_addr_end(addr, end); 490 ret = alloc_init_p4d(pgdp, addr, next, phys, prot, pgtable_alloc, 491 flags); 492 if (ret) 493 return ret; 494 phys += next - addr; 495 } while (pgdp++, addr = next, addr != end); 496 497 return 0; 498} 499 500static int __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, 501 unsigned long virt, phys_addr_t size, 502 pgprot_t prot, 503 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 504 int flags) 505{ 506 int ret; 507 508 mutex_lock(&fixmap_lock); 509 ret = __create_pgd_mapping_locked(pgdir, phys, virt, size, prot, 510 pgtable_alloc, flags); 511 mutex_unlock(&fixmap_lock); 512 513 return ret; 514} 515 516static void early_create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, 517 unsigned long virt, phys_addr_t size, 518 pgprot_t prot, 519 phys_addr_t (*pgtable_alloc)(enum pgtable_type), 520 int flags) 521{ 522 int ret; 523 524 ret = __create_pgd_mapping(pgdir, phys, virt, size, prot, pgtable_alloc, 525 flags); 526 if (ret) 527 panic("Failed to create page tables\n"); 528} 529 530static phys_addr_t __pgd_pgtable_alloc(struct mm_struct *mm, gfp_t gfp, 531 enum pgtable_type pgtable_type) 532{ 533 /* Page is zeroed by init_clear_pgtable() so don't duplicate effort. */ 534 struct ptdesc *ptdesc = pagetable_alloc(gfp & ~__GFP_ZERO, 0); 535 phys_addr_t pa; 536 537 if (!ptdesc) 538 return INVALID_PHYS_ADDR; 539 540 pa = page_to_phys(ptdesc_page(ptdesc)); 541 542 switch (pgtable_type) { 543 case TABLE_PTE: 544 BUG_ON(!pagetable_pte_ctor(mm, ptdesc)); 545 break; 546 case TABLE_PMD: 547 BUG_ON(!pagetable_pmd_ctor(mm, ptdesc)); 548 break; 549 case TABLE_PUD: 550 pagetable_pud_ctor(ptdesc); 551 break; 552 case TABLE_P4D: 553 pagetable_p4d_ctor(ptdesc); 554 break; 555 } 556 557 return pa; 558} 559 560static phys_addr_t 561pgd_pgtable_alloc_init_mm_gfp(enum pgtable_type pgtable_type, gfp_t gfp) 562{ 563 return __pgd_pgtable_alloc(&init_mm, gfp, pgtable_type); 564} 565 566static phys_addr_t __maybe_unused 567pgd_pgtable_alloc_init_mm(enum pgtable_type pgtable_type) 568{ 569 return pgd_pgtable_alloc_init_mm_gfp(pgtable_type, GFP_PGTABLE_KERNEL); 570} 571 572static phys_addr_t 573pgd_pgtable_alloc_special_mm(enum pgtable_type pgtable_type) 574{ 575 return __pgd_pgtable_alloc(NULL, GFP_PGTABLE_KERNEL, pgtable_type); 576} 577 578static void split_contpte(pte_t *ptep) 579{ 580 int i; 581 582 ptep = PTR_ALIGN_DOWN(ptep, sizeof(*ptep) * CONT_PTES); 583 for (i = 0; i < CONT_PTES; i++, ptep++) 584 __set_pte(ptep, pte_mknoncont(__ptep_get(ptep))); 585} 586 587static int split_pmd(pmd_t *pmdp, pmd_t pmd, gfp_t gfp, bool to_cont) 588{ 589 pmdval_t tableprot = PMD_TYPE_TABLE | PMD_TABLE_UXN | PMD_TABLE_AF; 590 unsigned long pfn = pmd_pfn(pmd); 591 pgprot_t prot = pmd_pgprot(pmd); 592 phys_addr_t pte_phys; 593 pte_t *ptep; 594 int i; 595 596 pte_phys = pgd_pgtable_alloc_init_mm_gfp(TABLE_PTE, gfp); 597 if (pte_phys == INVALID_PHYS_ADDR) 598 return -ENOMEM; 599 ptep = (pte_t *)phys_to_virt(pte_phys); 600 601 if (pgprot_val(prot) & PMD_SECT_PXN) 602 tableprot |= PMD_TABLE_PXN; 603 604 prot = __pgprot((pgprot_val(prot) & ~PTE_TYPE_MASK) | PTE_TYPE_PAGE); 605 if (!pmd_valid(pmd)) 606 prot = pte_pgprot(pte_mkinvalid(pfn_pte(0, prot))); 607 prot = __pgprot(pgprot_val(prot) & ~PTE_CONT); 608 if (to_cont) 609 prot = __pgprot(pgprot_val(prot) | PTE_CONT); 610 611 for (i = 0; i < PTRS_PER_PTE; i++, ptep++, pfn++) 612 __set_pte(ptep, pfn_pte(pfn, prot)); 613 614 /* 615 * Ensure the pte entries are visible to the table walker by the time 616 * the pmd entry that points to the ptes is visible. 617 */ 618 dsb(ishst); 619 __pmd_populate(pmdp, pte_phys, tableprot); 620 621 return 0; 622} 623 624static void split_contpmd(pmd_t *pmdp) 625{ 626 int i; 627 628 pmdp = PTR_ALIGN_DOWN(pmdp, sizeof(*pmdp) * CONT_PMDS); 629 for (i = 0; i < CONT_PMDS; i++, pmdp++) 630 set_pmd(pmdp, pmd_mknoncont(pmdp_get(pmdp))); 631} 632 633static int split_pud(pud_t *pudp, pud_t pud, gfp_t gfp, bool to_cont) 634{ 635 pudval_t tableprot = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF; 636 unsigned int step = PMD_SIZE >> PAGE_SHIFT; 637 unsigned long pfn = pud_pfn(pud); 638 pgprot_t prot = pud_pgprot(pud); 639 phys_addr_t pmd_phys; 640 pmd_t *pmdp; 641 int i; 642 643 pmd_phys = pgd_pgtable_alloc_init_mm_gfp(TABLE_PMD, gfp); 644 if (pmd_phys == INVALID_PHYS_ADDR) 645 return -ENOMEM; 646 pmdp = (pmd_t *)phys_to_virt(pmd_phys); 647 648 if (pgprot_val(prot) & PMD_SECT_PXN) 649 tableprot |= PUD_TABLE_PXN; 650 651 prot = __pgprot((pgprot_val(prot) & ~PMD_TYPE_MASK) | PMD_TYPE_SECT); 652 if (!pud_valid(pud)) 653 prot = pmd_pgprot(pmd_mkinvalid(pfn_pmd(0, prot))); 654 prot = __pgprot(pgprot_val(prot) & ~PTE_CONT); 655 if (to_cont) 656 prot = __pgprot(pgprot_val(prot) | PTE_CONT); 657 658 for (i = 0; i < PTRS_PER_PMD; i++, pmdp++, pfn += step) 659 set_pmd(pmdp, pfn_pmd(pfn, prot)); 660 661 /* 662 * Ensure the pmd entries are visible to the table walker by the time 663 * the pud entry that points to the pmds is visible. 664 */ 665 dsb(ishst); 666 __pud_populate(pudp, pmd_phys, tableprot); 667 668 return 0; 669} 670 671static int split_kernel_leaf_mapping_locked(unsigned long addr) 672{ 673 pgd_t *pgdp, pgd; 674 p4d_t *p4dp, p4d; 675 pud_t *pudp, pud; 676 pmd_t *pmdp, pmd; 677 pte_t *ptep, pte; 678 int ret = 0; 679 680 /* 681 * PGD: If addr is PGD aligned then addr already describes a leaf 682 * boundary. If not present then there is nothing to split. 683 */ 684 if (ALIGN_DOWN(addr, PGDIR_SIZE) == addr) 685 goto out; 686 pgdp = pgd_offset_k(addr); 687 pgd = pgdp_get(pgdp); 688 if (!pgd_present(pgd)) 689 goto out; 690 691 /* 692 * P4D: If addr is P4D aligned then addr already describes a leaf 693 * boundary. If not present then there is nothing to split. 694 */ 695 if (ALIGN_DOWN(addr, P4D_SIZE) == addr) 696 goto out; 697 p4dp = p4d_offset(pgdp, addr); 698 p4d = p4dp_get(p4dp); 699 if (!p4d_present(p4d)) 700 goto out; 701 702 /* 703 * PUD: If addr is PUD aligned then addr already describes a leaf 704 * boundary. If not present then there is nothing to split. Otherwise, 705 * if we have a pud leaf, split to contpmd. 706 */ 707 if (ALIGN_DOWN(addr, PUD_SIZE) == addr) 708 goto out; 709 pudp = pud_offset(p4dp, addr); 710 pud = pudp_get(pudp); 711 if (!pud_present(pud)) 712 goto out; 713 if (pud_leaf(pud)) { 714 ret = split_pud(pudp, pud, GFP_PGTABLE_KERNEL, true); 715 if (ret) 716 goto out; 717 } 718 719 /* 720 * CONTPMD: If addr is CONTPMD aligned then addr already describes a 721 * leaf boundary. If not present then there is nothing to split. 722 * Otherwise, if we have a contpmd leaf, split to pmd. 723 */ 724 if (ALIGN_DOWN(addr, CONT_PMD_SIZE) == addr) 725 goto out; 726 pmdp = pmd_offset(pudp, addr); 727 pmd = pmdp_get(pmdp); 728 if (!pmd_present(pmd)) 729 goto out; 730 if (pmd_leaf(pmd)) { 731 if (pmd_cont(pmd)) 732 split_contpmd(pmdp); 733 /* 734 * PMD: If addr is PMD aligned then addr already describes a 735 * leaf boundary. Otherwise, split to contpte. 736 */ 737 if (ALIGN_DOWN(addr, PMD_SIZE) == addr) 738 goto out; 739 ret = split_pmd(pmdp, pmd, GFP_PGTABLE_KERNEL, true); 740 if (ret) 741 goto out; 742 } 743 744 /* 745 * CONTPTE: If addr is CONTPTE aligned then addr already describes a 746 * leaf boundary. If not present then there is nothing to split. 747 * Otherwise, if we have a contpte leaf, split to pte. 748 */ 749 if (ALIGN_DOWN(addr, CONT_PTE_SIZE) == addr) 750 goto out; 751 ptep = pte_offset_kernel(pmdp, addr); 752 pte = __ptep_get(ptep); 753 if (!pte_present(pte)) 754 goto out; 755 if (pte_cont(pte)) 756 split_contpte(ptep); 757 758out: 759 return ret; 760} 761 762static inline bool force_pte_mapping(void) 763{ 764 const bool bbml2 = system_capabilities_finalized() ? 765 system_supports_bbml2_noabort() : cpu_supports_bbml2_noabort(); 766 767 if (debug_pagealloc_enabled()) 768 return true; 769 if (bbml2) 770 return false; 771 return rodata_full || arm64_kfence_can_set_direct_map() || is_realm_world(); 772} 773 774static DEFINE_MUTEX(pgtable_split_lock); 775 776int split_kernel_leaf_mapping(unsigned long start, unsigned long end) 777{ 778 int ret; 779 780 /* 781 * !BBML2_NOABORT systems should not be trying to change permissions on 782 * anything that is not pte-mapped in the first place. Just return early 783 * and let the permission change code raise a warning if not already 784 * pte-mapped. 785 */ 786 if (!system_supports_bbml2_noabort()) 787 return 0; 788 789 /* 790 * If the region is within a pte-mapped area, there is no need to try to 791 * split. Additionally, CONFIG_DEBUG_PAGEALLOC and CONFIG_KFENCE may 792 * change permissions from atomic context so for those cases (which are 793 * always pte-mapped), we must not go any further because taking the 794 * mutex below may sleep. 795 */ 796 if (force_pte_mapping() || is_kfence_address((void *)start)) 797 return 0; 798 799 /* 800 * Ensure start and end are at least page-aligned since this is the 801 * finest granularity we can split to. 802 */ 803 if (start != PAGE_ALIGN(start) || end != PAGE_ALIGN(end)) 804 return -EINVAL; 805 806 mutex_lock(&pgtable_split_lock); 807 arch_enter_lazy_mmu_mode(); 808 809 /* 810 * The split_kernel_leaf_mapping_locked() may sleep, it is not a 811 * problem for ARM64 since ARM64's lazy MMU implementation allows 812 * sleeping. 813 * 814 * Optimize for the common case of splitting out a single page from a 815 * larger mapping. Here we can just split on the "least aligned" of 816 * start and end and this will guarantee that there must also be a split 817 * on the more aligned address since the both addresses must be in the 818 * same contpte block and it must have been split to ptes. 819 */ 820 if (end - start == PAGE_SIZE) { 821 start = __ffs(start) < __ffs(end) ? start : end; 822 ret = split_kernel_leaf_mapping_locked(start); 823 } else { 824 ret = split_kernel_leaf_mapping_locked(start); 825 if (!ret) 826 ret = split_kernel_leaf_mapping_locked(end); 827 } 828 829 arch_leave_lazy_mmu_mode(); 830 mutex_unlock(&pgtable_split_lock); 831 return ret; 832} 833 834static int split_to_ptes_pud_entry(pud_t *pudp, unsigned long addr, 835 unsigned long next, struct mm_walk *walk) 836{ 837 gfp_t gfp = *(gfp_t *)walk->private; 838 pud_t pud = pudp_get(pudp); 839 int ret = 0; 840 841 if (pud_leaf(pud)) 842 ret = split_pud(pudp, pud, gfp, false); 843 844 return ret; 845} 846 847static int split_to_ptes_pmd_entry(pmd_t *pmdp, unsigned long addr, 848 unsigned long next, struct mm_walk *walk) 849{ 850 gfp_t gfp = *(gfp_t *)walk->private; 851 pmd_t pmd = pmdp_get(pmdp); 852 int ret = 0; 853 854 if (pmd_leaf(pmd)) { 855 if (pmd_cont(pmd)) 856 split_contpmd(pmdp); 857 ret = split_pmd(pmdp, pmd, gfp, false); 858 859 /* 860 * We have split the pmd directly to ptes so there is no need to 861 * visit each pte to check if they are contpte. 862 */ 863 walk->action = ACTION_CONTINUE; 864 } 865 866 return ret; 867} 868 869static int split_to_ptes_pte_entry(pte_t *ptep, unsigned long addr, 870 unsigned long next, struct mm_walk *walk) 871{ 872 pte_t pte = __ptep_get(ptep); 873 874 if (pte_cont(pte)) 875 split_contpte(ptep); 876 877 return 0; 878} 879 880static const struct mm_walk_ops split_to_ptes_ops = { 881 .pud_entry = split_to_ptes_pud_entry, 882 .pmd_entry = split_to_ptes_pmd_entry, 883 .pte_entry = split_to_ptes_pte_entry, 884}; 885 886static int range_split_to_ptes(unsigned long start, unsigned long end, gfp_t gfp) 887{ 888 int ret; 889 890 arch_enter_lazy_mmu_mode(); 891 ret = walk_kernel_page_table_range_lockless(start, end, 892 &split_to_ptes_ops, NULL, &gfp); 893 arch_leave_lazy_mmu_mode(); 894 895 return ret; 896} 897 898static bool linear_map_requires_bbml2 __initdata; 899 900u32 idmap_kpti_bbml2_flag; 901 902static void __init init_idmap_kpti_bbml2_flag(void) 903{ 904 WRITE_ONCE(idmap_kpti_bbml2_flag, 1); 905 /* Must be visible to other CPUs before stop_machine() is called. */ 906 smp_mb(); 907} 908 909static int __init linear_map_split_to_ptes(void *__unused) 910{ 911 /* 912 * Repainting the linear map must be done by CPU0 (the boot CPU) because 913 * that's the only CPU that we know supports BBML2. The other CPUs will 914 * be held in a waiting area with the idmap active. 915 */ 916 if (!smp_processor_id()) { 917 unsigned long lstart = _PAGE_OFFSET(vabits_actual); 918 unsigned long lend = PAGE_END; 919 unsigned long kstart = (unsigned long)lm_alias(_stext); 920 unsigned long kend = (unsigned long)lm_alias(__init_begin); 921 int ret; 922 923 /* 924 * Wait for all secondary CPUs to be put into the waiting area. 925 */ 926 smp_cond_load_acquire(&idmap_kpti_bbml2_flag, VAL == num_online_cpus()); 927 928 /* 929 * Walk all of the linear map [lstart, lend), except the kernel 930 * linear map alias [kstart, kend), and split all mappings to 931 * PTE. The kernel alias remains static throughout runtime so 932 * can continue to be safely mapped with large mappings. 933 */ 934 ret = range_split_to_ptes(lstart, kstart, GFP_ATOMIC); 935 if (!ret) 936 ret = range_split_to_ptes(kend, lend, GFP_ATOMIC); 937 if (ret) 938 panic("Failed to split linear map\n"); 939 flush_tlb_kernel_range(lstart, lend); 940 941 /* 942 * Relies on dsb in flush_tlb_kernel_range() to avoid reordering 943 * before any page table split operations. 944 */ 945 WRITE_ONCE(idmap_kpti_bbml2_flag, 0); 946 } else { 947 typedef void (wait_split_fn)(void); 948 extern wait_split_fn wait_linear_map_split_to_ptes; 949 wait_split_fn *wait_fn; 950 951 wait_fn = (void *)__pa_symbol(wait_linear_map_split_to_ptes); 952 953 /* 954 * At least one secondary CPU doesn't support BBML2 so cannot 955 * tolerate the size of the live mappings changing. So have the 956 * secondary CPUs wait for the boot CPU to make the changes 957 * with the idmap active and init_mm inactive. 958 */ 959 cpu_install_idmap(); 960 wait_fn(); 961 cpu_uninstall_idmap(); 962 } 963 964 return 0; 965} 966 967void __init linear_map_maybe_split_to_ptes(void) 968{ 969 if (linear_map_requires_bbml2 && !system_supports_bbml2_noabort()) { 970 init_idmap_kpti_bbml2_flag(); 971 stop_machine(linear_map_split_to_ptes, NULL, cpu_online_mask); 972 } 973} 974 975/* 976 * This function can only be used to modify existing table entries, 977 * without allocating new levels of table. Note that this permits the 978 * creation of new section or page entries. 979 */ 980void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt, 981 phys_addr_t size, pgprot_t prot) 982{ 983 if (virt < PAGE_OFFSET) { 984 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n", 985 &phys, virt); 986 return; 987 } 988 early_create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 989 NO_CONT_MAPPINGS); 990} 991 992void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys, 993 unsigned long virt, phys_addr_t size, 994 pgprot_t prot, bool page_mappings_only) 995{ 996 int flags = 0; 997 998 BUG_ON(mm == &init_mm); 999 1000 if (page_mappings_only) 1001 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1002 1003 early_create_pgd_mapping(mm->pgd, phys, virt, size, prot, 1004 pgd_pgtable_alloc_special_mm, flags); 1005} 1006 1007static void update_mapping_prot(phys_addr_t phys, unsigned long virt, 1008 phys_addr_t size, pgprot_t prot) 1009{ 1010 if (virt < PAGE_OFFSET) { 1011 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n", 1012 &phys, virt); 1013 return; 1014 } 1015 1016 early_create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL, 1017 NO_CONT_MAPPINGS); 1018 1019 /* flush the TLBs after updating live kernel mappings */ 1020 flush_tlb_kernel_range(virt, virt + size); 1021} 1022 1023static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start, 1024 phys_addr_t end, pgprot_t prot, int flags) 1025{ 1026 early_create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start, 1027 prot, early_pgtable_alloc, flags); 1028} 1029 1030void __init mark_linear_text_alias_ro(void) 1031{ 1032 /* 1033 * Remove the write permissions from the linear alias of .text/.rodata 1034 */ 1035 update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text), 1036 (unsigned long)__init_begin - (unsigned long)_text, 1037 PAGE_KERNEL_RO); 1038} 1039 1040#ifdef CONFIG_KFENCE 1041 1042bool __ro_after_init kfence_early_init = !!CONFIG_KFENCE_SAMPLE_INTERVAL; 1043 1044/* early_param() will be parsed before map_mem() below. */ 1045static int __init parse_kfence_early_init(char *arg) 1046{ 1047 int val; 1048 1049 if (get_option(&arg, &val)) 1050 kfence_early_init = !!val; 1051 return 0; 1052} 1053early_param("kfence.sample_interval", parse_kfence_early_init); 1054 1055static phys_addr_t __init arm64_kfence_alloc_pool(void) 1056{ 1057 phys_addr_t kfence_pool; 1058 1059 if (!kfence_early_init) 1060 return 0; 1061 1062 kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE); 1063 if (!kfence_pool) { 1064 pr_err("failed to allocate kfence pool\n"); 1065 kfence_early_init = false; 1066 return 0; 1067 } 1068 1069 /* Temporarily mark as NOMAP. */ 1070 memblock_mark_nomap(kfence_pool, KFENCE_POOL_SIZE); 1071 1072 return kfence_pool; 1073} 1074 1075static void __init arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) 1076{ 1077 if (!kfence_pool) 1078 return; 1079 1080 /* KFENCE pool needs page-level mapping. */ 1081 __map_memblock(pgdp, kfence_pool, kfence_pool + KFENCE_POOL_SIZE, 1082 pgprot_tagged(PAGE_KERNEL), 1083 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS); 1084 memblock_clear_nomap(kfence_pool, KFENCE_POOL_SIZE); 1085 __kfence_pool = phys_to_virt(kfence_pool); 1086} 1087 1088bool arch_kfence_init_pool(void) 1089{ 1090 unsigned long start = (unsigned long)__kfence_pool; 1091 unsigned long end = start + KFENCE_POOL_SIZE; 1092 int ret; 1093 1094 /* Exit early if we know the linear map is already pte-mapped. */ 1095 if (force_pte_mapping()) 1096 return true; 1097 1098 /* Kfence pool is already pte-mapped for the early init case. */ 1099 if (kfence_early_init) 1100 return true; 1101 1102 mutex_lock(&pgtable_split_lock); 1103 ret = range_split_to_ptes(start, end, GFP_PGTABLE_KERNEL); 1104 mutex_unlock(&pgtable_split_lock); 1105 1106 /* 1107 * Since the system supports bbml2_noabort, tlb invalidation is not 1108 * required here; the pgtable mappings have been split to pte but larger 1109 * entries may safely linger in the TLB. 1110 */ 1111 1112 return !ret; 1113} 1114#else /* CONFIG_KFENCE */ 1115 1116static inline phys_addr_t arm64_kfence_alloc_pool(void) { return 0; } 1117static inline void arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) { } 1118 1119#endif /* CONFIG_KFENCE */ 1120 1121static void __init map_mem(pgd_t *pgdp) 1122{ 1123 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN); 1124 phys_addr_t kernel_start = __pa_symbol(_text); 1125 phys_addr_t kernel_end = __pa_symbol(__init_begin); 1126 phys_addr_t start, end; 1127 phys_addr_t early_kfence_pool; 1128 int flags = NO_EXEC_MAPPINGS; 1129 u64 i; 1130 1131 /* 1132 * Setting hierarchical PXNTable attributes on table entries covering 1133 * the linear region is only possible if it is guaranteed that no table 1134 * entries at any level are being shared between the linear region and 1135 * the vmalloc region. Check whether this is true for the PGD level, in 1136 * which case it is guaranteed to be true for all other levels as well. 1137 * (Unless we are running with support for LPA2, in which case the 1138 * entire reduced VA space is covered by a single pgd_t which will have 1139 * been populated without the PXNTable attribute by the time we get here.) 1140 */ 1141 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end) && 1142 pgd_index(_PAGE_OFFSET(VA_BITS_MIN)) != PTRS_PER_PGD - 1); 1143 1144 early_kfence_pool = arm64_kfence_alloc_pool(); 1145 1146 linear_map_requires_bbml2 = !force_pte_mapping() && can_set_direct_map(); 1147 1148 if (force_pte_mapping()) 1149 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1150 1151 /* 1152 * Take care not to create a writable alias for the 1153 * read-only text and rodata sections of the kernel image. 1154 * So temporarily mark them as NOMAP to skip mappings in 1155 * the following for-loop 1156 */ 1157 memblock_mark_nomap(kernel_start, kernel_end - kernel_start); 1158 1159 /* map all the memory banks */ 1160 for_each_mem_range(i, &start, &end) { 1161 if (start >= end) 1162 break; 1163 /* 1164 * The linear map must allow allocation tags reading/writing 1165 * if MTE is present. Otherwise, it has the same attributes as 1166 * PAGE_KERNEL. 1167 */ 1168 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL), 1169 flags); 1170 } 1171 1172 /* 1173 * Map the linear alias of the [_text, __init_begin) interval 1174 * as non-executable now, and remove the write permission in 1175 * mark_linear_text_alias_ro() below (which will be called after 1176 * alternative patching has completed). This makes the contents 1177 * of the region accessible to subsystems such as hibernate, 1178 * but protects it from inadvertent modification or execution. 1179 * Note that contiguous mappings cannot be remapped in this way, 1180 * so we should avoid them here. 1181 */ 1182 __map_memblock(pgdp, kernel_start, kernel_end, 1183 PAGE_KERNEL, NO_CONT_MAPPINGS); 1184 memblock_clear_nomap(kernel_start, kernel_end - kernel_start); 1185 arm64_kfence_map_pool(early_kfence_pool, pgdp); 1186} 1187 1188void mark_rodata_ro(void) 1189{ 1190 unsigned long section_size; 1191 1192 /* 1193 * mark .rodata as read only. Use __init_begin rather than __end_rodata 1194 * to cover NOTES and EXCEPTION_TABLE. 1195 */ 1196 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata; 1197 WRITE_ONCE(rodata_is_rw, false); 1198 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata, 1199 section_size, PAGE_KERNEL_RO); 1200 /* mark the range between _text and _stext as read only. */ 1201 update_mapping_prot(__pa_symbol(_text), (unsigned long)_text, 1202 (unsigned long)_stext - (unsigned long)_text, 1203 PAGE_KERNEL_RO); 1204} 1205 1206static void __init declare_vma(struct vm_struct *vma, 1207 void *va_start, void *va_end, 1208 unsigned long vm_flags) 1209{ 1210 phys_addr_t pa_start = __pa_symbol(va_start); 1211 unsigned long size = va_end - va_start; 1212 1213 BUG_ON(!PAGE_ALIGNED(pa_start)); 1214 BUG_ON(!PAGE_ALIGNED(size)); 1215 1216 if (!(vm_flags & VM_NO_GUARD)) 1217 size += PAGE_SIZE; 1218 1219 vma->addr = va_start; 1220 vma->phys_addr = pa_start; 1221 vma->size = size; 1222 vma->flags = VM_MAP | vm_flags; 1223 vma->caller = __builtin_return_address(0); 1224 1225 vm_area_add_early(vma); 1226} 1227 1228#ifdef CONFIG_UNMAP_KERNEL_AT_EL0 1229#define KPTI_NG_TEMP_VA (-(1UL << PMD_SHIFT)) 1230 1231static phys_addr_t kpti_ng_temp_alloc __initdata; 1232 1233static phys_addr_t __init kpti_ng_pgd_alloc(enum pgtable_type type) 1234{ 1235 kpti_ng_temp_alloc -= PAGE_SIZE; 1236 return kpti_ng_temp_alloc; 1237} 1238 1239static int __init __kpti_install_ng_mappings(void *__unused) 1240{ 1241 typedef void (kpti_remap_fn)(int, int, phys_addr_t, unsigned long); 1242 extern kpti_remap_fn idmap_kpti_install_ng_mappings; 1243 kpti_remap_fn *remap_fn; 1244 1245 int cpu = smp_processor_id(); 1246 int levels = CONFIG_PGTABLE_LEVELS; 1247 int order = order_base_2(levels); 1248 u64 kpti_ng_temp_pgd_pa = 0; 1249 pgd_t *kpti_ng_temp_pgd; 1250 u64 alloc = 0; 1251 1252 if (levels == 5 && !pgtable_l5_enabled()) 1253 levels = 4; 1254 else if (levels == 4 && !pgtable_l4_enabled()) 1255 levels = 3; 1256 1257 remap_fn = (void *)__pa_symbol(idmap_kpti_install_ng_mappings); 1258 1259 if (!cpu) { 1260 int ret; 1261 1262 alloc = __get_free_pages(GFP_ATOMIC | __GFP_ZERO, order); 1263 kpti_ng_temp_pgd = (pgd_t *)(alloc + (levels - 1) * PAGE_SIZE); 1264 kpti_ng_temp_alloc = kpti_ng_temp_pgd_pa = __pa(kpti_ng_temp_pgd); 1265 1266 // 1267 // Create a minimal page table hierarchy that permits us to map 1268 // the swapper page tables temporarily as we traverse them. 1269 // 1270 // The physical pages are laid out as follows: 1271 // 1272 // +--------+-/-------+-/------ +-/------ +-\\\--------+ 1273 // : PTE[] : | PMD[] : | PUD[] : | P4D[] : ||| PGD[] : 1274 // +--------+-\-------+-\------ +-\------ +-///--------+ 1275 // ^ 1276 // The first page is mapped into this hierarchy at a PMD_SHIFT 1277 // aligned virtual address, so that we can manipulate the PTE 1278 // level entries while the mapping is active. The first entry 1279 // covers the PTE[] page itself, the remaining entries are free 1280 // to be used as a ad-hoc fixmap. 1281 // 1282 ret = __create_pgd_mapping_locked(kpti_ng_temp_pgd, __pa(alloc), 1283 KPTI_NG_TEMP_VA, PAGE_SIZE, PAGE_KERNEL, 1284 kpti_ng_pgd_alloc, 0); 1285 if (ret) 1286 panic("Failed to create page tables\n"); 1287 } 1288 1289 cpu_install_idmap(); 1290 remap_fn(cpu, num_online_cpus(), kpti_ng_temp_pgd_pa, KPTI_NG_TEMP_VA); 1291 cpu_uninstall_idmap(); 1292 1293 if (!cpu) { 1294 free_pages(alloc, order); 1295 arm64_use_ng_mappings = true; 1296 } 1297 1298 return 0; 1299} 1300 1301void __init kpti_install_ng_mappings(void) 1302{ 1303 /* Check whether KPTI is going to be used */ 1304 if (!arm64_kernel_unmapped_at_el0()) 1305 return; 1306 1307 /* 1308 * We don't need to rewrite the page-tables if either we've done 1309 * it already or we have KASLR enabled and therefore have not 1310 * created any global mappings at all. 1311 */ 1312 if (arm64_use_ng_mappings) 1313 return; 1314 1315 init_idmap_kpti_bbml2_flag(); 1316 stop_machine(__kpti_install_ng_mappings, NULL, cpu_online_mask); 1317} 1318 1319static pgprot_t __init kernel_exec_prot(void) 1320{ 1321 return rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC; 1322} 1323 1324static int __init map_entry_trampoline(void) 1325{ 1326 int i; 1327 1328 if (!arm64_kernel_unmapped_at_el0()) 1329 return 0; 1330 1331 pgprot_t prot = kernel_exec_prot(); 1332 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start); 1333 1334 /* The trampoline is always mapped and can therefore be global */ 1335 pgprot_val(prot) &= ~PTE_NG; 1336 1337 /* Map only the text into the trampoline page table */ 1338 memset(tramp_pg_dir, 0, PGD_SIZE); 1339 early_create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, 1340 entry_tramp_text_size(), prot, 1341 pgd_pgtable_alloc_init_mm, NO_BLOCK_MAPPINGS); 1342 1343 /* Map both the text and data into the kernel page table */ 1344 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++) 1345 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 1346 pa_start + i * PAGE_SIZE, prot); 1347 1348 if (IS_ENABLED(CONFIG_RELOCATABLE)) 1349 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i, 1350 pa_start + i * PAGE_SIZE, PAGE_KERNEL_RO); 1351 1352 return 0; 1353} 1354core_initcall(map_entry_trampoline); 1355#endif 1356 1357/* 1358 * Declare the VMA areas for the kernel 1359 */ 1360static void __init declare_kernel_vmas(void) 1361{ 1362 static struct vm_struct vmlinux_seg[KERNEL_SEGMENT_COUNT]; 1363 1364 declare_vma(&vmlinux_seg[0], _text, _etext, VM_NO_GUARD); 1365 declare_vma(&vmlinux_seg[1], __start_rodata, __inittext_begin, VM_NO_GUARD); 1366 declare_vma(&vmlinux_seg[2], __inittext_begin, __inittext_end, VM_NO_GUARD); 1367 declare_vma(&vmlinux_seg[3], __initdata_begin, __initdata_end, VM_NO_GUARD); 1368 declare_vma(&vmlinux_seg[4], _data, _end, 0); 1369} 1370 1371void __pi_map_range(phys_addr_t *pte, u64 start, u64 end, phys_addr_t pa, 1372 pgprot_t prot, int level, pte_t *tbl, bool may_use_cont, 1373 u64 va_offset); 1374 1375static u8 idmap_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init, 1376 kpti_bbml2_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init; 1377 1378static void __init create_idmap(void) 1379{ 1380 phys_addr_t start = __pa_symbol(__idmap_text_start); 1381 phys_addr_t end = __pa_symbol(__idmap_text_end); 1382 phys_addr_t ptep = __pa_symbol(idmap_ptes); 1383 1384 __pi_map_range(&ptep, start, end, start, PAGE_KERNEL_ROX, 1385 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 1386 __phys_to_virt(ptep) - ptep); 1387 1388 if (linear_map_requires_bbml2 || 1389 (IS_ENABLED(CONFIG_UNMAP_KERNEL_AT_EL0) && !arm64_use_ng_mappings)) { 1390 phys_addr_t pa = __pa_symbol(&idmap_kpti_bbml2_flag); 1391 1392 /* 1393 * The KPTI G-to-nG conversion code needs a read-write mapping 1394 * of its synchronization flag in the ID map. This is also used 1395 * when splitting the linear map to ptes if a secondary CPU 1396 * doesn't support bbml2. 1397 */ 1398 ptep = __pa_symbol(kpti_bbml2_ptes); 1399 __pi_map_range(&ptep, pa, pa + sizeof(u32), pa, PAGE_KERNEL, 1400 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false, 1401 __phys_to_virt(ptep) - ptep); 1402 } 1403} 1404 1405void __init paging_init(void) 1406{ 1407 map_mem(swapper_pg_dir); 1408 1409 memblock_allow_resize(); 1410 1411 create_idmap(); 1412 declare_kernel_vmas(); 1413} 1414 1415#ifdef CONFIG_MEMORY_HOTPLUG 1416static void free_hotplug_page_range(struct page *page, size_t size, 1417 struct vmem_altmap *altmap) 1418{ 1419 if (altmap) { 1420 vmem_altmap_free(altmap, size >> PAGE_SHIFT); 1421 } else { 1422 WARN_ON(PageReserved(page)); 1423 __free_pages(page, get_order(size)); 1424 } 1425} 1426 1427static void free_hotplug_pgtable_page(struct page *page) 1428{ 1429 free_hotplug_page_range(page, PAGE_SIZE, NULL); 1430} 1431 1432static bool pgtable_range_aligned(unsigned long start, unsigned long end, 1433 unsigned long floor, unsigned long ceiling, 1434 unsigned long mask) 1435{ 1436 start &= mask; 1437 if (start < floor) 1438 return false; 1439 1440 if (ceiling) { 1441 ceiling &= mask; 1442 if (!ceiling) 1443 return false; 1444 } 1445 1446 if (end - 1 > ceiling - 1) 1447 return false; 1448 return true; 1449} 1450 1451static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr, 1452 unsigned long end, bool free_mapped, 1453 struct vmem_altmap *altmap) 1454{ 1455 pte_t *ptep, pte; 1456 1457 do { 1458 ptep = pte_offset_kernel(pmdp, addr); 1459 pte = __ptep_get(ptep); 1460 if (pte_none(pte)) 1461 continue; 1462 1463 WARN_ON(!pte_present(pte)); 1464 __pte_clear(&init_mm, addr, ptep); 1465 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1466 if (free_mapped) 1467 free_hotplug_page_range(pte_page(pte), 1468 PAGE_SIZE, altmap); 1469 } while (addr += PAGE_SIZE, addr < end); 1470} 1471 1472static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr, 1473 unsigned long end, bool free_mapped, 1474 struct vmem_altmap *altmap) 1475{ 1476 unsigned long next; 1477 pmd_t *pmdp, pmd; 1478 1479 do { 1480 next = pmd_addr_end(addr, end); 1481 pmdp = pmd_offset(pudp, addr); 1482 pmd = READ_ONCE(*pmdp); 1483 if (pmd_none(pmd)) 1484 continue; 1485 1486 WARN_ON(!pmd_present(pmd)); 1487 if (pmd_sect(pmd)) { 1488 pmd_clear(pmdp); 1489 1490 /* 1491 * One TLBI should be sufficient here as the PMD_SIZE 1492 * range is mapped with a single block entry. 1493 */ 1494 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1495 if (free_mapped) 1496 free_hotplug_page_range(pmd_page(pmd), 1497 PMD_SIZE, altmap); 1498 continue; 1499 } 1500 WARN_ON(!pmd_table(pmd)); 1501 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap); 1502 } while (addr = next, addr < end); 1503} 1504 1505static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr, 1506 unsigned long end, bool free_mapped, 1507 struct vmem_altmap *altmap) 1508{ 1509 unsigned long next; 1510 pud_t *pudp, pud; 1511 1512 do { 1513 next = pud_addr_end(addr, end); 1514 pudp = pud_offset(p4dp, addr); 1515 pud = READ_ONCE(*pudp); 1516 if (pud_none(pud)) 1517 continue; 1518 1519 WARN_ON(!pud_present(pud)); 1520 if (pud_sect(pud)) { 1521 pud_clear(pudp); 1522 1523 /* 1524 * One TLBI should be sufficient here as the PUD_SIZE 1525 * range is mapped with a single block entry. 1526 */ 1527 flush_tlb_kernel_range(addr, addr + PAGE_SIZE); 1528 if (free_mapped) 1529 free_hotplug_page_range(pud_page(pud), 1530 PUD_SIZE, altmap); 1531 continue; 1532 } 1533 WARN_ON(!pud_table(pud)); 1534 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap); 1535 } while (addr = next, addr < end); 1536} 1537 1538static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr, 1539 unsigned long end, bool free_mapped, 1540 struct vmem_altmap *altmap) 1541{ 1542 unsigned long next; 1543 p4d_t *p4dp, p4d; 1544 1545 do { 1546 next = p4d_addr_end(addr, end); 1547 p4dp = p4d_offset(pgdp, addr); 1548 p4d = READ_ONCE(*p4dp); 1549 if (p4d_none(p4d)) 1550 continue; 1551 1552 WARN_ON(!p4d_present(p4d)); 1553 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap); 1554 } while (addr = next, addr < end); 1555} 1556 1557static void unmap_hotplug_range(unsigned long addr, unsigned long end, 1558 bool free_mapped, struct vmem_altmap *altmap) 1559{ 1560 unsigned long next; 1561 pgd_t *pgdp, pgd; 1562 1563 /* 1564 * altmap can only be used as vmemmap mapping backing memory. 1565 * In case the backing memory itself is not being freed, then 1566 * altmap is irrelevant. Warn about this inconsistency when 1567 * encountered. 1568 */ 1569 WARN_ON(!free_mapped && altmap); 1570 1571 do { 1572 next = pgd_addr_end(addr, end); 1573 pgdp = pgd_offset_k(addr); 1574 pgd = READ_ONCE(*pgdp); 1575 if (pgd_none(pgd)) 1576 continue; 1577 1578 WARN_ON(!pgd_present(pgd)); 1579 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap); 1580 } while (addr = next, addr < end); 1581} 1582 1583static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr, 1584 unsigned long end, unsigned long floor, 1585 unsigned long ceiling) 1586{ 1587 pte_t *ptep, pte; 1588 unsigned long i, start = addr; 1589 1590 do { 1591 ptep = pte_offset_kernel(pmdp, addr); 1592 pte = __ptep_get(ptep); 1593 1594 /* 1595 * This is just a sanity check here which verifies that 1596 * pte clearing has been done by earlier unmap loops. 1597 */ 1598 WARN_ON(!pte_none(pte)); 1599 } while (addr += PAGE_SIZE, addr < end); 1600 1601 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK)) 1602 return; 1603 1604 /* 1605 * Check whether we can free the pte page if the rest of the 1606 * entries are empty. Overlap with other regions have been 1607 * handled by the floor/ceiling check. 1608 */ 1609 ptep = pte_offset_kernel(pmdp, 0UL); 1610 for (i = 0; i < PTRS_PER_PTE; i++) { 1611 if (!pte_none(__ptep_get(&ptep[i]))) 1612 return; 1613 } 1614 1615 pmd_clear(pmdp); 1616 __flush_tlb_kernel_pgtable(start); 1617 free_hotplug_pgtable_page(virt_to_page(ptep)); 1618} 1619 1620static void free_empty_pmd_table(pud_t *pudp, unsigned long addr, 1621 unsigned long end, unsigned long floor, 1622 unsigned long ceiling) 1623{ 1624 pmd_t *pmdp, pmd; 1625 unsigned long i, next, start = addr; 1626 1627 do { 1628 next = pmd_addr_end(addr, end); 1629 pmdp = pmd_offset(pudp, addr); 1630 pmd = READ_ONCE(*pmdp); 1631 if (pmd_none(pmd)) 1632 continue; 1633 1634 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd)); 1635 free_empty_pte_table(pmdp, addr, next, floor, ceiling); 1636 } while (addr = next, addr < end); 1637 1638 if (CONFIG_PGTABLE_LEVELS <= 2) 1639 return; 1640 1641 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK)) 1642 return; 1643 1644 /* 1645 * Check whether we can free the pmd page if the rest of the 1646 * entries are empty. Overlap with other regions have been 1647 * handled by the floor/ceiling check. 1648 */ 1649 pmdp = pmd_offset(pudp, 0UL); 1650 for (i = 0; i < PTRS_PER_PMD; i++) { 1651 if (!pmd_none(READ_ONCE(pmdp[i]))) 1652 return; 1653 } 1654 1655 pud_clear(pudp); 1656 __flush_tlb_kernel_pgtable(start); 1657 free_hotplug_pgtable_page(virt_to_page(pmdp)); 1658} 1659 1660static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr, 1661 unsigned long end, unsigned long floor, 1662 unsigned long ceiling) 1663{ 1664 pud_t *pudp, pud; 1665 unsigned long i, next, start = addr; 1666 1667 do { 1668 next = pud_addr_end(addr, end); 1669 pudp = pud_offset(p4dp, addr); 1670 pud = READ_ONCE(*pudp); 1671 if (pud_none(pud)) 1672 continue; 1673 1674 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud)); 1675 free_empty_pmd_table(pudp, addr, next, floor, ceiling); 1676 } while (addr = next, addr < end); 1677 1678 if (!pgtable_l4_enabled()) 1679 return; 1680 1681 if (!pgtable_range_aligned(start, end, floor, ceiling, P4D_MASK)) 1682 return; 1683 1684 /* 1685 * Check whether we can free the pud page if the rest of the 1686 * entries are empty. Overlap with other regions have been 1687 * handled by the floor/ceiling check. 1688 */ 1689 pudp = pud_offset(p4dp, 0UL); 1690 for (i = 0; i < PTRS_PER_PUD; i++) { 1691 if (!pud_none(READ_ONCE(pudp[i]))) 1692 return; 1693 } 1694 1695 p4d_clear(p4dp); 1696 __flush_tlb_kernel_pgtable(start); 1697 free_hotplug_pgtable_page(virt_to_page(pudp)); 1698} 1699 1700static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr, 1701 unsigned long end, unsigned long floor, 1702 unsigned long ceiling) 1703{ 1704 p4d_t *p4dp, p4d; 1705 unsigned long i, next, start = addr; 1706 1707 do { 1708 next = p4d_addr_end(addr, end); 1709 p4dp = p4d_offset(pgdp, addr); 1710 p4d = READ_ONCE(*p4dp); 1711 if (p4d_none(p4d)) 1712 continue; 1713 1714 WARN_ON(!p4d_present(p4d)); 1715 free_empty_pud_table(p4dp, addr, next, floor, ceiling); 1716 } while (addr = next, addr < end); 1717 1718 if (!pgtable_l5_enabled()) 1719 return; 1720 1721 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK)) 1722 return; 1723 1724 /* 1725 * Check whether we can free the p4d page if the rest of the 1726 * entries are empty. Overlap with other regions have been 1727 * handled by the floor/ceiling check. 1728 */ 1729 p4dp = p4d_offset(pgdp, 0UL); 1730 for (i = 0; i < PTRS_PER_P4D; i++) { 1731 if (!p4d_none(READ_ONCE(p4dp[i]))) 1732 return; 1733 } 1734 1735 pgd_clear(pgdp); 1736 __flush_tlb_kernel_pgtable(start); 1737 free_hotplug_pgtable_page(virt_to_page(p4dp)); 1738} 1739 1740static void free_empty_tables(unsigned long addr, unsigned long end, 1741 unsigned long floor, unsigned long ceiling) 1742{ 1743 unsigned long next; 1744 pgd_t *pgdp, pgd; 1745 1746 do { 1747 next = pgd_addr_end(addr, end); 1748 pgdp = pgd_offset_k(addr); 1749 pgd = READ_ONCE(*pgdp); 1750 if (pgd_none(pgd)) 1751 continue; 1752 1753 WARN_ON(!pgd_present(pgd)); 1754 free_empty_p4d_table(pgdp, addr, next, floor, ceiling); 1755 } while (addr = next, addr < end); 1756} 1757#endif 1758 1759void __meminit vmemmap_set_pmd(pmd_t *pmdp, void *p, int node, 1760 unsigned long addr, unsigned long next) 1761{ 1762 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL)); 1763} 1764 1765int __meminit vmemmap_check_pmd(pmd_t *pmdp, int node, 1766 unsigned long addr, unsigned long next) 1767{ 1768 vmemmap_verify((pte_t *)pmdp, node, addr, next); 1769 1770 return pmd_sect(READ_ONCE(*pmdp)); 1771} 1772 1773int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 1774 struct vmem_altmap *altmap) 1775{ 1776 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1777 /* [start, end] should be within one section */ 1778 WARN_ON_ONCE(end - start > PAGES_PER_SECTION * sizeof(struct page)); 1779 1780 if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || 1781 (end - start < PAGES_PER_SECTION * sizeof(struct page))) 1782 return vmemmap_populate_basepages(start, end, node, altmap); 1783 else 1784 return vmemmap_populate_hugepages(start, end, node, altmap); 1785} 1786 1787#ifdef CONFIG_MEMORY_HOTPLUG 1788void vmemmap_free(unsigned long start, unsigned long end, 1789 struct vmem_altmap *altmap) 1790{ 1791 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END)); 1792 1793 unmap_hotplug_range(start, end, true, altmap); 1794 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END); 1795} 1796#endif /* CONFIG_MEMORY_HOTPLUG */ 1797 1798int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot) 1799{ 1800 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot)); 1801 1802 /* Only allow permission changes for now */ 1803 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)), 1804 pud_val(new_pud))) 1805 return 0; 1806 1807 VM_BUG_ON(phys & ~PUD_MASK); 1808 set_pud(pudp, new_pud); 1809 return 1; 1810} 1811 1812int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot) 1813{ 1814 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot)); 1815 1816 /* Only allow permission changes for now */ 1817 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)), 1818 pmd_val(new_pmd))) 1819 return 0; 1820 1821 VM_BUG_ON(phys & ~PMD_MASK); 1822 set_pmd(pmdp, new_pmd); 1823 return 1; 1824} 1825 1826#ifndef __PAGETABLE_P4D_FOLDED 1827void p4d_clear_huge(p4d_t *p4dp) 1828{ 1829} 1830#endif 1831 1832int pud_clear_huge(pud_t *pudp) 1833{ 1834 if (!pud_sect(READ_ONCE(*pudp))) 1835 return 0; 1836 pud_clear(pudp); 1837 return 1; 1838} 1839 1840int pmd_clear_huge(pmd_t *pmdp) 1841{ 1842 if (!pmd_sect(READ_ONCE(*pmdp))) 1843 return 0; 1844 pmd_clear(pmdp); 1845 return 1; 1846} 1847 1848static int __pmd_free_pte_page(pmd_t *pmdp, unsigned long addr, 1849 bool acquire_mmap_lock) 1850{ 1851 pte_t *table; 1852 pmd_t pmd; 1853 1854 pmd = READ_ONCE(*pmdp); 1855 1856 if (!pmd_table(pmd)) { 1857 VM_WARN_ON(1); 1858 return 1; 1859 } 1860 1861 /* See comment in pud_free_pmd_page for static key logic */ 1862 table = pte_offset_kernel(pmdp, addr); 1863 pmd_clear(pmdp); 1864 __flush_tlb_kernel_pgtable(addr); 1865 if (static_branch_unlikely(&arm64_ptdump_lock_key) && acquire_mmap_lock) { 1866 mmap_read_lock(&init_mm); 1867 mmap_read_unlock(&init_mm); 1868 } 1869 1870 pte_free_kernel(NULL, table); 1871 return 1; 1872} 1873 1874int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr) 1875{ 1876 /* If ptdump is walking the pagetables, acquire init_mm.mmap_lock */ 1877 return __pmd_free_pte_page(pmdp, addr, /* acquire_mmap_lock = */ true); 1878} 1879 1880int pud_free_pmd_page(pud_t *pudp, unsigned long addr) 1881{ 1882 pmd_t *table; 1883 pmd_t *pmdp; 1884 pud_t pud; 1885 unsigned long next, end; 1886 1887 pud = READ_ONCE(*pudp); 1888 1889 if (!pud_table(pud)) { 1890 VM_WARN_ON(1); 1891 return 1; 1892 } 1893 1894 table = pmd_offset(pudp, addr); 1895 1896 /* 1897 * Our objective is to prevent ptdump from reading a PMD table which has 1898 * been freed. In this race, if pud_free_pmd_page observes the key on 1899 * (which got flipped by ptdump) then the mmap lock sequence here will, 1900 * as a result of the mmap write lock/unlock sequence in ptdump, give 1901 * us the correct synchronization. If not, this means that ptdump has 1902 * yet not started walking the pagetables - the sequence of barriers 1903 * issued by __flush_tlb_kernel_pgtable() guarantees that ptdump will 1904 * observe an empty PUD. 1905 */ 1906 pud_clear(pudp); 1907 __flush_tlb_kernel_pgtable(addr); 1908 if (static_branch_unlikely(&arm64_ptdump_lock_key)) { 1909 mmap_read_lock(&init_mm); 1910 mmap_read_unlock(&init_mm); 1911 } 1912 1913 pmdp = table; 1914 next = addr; 1915 end = addr + PUD_SIZE; 1916 do { 1917 if (pmd_present(pmdp_get(pmdp))) 1918 /* 1919 * PMD has been isolated, so ptdump won't see it. No 1920 * need to acquire init_mm.mmap_lock. 1921 */ 1922 __pmd_free_pte_page(pmdp, next, /* acquire_mmap_lock = */ false); 1923 } while (pmdp++, next += PMD_SIZE, next != end); 1924 1925 pmd_free(NULL, table); 1926 return 1; 1927} 1928 1929#ifdef CONFIG_MEMORY_HOTPLUG 1930static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size) 1931{ 1932 unsigned long end = start + size; 1933 1934 WARN_ON(pgdir != init_mm.pgd); 1935 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END)); 1936 1937 unmap_hotplug_range(start, end, false, NULL); 1938 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END); 1939} 1940 1941struct range arch_get_mappable_range(void) 1942{ 1943 struct range mhp_range; 1944 phys_addr_t start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual)); 1945 phys_addr_t end_linear_pa = __pa(PAGE_END - 1); 1946 1947 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { 1948 /* 1949 * Check for a wrap, it is possible because of randomized linear 1950 * mapping the start physical address is actually bigger than 1951 * the end physical address. In this case set start to zero 1952 * because [0, end_linear_pa] range must still be able to cover 1953 * all addressable physical addresses. 1954 */ 1955 if (start_linear_pa > end_linear_pa) 1956 start_linear_pa = 0; 1957 } 1958 1959 WARN_ON(start_linear_pa > end_linear_pa); 1960 1961 /* 1962 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)] 1963 * accommodating both its ends but excluding PAGE_END. Max physical 1964 * range which can be mapped inside this linear mapping range, must 1965 * also be derived from its end points. 1966 */ 1967 mhp_range.start = start_linear_pa; 1968 mhp_range.end = end_linear_pa; 1969 1970 return mhp_range; 1971} 1972 1973int arch_add_memory(int nid, u64 start, u64 size, 1974 struct mhp_params *params) 1975{ 1976 int ret, flags = NO_EXEC_MAPPINGS; 1977 1978 VM_BUG_ON(!mhp_range_allowed(start, size, true)); 1979 1980 if (force_pte_mapping()) 1981 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; 1982 1983 ret = __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start), 1984 size, params->pgprot, pgd_pgtable_alloc_init_mm, 1985 flags); 1986 if (ret) 1987 goto err; 1988 1989 memblock_clear_nomap(start, size); 1990 1991 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, 1992 params); 1993 if (ret) 1994 goto err; 1995 1996 /* Address of hotplugged memory can be smaller */ 1997 max_pfn = max(max_pfn, PFN_UP(start + size)); 1998 max_low_pfn = max_pfn; 1999 2000 return 0; 2001 2002err: 2003 __remove_pgd_mapping(swapper_pg_dir, 2004 __phys_to_virt(start), size); 2005 return ret; 2006} 2007 2008void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap) 2009{ 2010 unsigned long start_pfn = start >> PAGE_SHIFT; 2011 unsigned long nr_pages = size >> PAGE_SHIFT; 2012 2013 __remove_pages(start_pfn, nr_pages, altmap); 2014 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size); 2015} 2016 2017/* 2018 * This memory hotplug notifier helps prevent boot memory from being 2019 * inadvertently removed as it blocks pfn range offlining process in 2020 * __offline_pages(). Hence this prevents both offlining as well as 2021 * removal process for boot memory which is initially always online. 2022 * In future if and when boot memory could be removed, this notifier 2023 * should be dropped and free_hotplug_page_range() should handle any 2024 * reserved pages allocated during boot. 2025 */ 2026static int prevent_bootmem_remove_notifier(struct notifier_block *nb, 2027 unsigned long action, void *data) 2028{ 2029 struct mem_section *ms; 2030 struct memory_notify *arg = data; 2031 unsigned long end_pfn = arg->start_pfn + arg->nr_pages; 2032 unsigned long pfn = arg->start_pfn; 2033 2034 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE)) 2035 return NOTIFY_OK; 2036 2037 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 2038 unsigned long start = PFN_PHYS(pfn); 2039 unsigned long end = start + (1UL << PA_SECTION_SHIFT); 2040 2041 ms = __pfn_to_section(pfn); 2042 if (!early_section(ms)) 2043 continue; 2044 2045 if (action == MEM_GOING_OFFLINE) { 2046 /* 2047 * Boot memory removal is not supported. Prevent 2048 * it via blocking any attempted offline request 2049 * for the boot memory and just report it. 2050 */ 2051 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end); 2052 return NOTIFY_BAD; 2053 } else if (action == MEM_OFFLINE) { 2054 /* 2055 * This should have never happened. Boot memory 2056 * offlining should have been prevented by this 2057 * very notifier. Probably some memory removal 2058 * procedure might have changed which would then 2059 * require further debug. 2060 */ 2061 pr_err("Boot memory [%lx %lx] offlined\n", start, end); 2062 2063 /* 2064 * Core memory hotplug does not process a return 2065 * code from the notifier for MEM_OFFLINE events. 2066 * The error condition has been reported. Return 2067 * from here as if ignored. 2068 */ 2069 return NOTIFY_DONE; 2070 } 2071 } 2072 return NOTIFY_OK; 2073} 2074 2075static struct notifier_block prevent_bootmem_remove_nb = { 2076 .notifier_call = prevent_bootmem_remove_notifier, 2077}; 2078 2079/* 2080 * This ensures that boot memory sections on the platform are online 2081 * from early boot. Memory sections could not be prevented from being 2082 * offlined, unless for some reason they are not online to begin with. 2083 * This helps validate the basic assumption on which the above memory 2084 * event notifier works to prevent boot memory section offlining and 2085 * its possible removal. 2086 */ 2087static void validate_bootmem_online(void) 2088{ 2089 phys_addr_t start, end, addr; 2090 struct mem_section *ms; 2091 u64 i; 2092 2093 /* 2094 * Scanning across all memblock might be expensive 2095 * on some big memory systems. Hence enable this 2096 * validation only with DEBUG_VM. 2097 */ 2098 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 2099 return; 2100 2101 for_each_mem_range(i, &start, &end) { 2102 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) { 2103 ms = __pfn_to_section(PHYS_PFN(addr)); 2104 2105 /* 2106 * All memory ranges in the system at this point 2107 * should have been marked as early sections. 2108 */ 2109 WARN_ON(!early_section(ms)); 2110 2111 /* 2112 * Memory notifier mechanism here to prevent boot 2113 * memory offlining depends on the fact that each 2114 * early section memory on the system is initially 2115 * online. Otherwise a given memory section which 2116 * is already offline will be overlooked and can 2117 * be removed completely. Call out such sections. 2118 */ 2119 if (!online_section(ms)) 2120 pr_err("Boot memory [%llx %llx] is offline, can be removed\n", 2121 addr, addr + (1UL << PA_SECTION_SHIFT)); 2122 } 2123 } 2124} 2125 2126static int __init prevent_bootmem_remove_init(void) 2127{ 2128 int ret = 0; 2129 2130 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) 2131 return ret; 2132 2133 validate_bootmem_online(); 2134 ret = register_memory_notifier(&prevent_bootmem_remove_nb); 2135 if (ret) 2136 pr_err("%s: Notifier registration failed %d\n", __func__, ret); 2137 2138 return ret; 2139} 2140early_initcall(prevent_bootmem_remove_init); 2141#endif 2142 2143pte_t modify_prot_start_ptes(struct vm_area_struct *vma, unsigned long addr, 2144 pte_t *ptep, unsigned int nr) 2145{ 2146 pte_t pte = get_and_clear_ptes(vma->vm_mm, addr, ptep, nr); 2147 2148 if (alternative_has_cap_unlikely(ARM64_WORKAROUND_2645198)) { 2149 /* 2150 * Break-before-make (BBM) is required for all user space mappings 2151 * when the permission changes from executable to non-executable 2152 * in cases where cpu is affected with errata #2645198. 2153 */ 2154 if (pte_accessible(vma->vm_mm, pte) && pte_user_exec(pte)) 2155 __flush_tlb_range(vma, addr, nr * PAGE_SIZE, 2156 PAGE_SIZE, true, 3); 2157 } 2158 2159 return pte; 2160} 2161 2162pte_t ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep) 2163{ 2164 return modify_prot_start_ptes(vma, addr, ptep, 1); 2165} 2166 2167void modify_prot_commit_ptes(struct vm_area_struct *vma, unsigned long addr, 2168 pte_t *ptep, pte_t old_pte, pte_t pte, 2169 unsigned int nr) 2170{ 2171 set_ptes(vma->vm_mm, addr, ptep, pte, nr); 2172} 2173 2174void ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, 2175 pte_t old_pte, pte_t pte) 2176{ 2177 modify_prot_commit_ptes(vma, addr, ptep, old_pte, pte, 1); 2178} 2179 2180/* 2181 * Atomically replaces the active TTBR1_EL1 PGD with a new VA-compatible PGD, 2182 * avoiding the possibility of conflicting TLB entries being allocated. 2183 */ 2184void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp) 2185{ 2186 typedef void (ttbr_replace_func)(phys_addr_t); 2187 extern ttbr_replace_func idmap_cpu_replace_ttbr1; 2188 ttbr_replace_func *replace_phys; 2189 unsigned long daif; 2190 2191 /* phys_to_ttbr() zeros lower 2 bits of ttbr with 52-bit PA */ 2192 phys_addr_t ttbr1 = phys_to_ttbr(virt_to_phys(pgdp)); 2193 2194 if (cnp) 2195 ttbr1 |= TTBR_CNP_BIT; 2196 2197 replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1); 2198 2199 cpu_install_idmap(); 2200 2201 /* 2202 * We really don't want to take *any* exceptions while TTBR1 is 2203 * in the process of being replaced so mask everything. 2204 */ 2205 daif = local_daif_save(); 2206 replace_phys(ttbr1); 2207 local_daif_restore(daif); 2208 2209 cpu_uninstall_idmap(); 2210} 2211 2212#ifdef CONFIG_ARCH_HAS_PKEYS 2213int arch_set_user_pkey_access(struct task_struct *tsk, int pkey, unsigned long init_val) 2214{ 2215 u64 new_por; 2216 u64 old_por; 2217 2218 if (!system_supports_poe()) 2219 return -ENOSPC; 2220 2221 /* 2222 * This code should only be called with valid 'pkey' 2223 * values originating from in-kernel users. Complain 2224 * if a bad value is observed. 2225 */ 2226 if (WARN_ON_ONCE(pkey >= arch_max_pkey())) 2227 return -EINVAL; 2228 2229 /* Set the bits we need in POR: */ 2230 new_por = POE_RWX; 2231 if (init_val & PKEY_DISABLE_WRITE) 2232 new_por &= ~POE_W; 2233 if (init_val & PKEY_DISABLE_ACCESS) 2234 new_por &= ~POE_RW; 2235 if (init_val & PKEY_DISABLE_READ) 2236 new_por &= ~POE_R; 2237 if (init_val & PKEY_DISABLE_EXECUTE) 2238 new_por &= ~POE_X; 2239 2240 /* Shift the bits in to the correct place in POR for pkey: */ 2241 new_por = POR_ELx_PERM_PREP(pkey, new_por); 2242 2243 /* Get old POR and mask off any old bits in place: */ 2244 old_por = read_sysreg_s(SYS_POR_EL0); 2245 old_por &= ~(POE_MASK << POR_ELx_PERM_SHIFT(pkey)); 2246 2247 /* Write old part along with new part: */ 2248 write_sysreg_s(old_por | new_por, SYS_POR_EL0); 2249 2250 return 0; 2251} 2252#endif
~/
Read-only file preview.