1 // SPDX-License-Identifier: GPL-2.0-only
3 * Based on arch/arm/mm/mmu.c
5 * Copyright (C) 1995-2005 Russell King
6 * Copyright (C) 2012 ARM Ltd.
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>
25 #include <linux/vmalloc.h>
26 #include <linux/set_memory.h>
27 #include <linux/kfence.h>
29 #include <asm/barrier.h>
30 #include <asm/cputype.h>
31 #include <asm/fixmap.h>
32 #include <asm/kasan.h>
33 #include <asm/kernel-pgtable.h>
34 #include <asm/sections.h>
35 #include <asm/setup.h>
36 #include <linux/sizes.h>
38 #include <asm/mmu_context.h>
39 #include <asm/ptdump.h>
40 #include <asm/tlbflush.h>
41 #include <asm/pgalloc.h>
42 #include <asm/kfence.h>
44 #define NO_BLOCK_MAPPINGS BIT(0)
45 #define NO_CONT_MAPPINGS BIT(1)
46 #define NO_EXEC_MAPPINGS BIT(2) /* assumes FEAT_HPDS is not used */
48 u64 kimage_voffset __ro_after_init;
49 EXPORT_SYMBOL(kimage_voffset);
51 u32 __boot_cpu_mode[] = { BOOT_CPU_MODE_EL2, BOOT_CPU_MODE_EL1 };
53 static bool rodata_is_rw __ro_after_init = true;
56 * The booting CPU updates the failed status @__early_cpu_boot_status,
57 * with MMU turned off.
59 long __section(".mmuoff.data.write") __early_cpu_boot_status;
62 * Empty_zero_page is a special page that is used for zero-initialized data
65 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
66 EXPORT_SYMBOL(empty_zero_page);
68 static DEFINE_SPINLOCK(swapper_pgdir_lock);
69 static DEFINE_MUTEX(fixmap_lock);
71 void noinstr set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
76 * Don't bother with the fixmap if swapper_pg_dir is still mapped
77 * writable in the kernel mapping.
80 WRITE_ONCE(*pgdp, pgd);
86 spin_lock(&swapper_pgdir_lock);
87 fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp));
88 WRITE_ONCE(*fixmap_pgdp, pgd);
90 * We need dsb(ishst) here to ensure the page-table-walker sees
91 * our new entry before set_p?d() returns. The fixmap's
92 * flush_tlb_kernel_range() via clear_fixmap() does this for us.
95 spin_unlock(&swapper_pgdir_lock);
98 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
99 unsigned long size, pgprot_t vma_prot)
101 if (!pfn_is_map_memory(pfn))
102 return pgprot_noncached(vma_prot);
103 else if (file->f_flags & O_SYNC)
104 return pgprot_writecombine(vma_prot);
107 EXPORT_SYMBOL(phys_mem_access_prot);
109 static phys_addr_t __init early_pgtable_alloc(int shift)
114 phys = memblock_phys_alloc_range(PAGE_SIZE, PAGE_SIZE, 0,
115 MEMBLOCK_ALLOC_NOLEAKTRACE);
117 panic("Failed to allocate page table page\n");
120 * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
121 * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
122 * any level of table.
124 ptr = pte_set_fixmap(phys);
126 memset(ptr, 0, PAGE_SIZE);
129 * Implicit barriers also ensure the zeroed page is visible to the page
137 bool pgattr_change_is_safe(u64 old, u64 new)
140 * The following mapping attributes may be updated in live
141 * kernel mappings without the need for break-before-make.
143 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG;
145 /* creating or taking down mappings is always safe */
146 if (!pte_valid(__pte(old)) || !pte_valid(__pte(new)))
149 /* A live entry's pfn should not change */
150 if (pte_pfn(__pte(old)) != pte_pfn(__pte(new)))
153 /* live contiguous mappings may not be manipulated at all */
154 if ((old | new) & PTE_CONT)
157 /* Transitioning from Non-Global to Global is unsafe */
158 if (old & ~new & PTE_NG)
162 * Changing the memory type between Normal and Normal-Tagged is safe
163 * since Tagged is considered a permission attribute from the
164 * mismatched attribute aliases perspective.
166 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
167 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) &&
168 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
169 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)))
170 mask |= PTE_ATTRINDX_MASK;
172 return ((old ^ new) & ~mask) == 0;
175 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end,
176 phys_addr_t phys, pgprot_t prot)
180 ptep = pte_set_fixmap_offset(pmdp, addr);
182 pte_t old_pte = READ_ONCE(*ptep);
184 set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot));
187 * After the PTE entry has been populated once, we
188 * only allow updates to the permission attributes.
190 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte),
191 READ_ONCE(pte_val(*ptep))));
194 } while (ptep++, addr += PAGE_SIZE, addr != end);
199 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
200 unsigned long end, phys_addr_t phys,
202 phys_addr_t (*pgtable_alloc)(int),
206 pmd_t pmd = READ_ONCE(*pmdp);
208 BUG_ON(pmd_sect(pmd));
210 pmdval_t pmdval = PMD_TYPE_TABLE | PMD_TABLE_UXN;
211 phys_addr_t pte_phys;
213 if (flags & NO_EXEC_MAPPINGS)
214 pmdval |= PMD_TABLE_PXN;
215 BUG_ON(!pgtable_alloc);
216 pte_phys = pgtable_alloc(PAGE_SHIFT);
217 __pmd_populate(pmdp, pte_phys, pmdval);
218 pmd = READ_ONCE(*pmdp);
220 BUG_ON(pmd_bad(pmd));
223 pgprot_t __prot = prot;
225 next = pte_cont_addr_end(addr, end);
227 /* use a contiguous mapping if the range is suitably aligned */
228 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) &&
229 (flags & NO_CONT_MAPPINGS) == 0)
230 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
232 init_pte(pmdp, addr, next, phys, __prot);
235 } while (addr = next, addr != end);
238 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end,
239 phys_addr_t phys, pgprot_t prot,
240 phys_addr_t (*pgtable_alloc)(int), int flags)
245 pmdp = pmd_set_fixmap_offset(pudp, addr);
247 pmd_t old_pmd = READ_ONCE(*pmdp);
249 next = pmd_addr_end(addr, end);
251 /* try section mapping first */
252 if (((addr | next | phys) & ~PMD_MASK) == 0 &&
253 (flags & NO_BLOCK_MAPPINGS) == 0) {
254 pmd_set_huge(pmdp, phys, prot);
257 * After the PMD entry has been populated once, we
258 * only allow updates to the permission attributes.
260 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
261 READ_ONCE(pmd_val(*pmdp))));
263 alloc_init_cont_pte(pmdp, addr, next, phys, prot,
264 pgtable_alloc, flags);
266 BUG_ON(pmd_val(old_pmd) != 0 &&
267 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp)));
270 } while (pmdp++, addr = next, addr != end);
275 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
276 unsigned long end, phys_addr_t phys,
278 phys_addr_t (*pgtable_alloc)(int), int flags)
281 pud_t pud = READ_ONCE(*pudp);
284 * Check for initial section mappings in the pgd/pud.
286 BUG_ON(pud_sect(pud));
288 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN;
289 phys_addr_t pmd_phys;
291 if (flags & NO_EXEC_MAPPINGS)
292 pudval |= PUD_TABLE_PXN;
293 BUG_ON(!pgtable_alloc);
294 pmd_phys = pgtable_alloc(PMD_SHIFT);
295 __pud_populate(pudp, pmd_phys, pudval);
296 pud = READ_ONCE(*pudp);
298 BUG_ON(pud_bad(pud));
301 pgprot_t __prot = prot;
303 next = pmd_cont_addr_end(addr, end);
305 /* use a contiguous mapping if the range is suitably aligned */
306 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) &&
307 (flags & NO_CONT_MAPPINGS) == 0)
308 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
310 init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags);
313 } while (addr = next, addr != end);
316 static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end,
317 phys_addr_t phys, pgprot_t prot,
318 phys_addr_t (*pgtable_alloc)(int),
323 p4d_t *p4dp = p4d_offset(pgdp, addr);
324 p4d_t p4d = READ_ONCE(*p4dp);
327 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN;
328 phys_addr_t pud_phys;
330 if (flags & NO_EXEC_MAPPINGS)
331 p4dval |= P4D_TABLE_PXN;
332 BUG_ON(!pgtable_alloc);
333 pud_phys = pgtable_alloc(PUD_SHIFT);
334 __p4d_populate(p4dp, pud_phys, p4dval);
335 p4d = READ_ONCE(*p4dp);
337 BUG_ON(p4d_bad(p4d));
339 pudp = pud_set_fixmap_offset(p4dp, addr);
341 pud_t old_pud = READ_ONCE(*pudp);
343 next = pud_addr_end(addr, end);
346 * For 4K granule only, attempt to put down a 1GB block
348 if (pud_sect_supported() &&
349 ((addr | next | phys) & ~PUD_MASK) == 0 &&
350 (flags & NO_BLOCK_MAPPINGS) == 0) {
351 pud_set_huge(pudp, phys, prot);
354 * After the PUD entry has been populated once, we
355 * only allow updates to the permission attributes.
357 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
358 READ_ONCE(pud_val(*pudp))));
360 alloc_init_cont_pmd(pudp, addr, next, phys, prot,
361 pgtable_alloc, flags);
363 BUG_ON(pud_val(old_pud) != 0 &&
364 pud_val(old_pud) != READ_ONCE(pud_val(*pudp)));
367 } while (pudp++, addr = next, addr != end);
372 static void __create_pgd_mapping_locked(pgd_t *pgdir, phys_addr_t phys,
373 unsigned long virt, phys_addr_t size,
375 phys_addr_t (*pgtable_alloc)(int),
378 unsigned long addr, end, next;
379 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt);
382 * If the virtual and physical address don't have the same offset
383 * within a page, we cannot map the region as the caller expects.
385 if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
389 addr = virt & PAGE_MASK;
390 end = PAGE_ALIGN(virt + size);
393 next = pgd_addr_end(addr, end);
394 alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc,
397 } while (pgdp++, addr = next, addr != end);
400 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
401 unsigned long virt, phys_addr_t size,
403 phys_addr_t (*pgtable_alloc)(int),
406 mutex_lock(&fixmap_lock);
407 __create_pgd_mapping_locked(pgdir, phys, virt, size, prot,
408 pgtable_alloc, flags);
409 mutex_unlock(&fixmap_lock);
412 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
413 extern __alias(__create_pgd_mapping_locked)
414 void create_kpti_ng_temp_pgd(pgd_t *pgdir, phys_addr_t phys, unsigned long virt,
415 phys_addr_t size, pgprot_t prot,
416 phys_addr_t (*pgtable_alloc)(int), int flags);
419 static phys_addr_t __pgd_pgtable_alloc(int shift)
421 void *ptr = (void *)__get_free_page(GFP_PGTABLE_KERNEL);
424 /* Ensure the zeroed page is visible to the page table walker */
429 static phys_addr_t pgd_pgtable_alloc(int shift)
431 phys_addr_t pa = __pgd_pgtable_alloc(shift);
432 struct ptdesc *ptdesc = page_ptdesc(phys_to_page(pa));
435 * Call proper page table ctor in case later we need to
436 * call core mm functions like apply_to_page_range() on
437 * this pre-allocated page table.
439 * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is
440 * folded, and if so pagetable_pte_ctor() becomes nop.
442 if (shift == PAGE_SHIFT)
443 BUG_ON(!pagetable_pte_ctor(ptdesc));
444 else if (shift == PMD_SHIFT)
445 BUG_ON(!pagetable_pmd_ctor(ptdesc));
451 * This function can only be used to modify existing table entries,
452 * without allocating new levels of table. Note that this permits the
453 * creation of new section or page entries.
455 void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
456 phys_addr_t size, pgprot_t prot)
458 if (virt < PAGE_OFFSET) {
459 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
463 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
467 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
468 unsigned long virt, phys_addr_t size,
469 pgprot_t prot, bool page_mappings_only)
473 BUG_ON(mm == &init_mm);
475 if (page_mappings_only)
476 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
478 __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
479 pgd_pgtable_alloc, flags);
482 static void update_mapping_prot(phys_addr_t phys, unsigned long virt,
483 phys_addr_t size, pgprot_t prot)
485 if (virt < PAGE_OFFSET) {
486 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n",
491 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
494 /* flush the TLBs after updating live kernel mappings */
495 flush_tlb_kernel_range(virt, virt + size);
498 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start,
499 phys_addr_t end, pgprot_t prot, int flags)
501 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start,
502 prot, early_pgtable_alloc, flags);
505 void __init mark_linear_text_alias_ro(void)
508 * Remove the write permissions from the linear alias of .text/.rodata
510 update_mapping_prot(__pa_symbol(_stext), (unsigned long)lm_alias(_stext),
511 (unsigned long)__init_begin - (unsigned long)_stext,
517 bool __ro_after_init kfence_early_init = !!CONFIG_KFENCE_SAMPLE_INTERVAL;
519 /* early_param() will be parsed before map_mem() below. */
520 static int __init parse_kfence_early_init(char *arg)
524 if (get_option(&arg, &val))
525 kfence_early_init = !!val;
528 early_param("kfence.sample_interval", parse_kfence_early_init);
530 static phys_addr_t __init arm64_kfence_alloc_pool(void)
532 phys_addr_t kfence_pool;
534 if (!kfence_early_init)
537 kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
539 pr_err("failed to allocate kfence pool\n");
540 kfence_early_init = false;
544 /* Temporarily mark as NOMAP. */
545 memblock_mark_nomap(kfence_pool, KFENCE_POOL_SIZE);
550 static void __init arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp)
555 /* KFENCE pool needs page-level mapping. */
556 __map_memblock(pgdp, kfence_pool, kfence_pool + KFENCE_POOL_SIZE,
557 pgprot_tagged(PAGE_KERNEL),
558 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
559 memblock_clear_nomap(kfence_pool, KFENCE_POOL_SIZE);
560 __kfence_pool = phys_to_virt(kfence_pool);
562 #else /* CONFIG_KFENCE */
564 static inline phys_addr_t arm64_kfence_alloc_pool(void) { return 0; }
565 static inline void arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) { }
567 #endif /* CONFIG_KFENCE */
569 static void __init map_mem(pgd_t *pgdp)
571 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN);
572 phys_addr_t kernel_start = __pa_symbol(_stext);
573 phys_addr_t kernel_end = __pa_symbol(__init_begin);
574 phys_addr_t start, end;
575 phys_addr_t early_kfence_pool;
576 int flags = NO_EXEC_MAPPINGS;
580 * Setting hierarchical PXNTable attributes on table entries covering
581 * the linear region is only possible if it is guaranteed that no table
582 * entries at any level are being shared between the linear region and
583 * the vmalloc region. Check whether this is true for the PGD level, in
584 * which case it is guaranteed to be true for all other levels as well.
586 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end));
588 early_kfence_pool = arm64_kfence_alloc_pool();
590 if (can_set_direct_map())
591 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
594 * Take care not to create a writable alias for the
595 * read-only text and rodata sections of the kernel image.
596 * So temporarily mark them as NOMAP to skip mappings in
597 * the following for-loop
599 memblock_mark_nomap(kernel_start, kernel_end - kernel_start);
601 /* map all the memory banks */
602 for_each_mem_range(i, &start, &end) {
606 * The linear map must allow allocation tags reading/writing
607 * if MTE is present. Otherwise, it has the same attributes as
610 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL),
615 * Map the linear alias of the [_stext, __init_begin) interval
616 * as non-executable now, and remove the write permission in
617 * mark_linear_text_alias_ro() below (which will be called after
618 * alternative patching has completed). This makes the contents
619 * of the region accessible to subsystems such as hibernate,
620 * but protects it from inadvertent modification or execution.
621 * Note that contiguous mappings cannot be remapped in this way,
622 * so we should avoid them here.
624 __map_memblock(pgdp, kernel_start, kernel_end,
625 PAGE_KERNEL, NO_CONT_MAPPINGS);
626 memblock_clear_nomap(kernel_start, kernel_end - kernel_start);
627 arm64_kfence_map_pool(early_kfence_pool, pgdp);
630 void mark_rodata_ro(void)
632 unsigned long section_size;
635 * mark .rodata as read only. Use __init_begin rather than __end_rodata
636 * to cover NOTES and EXCEPTION_TABLE.
638 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
639 WRITE_ONCE(rodata_is_rw, false);
640 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata,
641 section_size, PAGE_KERNEL_RO);
646 static void __init declare_vma(struct vm_struct *vma,
647 void *va_start, void *va_end,
648 unsigned long vm_flags)
650 phys_addr_t pa_start = __pa_symbol(va_start);
651 unsigned long size = va_end - va_start;
653 BUG_ON(!PAGE_ALIGNED(pa_start));
654 BUG_ON(!PAGE_ALIGNED(size));
656 if (!(vm_flags & VM_NO_GUARD))
659 vma->addr = va_start;
660 vma->phys_addr = pa_start;
662 vma->flags = VM_MAP | vm_flags;
663 vma->caller = __builtin_return_address(0);
665 vm_area_add_early(vma);
668 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
669 static pgprot_t kernel_exec_prot(void)
671 return rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
674 static int __init map_entry_trampoline(void)
678 if (!arm64_kernel_unmapped_at_el0())
681 pgprot_t prot = kernel_exec_prot();
682 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start);
684 /* The trampoline is always mapped and can therefore be global */
685 pgprot_val(prot) &= ~PTE_NG;
687 /* Map only the text into the trampoline page table */
688 memset(tramp_pg_dir, 0, PGD_SIZE);
689 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS,
690 entry_tramp_text_size(), prot,
691 __pgd_pgtable_alloc, NO_BLOCK_MAPPINGS);
693 /* Map both the text and data into the kernel page table */
694 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++)
695 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i,
696 pa_start + i * PAGE_SIZE, prot);
698 if (IS_ENABLED(CONFIG_RELOCATABLE))
699 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i,
700 pa_start + i * PAGE_SIZE, PAGE_KERNEL_RO);
704 core_initcall(map_entry_trampoline);
708 * Declare the VMA areas for the kernel
710 static void __init declare_kernel_vmas(void)
712 static struct vm_struct vmlinux_seg[KERNEL_SEGMENT_COUNT];
714 declare_vma(&vmlinux_seg[0], _stext, _etext, VM_NO_GUARD);
715 declare_vma(&vmlinux_seg[1], __start_rodata, __inittext_begin, VM_NO_GUARD);
716 declare_vma(&vmlinux_seg[2], __inittext_begin, __inittext_end, VM_NO_GUARD);
717 declare_vma(&vmlinux_seg[3], __initdata_begin, __initdata_end, VM_NO_GUARD);
718 declare_vma(&vmlinux_seg[4], _data, _end, 0);
721 void __pi_map_range(u64 *pgd, u64 start, u64 end, u64 pa, pgprot_t prot,
722 int level, pte_t *tbl, bool may_use_cont, u64 va_offset);
724 static u8 idmap_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init,
725 kpti_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init;
727 static void __init create_idmap(void)
729 u64 start = __pa_symbol(__idmap_text_start);
730 u64 end = __pa_symbol(__idmap_text_end);
731 u64 ptep = __pa_symbol(idmap_ptes);
733 __pi_map_range(&ptep, start, end, start, PAGE_KERNEL_ROX,
734 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false,
735 __phys_to_virt(ptep) - ptep);
737 if (IS_ENABLED(CONFIG_UNMAP_KERNEL_AT_EL0)) {
738 extern u32 __idmap_kpti_flag;
739 u64 pa = __pa_symbol(&__idmap_kpti_flag);
742 * The KPTI G-to-nG conversion code needs a read-write mapping
743 * of its synchronization flag in the ID map.
745 ptep = __pa_symbol(kpti_ptes);
746 __pi_map_range(&ptep, pa, pa + sizeof(u32), pa, PAGE_KERNEL,
747 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false,
748 __phys_to_virt(ptep) - ptep);
752 void __init paging_init(void)
754 map_mem(swapper_pg_dir);
756 memblock_allow_resize();
759 declare_kernel_vmas();
762 #ifdef CONFIG_MEMORY_HOTPLUG
763 static void free_hotplug_page_range(struct page *page, size_t size,
764 struct vmem_altmap *altmap)
767 vmem_altmap_free(altmap, size >> PAGE_SHIFT);
769 WARN_ON(PageReserved(page));
770 free_pages((unsigned long)page_address(page), get_order(size));
774 static void free_hotplug_pgtable_page(struct page *page)
776 free_hotplug_page_range(page, PAGE_SIZE, NULL);
779 static bool pgtable_range_aligned(unsigned long start, unsigned long end,
780 unsigned long floor, unsigned long ceiling,
793 if (end - 1 > ceiling - 1)
798 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr,
799 unsigned long end, bool free_mapped,
800 struct vmem_altmap *altmap)
805 ptep = pte_offset_kernel(pmdp, addr);
806 pte = READ_ONCE(*ptep);
810 WARN_ON(!pte_present(pte));
811 pte_clear(&init_mm, addr, ptep);
812 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
814 free_hotplug_page_range(pte_page(pte),
816 } while (addr += PAGE_SIZE, addr < end);
819 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr,
820 unsigned long end, bool free_mapped,
821 struct vmem_altmap *altmap)
827 next = pmd_addr_end(addr, end);
828 pmdp = pmd_offset(pudp, addr);
829 pmd = READ_ONCE(*pmdp);
833 WARN_ON(!pmd_present(pmd));
838 * One TLBI should be sufficient here as the PMD_SIZE
839 * range is mapped with a single block entry.
841 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
843 free_hotplug_page_range(pmd_page(pmd),
847 WARN_ON(!pmd_table(pmd));
848 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap);
849 } while (addr = next, addr < end);
852 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr,
853 unsigned long end, bool free_mapped,
854 struct vmem_altmap *altmap)
860 next = pud_addr_end(addr, end);
861 pudp = pud_offset(p4dp, addr);
862 pud = READ_ONCE(*pudp);
866 WARN_ON(!pud_present(pud));
871 * One TLBI should be sufficient here as the PUD_SIZE
872 * range is mapped with a single block entry.
874 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
876 free_hotplug_page_range(pud_page(pud),
880 WARN_ON(!pud_table(pud));
881 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap);
882 } while (addr = next, addr < end);
885 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr,
886 unsigned long end, bool free_mapped,
887 struct vmem_altmap *altmap)
893 next = p4d_addr_end(addr, end);
894 p4dp = p4d_offset(pgdp, addr);
895 p4d = READ_ONCE(*p4dp);
899 WARN_ON(!p4d_present(p4d));
900 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap);
901 } while (addr = next, addr < end);
904 static void unmap_hotplug_range(unsigned long addr, unsigned long end,
905 bool free_mapped, struct vmem_altmap *altmap)
911 * altmap can only be used as vmemmap mapping backing memory.
912 * In case the backing memory itself is not being freed, then
913 * altmap is irrelevant. Warn about this inconsistency when
916 WARN_ON(!free_mapped && altmap);
919 next = pgd_addr_end(addr, end);
920 pgdp = pgd_offset_k(addr);
921 pgd = READ_ONCE(*pgdp);
925 WARN_ON(!pgd_present(pgd));
926 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap);
927 } while (addr = next, addr < end);
930 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
931 unsigned long end, unsigned long floor,
932 unsigned long ceiling)
935 unsigned long i, start = addr;
938 ptep = pte_offset_kernel(pmdp, addr);
939 pte = READ_ONCE(*ptep);
942 * This is just a sanity check here which verifies that
943 * pte clearing has been done by earlier unmap loops.
945 WARN_ON(!pte_none(pte));
946 } while (addr += PAGE_SIZE, addr < end);
948 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK))
952 * Check whether we can free the pte page if the rest of the
953 * entries are empty. Overlap with other regions have been
954 * handled by the floor/ceiling check.
956 ptep = pte_offset_kernel(pmdp, 0UL);
957 for (i = 0; i < PTRS_PER_PTE; i++) {
958 if (!pte_none(READ_ONCE(ptep[i])))
963 __flush_tlb_kernel_pgtable(start);
964 free_hotplug_pgtable_page(virt_to_page(ptep));
967 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
968 unsigned long end, unsigned long floor,
969 unsigned long ceiling)
972 unsigned long i, next, start = addr;
975 next = pmd_addr_end(addr, end);
976 pmdp = pmd_offset(pudp, addr);
977 pmd = READ_ONCE(*pmdp);
981 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd));
982 free_empty_pte_table(pmdp, addr, next, floor, ceiling);
983 } while (addr = next, addr < end);
985 if (CONFIG_PGTABLE_LEVELS <= 2)
988 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK))
992 * Check whether we can free the pmd page if the rest of the
993 * entries are empty. Overlap with other regions have been
994 * handled by the floor/ceiling check.
996 pmdp = pmd_offset(pudp, 0UL);
997 for (i = 0; i < PTRS_PER_PMD; i++) {
998 if (!pmd_none(READ_ONCE(pmdp[i])))
1003 __flush_tlb_kernel_pgtable(start);
1004 free_hotplug_pgtable_page(virt_to_page(pmdp));
1007 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr,
1008 unsigned long end, unsigned long floor,
1009 unsigned long ceiling)
1012 unsigned long i, next, start = addr;
1015 next = pud_addr_end(addr, end);
1016 pudp = pud_offset(p4dp, addr);
1017 pud = READ_ONCE(*pudp);
1021 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud));
1022 free_empty_pmd_table(pudp, addr, next, floor, ceiling);
1023 } while (addr = next, addr < end);
1025 if (CONFIG_PGTABLE_LEVELS <= 3)
1028 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK))
1032 * Check whether we can free the pud page if the rest of the
1033 * entries are empty. Overlap with other regions have been
1034 * handled by the floor/ceiling check.
1036 pudp = pud_offset(p4dp, 0UL);
1037 for (i = 0; i < PTRS_PER_PUD; i++) {
1038 if (!pud_none(READ_ONCE(pudp[i])))
1043 __flush_tlb_kernel_pgtable(start);
1044 free_hotplug_pgtable_page(virt_to_page(pudp));
1047 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
1048 unsigned long end, unsigned long floor,
1049 unsigned long ceiling)
1055 next = p4d_addr_end(addr, end);
1056 p4dp = p4d_offset(pgdp, addr);
1057 p4d = READ_ONCE(*p4dp);
1061 WARN_ON(!p4d_present(p4d));
1062 free_empty_pud_table(p4dp, addr, next, floor, ceiling);
1063 } while (addr = next, addr < end);
1066 static void free_empty_tables(unsigned long addr, unsigned long end,
1067 unsigned long floor, unsigned long ceiling)
1073 next = pgd_addr_end(addr, end);
1074 pgdp = pgd_offset_k(addr);
1075 pgd = READ_ONCE(*pgdp);
1079 WARN_ON(!pgd_present(pgd));
1080 free_empty_p4d_table(pgdp, addr, next, floor, ceiling);
1081 } while (addr = next, addr < end);
1085 void __meminit vmemmap_set_pmd(pmd_t *pmdp, void *p, int node,
1086 unsigned long addr, unsigned long next)
1088 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL));
1091 int __meminit vmemmap_check_pmd(pmd_t *pmdp, int node,
1092 unsigned long addr, unsigned long next)
1094 vmemmap_verify((pte_t *)pmdp, node, addr, next);
1098 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1099 struct vmem_altmap *altmap)
1101 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1103 if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES))
1104 return vmemmap_populate_basepages(start, end, node, altmap);
1106 return vmemmap_populate_hugepages(start, end, node, altmap);
1109 #ifdef CONFIG_MEMORY_HOTPLUG
1110 void vmemmap_free(unsigned long start, unsigned long end,
1111 struct vmem_altmap *altmap)
1113 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1115 unmap_hotplug_range(start, end, true, altmap);
1116 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END);
1118 #endif /* CONFIG_MEMORY_HOTPLUG */
1120 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
1122 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot));
1124 /* Only allow permission changes for now */
1125 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)),
1129 VM_BUG_ON(phys & ~PUD_MASK);
1130 set_pud(pudp, new_pud);
1134 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
1136 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot));
1138 /* Only allow permission changes for now */
1139 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)),
1143 VM_BUG_ON(phys & ~PMD_MASK);
1144 set_pmd(pmdp, new_pmd);
1148 int pud_clear_huge(pud_t *pudp)
1150 if (!pud_sect(READ_ONCE(*pudp)))
1156 int pmd_clear_huge(pmd_t *pmdp)
1158 if (!pmd_sect(READ_ONCE(*pmdp)))
1164 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
1169 pmd = READ_ONCE(*pmdp);
1171 if (!pmd_table(pmd)) {
1176 table = pte_offset_kernel(pmdp, addr);
1178 __flush_tlb_kernel_pgtable(addr);
1179 pte_free_kernel(NULL, table);
1183 int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
1188 unsigned long next, end;
1190 pud = READ_ONCE(*pudp);
1192 if (!pud_table(pud)) {
1197 table = pmd_offset(pudp, addr);
1200 end = addr + PUD_SIZE;
1202 pmd_free_pte_page(pmdp, next);
1203 } while (pmdp++, next += PMD_SIZE, next != end);
1206 __flush_tlb_kernel_pgtable(addr);
1207 pmd_free(NULL, table);
1211 #ifdef CONFIG_MEMORY_HOTPLUG
1212 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size)
1214 unsigned long end = start + size;
1216 WARN_ON(pgdir != init_mm.pgd);
1217 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END));
1219 unmap_hotplug_range(start, end, false, NULL);
1220 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END);
1223 struct range arch_get_mappable_range(void)
1225 struct range mhp_range;
1226 u64 start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual));
1227 u64 end_linear_pa = __pa(PAGE_END - 1);
1229 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
1231 * Check for a wrap, it is possible because of randomized linear
1232 * mapping the start physical address is actually bigger than
1233 * the end physical address. In this case set start to zero
1234 * because [0, end_linear_pa] range must still be able to cover
1235 * all addressable physical addresses.
1237 if (start_linear_pa > end_linear_pa)
1238 start_linear_pa = 0;
1241 WARN_ON(start_linear_pa > end_linear_pa);
1244 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)]
1245 * accommodating both its ends but excluding PAGE_END. Max physical
1246 * range which can be mapped inside this linear mapping range, must
1247 * also be derived from its end points.
1249 mhp_range.start = start_linear_pa;
1250 mhp_range.end = end_linear_pa;
1255 int arch_add_memory(int nid, u64 start, u64 size,
1256 struct mhp_params *params)
1258 int ret, flags = NO_EXEC_MAPPINGS;
1260 VM_BUG_ON(!mhp_range_allowed(start, size, true));
1262 if (can_set_direct_map())
1263 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
1265 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
1266 size, params->pgprot, __pgd_pgtable_alloc,
1269 memblock_clear_nomap(start, size);
1271 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT,
1274 __remove_pgd_mapping(swapper_pg_dir,
1275 __phys_to_virt(start), size);
1277 max_pfn = PFN_UP(start + size);
1278 max_low_pfn = max_pfn;
1284 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
1286 unsigned long start_pfn = start >> PAGE_SHIFT;
1287 unsigned long nr_pages = size >> PAGE_SHIFT;
1289 __remove_pages(start_pfn, nr_pages, altmap);
1290 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size);
1294 * This memory hotplug notifier helps prevent boot memory from being
1295 * inadvertently removed as it blocks pfn range offlining process in
1296 * __offline_pages(). Hence this prevents both offlining as well as
1297 * removal process for boot memory which is initially always online.
1298 * In future if and when boot memory could be removed, this notifier
1299 * should be dropped and free_hotplug_page_range() should handle any
1300 * reserved pages allocated during boot.
1302 static int prevent_bootmem_remove_notifier(struct notifier_block *nb,
1303 unsigned long action, void *data)
1305 struct mem_section *ms;
1306 struct memory_notify *arg = data;
1307 unsigned long end_pfn = arg->start_pfn + arg->nr_pages;
1308 unsigned long pfn = arg->start_pfn;
1310 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE))
1313 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
1314 unsigned long start = PFN_PHYS(pfn);
1315 unsigned long end = start + (1UL << PA_SECTION_SHIFT);
1317 ms = __pfn_to_section(pfn);
1318 if (!early_section(ms))
1321 if (action == MEM_GOING_OFFLINE) {
1323 * Boot memory removal is not supported. Prevent
1324 * it via blocking any attempted offline request
1325 * for the boot memory and just report it.
1327 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end);
1329 } else if (action == MEM_OFFLINE) {
1331 * This should have never happened. Boot memory
1332 * offlining should have been prevented by this
1333 * very notifier. Probably some memory removal
1334 * procedure might have changed which would then
1335 * require further debug.
1337 pr_err("Boot memory [%lx %lx] offlined\n", start, end);
1340 * Core memory hotplug does not process a return
1341 * code from the notifier for MEM_OFFLINE events.
1342 * The error condition has been reported. Return
1343 * from here as if ignored.
1351 static struct notifier_block prevent_bootmem_remove_nb = {
1352 .notifier_call = prevent_bootmem_remove_notifier,
1356 * This ensures that boot memory sections on the platform are online
1357 * from early boot. Memory sections could not be prevented from being
1358 * offlined, unless for some reason they are not online to begin with.
1359 * This helps validate the basic assumption on which the above memory
1360 * event notifier works to prevent boot memory section offlining and
1361 * its possible removal.
1363 static void validate_bootmem_online(void)
1365 phys_addr_t start, end, addr;
1366 struct mem_section *ms;
1370 * Scanning across all memblock might be expensive
1371 * on some big memory systems. Hence enable this
1372 * validation only with DEBUG_VM.
1374 if (!IS_ENABLED(CONFIG_DEBUG_VM))
1377 for_each_mem_range(i, &start, &end) {
1378 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) {
1379 ms = __pfn_to_section(PHYS_PFN(addr));
1382 * All memory ranges in the system at this point
1383 * should have been marked as early sections.
1385 WARN_ON(!early_section(ms));
1388 * Memory notifier mechanism here to prevent boot
1389 * memory offlining depends on the fact that each
1390 * early section memory on the system is initially
1391 * online. Otherwise a given memory section which
1392 * is already offline will be overlooked and can
1393 * be removed completely. Call out such sections.
1395 if (!online_section(ms))
1396 pr_err("Boot memory [%llx %llx] is offline, can be removed\n",
1397 addr, addr + (1UL << PA_SECTION_SHIFT));
1402 static int __init prevent_bootmem_remove_init(void)
1406 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
1409 validate_bootmem_online();
1410 ret = register_memory_notifier(&prevent_bootmem_remove_nb);
1412 pr_err("%s: Notifier registration failed %d\n", __func__, ret);
1416 early_initcall(prevent_bootmem_remove_init);
1419 pte_t ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep)
1421 if (alternative_has_cap_unlikely(ARM64_WORKAROUND_2645198)) {
1423 * Break-before-make (BBM) is required for all user space mappings
1424 * when the permission changes from executable to non-executable
1425 * in cases where cpu is affected with errata #2645198.
1427 if (pte_user_exec(READ_ONCE(*ptep)))
1428 return ptep_clear_flush(vma, addr, ptep);
1430 return ptep_get_and_clear(vma->vm_mm, addr, ptep);
1433 void ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep,
1434 pte_t old_pte, pte_t pte)
1436 set_pte_at(vma->vm_mm, addr, ptep, pte);
1440 * Atomically replaces the active TTBR1_EL1 PGD with a new VA-compatible PGD,
1441 * avoiding the possibility of conflicting TLB entries being allocated.
1443 void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp)
1445 typedef void (ttbr_replace_func)(phys_addr_t);
1446 extern ttbr_replace_func idmap_cpu_replace_ttbr1;
1447 ttbr_replace_func *replace_phys;
1450 /* phys_to_ttbr() zeros lower 2 bits of ttbr with 52-bit PA */
1451 phys_addr_t ttbr1 = phys_to_ttbr(virt_to_phys(pgdp));
1454 ttbr1 |= TTBR_CNP_BIT;
1456 replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1);
1458 cpu_install_idmap();
1461 * We really don't want to take *any* exceptions while TTBR1 is
1462 * in the process of being replaced so mask everything.
1464 daif = local_daif_save();
1465 replace_phys(ttbr1);
1466 local_daif_restore(daif);
1468 cpu_uninstall_idmap();