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/memory.h>
24 #include <linux/vmalloc.h>
26 #include <asm/barrier.h>
27 #include <asm/cputype.h>
28 #include <asm/fixmap.h>
29 #include <asm/kasan.h>
30 #include <asm/kernel-pgtable.h>
31 #include <asm/sections.h>
32 #include <asm/setup.h>
33 #include <linux/sizes.h>
35 #include <asm/mmu_context.h>
36 #include <asm/ptdump.h>
37 #include <asm/tlbflush.h>
38 #include <asm/pgalloc.h>
40 #define NO_BLOCK_MAPPINGS BIT(0)
41 #define NO_CONT_MAPPINGS BIT(1)
43 u64 idmap_t0sz = TCR_T0SZ(VA_BITS);
44 u64 idmap_ptrs_per_pgd = PTRS_PER_PGD;
46 u64 __section(".mmuoff.data.write") vabits_actual;
47 EXPORT_SYMBOL(vabits_actual);
49 u64 kimage_voffset __ro_after_init;
50 EXPORT_SYMBOL(kimage_voffset);
53 * Empty_zero_page is a special page that is used for zero-initialized data
56 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
57 EXPORT_SYMBOL(empty_zero_page);
59 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
60 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused;
61 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused;
63 static DEFINE_SPINLOCK(swapper_pgdir_lock);
65 void set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
69 spin_lock(&swapper_pgdir_lock);
70 fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp));
71 WRITE_ONCE(*fixmap_pgdp, pgd);
73 * We need dsb(ishst) here to ensure the page-table-walker sees
74 * our new entry before set_p?d() returns. The fixmap's
75 * flush_tlb_kernel_range() via clear_fixmap() does this for us.
78 spin_unlock(&swapper_pgdir_lock);
81 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
82 unsigned long size, pgprot_t vma_prot)
85 return pgprot_noncached(vma_prot);
86 else if (file->f_flags & O_SYNC)
87 return pgprot_writecombine(vma_prot);
90 EXPORT_SYMBOL(phys_mem_access_prot);
92 static phys_addr_t __init early_pgtable_alloc(int shift)
97 phys = memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
99 panic("Failed to allocate page table page\n");
102 * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
103 * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
104 * any level of table.
106 ptr = pte_set_fixmap(phys);
108 memset(ptr, 0, PAGE_SIZE);
111 * Implicit barriers also ensure the zeroed page is visible to the page
119 static bool pgattr_change_is_safe(u64 old, u64 new)
122 * The following mapping attributes may be updated in live
123 * kernel mappings without the need for break-before-make.
125 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG;
127 /* creating or taking down mappings is always safe */
128 if (old == 0 || new == 0)
131 /* live contiguous mappings may not be manipulated at all */
132 if ((old | new) & PTE_CONT)
135 /* Transitioning from Non-Global to Global is unsafe */
136 if (old & ~new & PTE_NG)
140 * Changing the memory type between Normal and Normal-Tagged is safe
141 * since Tagged is considered a permission attribute from the
142 * mismatched attribute aliases perspective.
144 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
145 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) &&
146 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
147 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)))
148 mask |= PTE_ATTRINDX_MASK;
150 return ((old ^ new) & ~mask) == 0;
153 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end,
154 phys_addr_t phys, pgprot_t prot)
158 ptep = pte_set_fixmap_offset(pmdp, addr);
160 pte_t old_pte = READ_ONCE(*ptep);
162 set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot));
165 * After the PTE entry has been populated once, we
166 * only allow updates to the permission attributes.
168 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte),
169 READ_ONCE(pte_val(*ptep))));
172 } while (ptep++, addr += PAGE_SIZE, addr != end);
177 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
178 unsigned long end, phys_addr_t phys,
180 phys_addr_t (*pgtable_alloc)(int),
184 pmd_t pmd = READ_ONCE(*pmdp);
186 BUG_ON(pmd_sect(pmd));
188 phys_addr_t pte_phys;
189 BUG_ON(!pgtable_alloc);
190 pte_phys = pgtable_alloc(PAGE_SHIFT);
191 __pmd_populate(pmdp, pte_phys, PMD_TYPE_TABLE);
192 pmd = READ_ONCE(*pmdp);
194 BUG_ON(pmd_bad(pmd));
197 pgprot_t __prot = prot;
199 next = pte_cont_addr_end(addr, end);
201 /* use a contiguous mapping if the range is suitably aligned */
202 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) &&
203 (flags & NO_CONT_MAPPINGS) == 0)
204 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
206 init_pte(pmdp, addr, next, phys, __prot);
209 } while (addr = next, addr != end);
212 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end,
213 phys_addr_t phys, pgprot_t prot,
214 phys_addr_t (*pgtable_alloc)(int), int flags)
219 pmdp = pmd_set_fixmap_offset(pudp, addr);
221 pmd_t old_pmd = READ_ONCE(*pmdp);
223 next = pmd_addr_end(addr, end);
225 /* try section mapping first */
226 if (((addr | next | phys) & ~SECTION_MASK) == 0 &&
227 (flags & NO_BLOCK_MAPPINGS) == 0) {
228 pmd_set_huge(pmdp, phys, prot);
231 * After the PMD entry has been populated once, we
232 * only allow updates to the permission attributes.
234 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
235 READ_ONCE(pmd_val(*pmdp))));
237 alloc_init_cont_pte(pmdp, addr, next, phys, prot,
238 pgtable_alloc, flags);
240 BUG_ON(pmd_val(old_pmd) != 0 &&
241 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp)));
244 } while (pmdp++, addr = next, addr != end);
249 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
250 unsigned long end, phys_addr_t phys,
252 phys_addr_t (*pgtable_alloc)(int), int flags)
255 pud_t pud = READ_ONCE(*pudp);
258 * Check for initial section mappings in the pgd/pud.
260 BUG_ON(pud_sect(pud));
262 phys_addr_t pmd_phys;
263 BUG_ON(!pgtable_alloc);
264 pmd_phys = pgtable_alloc(PMD_SHIFT);
265 __pud_populate(pudp, pmd_phys, PUD_TYPE_TABLE);
266 pud = READ_ONCE(*pudp);
268 BUG_ON(pud_bad(pud));
271 pgprot_t __prot = prot;
273 next = pmd_cont_addr_end(addr, end);
275 /* use a contiguous mapping if the range is suitably aligned */
276 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) &&
277 (flags & NO_CONT_MAPPINGS) == 0)
278 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
280 init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags);
283 } while (addr = next, addr != end);
286 static inline bool use_1G_block(unsigned long addr, unsigned long next,
289 if (PAGE_SHIFT != 12)
292 if (((addr | next | phys) & ~PUD_MASK) != 0)
298 static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end,
299 phys_addr_t phys, pgprot_t prot,
300 phys_addr_t (*pgtable_alloc)(int),
305 p4d_t *p4dp = p4d_offset(pgdp, addr);
306 p4d_t p4d = READ_ONCE(*p4dp);
309 phys_addr_t pud_phys;
310 BUG_ON(!pgtable_alloc);
311 pud_phys = pgtable_alloc(PUD_SHIFT);
312 __p4d_populate(p4dp, pud_phys, PUD_TYPE_TABLE);
313 p4d = READ_ONCE(*p4dp);
315 BUG_ON(p4d_bad(p4d));
317 pudp = pud_set_fixmap_offset(p4dp, addr);
319 pud_t old_pud = READ_ONCE(*pudp);
321 next = pud_addr_end(addr, end);
324 * For 4K granule only, attempt to put down a 1GB block
326 if (use_1G_block(addr, next, phys) &&
327 (flags & NO_BLOCK_MAPPINGS) == 0) {
328 pud_set_huge(pudp, phys, prot);
331 * After the PUD entry has been populated once, we
332 * only allow updates to the permission attributes.
334 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
335 READ_ONCE(pud_val(*pudp))));
337 alloc_init_cont_pmd(pudp, addr, next, phys, prot,
338 pgtable_alloc, flags);
340 BUG_ON(pud_val(old_pud) != 0 &&
341 pud_val(old_pud) != READ_ONCE(pud_val(*pudp)));
344 } while (pudp++, addr = next, addr != end);
349 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
350 unsigned long virt, phys_addr_t size,
352 phys_addr_t (*pgtable_alloc)(int),
355 unsigned long addr, end, next;
356 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt);
359 * If the virtual and physical address don't have the same offset
360 * within a page, we cannot map the region as the caller expects.
362 if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
366 addr = virt & PAGE_MASK;
367 end = PAGE_ALIGN(virt + size);
370 next = pgd_addr_end(addr, end);
371 alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc,
374 } while (pgdp++, addr = next, addr != end);
377 static phys_addr_t __pgd_pgtable_alloc(int shift)
379 void *ptr = (void *)__get_free_page(GFP_PGTABLE_KERNEL);
382 /* Ensure the zeroed page is visible to the page table walker */
387 static phys_addr_t pgd_pgtable_alloc(int shift)
389 phys_addr_t pa = __pgd_pgtable_alloc(shift);
392 * Call proper page table ctor in case later we need to
393 * call core mm functions like apply_to_page_range() on
394 * this pre-allocated page table.
396 * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is
397 * folded, and if so pgtable_pmd_page_ctor() becomes nop.
399 if (shift == PAGE_SHIFT)
400 BUG_ON(!pgtable_pte_page_ctor(phys_to_page(pa)));
401 else if (shift == PMD_SHIFT)
402 BUG_ON(!pgtable_pmd_page_ctor(phys_to_page(pa)));
408 * This function can only be used to modify existing table entries,
409 * without allocating new levels of table. Note that this permits the
410 * creation of new section or page entries.
412 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
413 phys_addr_t size, pgprot_t prot)
415 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
416 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
420 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
424 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
425 unsigned long virt, phys_addr_t size,
426 pgprot_t prot, bool page_mappings_only)
430 BUG_ON(mm == &init_mm);
432 if (page_mappings_only)
433 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
435 __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
436 pgd_pgtable_alloc, flags);
439 static void update_mapping_prot(phys_addr_t phys, unsigned long virt,
440 phys_addr_t size, pgprot_t prot)
442 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
443 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n",
448 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
451 /* flush the TLBs after updating live kernel mappings */
452 flush_tlb_kernel_range(virt, virt + size);
455 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start,
456 phys_addr_t end, pgprot_t prot, int flags)
458 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start,
459 prot, early_pgtable_alloc, flags);
462 void __init mark_linear_text_alias_ro(void)
465 * Remove the write permissions from the linear alias of .text/.rodata
467 update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text),
468 (unsigned long)__init_begin - (unsigned long)_text,
472 static void __init map_mem(pgd_t *pgdp)
474 phys_addr_t kernel_start = __pa_symbol(_text);
475 phys_addr_t kernel_end = __pa_symbol(__init_begin);
476 phys_addr_t start, end;
480 if (rodata_full || debug_pagealloc_enabled())
481 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
484 * Take care not to create a writable alias for the
485 * read-only text and rodata sections of the kernel image.
486 * So temporarily mark them as NOMAP to skip mappings in
487 * the following for-loop
489 memblock_mark_nomap(kernel_start, kernel_end - kernel_start);
490 #ifdef CONFIG_KEXEC_CORE
492 memblock_mark_nomap(crashk_res.start,
493 resource_size(&crashk_res));
496 /* map all the memory banks */
497 for_each_mem_range(i, &start, &end) {
501 * The linear map must allow allocation tags reading/writing
502 * if MTE is present. Otherwise, it has the same attributes as
505 __map_memblock(pgdp, start, end, PAGE_KERNEL_TAGGED, flags);
509 * Map the linear alias of the [_text, __init_begin) interval
510 * as non-executable now, and remove the write permission in
511 * mark_linear_text_alias_ro() below (which will be called after
512 * alternative patching has completed). This makes the contents
513 * of the region accessible to subsystems such as hibernate,
514 * but protects it from inadvertent modification or execution.
515 * Note that contiguous mappings cannot be remapped in this way,
516 * so we should avoid them here.
518 __map_memblock(pgdp, kernel_start, kernel_end,
519 PAGE_KERNEL, NO_CONT_MAPPINGS);
520 memblock_clear_nomap(kernel_start, kernel_end - kernel_start);
522 #ifdef CONFIG_KEXEC_CORE
524 * Use page-level mappings here so that we can shrink the region
525 * in page granularity and put back unused memory to buddy system
526 * through /sys/kernel/kexec_crash_size interface.
528 if (crashk_res.end) {
529 __map_memblock(pgdp, crashk_res.start, crashk_res.end + 1,
531 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
532 memblock_clear_nomap(crashk_res.start,
533 resource_size(&crashk_res));
538 void mark_rodata_ro(void)
540 unsigned long section_size;
543 * mark .rodata as read only. Use __init_begin rather than __end_rodata
544 * to cover NOTES and EXCEPTION_TABLE.
546 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
547 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata,
548 section_size, PAGE_KERNEL_RO);
553 static void __init map_kernel_segment(pgd_t *pgdp, void *va_start, void *va_end,
554 pgprot_t prot, struct vm_struct *vma,
555 int flags, unsigned long vm_flags)
557 phys_addr_t pa_start = __pa_symbol(va_start);
558 unsigned long size = va_end - va_start;
560 BUG_ON(!PAGE_ALIGNED(pa_start));
561 BUG_ON(!PAGE_ALIGNED(size));
563 __create_pgd_mapping(pgdp, pa_start, (unsigned long)va_start, size, prot,
564 early_pgtable_alloc, flags);
566 if (!(vm_flags & VM_NO_GUARD))
569 vma->addr = va_start;
570 vma->phys_addr = pa_start;
572 vma->flags = VM_MAP | vm_flags;
573 vma->caller = __builtin_return_address(0);
575 vm_area_add_early(vma);
578 static int __init parse_rodata(char *arg)
580 int ret = strtobool(arg, &rodata_enabled);
586 /* permit 'full' in addition to boolean options */
587 if (strcmp(arg, "full"))
590 rodata_enabled = true;
594 early_param("rodata", parse_rodata);
596 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
597 static int __init map_entry_trampoline(void)
599 pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
600 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start);
602 /* The trampoline is always mapped and can therefore be global */
603 pgprot_val(prot) &= ~PTE_NG;
605 /* Map only the text into the trampoline page table */
606 memset(tramp_pg_dir, 0, PGD_SIZE);
607 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, PAGE_SIZE,
608 prot, __pgd_pgtable_alloc, 0);
610 /* Map both the text and data into the kernel page table */
611 __set_fixmap(FIX_ENTRY_TRAMP_TEXT, pa_start, prot);
612 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
613 extern char __entry_tramp_data_start[];
615 __set_fixmap(FIX_ENTRY_TRAMP_DATA,
616 __pa_symbol(__entry_tramp_data_start),
622 core_initcall(map_entry_trampoline);
626 * Open coded check for BTI, only for use to determine configuration
627 * for early mappings for before the cpufeature code has run.
629 static bool arm64_early_this_cpu_has_bti(void)
633 if (!IS_ENABLED(CONFIG_ARM64_BTI_KERNEL))
636 pfr1 = read_sysreg_s(SYS_ID_AA64PFR1_EL1);
637 return cpuid_feature_extract_unsigned_field(pfr1,
638 ID_AA64PFR1_BT_SHIFT);
642 * Create fine-grained mappings for the kernel.
644 static void __init map_kernel(pgd_t *pgdp)
646 static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext,
647 vmlinux_initdata, vmlinux_data;
650 * External debuggers may need to write directly to the text
651 * mapping to install SW breakpoints. Allow this (only) when
652 * explicitly requested with rodata=off.
654 pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
657 * If we have a CPU that supports BTI and a kernel built for
658 * BTI then mark the kernel executable text as guarded pages
659 * now so we don't have to rewrite the page tables later.
661 if (arm64_early_this_cpu_has_bti())
662 text_prot = __pgprot_modify(text_prot, PTE_GP, PTE_GP);
665 * Only rodata will be remapped with different permissions later on,
666 * all other segments are allowed to use contiguous mappings.
668 map_kernel_segment(pgdp, _text, _etext, text_prot, &vmlinux_text, 0,
670 map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL,
671 &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD);
672 map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot,
673 &vmlinux_inittext, 0, VM_NO_GUARD);
674 map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL,
675 &vmlinux_initdata, 0, VM_NO_GUARD);
676 map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0);
678 if (!READ_ONCE(pgd_val(*pgd_offset_pgd(pgdp, FIXADDR_START)))) {
680 * The fixmap falls in a separate pgd to the kernel, and doesn't
681 * live in the carveout for the swapper_pg_dir. We can simply
682 * re-use the existing dir for the fixmap.
684 set_pgd(pgd_offset_pgd(pgdp, FIXADDR_START),
685 READ_ONCE(*pgd_offset_k(FIXADDR_START)));
686 } else if (CONFIG_PGTABLE_LEVELS > 3) {
691 * The fixmap shares its top level pgd entry with the kernel
692 * mapping. This can really only occur when we are running
693 * with 16k/4 levels, so we can simply reuse the pud level
696 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
697 bm_pgdp = pgd_offset_pgd(pgdp, FIXADDR_START);
698 bm_p4dp = p4d_offset(bm_pgdp, FIXADDR_START);
699 bm_pudp = pud_set_fixmap_offset(bm_p4dp, FIXADDR_START);
700 pud_populate(&init_mm, bm_pudp, lm_alias(bm_pmd));
706 kasan_copy_shadow(pgdp);
709 void __init paging_init(void)
711 pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir));
718 cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
719 init_mm.pgd = swapper_pg_dir;
721 memblock_free(__pa_symbol(init_pg_dir),
722 __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir));
724 memblock_allow_resize();
728 * Check whether a kernel address is valid (derived from arch/x86/).
730 int kern_addr_valid(unsigned long addr)
738 addr = arch_kasan_reset_tag(addr);
739 if ((((long)addr) >> VA_BITS) != -1UL)
742 pgdp = pgd_offset_k(addr);
743 if (pgd_none(READ_ONCE(*pgdp)))
746 p4dp = p4d_offset(pgdp, addr);
747 if (p4d_none(READ_ONCE(*p4dp)))
750 pudp = pud_offset(p4dp, addr);
751 pud = READ_ONCE(*pudp);
756 return pfn_valid(pud_pfn(pud));
758 pmdp = pmd_offset(pudp, addr);
759 pmd = READ_ONCE(*pmdp);
764 return pfn_valid(pmd_pfn(pmd));
766 ptep = pte_offset_kernel(pmdp, addr);
767 pte = READ_ONCE(*ptep);
771 return pfn_valid(pte_pfn(pte));
774 #ifdef CONFIG_MEMORY_HOTPLUG
775 static void free_hotplug_page_range(struct page *page, size_t size,
776 struct vmem_altmap *altmap)
779 vmem_altmap_free(altmap, size >> PAGE_SHIFT);
781 WARN_ON(PageReserved(page));
782 free_pages((unsigned long)page_address(page), get_order(size));
786 static void free_hotplug_pgtable_page(struct page *page)
788 free_hotplug_page_range(page, PAGE_SIZE, NULL);
791 static bool pgtable_range_aligned(unsigned long start, unsigned long end,
792 unsigned long floor, unsigned long ceiling,
805 if (end - 1 > ceiling - 1)
810 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr,
811 unsigned long end, bool free_mapped,
812 struct vmem_altmap *altmap)
817 ptep = pte_offset_kernel(pmdp, addr);
818 pte = READ_ONCE(*ptep);
822 WARN_ON(!pte_present(pte));
823 pte_clear(&init_mm, addr, ptep);
824 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
826 free_hotplug_page_range(pte_page(pte),
828 } while (addr += PAGE_SIZE, addr < end);
831 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr,
832 unsigned long end, bool free_mapped,
833 struct vmem_altmap *altmap)
839 next = pmd_addr_end(addr, end);
840 pmdp = pmd_offset(pudp, addr);
841 pmd = READ_ONCE(*pmdp);
845 WARN_ON(!pmd_present(pmd));
850 * One TLBI should be sufficient here as the PMD_SIZE
851 * range is mapped with a single block entry.
853 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
855 free_hotplug_page_range(pmd_page(pmd),
859 WARN_ON(!pmd_table(pmd));
860 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap);
861 } while (addr = next, addr < end);
864 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr,
865 unsigned long end, bool free_mapped,
866 struct vmem_altmap *altmap)
872 next = pud_addr_end(addr, end);
873 pudp = pud_offset(p4dp, addr);
874 pud = READ_ONCE(*pudp);
878 WARN_ON(!pud_present(pud));
883 * One TLBI should be sufficient here as the PUD_SIZE
884 * range is mapped with a single block entry.
886 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
888 free_hotplug_page_range(pud_page(pud),
892 WARN_ON(!pud_table(pud));
893 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap);
894 } while (addr = next, addr < end);
897 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr,
898 unsigned long end, bool free_mapped,
899 struct vmem_altmap *altmap)
905 next = p4d_addr_end(addr, end);
906 p4dp = p4d_offset(pgdp, addr);
907 p4d = READ_ONCE(*p4dp);
911 WARN_ON(!p4d_present(p4d));
912 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap);
913 } while (addr = next, addr < end);
916 static void unmap_hotplug_range(unsigned long addr, unsigned long end,
917 bool free_mapped, struct vmem_altmap *altmap)
923 * altmap can only be used as vmemmap mapping backing memory.
924 * In case the backing memory itself is not being freed, then
925 * altmap is irrelevant. Warn about this inconsistency when
928 WARN_ON(!free_mapped && altmap);
931 next = pgd_addr_end(addr, end);
932 pgdp = pgd_offset_k(addr);
933 pgd = READ_ONCE(*pgdp);
937 WARN_ON(!pgd_present(pgd));
938 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap);
939 } while (addr = next, addr < end);
942 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
943 unsigned long end, unsigned long floor,
944 unsigned long ceiling)
947 unsigned long i, start = addr;
950 ptep = pte_offset_kernel(pmdp, addr);
951 pte = READ_ONCE(*ptep);
954 * This is just a sanity check here which verifies that
955 * pte clearing has been done by earlier unmap loops.
957 WARN_ON(!pte_none(pte));
958 } while (addr += PAGE_SIZE, addr < end);
960 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK))
964 * Check whether we can free the pte page if the rest of the
965 * entries are empty. Overlap with other regions have been
966 * handled by the floor/ceiling check.
968 ptep = pte_offset_kernel(pmdp, 0UL);
969 for (i = 0; i < PTRS_PER_PTE; i++) {
970 if (!pte_none(READ_ONCE(ptep[i])))
975 __flush_tlb_kernel_pgtable(start);
976 free_hotplug_pgtable_page(virt_to_page(ptep));
979 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
980 unsigned long end, unsigned long floor,
981 unsigned long ceiling)
984 unsigned long i, next, start = addr;
987 next = pmd_addr_end(addr, end);
988 pmdp = pmd_offset(pudp, addr);
989 pmd = READ_ONCE(*pmdp);
993 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd));
994 free_empty_pte_table(pmdp, addr, next, floor, ceiling);
995 } while (addr = next, addr < end);
997 if (CONFIG_PGTABLE_LEVELS <= 2)
1000 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK))
1004 * Check whether we can free the pmd page if the rest of the
1005 * entries are empty. Overlap with other regions have been
1006 * handled by the floor/ceiling check.
1008 pmdp = pmd_offset(pudp, 0UL);
1009 for (i = 0; i < PTRS_PER_PMD; i++) {
1010 if (!pmd_none(READ_ONCE(pmdp[i])))
1015 __flush_tlb_kernel_pgtable(start);
1016 free_hotplug_pgtable_page(virt_to_page(pmdp));
1019 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr,
1020 unsigned long end, unsigned long floor,
1021 unsigned long ceiling)
1024 unsigned long i, next, start = addr;
1027 next = pud_addr_end(addr, end);
1028 pudp = pud_offset(p4dp, addr);
1029 pud = READ_ONCE(*pudp);
1033 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud));
1034 free_empty_pmd_table(pudp, addr, next, floor, ceiling);
1035 } while (addr = next, addr < end);
1037 if (CONFIG_PGTABLE_LEVELS <= 3)
1040 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK))
1044 * Check whether we can free the pud page if the rest of the
1045 * entries are empty. Overlap with other regions have been
1046 * handled by the floor/ceiling check.
1048 pudp = pud_offset(p4dp, 0UL);
1049 for (i = 0; i < PTRS_PER_PUD; i++) {
1050 if (!pud_none(READ_ONCE(pudp[i])))
1055 __flush_tlb_kernel_pgtable(start);
1056 free_hotplug_pgtable_page(virt_to_page(pudp));
1059 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
1060 unsigned long end, unsigned long floor,
1061 unsigned long ceiling)
1067 next = p4d_addr_end(addr, end);
1068 p4dp = p4d_offset(pgdp, addr);
1069 p4d = READ_ONCE(*p4dp);
1073 WARN_ON(!p4d_present(p4d));
1074 free_empty_pud_table(p4dp, addr, next, floor, ceiling);
1075 } while (addr = next, addr < end);
1078 static void free_empty_tables(unsigned long addr, unsigned long end,
1079 unsigned long floor, unsigned long ceiling)
1085 next = pgd_addr_end(addr, end);
1086 pgdp = pgd_offset_k(addr);
1087 pgd = READ_ONCE(*pgdp);
1091 WARN_ON(!pgd_present(pgd));
1092 free_empty_p4d_table(pgdp, addr, next, floor, ceiling);
1093 } while (addr = next, addr < end);
1097 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1098 #if !ARM64_SWAPPER_USES_SECTION_MAPS
1099 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1100 struct vmem_altmap *altmap)
1102 return vmemmap_populate_basepages(start, end, node, altmap);
1104 #else /* !ARM64_SWAPPER_USES_SECTION_MAPS */
1105 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1106 struct vmem_altmap *altmap)
1108 unsigned long addr = start;
1116 next = pmd_addr_end(addr, end);
1118 pgdp = vmemmap_pgd_populate(addr, node);
1122 p4dp = vmemmap_p4d_populate(pgdp, addr, node);
1126 pudp = vmemmap_pud_populate(p4dp, addr, node);
1130 pmdp = pmd_offset(pudp, addr);
1131 if (pmd_none(READ_ONCE(*pmdp))) {
1134 p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
1138 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL));
1140 vmemmap_verify((pte_t *)pmdp, node, addr, next);
1141 } while (addr = next, addr != end);
1145 #endif /* !ARM64_SWAPPER_USES_SECTION_MAPS */
1146 void vmemmap_free(unsigned long start, unsigned long end,
1147 struct vmem_altmap *altmap)
1149 #ifdef CONFIG_MEMORY_HOTPLUG
1150 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1152 unmap_hotplug_range(start, end, true, altmap);
1153 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END);
1156 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
1158 static inline pud_t * fixmap_pud(unsigned long addr)
1160 pgd_t *pgdp = pgd_offset_k(addr);
1161 p4d_t *p4dp = p4d_offset(pgdp, addr);
1162 p4d_t p4d = READ_ONCE(*p4dp);
1164 BUG_ON(p4d_none(p4d) || p4d_bad(p4d));
1166 return pud_offset_kimg(p4dp, addr);
1169 static inline pmd_t * fixmap_pmd(unsigned long addr)
1171 pud_t *pudp = fixmap_pud(addr);
1172 pud_t pud = READ_ONCE(*pudp);
1174 BUG_ON(pud_none(pud) || pud_bad(pud));
1176 return pmd_offset_kimg(pudp, addr);
1179 static inline pte_t * fixmap_pte(unsigned long addr)
1181 return &bm_pte[pte_index(addr)];
1185 * The p*d_populate functions call virt_to_phys implicitly so they can't be used
1186 * directly on kernel symbols (bm_p*d). This function is called too early to use
1187 * lm_alias so __p*d_populate functions must be used to populate with the
1188 * physical address from __pa_symbol.
1190 void __init early_fixmap_init(void)
1196 unsigned long addr = FIXADDR_START;
1198 pgdp = pgd_offset_k(addr);
1199 p4dp = p4d_offset(pgdp, addr);
1200 p4d = READ_ONCE(*p4dp);
1201 if (CONFIG_PGTABLE_LEVELS > 3 &&
1202 !(p4d_none(p4d) || p4d_page_paddr(p4d) == __pa_symbol(bm_pud))) {
1204 * We only end up here if the kernel mapping and the fixmap
1205 * share the top level pgd entry, which should only happen on
1206 * 16k/4 levels configurations.
1208 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
1209 pudp = pud_offset_kimg(p4dp, addr);
1212 __p4d_populate(p4dp, __pa_symbol(bm_pud), PUD_TYPE_TABLE);
1213 pudp = fixmap_pud(addr);
1215 if (pud_none(READ_ONCE(*pudp)))
1216 __pud_populate(pudp, __pa_symbol(bm_pmd), PMD_TYPE_TABLE);
1217 pmdp = fixmap_pmd(addr);
1218 __pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE);
1221 * The boot-ioremap range spans multiple pmds, for which
1222 * we are not prepared:
1224 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
1225 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
1227 if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
1228 || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
1230 pr_warn("pmdp %p != %p, %p\n",
1231 pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
1232 fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
1233 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
1234 fix_to_virt(FIX_BTMAP_BEGIN));
1235 pr_warn("fix_to_virt(FIX_BTMAP_END): %08lx\n",
1236 fix_to_virt(FIX_BTMAP_END));
1238 pr_warn("FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
1239 pr_warn("FIX_BTMAP_BEGIN: %d\n", FIX_BTMAP_BEGIN);
1244 * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we
1245 * ever need to use IPIs for TLB broadcasting, then we're in trouble here.
1247 void __set_fixmap(enum fixed_addresses idx,
1248 phys_addr_t phys, pgprot_t flags)
1250 unsigned long addr = __fix_to_virt(idx);
1253 BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
1255 ptep = fixmap_pte(addr);
1257 if (pgprot_val(flags)) {
1258 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags));
1260 pte_clear(&init_mm, addr, ptep);
1261 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
1265 void *__init fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
1267 const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
1272 * Check whether the physical FDT address is set and meets the minimum
1273 * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
1274 * at least 8 bytes so that we can always access the magic and size
1275 * fields of the FDT header after mapping the first chunk, double check
1276 * here if that is indeed the case.
1278 BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
1279 if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
1283 * Make sure that the FDT region can be mapped without the need to
1284 * allocate additional translation table pages, so that it is safe
1285 * to call create_mapping_noalloc() this early.
1287 * On 64k pages, the FDT will be mapped using PTEs, so we need to
1288 * be in the same PMD as the rest of the fixmap.
1289 * On 4k pages, we'll use section mappings for the FDT so we only
1290 * have to be in the same PUD.
1292 BUILD_BUG_ON(dt_virt_base % SZ_2M);
1294 BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
1295 __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
1297 offset = dt_phys % SWAPPER_BLOCK_SIZE;
1298 dt_virt = (void *)dt_virt_base + offset;
1300 /* map the first chunk so we can read the size from the header */
1301 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
1302 dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
1304 if (fdt_magic(dt_virt) != FDT_MAGIC)
1307 *size = fdt_totalsize(dt_virt);
1308 if (*size > MAX_FDT_SIZE)
1311 if (offset + *size > SWAPPER_BLOCK_SIZE)
1312 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
1313 round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
1318 int __init arch_ioremap_p4d_supported(void)
1323 int __init arch_ioremap_pud_supported(void)
1326 * Only 4k granule supports level 1 block mappings.
1327 * SW table walks can't handle removal of intermediate entries.
1329 return IS_ENABLED(CONFIG_ARM64_4K_PAGES) &&
1330 !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
1333 int __init arch_ioremap_pmd_supported(void)
1335 /* See arch_ioremap_pud_supported() */
1336 return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
1339 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
1341 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot));
1343 /* Only allow permission changes for now */
1344 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)),
1348 VM_BUG_ON(phys & ~PUD_MASK);
1349 set_pud(pudp, new_pud);
1353 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
1355 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot));
1357 /* Only allow permission changes for now */
1358 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)),
1362 VM_BUG_ON(phys & ~PMD_MASK);
1363 set_pmd(pmdp, new_pmd);
1367 int pud_clear_huge(pud_t *pudp)
1369 if (!pud_sect(READ_ONCE(*pudp)))
1375 int pmd_clear_huge(pmd_t *pmdp)
1377 if (!pmd_sect(READ_ONCE(*pmdp)))
1383 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
1388 pmd = READ_ONCE(*pmdp);
1390 if (!pmd_table(pmd)) {
1395 table = pte_offset_kernel(pmdp, addr);
1397 __flush_tlb_kernel_pgtable(addr);
1398 pte_free_kernel(NULL, table);
1402 int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
1407 unsigned long next, end;
1409 pud = READ_ONCE(*pudp);
1411 if (!pud_table(pud)) {
1416 table = pmd_offset(pudp, addr);
1419 end = addr + PUD_SIZE;
1421 pmd_free_pte_page(pmdp, next);
1422 } while (pmdp++, next += PMD_SIZE, next != end);
1425 __flush_tlb_kernel_pgtable(addr);
1426 pmd_free(NULL, table);
1430 int p4d_free_pud_page(p4d_t *p4d, unsigned long addr)
1432 return 0; /* Don't attempt a block mapping */
1435 #ifdef CONFIG_MEMORY_HOTPLUG
1436 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size)
1438 unsigned long end = start + size;
1440 WARN_ON(pgdir != init_mm.pgd);
1441 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END));
1443 unmap_hotplug_range(start, end, false, NULL);
1444 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END);
1447 static bool inside_linear_region(u64 start, u64 size)
1450 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)]
1451 * accommodating both its ends but excluding PAGE_END. Max physical
1452 * range which can be mapped inside this linear mapping range, must
1453 * also be derived from its end points.
1455 return start >= __pa(_PAGE_OFFSET(vabits_actual)) &&
1456 (start + size - 1) <= __pa(PAGE_END - 1);
1459 int arch_add_memory(int nid, u64 start, u64 size,
1460 struct mhp_params *params)
1464 if (!inside_linear_region(start, size)) {
1465 pr_err("[%llx %llx] is outside linear mapping region\n", start, start + size);
1469 if (rodata_full || debug_pagealloc_enabled())
1470 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
1472 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
1473 size, params->pgprot, __pgd_pgtable_alloc,
1476 memblock_clear_nomap(start, size);
1478 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT,
1481 __remove_pgd_mapping(swapper_pg_dir,
1482 __phys_to_virt(start), size);
1486 void arch_remove_memory(int nid, u64 start, u64 size,
1487 struct vmem_altmap *altmap)
1489 unsigned long start_pfn = start >> PAGE_SHIFT;
1490 unsigned long nr_pages = size >> PAGE_SHIFT;
1492 __remove_pages(start_pfn, nr_pages, altmap);
1493 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size);
1497 * This memory hotplug notifier helps prevent boot memory from being
1498 * inadvertently removed as it blocks pfn range offlining process in
1499 * __offline_pages(). Hence this prevents both offlining as well as
1500 * removal process for boot memory which is initially always online.
1501 * In future if and when boot memory could be removed, this notifier
1502 * should be dropped and free_hotplug_page_range() should handle any
1503 * reserved pages allocated during boot.
1505 static int prevent_bootmem_remove_notifier(struct notifier_block *nb,
1506 unsigned long action, void *data)
1508 struct mem_section *ms;
1509 struct memory_notify *arg = data;
1510 unsigned long end_pfn = arg->start_pfn + arg->nr_pages;
1511 unsigned long pfn = arg->start_pfn;
1513 if (action != MEM_GOING_OFFLINE)
1516 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
1517 ms = __pfn_to_section(pfn);
1518 if (early_section(ms))
1524 static struct notifier_block prevent_bootmem_remove_nb = {
1525 .notifier_call = prevent_bootmem_remove_notifier,
1528 static int __init prevent_bootmem_remove_init(void)
1530 return register_memory_notifier(&prevent_bootmem_remove_nb);
1532 device_initcall(prevent_bootmem_remove_init);