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>
39 #define NO_BLOCK_MAPPINGS BIT(0)
40 #define NO_CONT_MAPPINGS BIT(1)
42 u64 idmap_t0sz = TCR_T0SZ(VA_BITS);
43 u64 idmap_ptrs_per_pgd = PTRS_PER_PGD;
45 u64 __section(".mmuoff.data.write") vabits_actual;
46 EXPORT_SYMBOL(vabits_actual);
48 u64 kimage_voffset __ro_after_init;
49 EXPORT_SYMBOL(kimage_voffset);
52 * Empty_zero_page is a special page that is used for zero-initialized data
55 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
56 EXPORT_SYMBOL(empty_zero_page);
58 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
59 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused;
60 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused;
62 static DEFINE_SPINLOCK(swapper_pgdir_lock);
64 void set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
68 spin_lock(&swapper_pgdir_lock);
69 fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp));
70 WRITE_ONCE(*fixmap_pgdp, pgd);
72 * We need dsb(ishst) here to ensure the page-table-walker sees
73 * our new entry before set_p?d() returns. The fixmap's
74 * flush_tlb_kernel_range() via clear_fixmap() does this for us.
77 spin_unlock(&swapper_pgdir_lock);
80 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
81 unsigned long size, pgprot_t vma_prot)
84 return pgprot_noncached(vma_prot);
85 else if (file->f_flags & O_SYNC)
86 return pgprot_writecombine(vma_prot);
89 EXPORT_SYMBOL(phys_mem_access_prot);
91 static phys_addr_t __init early_pgtable_alloc(int shift)
96 phys = memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
98 panic("Failed to allocate page table page\n");
101 * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
102 * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
103 * any level of table.
105 ptr = pte_set_fixmap(phys);
107 memset(ptr, 0, PAGE_SIZE);
110 * Implicit barriers also ensure the zeroed page is visible to the page
118 static bool pgattr_change_is_safe(u64 old, u64 new)
121 * The following mapping attributes may be updated in live
122 * kernel mappings without the need for break-before-make.
124 static const pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG;
126 /* creating or taking down mappings is always safe */
127 if (old == 0 || new == 0)
130 /* live contiguous mappings may not be manipulated at all */
131 if ((old | new) & PTE_CONT)
134 /* Transitioning from Non-Global to Global is unsafe */
135 if (old & ~new & PTE_NG)
138 return ((old ^ new) & ~mask) == 0;
141 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end,
142 phys_addr_t phys, pgprot_t prot)
146 ptep = pte_set_fixmap_offset(pmdp, addr);
148 pte_t old_pte = READ_ONCE(*ptep);
150 set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot));
153 * After the PTE entry has been populated once, we
154 * only allow updates to the permission attributes.
156 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte),
157 READ_ONCE(pte_val(*ptep))));
160 } while (ptep++, addr += PAGE_SIZE, addr != end);
165 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
166 unsigned long end, phys_addr_t phys,
168 phys_addr_t (*pgtable_alloc)(int),
172 pmd_t pmd = READ_ONCE(*pmdp);
174 BUG_ON(pmd_sect(pmd));
176 phys_addr_t pte_phys;
177 BUG_ON(!pgtable_alloc);
178 pte_phys = pgtable_alloc(PAGE_SHIFT);
179 __pmd_populate(pmdp, pte_phys, PMD_TYPE_TABLE);
180 pmd = READ_ONCE(*pmdp);
182 BUG_ON(pmd_bad(pmd));
185 pgprot_t __prot = prot;
187 next = pte_cont_addr_end(addr, end);
189 /* use a contiguous mapping if the range is suitably aligned */
190 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) &&
191 (flags & NO_CONT_MAPPINGS) == 0)
192 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
194 init_pte(pmdp, addr, next, phys, __prot);
197 } while (addr = next, addr != end);
200 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end,
201 phys_addr_t phys, pgprot_t prot,
202 phys_addr_t (*pgtable_alloc)(int), int flags)
207 pmdp = pmd_set_fixmap_offset(pudp, addr);
209 pmd_t old_pmd = READ_ONCE(*pmdp);
211 next = pmd_addr_end(addr, end);
213 /* try section mapping first */
214 if (((addr | next | phys) & ~SECTION_MASK) == 0 &&
215 (flags & NO_BLOCK_MAPPINGS) == 0) {
216 pmd_set_huge(pmdp, phys, prot);
219 * After the PMD entry has been populated once, we
220 * only allow updates to the permission attributes.
222 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
223 READ_ONCE(pmd_val(*pmdp))));
225 alloc_init_cont_pte(pmdp, addr, next, phys, prot,
226 pgtable_alloc, flags);
228 BUG_ON(pmd_val(old_pmd) != 0 &&
229 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp)));
232 } while (pmdp++, addr = next, addr != end);
237 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
238 unsigned long end, phys_addr_t phys,
240 phys_addr_t (*pgtable_alloc)(int), int flags)
243 pud_t pud = READ_ONCE(*pudp);
246 * Check for initial section mappings in the pgd/pud.
248 BUG_ON(pud_sect(pud));
250 phys_addr_t pmd_phys;
251 BUG_ON(!pgtable_alloc);
252 pmd_phys = pgtable_alloc(PMD_SHIFT);
253 __pud_populate(pudp, pmd_phys, PUD_TYPE_TABLE);
254 pud = READ_ONCE(*pudp);
256 BUG_ON(pud_bad(pud));
259 pgprot_t __prot = prot;
261 next = pmd_cont_addr_end(addr, end);
263 /* use a contiguous mapping if the range is suitably aligned */
264 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) &&
265 (flags & NO_CONT_MAPPINGS) == 0)
266 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
268 init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags);
271 } while (addr = next, addr != end);
274 static inline bool use_1G_block(unsigned long addr, unsigned long next,
277 if (PAGE_SHIFT != 12)
280 if (((addr | next | phys) & ~PUD_MASK) != 0)
286 static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end,
287 phys_addr_t phys, pgprot_t prot,
288 phys_addr_t (*pgtable_alloc)(int),
293 p4d_t *p4dp = p4d_offset(pgdp, addr);
294 p4d_t p4d = READ_ONCE(*p4dp);
297 phys_addr_t pud_phys;
298 BUG_ON(!pgtable_alloc);
299 pud_phys = pgtable_alloc(PUD_SHIFT);
300 __p4d_populate(p4dp, pud_phys, PUD_TYPE_TABLE);
301 p4d = READ_ONCE(*p4dp);
303 BUG_ON(p4d_bad(p4d));
305 pudp = pud_set_fixmap_offset(p4dp, addr);
307 pud_t old_pud = READ_ONCE(*pudp);
309 next = pud_addr_end(addr, end);
312 * For 4K granule only, attempt to put down a 1GB block
314 if (use_1G_block(addr, next, phys) &&
315 (flags & NO_BLOCK_MAPPINGS) == 0) {
316 pud_set_huge(pudp, phys, prot);
319 * After the PUD entry has been populated once, we
320 * only allow updates to the permission attributes.
322 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
323 READ_ONCE(pud_val(*pudp))));
325 alloc_init_cont_pmd(pudp, addr, next, phys, prot,
326 pgtable_alloc, flags);
328 BUG_ON(pud_val(old_pud) != 0 &&
329 pud_val(old_pud) != READ_ONCE(pud_val(*pudp)));
332 } while (pudp++, addr = next, addr != end);
337 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
338 unsigned long virt, phys_addr_t size,
340 phys_addr_t (*pgtable_alloc)(int),
343 unsigned long addr, end, next;
344 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt);
347 * If the virtual and physical address don't have the same offset
348 * within a page, we cannot map the region as the caller expects.
350 if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
354 addr = virt & PAGE_MASK;
355 end = PAGE_ALIGN(virt + size);
358 next = pgd_addr_end(addr, end);
359 alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc,
362 } while (pgdp++, addr = next, addr != end);
365 static phys_addr_t __pgd_pgtable_alloc(int shift)
367 void *ptr = (void *)__get_free_page(GFP_PGTABLE_KERNEL);
370 /* Ensure the zeroed page is visible to the page table walker */
375 static phys_addr_t pgd_pgtable_alloc(int shift)
377 phys_addr_t pa = __pgd_pgtable_alloc(shift);
380 * Call proper page table ctor in case later we need to
381 * call core mm functions like apply_to_page_range() on
382 * this pre-allocated page table.
384 * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is
385 * folded, and if so pgtable_pmd_page_ctor() becomes nop.
387 if (shift == PAGE_SHIFT)
388 BUG_ON(!pgtable_pte_page_ctor(phys_to_page(pa)));
389 else if (shift == PMD_SHIFT)
390 BUG_ON(!pgtable_pmd_page_ctor(phys_to_page(pa)));
396 * This function can only be used to modify existing table entries,
397 * without allocating new levels of table. Note that this permits the
398 * creation of new section or page entries.
400 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
401 phys_addr_t size, pgprot_t prot)
403 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
404 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
408 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
412 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
413 unsigned long virt, phys_addr_t size,
414 pgprot_t prot, bool page_mappings_only)
418 BUG_ON(mm == &init_mm);
420 if (page_mappings_only)
421 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
423 __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
424 pgd_pgtable_alloc, flags);
427 static void update_mapping_prot(phys_addr_t phys, unsigned long virt,
428 phys_addr_t size, pgprot_t prot)
430 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
431 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n",
436 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
439 /* flush the TLBs after updating live kernel mappings */
440 flush_tlb_kernel_range(virt, virt + size);
443 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start,
444 phys_addr_t end, pgprot_t prot, int flags)
446 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start,
447 prot, early_pgtable_alloc, flags);
450 void __init mark_linear_text_alias_ro(void)
453 * Remove the write permissions from the linear alias of .text/.rodata
455 update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text),
456 (unsigned long)__init_begin - (unsigned long)_text,
460 static void __init map_mem(pgd_t *pgdp)
462 phys_addr_t kernel_start = __pa_symbol(_text);
463 phys_addr_t kernel_end = __pa_symbol(__init_begin);
464 struct memblock_region *reg;
467 if (rodata_full || debug_pagealloc_enabled())
468 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
471 * Take care not to create a writable alias for the
472 * read-only text and rodata sections of the kernel image.
473 * So temporarily mark them as NOMAP to skip mappings in
474 * the following for-loop
476 memblock_mark_nomap(kernel_start, kernel_end - kernel_start);
477 #ifdef CONFIG_KEXEC_CORE
479 memblock_mark_nomap(crashk_res.start,
480 resource_size(&crashk_res));
483 /* map all the memory banks */
484 for_each_memblock(memory, reg) {
485 phys_addr_t start = reg->base;
486 phys_addr_t end = start + reg->size;
490 if (memblock_is_nomap(reg))
493 __map_memblock(pgdp, start, end, PAGE_KERNEL, flags);
497 * Map the linear alias of the [_text, __init_begin) interval
498 * as non-executable now, and remove the write permission in
499 * mark_linear_text_alias_ro() below (which will be called after
500 * alternative patching has completed). This makes the contents
501 * of the region accessible to subsystems such as hibernate,
502 * but protects it from inadvertent modification or execution.
503 * Note that contiguous mappings cannot be remapped in this way,
504 * so we should avoid them here.
506 __map_memblock(pgdp, kernel_start, kernel_end,
507 PAGE_KERNEL, NO_CONT_MAPPINGS);
508 memblock_clear_nomap(kernel_start, kernel_end - kernel_start);
510 #ifdef CONFIG_KEXEC_CORE
512 * Use page-level mappings here so that we can shrink the region
513 * in page granularity and put back unused memory to buddy system
514 * through /sys/kernel/kexec_crash_size interface.
516 if (crashk_res.end) {
517 __map_memblock(pgdp, crashk_res.start, crashk_res.end + 1,
519 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
520 memblock_clear_nomap(crashk_res.start,
521 resource_size(&crashk_res));
526 void mark_rodata_ro(void)
528 unsigned long section_size;
531 * mark .rodata as read only. Use __init_begin rather than __end_rodata
532 * to cover NOTES and EXCEPTION_TABLE.
534 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
535 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata,
536 section_size, PAGE_KERNEL_RO);
541 static void __init map_kernel_segment(pgd_t *pgdp, void *va_start, void *va_end,
542 pgprot_t prot, struct vm_struct *vma,
543 int flags, unsigned long vm_flags)
545 phys_addr_t pa_start = __pa_symbol(va_start);
546 unsigned long size = va_end - va_start;
548 BUG_ON(!PAGE_ALIGNED(pa_start));
549 BUG_ON(!PAGE_ALIGNED(size));
551 __create_pgd_mapping(pgdp, pa_start, (unsigned long)va_start, size, prot,
552 early_pgtable_alloc, flags);
554 if (!(vm_flags & VM_NO_GUARD))
557 vma->addr = va_start;
558 vma->phys_addr = pa_start;
560 vma->flags = VM_MAP | vm_flags;
561 vma->caller = __builtin_return_address(0);
563 vm_area_add_early(vma);
566 static int __init parse_rodata(char *arg)
568 int ret = strtobool(arg, &rodata_enabled);
574 /* permit 'full' in addition to boolean options */
575 if (strcmp(arg, "full"))
578 rodata_enabled = true;
582 early_param("rodata", parse_rodata);
584 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
585 static int __init map_entry_trampoline(void)
587 pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
588 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start);
590 /* The trampoline is always mapped and can therefore be global */
591 pgprot_val(prot) &= ~PTE_NG;
593 /* Map only the text into the trampoline page table */
594 memset(tramp_pg_dir, 0, PGD_SIZE);
595 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, PAGE_SIZE,
596 prot, __pgd_pgtable_alloc, 0);
598 /* Map both the text and data into the kernel page table */
599 __set_fixmap(FIX_ENTRY_TRAMP_TEXT, pa_start, prot);
600 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
601 extern char __entry_tramp_data_start[];
603 __set_fixmap(FIX_ENTRY_TRAMP_DATA,
604 __pa_symbol(__entry_tramp_data_start),
610 core_initcall(map_entry_trampoline);
614 * Open coded check for BTI, only for use to determine configuration
615 * for early mappings for before the cpufeature code has run.
617 static bool arm64_early_this_cpu_has_bti(void)
621 if (!IS_ENABLED(CONFIG_ARM64_BTI_KERNEL))
624 pfr1 = read_sysreg_s(SYS_ID_AA64PFR1_EL1);
625 return cpuid_feature_extract_unsigned_field(pfr1,
626 ID_AA64PFR1_BT_SHIFT);
630 * Create fine-grained mappings for the kernel.
632 static void __init map_kernel(pgd_t *pgdp)
634 static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext,
635 vmlinux_initdata, vmlinux_data;
638 * External debuggers may need to write directly to the text
639 * mapping to install SW breakpoints. Allow this (only) when
640 * explicitly requested with rodata=off.
642 pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
645 * If we have a CPU that supports BTI and a kernel built for
646 * BTI then mark the kernel executable text as guarded pages
647 * now so we don't have to rewrite the page tables later.
649 if (arm64_early_this_cpu_has_bti())
650 text_prot = __pgprot_modify(text_prot, PTE_GP, PTE_GP);
653 * Only rodata will be remapped with different permissions later on,
654 * all other segments are allowed to use contiguous mappings.
656 map_kernel_segment(pgdp, _text, _etext, text_prot, &vmlinux_text, 0,
658 map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL,
659 &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD);
660 map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot,
661 &vmlinux_inittext, 0, VM_NO_GUARD);
662 map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL,
663 &vmlinux_initdata, 0, VM_NO_GUARD);
664 map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0);
666 if (!READ_ONCE(pgd_val(*pgd_offset_pgd(pgdp, FIXADDR_START)))) {
668 * The fixmap falls in a separate pgd to the kernel, and doesn't
669 * live in the carveout for the swapper_pg_dir. We can simply
670 * re-use the existing dir for the fixmap.
672 set_pgd(pgd_offset_pgd(pgdp, FIXADDR_START),
673 READ_ONCE(*pgd_offset_k(FIXADDR_START)));
674 } else if (CONFIG_PGTABLE_LEVELS > 3) {
679 * The fixmap shares its top level pgd entry with the kernel
680 * mapping. This can really only occur when we are running
681 * with 16k/4 levels, so we can simply reuse the pud level
684 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
685 bm_pgdp = pgd_offset_pgd(pgdp, FIXADDR_START);
686 bm_p4dp = p4d_offset(bm_pgdp, FIXADDR_START);
687 bm_pudp = pud_set_fixmap_offset(bm_p4dp, FIXADDR_START);
688 pud_populate(&init_mm, bm_pudp, lm_alias(bm_pmd));
694 kasan_copy_shadow(pgdp);
697 void __init paging_init(void)
699 pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir));
706 cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
707 init_mm.pgd = swapper_pg_dir;
709 memblock_free(__pa_symbol(init_pg_dir),
710 __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir));
712 memblock_allow_resize();
716 * Check whether a kernel address is valid (derived from arch/x86/).
718 int kern_addr_valid(unsigned long addr)
726 addr = arch_kasan_reset_tag(addr);
727 if ((((long)addr) >> VA_BITS) != -1UL)
730 pgdp = pgd_offset_k(addr);
731 if (pgd_none(READ_ONCE(*pgdp)))
734 p4dp = p4d_offset(pgdp, addr);
735 if (p4d_none(READ_ONCE(*p4dp)))
738 pudp = pud_offset(p4dp, addr);
739 pud = READ_ONCE(*pudp);
744 return pfn_valid(pud_pfn(pud));
746 pmdp = pmd_offset(pudp, addr);
747 pmd = READ_ONCE(*pmdp);
752 return pfn_valid(pmd_pfn(pmd));
754 ptep = pte_offset_kernel(pmdp, addr);
755 pte = READ_ONCE(*ptep);
759 return pfn_valid(pte_pfn(pte));
762 #ifdef CONFIG_MEMORY_HOTPLUG
763 static void free_hotplug_page_range(struct page *page, size_t size)
765 WARN_ON(PageReserved(page));
766 free_pages((unsigned long)page_address(page), get_order(size));
769 static void free_hotplug_pgtable_page(struct page *page)
771 free_hotplug_page_range(page, PAGE_SIZE);
774 static bool pgtable_range_aligned(unsigned long start, unsigned long end,
775 unsigned long floor, unsigned long ceiling,
788 if (end - 1 > ceiling - 1)
793 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr,
794 unsigned long end, bool free_mapped)
799 ptep = pte_offset_kernel(pmdp, addr);
800 pte = READ_ONCE(*ptep);
804 WARN_ON(!pte_present(pte));
805 pte_clear(&init_mm, addr, ptep);
806 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
808 free_hotplug_page_range(pte_page(pte), PAGE_SIZE);
809 } while (addr += PAGE_SIZE, addr < end);
812 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr,
813 unsigned long end, bool free_mapped)
819 next = pmd_addr_end(addr, end);
820 pmdp = pmd_offset(pudp, addr);
821 pmd = READ_ONCE(*pmdp);
825 WARN_ON(!pmd_present(pmd));
830 * One TLBI should be sufficient here as the PMD_SIZE
831 * range is mapped with a single block entry.
833 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
835 free_hotplug_page_range(pmd_page(pmd),
839 WARN_ON(!pmd_table(pmd));
840 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped);
841 } while (addr = next, addr < end);
844 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr,
845 unsigned long end, bool free_mapped)
851 next = pud_addr_end(addr, end);
852 pudp = pud_offset(p4dp, addr);
853 pud = READ_ONCE(*pudp);
857 WARN_ON(!pud_present(pud));
862 * One TLBI should be sufficient here as the PUD_SIZE
863 * range is mapped with a single block entry.
865 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
867 free_hotplug_page_range(pud_page(pud),
871 WARN_ON(!pud_table(pud));
872 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped);
873 } while (addr = next, addr < end);
876 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr,
877 unsigned long end, bool free_mapped)
883 next = p4d_addr_end(addr, end);
884 p4dp = p4d_offset(pgdp, addr);
885 p4d = READ_ONCE(*p4dp);
889 WARN_ON(!p4d_present(p4d));
890 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped);
891 } while (addr = next, addr < end);
894 static void unmap_hotplug_range(unsigned long addr, unsigned long end,
901 next = pgd_addr_end(addr, end);
902 pgdp = pgd_offset_k(addr);
903 pgd = READ_ONCE(*pgdp);
907 WARN_ON(!pgd_present(pgd));
908 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped);
909 } while (addr = next, addr < end);
912 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
913 unsigned long end, unsigned long floor,
914 unsigned long ceiling)
917 unsigned long i, start = addr;
920 ptep = pte_offset_kernel(pmdp, addr);
921 pte = READ_ONCE(*ptep);
924 * This is just a sanity check here which verifies that
925 * pte clearing has been done by earlier unmap loops.
927 WARN_ON(!pte_none(pte));
928 } while (addr += PAGE_SIZE, addr < end);
930 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK))
934 * Check whether we can free the pte page if the rest of the
935 * entries are empty. Overlap with other regions have been
936 * handled by the floor/ceiling check.
938 ptep = pte_offset_kernel(pmdp, 0UL);
939 for (i = 0; i < PTRS_PER_PTE; i++) {
940 if (!pte_none(READ_ONCE(ptep[i])))
945 __flush_tlb_kernel_pgtable(start);
946 free_hotplug_pgtable_page(virt_to_page(ptep));
949 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
950 unsigned long end, unsigned long floor,
951 unsigned long ceiling)
954 unsigned long i, next, start = addr;
957 next = pmd_addr_end(addr, end);
958 pmdp = pmd_offset(pudp, addr);
959 pmd = READ_ONCE(*pmdp);
963 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd));
964 free_empty_pte_table(pmdp, addr, next, floor, ceiling);
965 } while (addr = next, addr < end);
967 if (CONFIG_PGTABLE_LEVELS <= 2)
970 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK))
974 * Check whether we can free the pmd page if the rest of the
975 * entries are empty. Overlap with other regions have been
976 * handled by the floor/ceiling check.
978 pmdp = pmd_offset(pudp, 0UL);
979 for (i = 0; i < PTRS_PER_PMD; i++) {
980 if (!pmd_none(READ_ONCE(pmdp[i])))
985 __flush_tlb_kernel_pgtable(start);
986 free_hotplug_pgtable_page(virt_to_page(pmdp));
989 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr,
990 unsigned long end, unsigned long floor,
991 unsigned long ceiling)
994 unsigned long i, next, start = addr;
997 next = pud_addr_end(addr, end);
998 pudp = pud_offset(p4dp, addr);
999 pud = READ_ONCE(*pudp);
1003 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud));
1004 free_empty_pmd_table(pudp, addr, next, floor, ceiling);
1005 } while (addr = next, addr < end);
1007 if (CONFIG_PGTABLE_LEVELS <= 3)
1010 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK))
1014 * Check whether we can free the pud page if the rest of the
1015 * entries are empty. Overlap with other regions have been
1016 * handled by the floor/ceiling check.
1018 pudp = pud_offset(p4dp, 0UL);
1019 for (i = 0; i < PTRS_PER_PUD; i++) {
1020 if (!pud_none(READ_ONCE(pudp[i])))
1025 __flush_tlb_kernel_pgtable(start);
1026 free_hotplug_pgtable_page(virt_to_page(pudp));
1029 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
1030 unsigned long end, unsigned long floor,
1031 unsigned long ceiling)
1037 next = p4d_addr_end(addr, end);
1038 p4dp = p4d_offset(pgdp, addr);
1039 p4d = READ_ONCE(*p4dp);
1043 WARN_ON(!p4d_present(p4d));
1044 free_empty_pud_table(p4dp, addr, next, floor, ceiling);
1045 } while (addr = next, addr < end);
1048 static void free_empty_tables(unsigned long addr, unsigned long end,
1049 unsigned long floor, unsigned long ceiling)
1055 next = pgd_addr_end(addr, end);
1056 pgdp = pgd_offset_k(addr);
1057 pgd = READ_ONCE(*pgdp);
1061 WARN_ON(!pgd_present(pgd));
1062 free_empty_p4d_table(pgdp, addr, next, floor, ceiling);
1063 } while (addr = next, addr < end);
1067 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1068 #if !ARM64_SWAPPER_USES_SECTION_MAPS
1069 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1070 struct vmem_altmap *altmap)
1072 return vmemmap_populate_basepages(start, end, node);
1074 #else /* !ARM64_SWAPPER_USES_SECTION_MAPS */
1075 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1076 struct vmem_altmap *altmap)
1078 unsigned long addr = start;
1086 next = pmd_addr_end(addr, end);
1088 pgdp = vmemmap_pgd_populate(addr, node);
1092 p4dp = vmemmap_p4d_populate(pgdp, addr, node);
1096 pudp = vmemmap_pud_populate(p4dp, addr, node);
1100 pmdp = pmd_offset(pudp, addr);
1101 if (pmd_none(READ_ONCE(*pmdp))) {
1104 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1108 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL));
1110 vmemmap_verify((pte_t *)pmdp, node, addr, next);
1111 } while (addr = next, addr != end);
1115 #endif /* !ARM64_SWAPPER_USES_SECTION_MAPS */
1116 void vmemmap_free(unsigned long start, unsigned long end,
1117 struct vmem_altmap *altmap)
1119 #ifdef CONFIG_MEMORY_HOTPLUG
1120 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1122 unmap_hotplug_range(start, end, true);
1123 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END);
1126 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
1128 static inline pud_t * fixmap_pud(unsigned long addr)
1130 pgd_t *pgdp = pgd_offset_k(addr);
1131 p4d_t *p4dp = p4d_offset(pgdp, addr);
1132 p4d_t p4d = READ_ONCE(*p4dp);
1134 BUG_ON(p4d_none(p4d) || p4d_bad(p4d));
1136 return pud_offset_kimg(p4dp, addr);
1139 static inline pmd_t * fixmap_pmd(unsigned long addr)
1141 pud_t *pudp = fixmap_pud(addr);
1142 pud_t pud = READ_ONCE(*pudp);
1144 BUG_ON(pud_none(pud) || pud_bad(pud));
1146 return pmd_offset_kimg(pudp, addr);
1149 static inline pte_t * fixmap_pte(unsigned long addr)
1151 return &bm_pte[pte_index(addr)];
1155 * The p*d_populate functions call virt_to_phys implicitly so they can't be used
1156 * directly on kernel symbols (bm_p*d). This function is called too early to use
1157 * lm_alias so __p*d_populate functions must be used to populate with the
1158 * physical address from __pa_symbol.
1160 void __init early_fixmap_init(void)
1166 unsigned long addr = FIXADDR_START;
1168 pgdp = pgd_offset_k(addr);
1169 p4dp = p4d_offset(pgdp, addr);
1170 p4d = READ_ONCE(*p4dp);
1171 if (CONFIG_PGTABLE_LEVELS > 3 &&
1172 !(p4d_none(p4d) || p4d_page_paddr(p4d) == __pa_symbol(bm_pud))) {
1174 * We only end up here if the kernel mapping and the fixmap
1175 * share the top level pgd entry, which should only happen on
1176 * 16k/4 levels configurations.
1178 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
1179 pudp = pud_offset_kimg(p4dp, addr);
1182 __p4d_populate(p4dp, __pa_symbol(bm_pud), PUD_TYPE_TABLE);
1183 pudp = fixmap_pud(addr);
1185 if (pud_none(READ_ONCE(*pudp)))
1186 __pud_populate(pudp, __pa_symbol(bm_pmd), PMD_TYPE_TABLE);
1187 pmdp = fixmap_pmd(addr);
1188 __pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE);
1191 * The boot-ioremap range spans multiple pmds, for which
1192 * we are not prepared:
1194 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
1195 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
1197 if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
1198 || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
1200 pr_warn("pmdp %p != %p, %p\n",
1201 pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
1202 fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
1203 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
1204 fix_to_virt(FIX_BTMAP_BEGIN));
1205 pr_warn("fix_to_virt(FIX_BTMAP_END): %08lx\n",
1206 fix_to_virt(FIX_BTMAP_END));
1208 pr_warn("FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
1209 pr_warn("FIX_BTMAP_BEGIN: %d\n", FIX_BTMAP_BEGIN);
1214 * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we
1215 * ever need to use IPIs for TLB broadcasting, then we're in trouble here.
1217 void __set_fixmap(enum fixed_addresses idx,
1218 phys_addr_t phys, pgprot_t flags)
1220 unsigned long addr = __fix_to_virt(idx);
1223 BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
1225 ptep = fixmap_pte(addr);
1227 if (pgprot_val(flags)) {
1228 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags));
1230 pte_clear(&init_mm, addr, ptep);
1231 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
1235 void *__init fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
1237 const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
1242 * Check whether the physical FDT address is set and meets the minimum
1243 * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
1244 * at least 8 bytes so that we can always access the magic and size
1245 * fields of the FDT header after mapping the first chunk, double check
1246 * here if that is indeed the case.
1248 BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
1249 if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
1253 * Make sure that the FDT region can be mapped without the need to
1254 * allocate additional translation table pages, so that it is safe
1255 * to call create_mapping_noalloc() this early.
1257 * On 64k pages, the FDT will be mapped using PTEs, so we need to
1258 * be in the same PMD as the rest of the fixmap.
1259 * On 4k pages, we'll use section mappings for the FDT so we only
1260 * have to be in the same PUD.
1262 BUILD_BUG_ON(dt_virt_base % SZ_2M);
1264 BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
1265 __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
1267 offset = dt_phys % SWAPPER_BLOCK_SIZE;
1268 dt_virt = (void *)dt_virt_base + offset;
1270 /* map the first chunk so we can read the size from the header */
1271 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
1272 dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
1274 if (fdt_magic(dt_virt) != FDT_MAGIC)
1277 *size = fdt_totalsize(dt_virt);
1278 if (*size > MAX_FDT_SIZE)
1281 if (offset + *size > SWAPPER_BLOCK_SIZE)
1282 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
1283 round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
1288 int __init arch_ioremap_p4d_supported(void)
1293 int __init arch_ioremap_pud_supported(void)
1296 * Only 4k granule supports level 1 block mappings.
1297 * SW table walks can't handle removal of intermediate entries.
1299 return IS_ENABLED(CONFIG_ARM64_4K_PAGES) &&
1300 !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
1303 int __init arch_ioremap_pmd_supported(void)
1305 /* See arch_ioremap_pud_supported() */
1306 return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
1309 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
1311 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot));
1313 /* Only allow permission changes for now */
1314 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)),
1318 VM_BUG_ON(phys & ~PUD_MASK);
1319 set_pud(pudp, new_pud);
1323 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
1325 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot));
1327 /* Only allow permission changes for now */
1328 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)),
1332 VM_BUG_ON(phys & ~PMD_MASK);
1333 set_pmd(pmdp, new_pmd);
1337 int pud_clear_huge(pud_t *pudp)
1339 if (!pud_sect(READ_ONCE(*pudp)))
1345 int pmd_clear_huge(pmd_t *pmdp)
1347 if (!pmd_sect(READ_ONCE(*pmdp)))
1353 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
1358 pmd = READ_ONCE(*pmdp);
1360 if (!pmd_table(pmd)) {
1365 table = pte_offset_kernel(pmdp, addr);
1367 __flush_tlb_kernel_pgtable(addr);
1368 pte_free_kernel(NULL, table);
1372 int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
1377 unsigned long next, end;
1379 pud = READ_ONCE(*pudp);
1381 if (!pud_table(pud)) {
1386 table = pmd_offset(pudp, addr);
1389 end = addr + PUD_SIZE;
1391 pmd_free_pte_page(pmdp, next);
1392 } while (pmdp++, next += PMD_SIZE, next != end);
1395 __flush_tlb_kernel_pgtable(addr);
1396 pmd_free(NULL, table);
1400 int p4d_free_pud_page(p4d_t *p4d, unsigned long addr)
1402 return 0; /* Don't attempt a block mapping */
1405 #ifdef CONFIG_MEMORY_HOTPLUG
1406 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size)
1408 unsigned long end = start + size;
1410 WARN_ON(pgdir != init_mm.pgd);
1411 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END));
1413 unmap_hotplug_range(start, end, false);
1414 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END);
1417 int arch_add_memory(int nid, u64 start, u64 size,
1418 struct mhp_params *params)
1422 if (rodata_full || debug_pagealloc_enabled())
1423 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
1425 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
1426 size, params->pgprot, __pgd_pgtable_alloc,
1429 memblock_clear_nomap(start, size);
1431 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT,
1434 __remove_pgd_mapping(swapper_pg_dir,
1435 __phys_to_virt(start), size);
1439 void arch_remove_memory(int nid, u64 start, u64 size,
1440 struct vmem_altmap *altmap)
1442 unsigned long start_pfn = start >> PAGE_SHIFT;
1443 unsigned long nr_pages = size >> PAGE_SHIFT;
1445 __remove_pages(start_pfn, nr_pages, altmap);
1446 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size);
1450 * This memory hotplug notifier helps prevent boot memory from being
1451 * inadvertently removed as it blocks pfn range offlining process in
1452 * __offline_pages(). Hence this prevents both offlining as well as
1453 * removal process for boot memory which is initially always online.
1454 * In future if and when boot memory could be removed, this notifier
1455 * should be dropped and free_hotplug_page_range() should handle any
1456 * reserved pages allocated during boot.
1458 static int prevent_bootmem_remove_notifier(struct notifier_block *nb,
1459 unsigned long action, void *data)
1461 struct mem_section *ms;
1462 struct memory_notify *arg = data;
1463 unsigned long end_pfn = arg->start_pfn + arg->nr_pages;
1464 unsigned long pfn = arg->start_pfn;
1466 if (action != MEM_GOING_OFFLINE)
1469 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
1470 ms = __pfn_to_section(pfn);
1471 if (early_section(ms))
1477 static struct notifier_block prevent_bootmem_remove_nb = {
1478 .notifier_call = prevent_bootmem_remove_notifier,
1481 static int __init prevent_bootmem_remove_init(void)
1483 return register_memory_notifier(&prevent_bootmem_remove_nb);
1485 device_initcall(prevent_bootmem_remove_init);