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>
25 #include <linux/set_memory.h>
27 #include <asm/barrier.h>
28 #include <asm/cputype.h>
29 #include <asm/fixmap.h>
30 #include <asm/kasan.h>
31 #include <asm/kernel-pgtable.h>
32 #include <asm/sections.h>
33 #include <asm/setup.h>
34 #include <linux/sizes.h>
36 #include <asm/mmu_context.h>
37 #include <asm/ptdump.h>
38 #include <asm/tlbflush.h>
39 #include <asm/pgalloc.h>
41 #define NO_BLOCK_MAPPINGS BIT(0)
42 #define NO_CONT_MAPPINGS BIT(1)
43 #define NO_EXEC_MAPPINGS BIT(2) /* assumes FEAT_HPDS is not used */
45 u64 idmap_t0sz = TCR_T0SZ(VA_BITS_MIN);
46 u64 idmap_ptrs_per_pgd = PTRS_PER_PGD;
48 u64 __section(".mmuoff.data.write") vabits_actual;
49 EXPORT_SYMBOL(vabits_actual);
51 u64 kimage_voffset __ro_after_init;
52 EXPORT_SYMBOL(kimage_voffset);
55 * Empty_zero_page is a special page that is used for zero-initialized data
58 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
59 EXPORT_SYMBOL(empty_zero_page);
61 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
62 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused;
63 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused;
65 static DEFINE_SPINLOCK(swapper_pgdir_lock);
66 static DEFINE_MUTEX(fixmap_lock);
68 void set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
72 spin_lock(&swapper_pgdir_lock);
73 fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp));
74 WRITE_ONCE(*fixmap_pgdp, pgd);
76 * We need dsb(ishst) here to ensure the page-table-walker sees
77 * our new entry before set_p?d() returns. The fixmap's
78 * flush_tlb_kernel_range() via clear_fixmap() does this for us.
81 spin_unlock(&swapper_pgdir_lock);
84 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
85 unsigned long size, pgprot_t vma_prot)
87 if (!pfn_is_map_memory(pfn))
88 return pgprot_noncached(vma_prot);
89 else if (file->f_flags & O_SYNC)
90 return pgprot_writecombine(vma_prot);
93 EXPORT_SYMBOL(phys_mem_access_prot);
95 static phys_addr_t __init early_pgtable_alloc(int shift)
100 phys = memblock_phys_alloc_range(PAGE_SIZE, PAGE_SIZE, 0,
101 MEMBLOCK_ALLOC_NOLEAKTRACE);
103 panic("Failed to allocate page table page\n");
106 * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
107 * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
108 * any level of table.
110 ptr = pte_set_fixmap(phys);
112 memset(ptr, 0, PAGE_SIZE);
115 * Implicit barriers also ensure the zeroed page is visible to the page
123 static bool pgattr_change_is_safe(u64 old, u64 new)
126 * The following mapping attributes may be updated in live
127 * kernel mappings without the need for break-before-make.
129 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG;
131 /* creating or taking down mappings is always safe */
132 if (old == 0 || new == 0)
135 /* live contiguous mappings may not be manipulated at all */
136 if ((old | new) & PTE_CONT)
139 /* Transitioning from Non-Global to Global is unsafe */
140 if (old & ~new & PTE_NG)
144 * Changing the memory type between Normal and Normal-Tagged is safe
145 * since Tagged is considered a permission attribute from the
146 * mismatched attribute aliases perspective.
148 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
149 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) &&
150 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
151 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)))
152 mask |= PTE_ATTRINDX_MASK;
154 return ((old ^ new) & ~mask) == 0;
157 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end,
158 phys_addr_t phys, pgprot_t prot)
162 ptep = pte_set_fixmap_offset(pmdp, addr);
164 pte_t old_pte = READ_ONCE(*ptep);
166 set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot));
169 * After the PTE entry has been populated once, we
170 * only allow updates to the permission attributes.
172 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte),
173 READ_ONCE(pte_val(*ptep))));
176 } while (ptep++, addr += PAGE_SIZE, addr != end);
181 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
182 unsigned long end, phys_addr_t phys,
184 phys_addr_t (*pgtable_alloc)(int),
188 pmd_t pmd = READ_ONCE(*pmdp);
190 BUG_ON(pmd_sect(pmd));
192 pmdval_t pmdval = PMD_TYPE_TABLE | PMD_TABLE_UXN;
193 phys_addr_t pte_phys;
195 if (flags & NO_EXEC_MAPPINGS)
196 pmdval |= PMD_TABLE_PXN;
197 BUG_ON(!pgtable_alloc);
198 pte_phys = pgtable_alloc(PAGE_SHIFT);
199 __pmd_populate(pmdp, pte_phys, pmdval);
200 pmd = READ_ONCE(*pmdp);
202 BUG_ON(pmd_bad(pmd));
205 pgprot_t __prot = prot;
207 next = pte_cont_addr_end(addr, end);
209 /* use a contiguous mapping if the range is suitably aligned */
210 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) &&
211 (flags & NO_CONT_MAPPINGS) == 0)
212 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
214 init_pte(pmdp, addr, next, phys, __prot);
217 } while (addr = next, addr != end);
220 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end,
221 phys_addr_t phys, pgprot_t prot,
222 phys_addr_t (*pgtable_alloc)(int), int flags)
227 pmdp = pmd_set_fixmap_offset(pudp, addr);
229 pmd_t old_pmd = READ_ONCE(*pmdp);
231 next = pmd_addr_end(addr, end);
233 /* try section mapping first */
234 if (((addr | next | phys) & ~PMD_MASK) == 0 &&
235 (flags & NO_BLOCK_MAPPINGS) == 0) {
236 pmd_set_huge(pmdp, phys, prot);
239 * After the PMD entry has been populated once, we
240 * only allow updates to the permission attributes.
242 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
243 READ_ONCE(pmd_val(*pmdp))));
245 alloc_init_cont_pte(pmdp, addr, next, phys, prot,
246 pgtable_alloc, flags);
248 BUG_ON(pmd_val(old_pmd) != 0 &&
249 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp)));
252 } while (pmdp++, addr = next, addr != end);
257 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
258 unsigned long end, phys_addr_t phys,
260 phys_addr_t (*pgtable_alloc)(int), int flags)
263 pud_t pud = READ_ONCE(*pudp);
266 * Check for initial section mappings in the pgd/pud.
268 BUG_ON(pud_sect(pud));
270 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN;
271 phys_addr_t pmd_phys;
273 if (flags & NO_EXEC_MAPPINGS)
274 pudval |= PUD_TABLE_PXN;
275 BUG_ON(!pgtable_alloc);
276 pmd_phys = pgtable_alloc(PMD_SHIFT);
277 __pud_populate(pudp, pmd_phys, pudval);
278 pud = READ_ONCE(*pudp);
280 BUG_ON(pud_bad(pud));
283 pgprot_t __prot = prot;
285 next = pmd_cont_addr_end(addr, end);
287 /* use a contiguous mapping if the range is suitably aligned */
288 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) &&
289 (flags & NO_CONT_MAPPINGS) == 0)
290 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
292 init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags);
295 } while (addr = next, addr != end);
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 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN;
310 phys_addr_t pud_phys;
312 if (flags & NO_EXEC_MAPPINGS)
313 p4dval |= P4D_TABLE_PXN;
314 BUG_ON(!pgtable_alloc);
315 pud_phys = pgtable_alloc(PUD_SHIFT);
316 __p4d_populate(p4dp, pud_phys, p4dval);
317 p4d = READ_ONCE(*p4dp);
319 BUG_ON(p4d_bad(p4d));
322 * No need for locking during early boot. And it doesn't work as
323 * expected with KASLR enabled.
325 if (system_state != SYSTEM_BOOTING)
326 mutex_lock(&fixmap_lock);
327 pudp = pud_set_fixmap_offset(p4dp, addr);
329 pud_t old_pud = READ_ONCE(*pudp);
331 next = pud_addr_end(addr, end);
334 * For 4K granule only, attempt to put down a 1GB block
336 if (pud_sect_supported() &&
337 ((addr | next | phys) & ~PUD_MASK) == 0 &&
338 (flags & NO_BLOCK_MAPPINGS) == 0) {
339 pud_set_huge(pudp, phys, prot);
342 * After the PUD entry has been populated once, we
343 * only allow updates to the permission attributes.
345 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
346 READ_ONCE(pud_val(*pudp))));
348 alloc_init_cont_pmd(pudp, addr, next, phys, prot,
349 pgtable_alloc, flags);
351 BUG_ON(pud_val(old_pud) != 0 &&
352 pud_val(old_pud) != READ_ONCE(pud_val(*pudp)));
355 } while (pudp++, addr = next, addr != end);
358 if (system_state != SYSTEM_BOOTING)
359 mutex_unlock(&fixmap_lock);
362 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
363 unsigned long virt, phys_addr_t size,
365 phys_addr_t (*pgtable_alloc)(int),
368 unsigned long addr, end, next;
369 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt);
372 * If the virtual and physical address don't have the same offset
373 * within a page, we cannot map the region as the caller expects.
375 if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
379 addr = virt & PAGE_MASK;
380 end = PAGE_ALIGN(virt + size);
383 next = pgd_addr_end(addr, end);
384 alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc,
387 } while (pgdp++, addr = next, addr != end);
390 static phys_addr_t __pgd_pgtable_alloc(int shift)
392 void *ptr = (void *)__get_free_page(GFP_PGTABLE_KERNEL);
395 /* Ensure the zeroed page is visible to the page table walker */
400 static phys_addr_t pgd_pgtable_alloc(int shift)
402 phys_addr_t pa = __pgd_pgtable_alloc(shift);
405 * Call proper page table ctor in case later we need to
406 * call core mm functions like apply_to_page_range() on
407 * this pre-allocated page table.
409 * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is
410 * folded, and if so pgtable_pmd_page_ctor() becomes nop.
412 if (shift == PAGE_SHIFT)
413 BUG_ON(!pgtable_pte_page_ctor(phys_to_page(pa)));
414 else if (shift == PMD_SHIFT)
415 BUG_ON(!pgtable_pmd_page_ctor(phys_to_page(pa)));
421 * This function can only be used to modify existing table entries,
422 * without allocating new levels of table. Note that this permits the
423 * creation of new section or page entries.
425 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
426 phys_addr_t size, pgprot_t prot)
428 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
429 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
433 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
437 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
438 unsigned long virt, phys_addr_t size,
439 pgprot_t prot, bool page_mappings_only)
443 BUG_ON(mm == &init_mm);
445 if (page_mappings_only)
446 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
448 __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
449 pgd_pgtable_alloc, flags);
452 static void update_mapping_prot(phys_addr_t phys, unsigned long virt,
453 phys_addr_t size, pgprot_t prot)
455 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
456 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n",
461 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
464 /* flush the TLBs after updating live kernel mappings */
465 flush_tlb_kernel_range(virt, virt + size);
468 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start,
469 phys_addr_t end, pgprot_t prot, int flags)
471 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start,
472 prot, early_pgtable_alloc, flags);
475 void __init mark_linear_text_alias_ro(void)
478 * Remove the write permissions from the linear alias of .text/.rodata
480 update_mapping_prot(__pa_symbol(_stext), (unsigned long)lm_alias(_stext),
481 (unsigned long)__init_begin - (unsigned long)_stext,
485 static bool crash_mem_map __initdata;
487 static int __init enable_crash_mem_map(char *arg)
490 * Proper parameter parsing is done by reserve_crashkernel(). We only
491 * need to know if the linear map has to avoid block mappings so that
492 * the crashkernel reservations can be unmapped later.
494 crash_mem_map = true;
498 early_param("crashkernel", enable_crash_mem_map);
500 static void __init map_mem(pgd_t *pgdp)
502 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN);
503 phys_addr_t kernel_start = __pa_symbol(_stext);
504 phys_addr_t kernel_end = __pa_symbol(__init_begin);
505 phys_addr_t start, end;
506 int flags = NO_EXEC_MAPPINGS;
510 * Setting hierarchical PXNTable attributes on table entries covering
511 * the linear region is only possible if it is guaranteed that no table
512 * entries at any level are being shared between the linear region and
513 * the vmalloc region. Check whether this is true for the PGD level, in
514 * which case it is guaranteed to be true for all other levels as well.
516 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end));
518 if (can_set_direct_map() || IS_ENABLED(CONFIG_KFENCE))
519 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
522 * Take care not to create a writable alias for the
523 * read-only text and rodata sections of the kernel image.
524 * So temporarily mark them as NOMAP to skip mappings in
525 * the following for-loop
527 memblock_mark_nomap(kernel_start, kernel_end - kernel_start);
529 #ifdef CONFIG_KEXEC_CORE
531 if (IS_ENABLED(CONFIG_ZONE_DMA) ||
532 IS_ENABLED(CONFIG_ZONE_DMA32))
533 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
534 else if (crashk_res.end)
535 memblock_mark_nomap(crashk_res.start,
536 resource_size(&crashk_res));
540 /* map all the memory banks */
541 for_each_mem_range(i, &start, &end) {
545 * The linear map must allow allocation tags reading/writing
546 * if MTE is present. Otherwise, it has the same attributes as
549 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL),
554 * Map the linear alias of the [_stext, __init_begin) interval
555 * as non-executable now, and remove the write permission in
556 * mark_linear_text_alias_ro() below (which will be called after
557 * alternative patching has completed). This makes the contents
558 * of the region accessible to subsystems such as hibernate,
559 * but protects it from inadvertent modification or execution.
560 * Note that contiguous mappings cannot be remapped in this way,
561 * so we should avoid them here.
563 __map_memblock(pgdp, kernel_start, kernel_end,
564 PAGE_KERNEL, NO_CONT_MAPPINGS);
565 memblock_clear_nomap(kernel_start, kernel_end - kernel_start);
568 * Use page-level mappings here so that we can shrink the region
569 * in page granularity and put back unused memory to buddy system
570 * through /sys/kernel/kexec_crash_size interface.
572 #ifdef CONFIG_KEXEC_CORE
574 !IS_ENABLED(CONFIG_ZONE_DMA) && !IS_ENABLED(CONFIG_ZONE_DMA32)) {
575 if (crashk_res.end) {
576 __map_memblock(pgdp, crashk_res.start,
579 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
580 memblock_clear_nomap(crashk_res.start,
581 resource_size(&crashk_res));
587 void mark_rodata_ro(void)
589 unsigned long section_size;
592 * mark .rodata as read only. Use __init_begin rather than __end_rodata
593 * to cover NOTES and EXCEPTION_TABLE.
595 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
596 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata,
597 section_size, PAGE_KERNEL_RO);
602 static void __init map_kernel_segment(pgd_t *pgdp, void *va_start, void *va_end,
603 pgprot_t prot, struct vm_struct *vma,
604 int flags, unsigned long vm_flags)
606 phys_addr_t pa_start = __pa_symbol(va_start);
607 unsigned long size = va_end - va_start;
609 BUG_ON(!PAGE_ALIGNED(pa_start));
610 BUG_ON(!PAGE_ALIGNED(size));
612 __create_pgd_mapping(pgdp, pa_start, (unsigned long)va_start, size, prot,
613 early_pgtable_alloc, flags);
615 if (!(vm_flags & VM_NO_GUARD))
618 vma->addr = va_start;
619 vma->phys_addr = pa_start;
621 vma->flags = VM_MAP | vm_flags;
622 vma->caller = __builtin_return_address(0);
624 vm_area_add_early(vma);
627 static int __init parse_rodata(char *arg)
629 int ret = strtobool(arg, &rodata_enabled);
635 /* permit 'full' in addition to boolean options */
636 if (strcmp(arg, "full"))
639 rodata_enabled = true;
643 early_param("rodata", parse_rodata);
645 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
646 static int __init map_entry_trampoline(void)
650 pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
651 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start);
653 /* The trampoline is always mapped and can therefore be global */
654 pgprot_val(prot) &= ~PTE_NG;
656 /* Map only the text into the trampoline page table */
657 memset(tramp_pg_dir, 0, PGD_SIZE);
658 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS,
659 entry_tramp_text_size(), prot,
660 __pgd_pgtable_alloc, NO_BLOCK_MAPPINGS);
662 /* Map both the text and data into the kernel page table */
663 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++)
664 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i,
665 pa_start + i * PAGE_SIZE, prot);
667 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
668 extern char __entry_tramp_data_start[];
670 __set_fixmap(FIX_ENTRY_TRAMP_DATA,
671 __pa_symbol(__entry_tramp_data_start),
677 core_initcall(map_entry_trampoline);
681 * Open coded check for BTI, only for use to determine configuration
682 * for early mappings for before the cpufeature code has run.
684 static bool arm64_early_this_cpu_has_bti(void)
688 if (!IS_ENABLED(CONFIG_ARM64_BTI_KERNEL))
691 pfr1 = __read_sysreg_by_encoding(SYS_ID_AA64PFR1_EL1);
692 return cpuid_feature_extract_unsigned_field(pfr1,
693 ID_AA64PFR1_BT_SHIFT);
697 * Create fine-grained mappings for the kernel.
699 static void __init map_kernel(pgd_t *pgdp)
701 static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext,
702 vmlinux_initdata, vmlinux_data;
705 * External debuggers may need to write directly to the text
706 * mapping to install SW breakpoints. Allow this (only) when
707 * explicitly requested with rodata=off.
709 pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
712 * If we have a CPU that supports BTI and a kernel built for
713 * BTI then mark the kernel executable text as guarded pages
714 * now so we don't have to rewrite the page tables later.
716 if (arm64_early_this_cpu_has_bti())
717 text_prot = __pgprot_modify(text_prot, PTE_GP, PTE_GP);
720 * Only rodata will be remapped with different permissions later on,
721 * all other segments are allowed to use contiguous mappings.
723 map_kernel_segment(pgdp, _stext, _etext, text_prot, &vmlinux_text, 0,
725 map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL,
726 &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD);
727 map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot,
728 &vmlinux_inittext, 0, VM_NO_GUARD);
729 map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL,
730 &vmlinux_initdata, 0, VM_NO_GUARD);
731 map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0);
733 if (!READ_ONCE(pgd_val(*pgd_offset_pgd(pgdp, FIXADDR_START)))) {
735 * The fixmap falls in a separate pgd to the kernel, and doesn't
736 * live in the carveout for the swapper_pg_dir. We can simply
737 * re-use the existing dir for the fixmap.
739 set_pgd(pgd_offset_pgd(pgdp, FIXADDR_START),
740 READ_ONCE(*pgd_offset_k(FIXADDR_START)));
741 } else if (CONFIG_PGTABLE_LEVELS > 3) {
746 * The fixmap shares its top level pgd entry with the kernel
747 * mapping. This can really only occur when we are running
748 * with 16k/4 levels, so we can simply reuse the pud level
751 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
752 bm_pgdp = pgd_offset_pgd(pgdp, FIXADDR_START);
753 bm_p4dp = p4d_offset(bm_pgdp, FIXADDR_START);
754 bm_pudp = pud_set_fixmap_offset(bm_p4dp, FIXADDR_START);
755 pud_populate(&init_mm, bm_pudp, lm_alias(bm_pmd));
761 kasan_copy_shadow(pgdp);
764 void __init paging_init(void)
766 pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir));
773 cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
774 init_mm.pgd = swapper_pg_dir;
776 memblock_phys_free(__pa_symbol(init_pg_dir),
777 __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir));
779 memblock_allow_resize();
783 * Check whether a kernel address is valid (derived from arch/x86/).
785 int kern_addr_valid(unsigned long addr)
793 addr = arch_kasan_reset_tag(addr);
794 if ((((long)addr) >> VA_BITS) != -1UL)
797 pgdp = pgd_offset_k(addr);
798 if (pgd_none(READ_ONCE(*pgdp)))
801 p4dp = p4d_offset(pgdp, addr);
802 if (p4d_none(READ_ONCE(*p4dp)))
805 pudp = pud_offset(p4dp, addr);
806 pud = READ_ONCE(*pudp);
811 return pfn_valid(pud_pfn(pud));
813 pmdp = pmd_offset(pudp, addr);
814 pmd = READ_ONCE(*pmdp);
819 return pfn_valid(pmd_pfn(pmd));
821 ptep = pte_offset_kernel(pmdp, addr);
822 pte = READ_ONCE(*ptep);
826 return pfn_valid(pte_pfn(pte));
829 #ifdef CONFIG_MEMORY_HOTPLUG
830 static void free_hotplug_page_range(struct page *page, size_t size,
831 struct vmem_altmap *altmap)
834 vmem_altmap_free(altmap, size >> PAGE_SHIFT);
836 WARN_ON(PageReserved(page));
837 free_pages((unsigned long)page_address(page), get_order(size));
841 static void free_hotplug_pgtable_page(struct page *page)
843 free_hotplug_page_range(page, PAGE_SIZE, NULL);
846 static bool pgtable_range_aligned(unsigned long start, unsigned long end,
847 unsigned long floor, unsigned long ceiling,
860 if (end - 1 > ceiling - 1)
865 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr,
866 unsigned long end, bool free_mapped,
867 struct vmem_altmap *altmap)
872 ptep = pte_offset_kernel(pmdp, addr);
873 pte = READ_ONCE(*ptep);
877 WARN_ON(!pte_present(pte));
878 pte_clear(&init_mm, addr, ptep);
879 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
881 free_hotplug_page_range(pte_page(pte),
883 } while (addr += PAGE_SIZE, addr < end);
886 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr,
887 unsigned long end, bool free_mapped,
888 struct vmem_altmap *altmap)
894 next = pmd_addr_end(addr, end);
895 pmdp = pmd_offset(pudp, addr);
896 pmd = READ_ONCE(*pmdp);
900 WARN_ON(!pmd_present(pmd));
905 * One TLBI should be sufficient here as the PMD_SIZE
906 * range is mapped with a single block entry.
908 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
910 free_hotplug_page_range(pmd_page(pmd),
914 WARN_ON(!pmd_table(pmd));
915 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap);
916 } while (addr = next, addr < end);
919 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr,
920 unsigned long end, bool free_mapped,
921 struct vmem_altmap *altmap)
927 next = pud_addr_end(addr, end);
928 pudp = pud_offset(p4dp, addr);
929 pud = READ_ONCE(*pudp);
933 WARN_ON(!pud_present(pud));
938 * One TLBI should be sufficient here as the PUD_SIZE
939 * range is mapped with a single block entry.
941 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
943 free_hotplug_page_range(pud_page(pud),
947 WARN_ON(!pud_table(pud));
948 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap);
949 } while (addr = next, addr < end);
952 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr,
953 unsigned long end, bool free_mapped,
954 struct vmem_altmap *altmap)
960 next = p4d_addr_end(addr, end);
961 p4dp = p4d_offset(pgdp, addr);
962 p4d = READ_ONCE(*p4dp);
966 WARN_ON(!p4d_present(p4d));
967 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap);
968 } while (addr = next, addr < end);
971 static void unmap_hotplug_range(unsigned long addr, unsigned long end,
972 bool free_mapped, struct vmem_altmap *altmap)
978 * altmap can only be used as vmemmap mapping backing memory.
979 * In case the backing memory itself is not being freed, then
980 * altmap is irrelevant. Warn about this inconsistency when
983 WARN_ON(!free_mapped && altmap);
986 next = pgd_addr_end(addr, end);
987 pgdp = pgd_offset_k(addr);
988 pgd = READ_ONCE(*pgdp);
992 WARN_ON(!pgd_present(pgd));
993 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap);
994 } while (addr = next, addr < end);
997 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
998 unsigned long end, unsigned long floor,
999 unsigned long ceiling)
1002 unsigned long i, start = addr;
1005 ptep = pte_offset_kernel(pmdp, addr);
1006 pte = READ_ONCE(*ptep);
1009 * This is just a sanity check here which verifies that
1010 * pte clearing has been done by earlier unmap loops.
1012 WARN_ON(!pte_none(pte));
1013 } while (addr += PAGE_SIZE, addr < end);
1015 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK))
1019 * Check whether we can free the pte page if the rest of the
1020 * entries are empty. Overlap with other regions have been
1021 * handled by the floor/ceiling check.
1023 ptep = pte_offset_kernel(pmdp, 0UL);
1024 for (i = 0; i < PTRS_PER_PTE; i++) {
1025 if (!pte_none(READ_ONCE(ptep[i])))
1030 __flush_tlb_kernel_pgtable(start);
1031 free_hotplug_pgtable_page(virt_to_page(ptep));
1034 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
1035 unsigned long end, unsigned long floor,
1036 unsigned long ceiling)
1039 unsigned long i, next, start = addr;
1042 next = pmd_addr_end(addr, end);
1043 pmdp = pmd_offset(pudp, addr);
1044 pmd = READ_ONCE(*pmdp);
1048 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd));
1049 free_empty_pte_table(pmdp, addr, next, floor, ceiling);
1050 } while (addr = next, addr < end);
1052 if (CONFIG_PGTABLE_LEVELS <= 2)
1055 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK))
1059 * Check whether we can free the pmd page if the rest of the
1060 * entries are empty. Overlap with other regions have been
1061 * handled by the floor/ceiling check.
1063 pmdp = pmd_offset(pudp, 0UL);
1064 for (i = 0; i < PTRS_PER_PMD; i++) {
1065 if (!pmd_none(READ_ONCE(pmdp[i])))
1070 __flush_tlb_kernel_pgtable(start);
1071 free_hotplug_pgtable_page(virt_to_page(pmdp));
1074 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr,
1075 unsigned long end, unsigned long floor,
1076 unsigned long ceiling)
1079 unsigned long i, next, start = addr;
1082 next = pud_addr_end(addr, end);
1083 pudp = pud_offset(p4dp, addr);
1084 pud = READ_ONCE(*pudp);
1088 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud));
1089 free_empty_pmd_table(pudp, addr, next, floor, ceiling);
1090 } while (addr = next, addr < end);
1092 if (CONFIG_PGTABLE_LEVELS <= 3)
1095 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK))
1099 * Check whether we can free the pud page if the rest of the
1100 * entries are empty. Overlap with other regions have been
1101 * handled by the floor/ceiling check.
1103 pudp = pud_offset(p4dp, 0UL);
1104 for (i = 0; i < PTRS_PER_PUD; i++) {
1105 if (!pud_none(READ_ONCE(pudp[i])))
1110 __flush_tlb_kernel_pgtable(start);
1111 free_hotplug_pgtable_page(virt_to_page(pudp));
1114 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
1115 unsigned long end, unsigned long floor,
1116 unsigned long ceiling)
1122 next = p4d_addr_end(addr, end);
1123 p4dp = p4d_offset(pgdp, addr);
1124 p4d = READ_ONCE(*p4dp);
1128 WARN_ON(!p4d_present(p4d));
1129 free_empty_pud_table(p4dp, addr, next, floor, ceiling);
1130 } while (addr = next, addr < end);
1133 static void free_empty_tables(unsigned long addr, unsigned long end,
1134 unsigned long floor, unsigned long ceiling)
1140 next = pgd_addr_end(addr, end);
1141 pgdp = pgd_offset_k(addr);
1142 pgd = READ_ONCE(*pgdp);
1146 WARN_ON(!pgd_present(pgd));
1147 free_empty_p4d_table(pgdp, addr, next, floor, ceiling);
1148 } while (addr = next, addr < end);
1152 #if !ARM64_KERNEL_USES_PMD_MAPS
1153 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1154 struct vmem_altmap *altmap)
1156 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1157 return vmemmap_populate_basepages(start, end, node, altmap);
1159 #else /* !ARM64_KERNEL_USES_PMD_MAPS */
1160 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1161 struct vmem_altmap *altmap)
1163 unsigned long addr = start;
1170 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1172 next = pmd_addr_end(addr, end);
1174 pgdp = vmemmap_pgd_populate(addr, node);
1178 p4dp = vmemmap_p4d_populate(pgdp, addr, node);
1182 pudp = vmemmap_pud_populate(p4dp, addr, node);
1186 pmdp = pmd_offset(pudp, addr);
1187 if (pmd_none(READ_ONCE(*pmdp))) {
1190 p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
1192 if (vmemmap_populate_basepages(addr, next, node, altmap))
1197 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL));
1199 vmemmap_verify((pte_t *)pmdp, node, addr, next);
1200 } while (addr = next, addr != end);
1204 #endif /* !ARM64_KERNEL_USES_PMD_MAPS */
1206 #ifdef CONFIG_MEMORY_HOTPLUG
1207 void vmemmap_free(unsigned long start, unsigned long end,
1208 struct vmem_altmap *altmap)
1210 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1212 unmap_hotplug_range(start, end, true, altmap);
1213 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END);
1215 #endif /* CONFIG_MEMORY_HOTPLUG */
1217 static inline pud_t *fixmap_pud(unsigned long addr)
1219 pgd_t *pgdp = pgd_offset_k(addr);
1220 p4d_t *p4dp = p4d_offset(pgdp, addr);
1221 p4d_t p4d = READ_ONCE(*p4dp);
1223 BUG_ON(p4d_none(p4d) || p4d_bad(p4d));
1225 return pud_offset_kimg(p4dp, addr);
1228 static inline pmd_t *fixmap_pmd(unsigned long addr)
1230 pud_t *pudp = fixmap_pud(addr);
1231 pud_t pud = READ_ONCE(*pudp);
1233 BUG_ON(pud_none(pud) || pud_bad(pud));
1235 return pmd_offset_kimg(pudp, addr);
1238 static inline pte_t *fixmap_pte(unsigned long addr)
1240 return &bm_pte[pte_index(addr)];
1244 * The p*d_populate functions call virt_to_phys implicitly so they can't be used
1245 * directly on kernel symbols (bm_p*d). This function is called too early to use
1246 * lm_alias so __p*d_populate functions must be used to populate with the
1247 * physical address from __pa_symbol.
1249 void __init early_fixmap_init(void)
1255 unsigned long addr = FIXADDR_START;
1257 pgdp = pgd_offset_k(addr);
1258 p4dp = p4d_offset(pgdp, addr);
1259 p4d = READ_ONCE(*p4dp);
1260 if (CONFIG_PGTABLE_LEVELS > 3 &&
1261 !(p4d_none(p4d) || p4d_page_paddr(p4d) == __pa_symbol(bm_pud))) {
1263 * We only end up here if the kernel mapping and the fixmap
1264 * share the top level pgd entry, which should only happen on
1265 * 16k/4 levels configurations.
1267 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
1268 pudp = pud_offset_kimg(p4dp, addr);
1271 __p4d_populate(p4dp, __pa_symbol(bm_pud), P4D_TYPE_TABLE);
1272 pudp = fixmap_pud(addr);
1274 if (pud_none(READ_ONCE(*pudp)))
1275 __pud_populate(pudp, __pa_symbol(bm_pmd), PUD_TYPE_TABLE);
1276 pmdp = fixmap_pmd(addr);
1277 __pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE);
1280 * The boot-ioremap range spans multiple pmds, for which
1281 * we are not prepared:
1283 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
1284 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
1286 if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
1287 || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
1289 pr_warn("pmdp %p != %p, %p\n",
1290 pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
1291 fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
1292 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
1293 fix_to_virt(FIX_BTMAP_BEGIN));
1294 pr_warn("fix_to_virt(FIX_BTMAP_END): %08lx\n",
1295 fix_to_virt(FIX_BTMAP_END));
1297 pr_warn("FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
1298 pr_warn("FIX_BTMAP_BEGIN: %d\n", FIX_BTMAP_BEGIN);
1303 * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we
1304 * ever need to use IPIs for TLB broadcasting, then we're in trouble here.
1306 void __set_fixmap(enum fixed_addresses idx,
1307 phys_addr_t phys, pgprot_t flags)
1309 unsigned long addr = __fix_to_virt(idx);
1312 BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
1314 ptep = fixmap_pte(addr);
1316 if (pgprot_val(flags)) {
1317 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags));
1319 pte_clear(&init_mm, addr, ptep);
1320 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
1324 void *__init fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
1326 const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
1331 * Check whether the physical FDT address is set and meets the minimum
1332 * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
1333 * at least 8 bytes so that we can always access the magic and size
1334 * fields of the FDT header after mapping the first chunk, double check
1335 * here if that is indeed the case.
1337 BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
1338 if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
1342 * Make sure that the FDT region can be mapped without the need to
1343 * allocate additional translation table pages, so that it is safe
1344 * to call create_mapping_noalloc() this early.
1346 * On 64k pages, the FDT will be mapped using PTEs, so we need to
1347 * be in the same PMD as the rest of the fixmap.
1348 * On 4k pages, we'll use section mappings for the FDT so we only
1349 * have to be in the same PUD.
1351 BUILD_BUG_ON(dt_virt_base % SZ_2M);
1353 BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
1354 __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
1356 offset = dt_phys % SWAPPER_BLOCK_SIZE;
1357 dt_virt = (void *)dt_virt_base + offset;
1359 /* map the first chunk so we can read the size from the header */
1360 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
1361 dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
1363 if (fdt_magic(dt_virt) != FDT_MAGIC)
1366 *size = fdt_totalsize(dt_virt);
1367 if (*size > MAX_FDT_SIZE)
1370 if (offset + *size > SWAPPER_BLOCK_SIZE)
1371 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
1372 round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
1377 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
1379 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot));
1381 /* Only allow permission changes for now */
1382 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)),
1386 VM_BUG_ON(phys & ~PUD_MASK);
1387 set_pud(pudp, new_pud);
1391 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
1393 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot));
1395 /* Only allow permission changes for now */
1396 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)),
1400 VM_BUG_ON(phys & ~PMD_MASK);
1401 set_pmd(pmdp, new_pmd);
1405 int pud_clear_huge(pud_t *pudp)
1407 if (!pud_sect(READ_ONCE(*pudp)))
1413 int pmd_clear_huge(pmd_t *pmdp)
1415 if (!pmd_sect(READ_ONCE(*pmdp)))
1421 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
1426 pmd = READ_ONCE(*pmdp);
1428 if (!pmd_table(pmd)) {
1433 table = pte_offset_kernel(pmdp, addr);
1435 __flush_tlb_kernel_pgtable(addr);
1436 pte_free_kernel(NULL, table);
1440 int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
1445 unsigned long next, end;
1447 pud = READ_ONCE(*pudp);
1449 if (!pud_table(pud)) {
1454 table = pmd_offset(pudp, addr);
1457 end = addr + PUD_SIZE;
1459 pmd_free_pte_page(pmdp, next);
1460 } while (pmdp++, next += PMD_SIZE, next != end);
1463 __flush_tlb_kernel_pgtable(addr);
1464 pmd_free(NULL, table);
1468 #ifdef CONFIG_MEMORY_HOTPLUG
1469 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size)
1471 unsigned long end = start + size;
1473 WARN_ON(pgdir != init_mm.pgd);
1474 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END));
1476 unmap_hotplug_range(start, end, false, NULL);
1477 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END);
1480 struct range arch_get_mappable_range(void)
1482 struct range mhp_range;
1483 u64 start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual));
1484 u64 end_linear_pa = __pa(PAGE_END - 1);
1486 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
1488 * Check for a wrap, it is possible because of randomized linear
1489 * mapping the start physical address is actually bigger than
1490 * the end physical address. In this case set start to zero
1491 * because [0, end_linear_pa] range must still be able to cover
1492 * all addressable physical addresses.
1494 if (start_linear_pa > end_linear_pa)
1495 start_linear_pa = 0;
1498 WARN_ON(start_linear_pa > end_linear_pa);
1501 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)]
1502 * accommodating both its ends but excluding PAGE_END. Max physical
1503 * range which can be mapped inside this linear mapping range, must
1504 * also be derived from its end points.
1506 mhp_range.start = start_linear_pa;
1507 mhp_range.end = end_linear_pa;
1512 int arch_add_memory(int nid, u64 start, u64 size,
1513 struct mhp_params *params)
1515 int ret, flags = NO_EXEC_MAPPINGS;
1517 VM_BUG_ON(!mhp_range_allowed(start, size, true));
1520 * KFENCE requires linear map to be mapped at page granularity, so that
1521 * it is possible to protect/unprotect single pages in the KFENCE pool.
1523 if (can_set_direct_map() || IS_ENABLED(CONFIG_KFENCE))
1524 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
1526 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
1527 size, params->pgprot, __pgd_pgtable_alloc,
1530 memblock_clear_nomap(start, size);
1532 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT,
1535 __remove_pgd_mapping(swapper_pg_dir,
1536 __phys_to_virt(start), size);
1538 max_pfn = PFN_UP(start + size);
1539 max_low_pfn = max_pfn;
1545 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
1547 unsigned long start_pfn = start >> PAGE_SHIFT;
1548 unsigned long nr_pages = size >> PAGE_SHIFT;
1550 __remove_pages(start_pfn, nr_pages, altmap);
1551 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size);
1555 * This memory hotplug notifier helps prevent boot memory from being
1556 * inadvertently removed as it blocks pfn range offlining process in
1557 * __offline_pages(). Hence this prevents both offlining as well as
1558 * removal process for boot memory which is initially always online.
1559 * In future if and when boot memory could be removed, this notifier
1560 * should be dropped and free_hotplug_page_range() should handle any
1561 * reserved pages allocated during boot.
1563 static int prevent_bootmem_remove_notifier(struct notifier_block *nb,
1564 unsigned long action, void *data)
1566 struct mem_section *ms;
1567 struct memory_notify *arg = data;
1568 unsigned long end_pfn = arg->start_pfn + arg->nr_pages;
1569 unsigned long pfn = arg->start_pfn;
1571 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE))
1574 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
1575 unsigned long start = PFN_PHYS(pfn);
1576 unsigned long end = start + (1UL << PA_SECTION_SHIFT);
1578 ms = __pfn_to_section(pfn);
1579 if (!early_section(ms))
1582 if (action == MEM_GOING_OFFLINE) {
1584 * Boot memory removal is not supported. Prevent
1585 * it via blocking any attempted offline request
1586 * for the boot memory and just report it.
1588 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end);
1590 } else if (action == MEM_OFFLINE) {
1592 * This should have never happened. Boot memory
1593 * offlining should have been prevented by this
1594 * very notifier. Probably some memory removal
1595 * procedure might have changed which would then
1596 * require further debug.
1598 pr_err("Boot memory [%lx %lx] offlined\n", start, end);
1601 * Core memory hotplug does not process a return
1602 * code from the notifier for MEM_OFFLINE events.
1603 * The error condition has been reported. Return
1604 * from here as if ignored.
1612 static struct notifier_block prevent_bootmem_remove_nb = {
1613 .notifier_call = prevent_bootmem_remove_notifier,
1617 * This ensures that boot memory sections on the platform are online
1618 * from early boot. Memory sections could not be prevented from being
1619 * offlined, unless for some reason they are not online to begin with.
1620 * This helps validate the basic assumption on which the above memory
1621 * event notifier works to prevent boot memory section offlining and
1622 * its possible removal.
1624 static void validate_bootmem_online(void)
1626 phys_addr_t start, end, addr;
1627 struct mem_section *ms;
1631 * Scanning across all memblock might be expensive
1632 * on some big memory systems. Hence enable this
1633 * validation only with DEBUG_VM.
1635 if (!IS_ENABLED(CONFIG_DEBUG_VM))
1638 for_each_mem_range(i, &start, &end) {
1639 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) {
1640 ms = __pfn_to_section(PHYS_PFN(addr));
1643 * All memory ranges in the system at this point
1644 * should have been marked as early sections.
1646 WARN_ON(!early_section(ms));
1649 * Memory notifier mechanism here to prevent boot
1650 * memory offlining depends on the fact that each
1651 * early section memory on the system is initially
1652 * online. Otherwise a given memory section which
1653 * is already offline will be overlooked and can
1654 * be removed completely. Call out such sections.
1656 if (!online_section(ms))
1657 pr_err("Boot memory [%llx %llx] is offline, can be removed\n",
1658 addr, addr + (1UL << PA_SECTION_SHIFT));
1663 static int __init prevent_bootmem_remove_init(void)
1667 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
1670 validate_bootmem_online();
1671 ret = register_memory_notifier(&prevent_bootmem_remove_nb);
1673 pr_err("%s: Notifier registration failed %d\n", __func__, ret);
1677 early_initcall(prevent_bootmem_remove_init);