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(PAGE_SIZE, PAGE_SIZE);
102 panic("Failed to allocate page table page\n");
105 * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
106 * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
107 * any level of table.
109 ptr = pte_set_fixmap(phys);
111 memset(ptr, 0, PAGE_SIZE);
114 * Implicit barriers also ensure the zeroed page is visible to the page
122 static bool pgattr_change_is_safe(u64 old, u64 new)
125 * The following mapping attributes may be updated in live
126 * kernel mappings without the need for break-before-make.
128 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG;
130 /* creating or taking down mappings is always safe */
131 if (old == 0 || new == 0)
134 /* live contiguous mappings may not be manipulated at all */
135 if ((old | new) & PTE_CONT)
138 /* Transitioning from Non-Global to Global is unsafe */
139 if (old & ~new & PTE_NG)
143 * Changing the memory type between Normal and Normal-Tagged is safe
144 * since Tagged is considered a permission attribute from the
145 * mismatched attribute aliases perspective.
147 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
148 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) &&
149 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
150 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)))
151 mask |= PTE_ATTRINDX_MASK;
153 return ((old ^ new) & ~mask) == 0;
156 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end,
157 phys_addr_t phys, pgprot_t prot)
161 ptep = pte_set_fixmap_offset(pmdp, addr);
163 pte_t old_pte = READ_ONCE(*ptep);
165 set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot));
168 * After the PTE entry has been populated once, we
169 * only allow updates to the permission attributes.
171 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte),
172 READ_ONCE(pte_val(*ptep))));
175 } while (ptep++, addr += PAGE_SIZE, addr != end);
180 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
181 unsigned long end, phys_addr_t phys,
183 phys_addr_t (*pgtable_alloc)(int),
187 pmd_t pmd = READ_ONCE(*pmdp);
189 BUG_ON(pmd_sect(pmd));
191 pmdval_t pmdval = PMD_TYPE_TABLE | PMD_TABLE_UXN;
192 phys_addr_t pte_phys;
194 if (flags & NO_EXEC_MAPPINGS)
195 pmdval |= PMD_TABLE_PXN;
196 BUG_ON(!pgtable_alloc);
197 pte_phys = pgtable_alloc(PAGE_SHIFT);
198 __pmd_populate(pmdp, pte_phys, pmdval);
199 pmd = READ_ONCE(*pmdp);
201 BUG_ON(pmd_bad(pmd));
204 pgprot_t __prot = prot;
206 next = pte_cont_addr_end(addr, end);
208 /* use a contiguous mapping if the range is suitably aligned */
209 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) &&
210 (flags & NO_CONT_MAPPINGS) == 0)
211 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
213 init_pte(pmdp, addr, next, phys, __prot);
216 } while (addr = next, addr != end);
219 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end,
220 phys_addr_t phys, pgprot_t prot,
221 phys_addr_t (*pgtable_alloc)(int), int flags)
226 pmdp = pmd_set_fixmap_offset(pudp, addr);
228 pmd_t old_pmd = READ_ONCE(*pmdp);
230 next = pmd_addr_end(addr, end);
232 /* try section mapping first */
233 if (((addr | next | phys) & ~PMD_MASK) == 0 &&
234 (flags & NO_BLOCK_MAPPINGS) == 0) {
235 pmd_set_huge(pmdp, phys, prot);
238 * After the PMD entry has been populated once, we
239 * only allow updates to the permission attributes.
241 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
242 READ_ONCE(pmd_val(*pmdp))));
244 alloc_init_cont_pte(pmdp, addr, next, phys, prot,
245 pgtable_alloc, flags);
247 BUG_ON(pmd_val(old_pmd) != 0 &&
248 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp)));
251 } while (pmdp++, addr = next, addr != end);
256 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
257 unsigned long end, phys_addr_t phys,
259 phys_addr_t (*pgtable_alloc)(int), int flags)
262 pud_t pud = READ_ONCE(*pudp);
265 * Check for initial section mappings in the pgd/pud.
267 BUG_ON(pud_sect(pud));
269 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN;
270 phys_addr_t pmd_phys;
272 if (flags & NO_EXEC_MAPPINGS)
273 pudval |= PUD_TABLE_PXN;
274 BUG_ON(!pgtable_alloc);
275 pmd_phys = pgtable_alloc(PMD_SHIFT);
276 __pud_populate(pudp, pmd_phys, pudval);
277 pud = READ_ONCE(*pudp);
279 BUG_ON(pud_bad(pud));
282 pgprot_t __prot = prot;
284 next = pmd_cont_addr_end(addr, end);
286 /* use a contiguous mapping if the range is suitably aligned */
287 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) &&
288 (flags & NO_CONT_MAPPINGS) == 0)
289 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
291 init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags);
294 } while (addr = next, addr != end);
297 static inline bool use_1G_block(unsigned long addr, unsigned long next,
300 if (PAGE_SHIFT != 12)
303 if (((addr | next | phys) & ~PUD_MASK) != 0)
309 static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end,
310 phys_addr_t phys, pgprot_t prot,
311 phys_addr_t (*pgtable_alloc)(int),
316 p4d_t *p4dp = p4d_offset(pgdp, addr);
317 p4d_t p4d = READ_ONCE(*p4dp);
320 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN;
321 phys_addr_t pud_phys;
323 if (flags & NO_EXEC_MAPPINGS)
324 p4dval |= P4D_TABLE_PXN;
325 BUG_ON(!pgtable_alloc);
326 pud_phys = pgtable_alloc(PUD_SHIFT);
327 __p4d_populate(p4dp, pud_phys, p4dval);
328 p4d = READ_ONCE(*p4dp);
330 BUG_ON(p4d_bad(p4d));
333 * No need for locking during early boot. And it doesn't work as
334 * expected with KASLR enabled.
336 if (system_state != SYSTEM_BOOTING)
337 mutex_lock(&fixmap_lock);
338 pudp = pud_set_fixmap_offset(p4dp, addr);
340 pud_t old_pud = READ_ONCE(*pudp);
342 next = pud_addr_end(addr, end);
345 * For 4K granule only, attempt to put down a 1GB block
347 if (use_1G_block(addr, next, phys) &&
348 (flags & NO_BLOCK_MAPPINGS) == 0) {
349 pud_set_huge(pudp, phys, prot);
352 * After the PUD entry has been populated once, we
353 * only allow updates to the permission attributes.
355 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
356 READ_ONCE(pud_val(*pudp))));
358 alloc_init_cont_pmd(pudp, addr, next, phys, prot,
359 pgtable_alloc, flags);
361 BUG_ON(pud_val(old_pud) != 0 &&
362 pud_val(old_pud) != READ_ONCE(pud_val(*pudp)));
365 } while (pudp++, addr = next, addr != end);
368 if (system_state != SYSTEM_BOOTING)
369 mutex_unlock(&fixmap_lock);
372 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
373 unsigned long virt, phys_addr_t size,
375 phys_addr_t (*pgtable_alloc)(int),
378 unsigned long addr, end, next;
379 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt);
382 * If the virtual and physical address don't have the same offset
383 * within a page, we cannot map the region as the caller expects.
385 if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
389 addr = virt & PAGE_MASK;
390 end = PAGE_ALIGN(virt + size);
393 next = pgd_addr_end(addr, end);
394 alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc,
397 } while (pgdp++, addr = next, addr != end);
400 static phys_addr_t __pgd_pgtable_alloc(int shift)
402 void *ptr = (void *)__get_free_page(GFP_PGTABLE_KERNEL);
405 /* Ensure the zeroed page is visible to the page table walker */
410 static phys_addr_t pgd_pgtable_alloc(int shift)
412 phys_addr_t pa = __pgd_pgtable_alloc(shift);
415 * Call proper page table ctor in case later we need to
416 * call core mm functions like apply_to_page_range() on
417 * this pre-allocated page table.
419 * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is
420 * folded, and if so pgtable_pmd_page_ctor() becomes nop.
422 if (shift == PAGE_SHIFT)
423 BUG_ON(!pgtable_pte_page_ctor(phys_to_page(pa)));
424 else if (shift == PMD_SHIFT)
425 BUG_ON(!pgtable_pmd_page_ctor(phys_to_page(pa)));
431 * This function can only be used to modify existing table entries,
432 * without allocating new levels of table. Note that this permits the
433 * creation of new section or page entries.
435 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
436 phys_addr_t size, pgprot_t prot)
438 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
439 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
443 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
447 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
448 unsigned long virt, phys_addr_t size,
449 pgprot_t prot, bool page_mappings_only)
453 BUG_ON(mm == &init_mm);
455 if (page_mappings_only)
456 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
458 __create_pgd_mapping(mm->pgd, phys, virt, size, prot,
459 pgd_pgtable_alloc, flags);
462 static void update_mapping_prot(phys_addr_t phys, unsigned long virt,
463 phys_addr_t size, pgprot_t prot)
465 if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
466 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n",
471 __create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
474 /* flush the TLBs after updating live kernel mappings */
475 flush_tlb_kernel_range(virt, virt + size);
478 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start,
479 phys_addr_t end, pgprot_t prot, int flags)
481 __create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start,
482 prot, early_pgtable_alloc, flags);
485 void __init mark_linear_text_alias_ro(void)
488 * Remove the write permissions from the linear alias of .text/.rodata
490 update_mapping_prot(__pa_symbol(_stext), (unsigned long)lm_alias(_stext),
491 (unsigned long)__init_begin - (unsigned long)_stext,
495 static bool crash_mem_map __initdata;
497 static int __init enable_crash_mem_map(char *arg)
500 * Proper parameter parsing is done by reserve_crashkernel(). We only
501 * need to know if the linear map has to avoid block mappings so that
502 * the crashkernel reservations can be unmapped later.
504 crash_mem_map = true;
508 early_param("crashkernel", enable_crash_mem_map);
510 static void __init map_mem(pgd_t *pgdp)
512 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN);
513 phys_addr_t kernel_start = __pa_symbol(_stext);
514 phys_addr_t kernel_end = __pa_symbol(__init_begin);
515 phys_addr_t start, end;
516 int flags = NO_EXEC_MAPPINGS;
520 * Setting hierarchical PXNTable attributes on table entries covering
521 * the linear region is only possible if it is guaranteed that no table
522 * entries at any level are being shared between the linear region and
523 * the vmalloc region. Check whether this is true for the PGD level, in
524 * which case it is guaranteed to be true for all other levels as well.
526 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end));
528 if (can_set_direct_map() || IS_ENABLED(CONFIG_KFENCE))
529 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
532 * Take care not to create a writable alias for the
533 * read-only text and rodata sections of the kernel image.
534 * So temporarily mark them as NOMAP to skip mappings in
535 * the following for-loop
537 memblock_mark_nomap(kernel_start, kernel_end - kernel_start);
539 #ifdef CONFIG_KEXEC_CORE
541 if (IS_ENABLED(CONFIG_ZONE_DMA) ||
542 IS_ENABLED(CONFIG_ZONE_DMA32))
543 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
544 else if (crashk_res.end)
545 memblock_mark_nomap(crashk_res.start,
546 resource_size(&crashk_res));
550 /* map all the memory banks */
551 for_each_mem_range(i, &start, &end) {
555 * The linear map must allow allocation tags reading/writing
556 * if MTE is present. Otherwise, it has the same attributes as
559 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL),
564 * Map the linear alias of the [_stext, __init_begin) interval
565 * as non-executable now, and remove the write permission in
566 * mark_linear_text_alias_ro() below (which will be called after
567 * alternative patching has completed). This makes the contents
568 * of the region accessible to subsystems such as hibernate,
569 * but protects it from inadvertent modification or execution.
570 * Note that contiguous mappings cannot be remapped in this way,
571 * so we should avoid them here.
573 __map_memblock(pgdp, kernel_start, kernel_end,
574 PAGE_KERNEL, NO_CONT_MAPPINGS);
575 memblock_clear_nomap(kernel_start, kernel_end - kernel_start);
578 * Use page-level mappings here so that we can shrink the region
579 * in page granularity and put back unused memory to buddy system
580 * through /sys/kernel/kexec_crash_size interface.
582 #ifdef CONFIG_KEXEC_CORE
584 !IS_ENABLED(CONFIG_ZONE_DMA) && !IS_ENABLED(CONFIG_ZONE_DMA32)) {
585 if (crashk_res.end) {
586 __map_memblock(pgdp, crashk_res.start,
589 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
590 memblock_clear_nomap(crashk_res.start,
591 resource_size(&crashk_res));
597 void mark_rodata_ro(void)
599 unsigned long section_size;
602 * mark .rodata as read only. Use __init_begin rather than __end_rodata
603 * to cover NOTES and EXCEPTION_TABLE.
605 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
606 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata,
607 section_size, PAGE_KERNEL_RO);
612 static void __init map_kernel_segment(pgd_t *pgdp, void *va_start, void *va_end,
613 pgprot_t prot, struct vm_struct *vma,
614 int flags, unsigned long vm_flags)
616 phys_addr_t pa_start = __pa_symbol(va_start);
617 unsigned long size = va_end - va_start;
619 BUG_ON(!PAGE_ALIGNED(pa_start));
620 BUG_ON(!PAGE_ALIGNED(size));
622 __create_pgd_mapping(pgdp, pa_start, (unsigned long)va_start, size, prot,
623 early_pgtable_alloc, flags);
625 if (!(vm_flags & VM_NO_GUARD))
628 vma->addr = va_start;
629 vma->phys_addr = pa_start;
631 vma->flags = VM_MAP | vm_flags;
632 vma->caller = __builtin_return_address(0);
634 vm_area_add_early(vma);
637 static int __init parse_rodata(char *arg)
639 int ret = strtobool(arg, &rodata_enabled);
645 /* permit 'full' in addition to boolean options */
646 if (strcmp(arg, "full"))
649 rodata_enabled = true;
653 early_param("rodata", parse_rodata);
655 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
656 static int __init map_entry_trampoline(void)
660 pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
661 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start);
663 /* The trampoline is always mapped and can therefore be global */
664 pgprot_val(prot) &= ~PTE_NG;
666 /* Map only the text into the trampoline page table */
667 memset(tramp_pg_dir, 0, PGD_SIZE);
668 __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS,
669 entry_tramp_text_size(), prot,
670 __pgd_pgtable_alloc, NO_BLOCK_MAPPINGS);
672 /* Map both the text and data into the kernel page table */
673 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++)
674 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i,
675 pa_start + i * PAGE_SIZE, prot);
677 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
678 extern char __entry_tramp_data_start[];
680 __set_fixmap(FIX_ENTRY_TRAMP_DATA,
681 __pa_symbol(__entry_tramp_data_start),
687 core_initcall(map_entry_trampoline);
691 * Open coded check for BTI, only for use to determine configuration
692 * for early mappings for before the cpufeature code has run.
694 static bool arm64_early_this_cpu_has_bti(void)
698 if (!IS_ENABLED(CONFIG_ARM64_BTI_KERNEL))
701 pfr1 = __read_sysreg_by_encoding(SYS_ID_AA64PFR1_EL1);
702 return cpuid_feature_extract_unsigned_field(pfr1,
703 ID_AA64PFR1_BT_SHIFT);
707 * Create fine-grained mappings for the kernel.
709 static void __init map_kernel(pgd_t *pgdp)
711 static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext,
712 vmlinux_initdata, vmlinux_data;
715 * External debuggers may need to write directly to the text
716 * mapping to install SW breakpoints. Allow this (only) when
717 * explicitly requested with rodata=off.
719 pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
722 * If we have a CPU that supports BTI and a kernel built for
723 * BTI then mark the kernel executable text as guarded pages
724 * now so we don't have to rewrite the page tables later.
726 if (arm64_early_this_cpu_has_bti())
727 text_prot = __pgprot_modify(text_prot, PTE_GP, PTE_GP);
730 * Only rodata will be remapped with different permissions later on,
731 * all other segments are allowed to use contiguous mappings.
733 map_kernel_segment(pgdp, _stext, _etext, text_prot, &vmlinux_text, 0,
735 map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL,
736 &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD);
737 map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot,
738 &vmlinux_inittext, 0, VM_NO_GUARD);
739 map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL,
740 &vmlinux_initdata, 0, VM_NO_GUARD);
741 map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0);
743 if (!READ_ONCE(pgd_val(*pgd_offset_pgd(pgdp, FIXADDR_START)))) {
745 * The fixmap falls in a separate pgd to the kernel, and doesn't
746 * live in the carveout for the swapper_pg_dir. We can simply
747 * re-use the existing dir for the fixmap.
749 set_pgd(pgd_offset_pgd(pgdp, FIXADDR_START),
750 READ_ONCE(*pgd_offset_k(FIXADDR_START)));
751 } else if (CONFIG_PGTABLE_LEVELS > 3) {
756 * The fixmap shares its top level pgd entry with the kernel
757 * mapping. This can really only occur when we are running
758 * with 16k/4 levels, so we can simply reuse the pud level
761 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
762 bm_pgdp = pgd_offset_pgd(pgdp, FIXADDR_START);
763 bm_p4dp = p4d_offset(bm_pgdp, FIXADDR_START);
764 bm_pudp = pud_set_fixmap_offset(bm_p4dp, FIXADDR_START);
765 pud_populate(&init_mm, bm_pudp, lm_alias(bm_pmd));
771 kasan_copy_shadow(pgdp);
774 void __init paging_init(void)
776 pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir));
783 cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
784 init_mm.pgd = swapper_pg_dir;
786 memblock_free(__pa_symbol(init_pg_dir),
787 __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir));
789 memblock_allow_resize();
793 * Check whether a kernel address is valid (derived from arch/x86/).
795 int kern_addr_valid(unsigned long addr)
803 addr = arch_kasan_reset_tag(addr);
804 if ((((long)addr) >> VA_BITS) != -1UL)
807 pgdp = pgd_offset_k(addr);
808 if (pgd_none(READ_ONCE(*pgdp)))
811 p4dp = p4d_offset(pgdp, addr);
812 if (p4d_none(READ_ONCE(*p4dp)))
815 pudp = pud_offset(p4dp, addr);
816 pud = READ_ONCE(*pudp);
821 return pfn_valid(pud_pfn(pud));
823 pmdp = pmd_offset(pudp, addr);
824 pmd = READ_ONCE(*pmdp);
829 return pfn_valid(pmd_pfn(pmd));
831 ptep = pte_offset_kernel(pmdp, addr);
832 pte = READ_ONCE(*ptep);
836 return pfn_valid(pte_pfn(pte));
839 #ifdef CONFIG_MEMORY_HOTPLUG
840 static void free_hotplug_page_range(struct page *page, size_t size,
841 struct vmem_altmap *altmap)
844 vmem_altmap_free(altmap, size >> PAGE_SHIFT);
846 WARN_ON(PageReserved(page));
847 free_pages((unsigned long)page_address(page), get_order(size));
851 static void free_hotplug_pgtable_page(struct page *page)
853 free_hotplug_page_range(page, PAGE_SIZE, NULL);
856 static bool pgtable_range_aligned(unsigned long start, unsigned long end,
857 unsigned long floor, unsigned long ceiling,
870 if (end - 1 > ceiling - 1)
875 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr,
876 unsigned long end, bool free_mapped,
877 struct vmem_altmap *altmap)
882 ptep = pte_offset_kernel(pmdp, addr);
883 pte = READ_ONCE(*ptep);
887 WARN_ON(!pte_present(pte));
888 pte_clear(&init_mm, addr, ptep);
889 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
891 free_hotplug_page_range(pte_page(pte),
893 } while (addr += PAGE_SIZE, addr < end);
896 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr,
897 unsigned long end, bool free_mapped,
898 struct vmem_altmap *altmap)
904 next = pmd_addr_end(addr, end);
905 pmdp = pmd_offset(pudp, addr);
906 pmd = READ_ONCE(*pmdp);
910 WARN_ON(!pmd_present(pmd));
915 * One TLBI should be sufficient here as the PMD_SIZE
916 * range is mapped with a single block entry.
918 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
920 free_hotplug_page_range(pmd_page(pmd),
924 WARN_ON(!pmd_table(pmd));
925 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap);
926 } while (addr = next, addr < end);
929 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr,
930 unsigned long end, bool free_mapped,
931 struct vmem_altmap *altmap)
937 next = pud_addr_end(addr, end);
938 pudp = pud_offset(p4dp, addr);
939 pud = READ_ONCE(*pudp);
943 WARN_ON(!pud_present(pud));
948 * One TLBI should be sufficient here as the PUD_SIZE
949 * range is mapped with a single block entry.
951 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
953 free_hotplug_page_range(pud_page(pud),
957 WARN_ON(!pud_table(pud));
958 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap);
959 } while (addr = next, addr < end);
962 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr,
963 unsigned long end, bool free_mapped,
964 struct vmem_altmap *altmap)
970 next = p4d_addr_end(addr, end);
971 p4dp = p4d_offset(pgdp, addr);
972 p4d = READ_ONCE(*p4dp);
976 WARN_ON(!p4d_present(p4d));
977 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap);
978 } while (addr = next, addr < end);
981 static void unmap_hotplug_range(unsigned long addr, unsigned long end,
982 bool free_mapped, struct vmem_altmap *altmap)
988 * altmap can only be used as vmemmap mapping backing memory.
989 * In case the backing memory itself is not being freed, then
990 * altmap is irrelevant. Warn about this inconsistency when
993 WARN_ON(!free_mapped && altmap);
996 next = pgd_addr_end(addr, end);
997 pgdp = pgd_offset_k(addr);
998 pgd = READ_ONCE(*pgdp);
1002 WARN_ON(!pgd_present(pgd));
1003 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap);
1004 } while (addr = next, addr < end);
1007 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
1008 unsigned long end, unsigned long floor,
1009 unsigned long ceiling)
1012 unsigned long i, start = addr;
1015 ptep = pte_offset_kernel(pmdp, addr);
1016 pte = READ_ONCE(*ptep);
1019 * This is just a sanity check here which verifies that
1020 * pte clearing has been done by earlier unmap loops.
1022 WARN_ON(!pte_none(pte));
1023 } while (addr += PAGE_SIZE, addr < end);
1025 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK))
1029 * Check whether we can free the pte page if the rest of the
1030 * entries are empty. Overlap with other regions have been
1031 * handled by the floor/ceiling check.
1033 ptep = pte_offset_kernel(pmdp, 0UL);
1034 for (i = 0; i < PTRS_PER_PTE; i++) {
1035 if (!pte_none(READ_ONCE(ptep[i])))
1040 __flush_tlb_kernel_pgtable(start);
1041 free_hotplug_pgtable_page(virt_to_page(ptep));
1044 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
1045 unsigned long end, unsigned long floor,
1046 unsigned long ceiling)
1049 unsigned long i, next, start = addr;
1052 next = pmd_addr_end(addr, end);
1053 pmdp = pmd_offset(pudp, addr);
1054 pmd = READ_ONCE(*pmdp);
1058 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd));
1059 free_empty_pte_table(pmdp, addr, next, floor, ceiling);
1060 } while (addr = next, addr < end);
1062 if (CONFIG_PGTABLE_LEVELS <= 2)
1065 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK))
1069 * Check whether we can free the pmd page if the rest of the
1070 * entries are empty. Overlap with other regions have been
1071 * handled by the floor/ceiling check.
1073 pmdp = pmd_offset(pudp, 0UL);
1074 for (i = 0; i < PTRS_PER_PMD; i++) {
1075 if (!pmd_none(READ_ONCE(pmdp[i])))
1080 __flush_tlb_kernel_pgtable(start);
1081 free_hotplug_pgtable_page(virt_to_page(pmdp));
1084 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr,
1085 unsigned long end, unsigned long floor,
1086 unsigned long ceiling)
1089 unsigned long i, next, start = addr;
1092 next = pud_addr_end(addr, end);
1093 pudp = pud_offset(p4dp, addr);
1094 pud = READ_ONCE(*pudp);
1098 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud));
1099 free_empty_pmd_table(pudp, addr, next, floor, ceiling);
1100 } while (addr = next, addr < end);
1102 if (CONFIG_PGTABLE_LEVELS <= 3)
1105 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK))
1109 * Check whether we can free the pud page if the rest of the
1110 * entries are empty. Overlap with other regions have been
1111 * handled by the floor/ceiling check.
1113 pudp = pud_offset(p4dp, 0UL);
1114 for (i = 0; i < PTRS_PER_PUD; i++) {
1115 if (!pud_none(READ_ONCE(pudp[i])))
1120 __flush_tlb_kernel_pgtable(start);
1121 free_hotplug_pgtable_page(virt_to_page(pudp));
1124 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
1125 unsigned long end, unsigned long floor,
1126 unsigned long ceiling)
1132 next = p4d_addr_end(addr, end);
1133 p4dp = p4d_offset(pgdp, addr);
1134 p4d = READ_ONCE(*p4dp);
1138 WARN_ON(!p4d_present(p4d));
1139 free_empty_pud_table(p4dp, addr, next, floor, ceiling);
1140 } while (addr = next, addr < end);
1143 static void free_empty_tables(unsigned long addr, unsigned long end,
1144 unsigned long floor, unsigned long ceiling)
1150 next = pgd_addr_end(addr, end);
1151 pgdp = pgd_offset_k(addr);
1152 pgd = READ_ONCE(*pgdp);
1156 WARN_ON(!pgd_present(pgd));
1157 free_empty_p4d_table(pgdp, addr, next, floor, ceiling);
1158 } while (addr = next, addr < end);
1162 #if !ARM64_KERNEL_USES_PMD_MAPS
1163 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1164 struct vmem_altmap *altmap)
1166 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1167 return vmemmap_populate_basepages(start, end, node, altmap);
1169 #else /* !ARM64_KERNEL_USES_PMD_MAPS */
1170 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1171 struct vmem_altmap *altmap)
1173 unsigned long addr = start;
1180 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1182 next = pmd_addr_end(addr, end);
1184 pgdp = vmemmap_pgd_populate(addr, node);
1188 p4dp = vmemmap_p4d_populate(pgdp, addr, node);
1192 pudp = vmemmap_pud_populate(p4dp, addr, node);
1196 pmdp = pmd_offset(pudp, addr);
1197 if (pmd_none(READ_ONCE(*pmdp))) {
1200 p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
1202 if (vmemmap_populate_basepages(addr, next, node, altmap))
1207 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL));
1209 vmemmap_verify((pte_t *)pmdp, node, addr, next);
1210 } while (addr = next, addr != end);
1214 #endif /* !ARM64_KERNEL_USES_PMD_MAPS */
1216 #ifdef CONFIG_MEMORY_HOTPLUG
1217 void vmemmap_free(unsigned long start, unsigned long end,
1218 struct vmem_altmap *altmap)
1220 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1222 unmap_hotplug_range(start, end, true, altmap);
1223 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END);
1225 #endif /* CONFIG_MEMORY_HOTPLUG */
1227 static inline pud_t *fixmap_pud(unsigned long addr)
1229 pgd_t *pgdp = pgd_offset_k(addr);
1230 p4d_t *p4dp = p4d_offset(pgdp, addr);
1231 p4d_t p4d = READ_ONCE(*p4dp);
1233 BUG_ON(p4d_none(p4d) || p4d_bad(p4d));
1235 return pud_offset_kimg(p4dp, addr);
1238 static inline pmd_t *fixmap_pmd(unsigned long addr)
1240 pud_t *pudp = fixmap_pud(addr);
1241 pud_t pud = READ_ONCE(*pudp);
1243 BUG_ON(pud_none(pud) || pud_bad(pud));
1245 return pmd_offset_kimg(pudp, addr);
1248 static inline pte_t *fixmap_pte(unsigned long addr)
1250 return &bm_pte[pte_index(addr)];
1254 * The p*d_populate functions call virt_to_phys implicitly so they can't be used
1255 * directly on kernel symbols (bm_p*d). This function is called too early to use
1256 * lm_alias so __p*d_populate functions must be used to populate with the
1257 * physical address from __pa_symbol.
1259 void __init early_fixmap_init(void)
1265 unsigned long addr = FIXADDR_START;
1267 pgdp = pgd_offset_k(addr);
1268 p4dp = p4d_offset(pgdp, addr);
1269 p4d = READ_ONCE(*p4dp);
1270 if (CONFIG_PGTABLE_LEVELS > 3 &&
1271 !(p4d_none(p4d) || p4d_page_paddr(p4d) == __pa_symbol(bm_pud))) {
1273 * We only end up here if the kernel mapping and the fixmap
1274 * share the top level pgd entry, which should only happen on
1275 * 16k/4 levels configurations.
1277 BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
1278 pudp = pud_offset_kimg(p4dp, addr);
1281 __p4d_populate(p4dp, __pa_symbol(bm_pud), P4D_TYPE_TABLE);
1282 pudp = fixmap_pud(addr);
1284 if (pud_none(READ_ONCE(*pudp)))
1285 __pud_populate(pudp, __pa_symbol(bm_pmd), PUD_TYPE_TABLE);
1286 pmdp = fixmap_pmd(addr);
1287 __pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE);
1290 * The boot-ioremap range spans multiple pmds, for which
1291 * we are not prepared:
1293 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
1294 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
1296 if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
1297 || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
1299 pr_warn("pmdp %p != %p, %p\n",
1300 pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
1301 fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
1302 pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
1303 fix_to_virt(FIX_BTMAP_BEGIN));
1304 pr_warn("fix_to_virt(FIX_BTMAP_END): %08lx\n",
1305 fix_to_virt(FIX_BTMAP_END));
1307 pr_warn("FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
1308 pr_warn("FIX_BTMAP_BEGIN: %d\n", FIX_BTMAP_BEGIN);
1313 * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we
1314 * ever need to use IPIs for TLB broadcasting, then we're in trouble here.
1316 void __set_fixmap(enum fixed_addresses idx,
1317 phys_addr_t phys, pgprot_t flags)
1319 unsigned long addr = __fix_to_virt(idx);
1322 BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
1324 ptep = fixmap_pte(addr);
1326 if (pgprot_val(flags)) {
1327 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags));
1329 pte_clear(&init_mm, addr, ptep);
1330 flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
1334 void *__init fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
1336 const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
1341 * Check whether the physical FDT address is set and meets the minimum
1342 * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
1343 * at least 8 bytes so that we can always access the magic and size
1344 * fields of the FDT header after mapping the first chunk, double check
1345 * here if that is indeed the case.
1347 BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
1348 if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
1352 * Make sure that the FDT region can be mapped without the need to
1353 * allocate additional translation table pages, so that it is safe
1354 * to call create_mapping_noalloc() this early.
1356 * On 64k pages, the FDT will be mapped using PTEs, so we need to
1357 * be in the same PMD as the rest of the fixmap.
1358 * On 4k pages, we'll use section mappings for the FDT so we only
1359 * have to be in the same PUD.
1361 BUILD_BUG_ON(dt_virt_base % SZ_2M);
1363 BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
1364 __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
1366 offset = dt_phys % SWAPPER_BLOCK_SIZE;
1367 dt_virt = (void *)dt_virt_base + offset;
1369 /* map the first chunk so we can read the size from the header */
1370 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
1371 dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
1373 if (fdt_magic(dt_virt) != FDT_MAGIC)
1376 *size = fdt_totalsize(dt_virt);
1377 if (*size > MAX_FDT_SIZE)
1380 if (offset + *size > SWAPPER_BLOCK_SIZE)
1381 create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
1382 round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
1387 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
1389 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot));
1391 /* Only allow permission changes for now */
1392 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)),
1396 VM_BUG_ON(phys & ~PUD_MASK);
1397 set_pud(pudp, new_pud);
1401 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
1403 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot));
1405 /* Only allow permission changes for now */
1406 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)),
1410 VM_BUG_ON(phys & ~PMD_MASK);
1411 set_pmd(pmdp, new_pmd);
1415 int pud_clear_huge(pud_t *pudp)
1417 if (!pud_sect(READ_ONCE(*pudp)))
1423 int pmd_clear_huge(pmd_t *pmdp)
1425 if (!pmd_sect(READ_ONCE(*pmdp)))
1431 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
1436 pmd = READ_ONCE(*pmdp);
1438 if (!pmd_table(pmd)) {
1443 table = pte_offset_kernel(pmdp, addr);
1445 __flush_tlb_kernel_pgtable(addr);
1446 pte_free_kernel(NULL, table);
1450 int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
1455 unsigned long next, end;
1457 pud = READ_ONCE(*pudp);
1459 if (!pud_table(pud)) {
1464 table = pmd_offset(pudp, addr);
1467 end = addr + PUD_SIZE;
1469 pmd_free_pte_page(pmdp, next);
1470 } while (pmdp++, next += PMD_SIZE, next != end);
1473 __flush_tlb_kernel_pgtable(addr);
1474 pmd_free(NULL, table);
1478 #ifdef CONFIG_MEMORY_HOTPLUG
1479 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size)
1481 unsigned long end = start + size;
1483 WARN_ON(pgdir != init_mm.pgd);
1484 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END));
1486 unmap_hotplug_range(start, end, false, NULL);
1487 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END);
1490 struct range arch_get_mappable_range(void)
1492 struct range mhp_range;
1493 u64 start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual));
1494 u64 end_linear_pa = __pa(PAGE_END - 1);
1496 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
1498 * Check for a wrap, it is possible because of randomized linear
1499 * mapping the start physical address is actually bigger than
1500 * the end physical address. In this case set start to zero
1501 * because [0, end_linear_pa] range must still be able to cover
1502 * all addressable physical addresses.
1504 if (start_linear_pa > end_linear_pa)
1505 start_linear_pa = 0;
1508 WARN_ON(start_linear_pa > end_linear_pa);
1511 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)]
1512 * accommodating both its ends but excluding PAGE_END. Max physical
1513 * range which can be mapped inside this linear mapping range, must
1514 * also be derived from its end points.
1516 mhp_range.start = start_linear_pa;
1517 mhp_range.end = end_linear_pa;
1522 int arch_add_memory(int nid, u64 start, u64 size,
1523 struct mhp_params *params)
1525 int ret, flags = NO_EXEC_MAPPINGS;
1527 VM_BUG_ON(!mhp_range_allowed(start, size, true));
1530 * KFENCE requires linear map to be mapped at page granularity, so that
1531 * it is possible to protect/unprotect single pages in the KFENCE pool.
1533 if (can_set_direct_map() || IS_ENABLED(CONFIG_KFENCE))
1534 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
1536 __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
1537 size, params->pgprot, __pgd_pgtable_alloc,
1540 memblock_clear_nomap(start, size);
1542 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT,
1545 __remove_pgd_mapping(swapper_pg_dir,
1546 __phys_to_virt(start), size);
1548 max_pfn = PFN_UP(start + size);
1549 max_low_pfn = max_pfn;
1555 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
1557 unsigned long start_pfn = start >> PAGE_SHIFT;
1558 unsigned long nr_pages = size >> PAGE_SHIFT;
1560 __remove_pages(start_pfn, nr_pages, altmap);
1561 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size);
1565 * This memory hotplug notifier helps prevent boot memory from being
1566 * inadvertently removed as it blocks pfn range offlining process in
1567 * __offline_pages(). Hence this prevents both offlining as well as
1568 * removal process for boot memory which is initially always online.
1569 * In future if and when boot memory could be removed, this notifier
1570 * should be dropped and free_hotplug_page_range() should handle any
1571 * reserved pages allocated during boot.
1573 static int prevent_bootmem_remove_notifier(struct notifier_block *nb,
1574 unsigned long action, void *data)
1576 struct mem_section *ms;
1577 struct memory_notify *arg = data;
1578 unsigned long end_pfn = arg->start_pfn + arg->nr_pages;
1579 unsigned long pfn = arg->start_pfn;
1581 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE))
1584 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
1585 unsigned long start = PFN_PHYS(pfn);
1586 unsigned long end = start + (1UL << PA_SECTION_SHIFT);
1588 ms = __pfn_to_section(pfn);
1589 if (!early_section(ms))
1592 if (action == MEM_GOING_OFFLINE) {
1594 * Boot memory removal is not supported. Prevent
1595 * it via blocking any attempted offline request
1596 * for the boot memory and just report it.
1598 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end);
1600 } else if (action == MEM_OFFLINE) {
1602 * This should have never happened. Boot memory
1603 * offlining should have been prevented by this
1604 * very notifier. Probably some memory removal
1605 * procedure might have changed which would then
1606 * require further debug.
1608 pr_err("Boot memory [%lx %lx] offlined\n", start, end);
1611 * Core memory hotplug does not process a return
1612 * code from the notifier for MEM_OFFLINE events.
1613 * The error condition has been reported. Return
1614 * from here as if ignored.
1622 static struct notifier_block prevent_bootmem_remove_nb = {
1623 .notifier_call = prevent_bootmem_remove_notifier,
1627 * This ensures that boot memory sections on the platform are online
1628 * from early boot. Memory sections could not be prevented from being
1629 * offlined, unless for some reason they are not online to begin with.
1630 * This helps validate the basic assumption on which the above memory
1631 * event notifier works to prevent boot memory section offlining and
1632 * its possible removal.
1634 static void validate_bootmem_online(void)
1636 phys_addr_t start, end, addr;
1637 struct mem_section *ms;
1641 * Scanning across all memblock might be expensive
1642 * on some big memory systems. Hence enable this
1643 * validation only with DEBUG_VM.
1645 if (!IS_ENABLED(CONFIG_DEBUG_VM))
1648 for_each_mem_range(i, &start, &end) {
1649 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) {
1650 ms = __pfn_to_section(PHYS_PFN(addr));
1653 * All memory ranges in the system at this point
1654 * should have been marked as early sections.
1656 WARN_ON(!early_section(ms));
1659 * Memory notifier mechanism here to prevent boot
1660 * memory offlining depends on the fact that each
1661 * early section memory on the system is initially
1662 * online. Otherwise a given memory section which
1663 * is already offline will be overlooked and can
1664 * be removed completely. Call out such sections.
1666 if (!online_section(ms))
1667 pr_err("Boot memory [%llx %llx] is offline, can be removed\n",
1668 addr, addr + (1UL << PA_SECTION_SHIFT));
1673 static int __init prevent_bootmem_remove_init(void)
1677 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
1680 validate_bootmem_online();
1681 ret = register_memory_notifier(&prevent_bootmem_remove_nb);
1683 pr_err("%s: Notifier registration failed %d\n", __func__, ret);
1687 early_initcall(prevent_bootmem_remove_init);