[PATCH V4 3/4] arm64/mm: Use p4dp_get() for P4D accesses

1 view
Skip to first unread message

Anshuman Khandual

unread,
Nov 4, 2026, 11:34:22 PMNov 4
to linux-ar...@lists.infradead.org, catalin...@arm.com, wi...@kernel.org, Anshuman Khandual, Ryan Roberts, Mark Rutland, Lorenzo Stoakes, Andrew Morton, David Hildenbrand, Mike Rapoport, Linu Cherian, linux-...@vger.kernel.org, linu...@kvack.org, kasa...@googlegroups.com
Replace READ_ONCE() with p4dp_get() for P4D accesses in preparation for
supporting both D64 and D128 translation table formats.

READ_ONCE() cannot currently be used for 128-bit page table entries on
arm64 because it does not provide the required 128-bit single-copy
atomicity, causing builds to fail for accesses wider than 64 bits.

Although LDP/STP provide the required atomicity when FEAT_LSE is
available (as required by FEAT_D128), extending READ_ONCE() to support
128-bit accesses is undesirable. READ_ONCE() is a general-purpose API,
so doing so could encourage other 128-bit users that would either fail
to build in configurations without D128 support or, if D128 becomes a
runtime option, silently permit tearing on systems without the required
hardware support.

Instead, standardize P4D accesses on the existing page-table helpers.
These can be overridden on arm64 to provide 128-bit single-copy
atomicity when required.

Both READ_ONCE() instances in pud_offset_[phys]() could not be converted
into p4pd_get() because of cyclic dependency between <linux/pgtable.h> and
<asm/pgtable.h>. Move the function inside <arch/arm64/mm/mmu.c> instead to
workaround the problem.

No functional change intended.

Cc: Catalin Marinas <catalin...@arm.com>
Cc: Will Deacon <wi...@kernel.org>
Cc: Ryan Roberts <ryan.r...@arm.com>
Cc: Mark Rutland <mark.r...@arm.com>
Cc: linux-ar...@lists.infradead.org
Cc: linux-...@vger.kernel.org
Cc: kasa...@googlegroups.com
Reviewed-by: Ryan Roberts <ryan.r...@arm.com>
Signed-off-by: Anshuman Khandual <anshuman...@arm.com>
---
Changes in V4:

- Dropped p4dp_get() platform override
- Moved both pud_offset[_phys]() inside arch/arm64/mm/mmu.c

arch/arm64/include/asm/pgtable.h | 12 ++----------
arch/arm64/mm/fault.c | 2 +-
arch/arm64/mm/fixmap.c | 2 +-
arch/arm64/mm/hugetlbpage.c | 2 +-
arch/arm64/mm/kasan_init.c | 4 ++--
arch/arm64/mm/mmu.c | 26 ++++++++++++++++++++------
arch/arm64/mm/pageattr.c | 2 +-
arch/arm64/mm/trans_pgd.c | 4 ++--
8 files changed, 30 insertions(+), 24 deletions(-)

diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index 96cfcc70acaa..8dffa7f787ac 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -960,12 +960,7 @@ static inline pud_t *p4d_pgtable(p4d_t p4d)
return (pud_t *)__va(p4d_page_paddr(p4d));
}

-static inline phys_addr_t pud_offset_phys(p4d_t *p4dp, unsigned long addr)
-{
- VM_WARN_ON_ONCE(!pgtable_l4_enabled());
-
- return p4d_page_paddr(READ_ONCE(*p4dp)) + pud_index(addr) * sizeof(pud_t);
-}
+phys_addr_t pud_offset_phys(p4d_t *p4dp, unsigned long addr);

static inline
pud_t *pud_offset_lockless(p4d_t *p4dp, p4d_t p4d, unsigned long addr)
@@ -976,10 +971,7 @@ pud_t *pud_offset_lockless(p4d_t *p4dp, p4d_t p4d, unsigned long addr)
}
#define pud_offset_lockless pud_offset_lockless

-static inline pud_t *pud_offset(p4d_t *p4dp, unsigned long addr)
-{
- return pud_offset_lockless(p4dp, READ_ONCE(*p4dp), addr);
-}
+pud_t *pud_offset(p4d_t *p4dp, unsigned long addr);
#define pud_offset pud_offset

static inline pud_t *pud_set_fixmap(unsigned long addr)
diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index ea0cdc730768..934b534c5b89 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c
@@ -174,7 +174,7 @@ static void show_pte(unsigned long addr)
break;

p4dp = p4d_offset_lockless(pgdp, pgd, addr);
- p4d = READ_ONCE(*p4dp);
+ p4d = p4dp_get(p4dp);
ptval_to_str(pxd_str, p4d_val(p4d));
pr_cont(", p4d=%s", pxd_str);
if (p4d_none(p4d) || p4d_bad(p4d))
diff --git a/arch/arm64/mm/fixmap.c b/arch/arm64/mm/fixmap.c
index d9a870836faf..2a0fe0b0f771 100644
--- a/arch/arm64/mm/fixmap.c
+++ b/arch/arm64/mm/fixmap.c
@@ -74,7 +74,7 @@ static void __init early_fixmap_init_pmd(pud_t *pudp, unsigned long addr,
static void __init early_fixmap_init_pud(p4d_t *p4dp, unsigned long addr,
unsigned long end)
{
- p4d_t p4d = READ_ONCE(*p4dp);
+ p4d_t p4d = p4dp_get(p4dp);
pud_t *pudp;

if (CONFIG_PGTABLE_LEVELS > 3 && !p4d_none(p4d) &&
diff --git a/arch/arm64/mm/hugetlbpage.c b/arch/arm64/mm/hugetlbpage.c
index c9ad5e75b073..5771bf49e1fe 100644
--- a/arch/arm64/mm/hugetlbpage.c
+++ b/arch/arm64/mm/hugetlbpage.c
@@ -288,7 +288,7 @@ pte_t *huge_pte_offset(struct mm_struct *mm,
return NULL;

p4dp = p4d_offset(pgdp, addr);
- if (!p4d_present(READ_ONCE(*p4dp)))
+ if (!p4d_present(p4dp_get(p4dp)))
return NULL;

pudp = pud_offset(p4dp, addr);
diff --git a/arch/arm64/mm/kasan_init.c b/arch/arm64/mm/kasan_init.c
index aad29bcc7622..4dd8c1186471 100644
--- a/arch/arm64/mm/kasan_init.c
+++ b/arch/arm64/mm/kasan_init.c
@@ -89,7 +89,7 @@ static pmd_t *__init kasan_pmd_offset(pud_t *pudp, unsigned long addr, int node,
static pud_t *__init kasan_pud_offset(p4d_t *p4dp, unsigned long addr, int node,
bool early)
{
- if (p4d_none(READ_ONCE(*p4dp))) {
+ if (p4d_none(p4dp_get(p4dp))) {
phys_addr_t pud_phys = early ?
__pa_symbol(kasan_early_shadow_pud)
: kasan_alloc_zeroed_page(node);
@@ -162,7 +162,7 @@ static void __init kasan_p4d_populate(pgd_t *pgdp, unsigned long addr,
do {
next = p4d_addr_end(addr, end);
kasan_pud_populate(p4dp, addr, next, node, early);
- } while (p4dp++, addr = next, addr != end && p4d_none(READ_ONCE(*p4dp)));
+ } while (p4dp++, addr = next, addr != end && p4d_none(p4dp_get(p4dp)));
}

static void __init kasan_pgd_populate(unsigned long addr, unsigned long end,
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 7aa8793f13c5..711d734a63f7 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -367,7 +367,7 @@ static int alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end,
{
int ret = 0;
unsigned long next;
- p4d_t p4d = READ_ONCE(*p4dp);
+ p4d_t p4d = p4dp_get(p4dp);
pud_t *pudp;

if (p4d_none(p4d)) {
@@ -457,7 +457,7 @@ static int alloc_init_p4d(pgd_t *pgdp, unsigned long addr, unsigned long end,
}

do {
- p4d_t old_p4d = READ_ONCE(*p4dp);
+ p4d_t old_p4d = p4dp_get(p4dp);

next = p4d_addr_end(addr, end);

@@ -467,7 +467,7 @@ static int alloc_init_p4d(pgd_t *pgdp, unsigned long addr, unsigned long end,
goto out;

VM_WARN_ON_ONCE(p4d_val(old_p4d) != 0 &&
- p4d_val(old_p4d) != READ_ONCE(p4d_val(*p4dp)));
+ p4d_val(old_p4d) != (p4d_val(p4dp_get(p4dp))));

phys += next - addr;
} while (p4dp++, addr = next, addr != end);
@@ -1622,7 +1622,7 @@ static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr,
do {
next = p4d_addr_end(addr, end);
p4dp = p4d_offset(pgdp, addr);
- p4d = READ_ONCE(*p4dp);
+ p4d = p4dp_get(p4dp);
if (p4d_none(p4d))
continue;

@@ -1788,7 +1788,7 @@ static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
do {
next = p4d_addr_end(addr, end);
p4dp = p4d_offset(pgdp, addr);
- p4d = READ_ONCE(*p4dp);
+ p4d = p4dp_get(p4dp);
if (p4d_none(p4d))
continue;

@@ -1809,7 +1809,7 @@ static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
*/
p4dp = p4d_offset(pgdp, 0UL);
for (i = 0; i < PTRS_PER_P4D; i++) {
- if (!p4d_none(READ_ONCE(p4dp[i])))
+ if (!p4d_none(p4dp_get(p4dp + i)))
return;
}

@@ -2450,3 +2450,17 @@ bool pmdp_test_and_clear_young(struct vm_area_struct *vma, unsigned long address
return __ptep_test_and_clear_young(vma, address, (pte_t *)pmdp);
}
#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG */
+
+#if CONFIG_PGTABLE_LEVELS > 3
+phys_addr_t pud_offset_phys(p4d_t *p4dp, unsigned long addr)
+{
+ VM_WARN_ON_ONCE(!pgtable_l4_enabled());
+
+ return p4d_page_paddr(p4dp_get(p4dp)) + pud_index(addr) * sizeof(pud_t);
+}
+
+pud_t *pud_offset(p4d_t *p4dp, unsigned long addr)
+{
+ return pud_offset_lockless(p4dp, p4dp_get(p4dp), addr);
+}
+#endif /* CONFIG_PGTABLE_LEVELS > 3 */
diff --git a/arch/arm64/mm/pageattr.c b/arch/arm64/mm/pageattr.c
index c544a61a5afd..172bc516164c 100644
--- a/arch/arm64/mm/pageattr.c
+++ b/arch/arm64/mm/pageattr.c
@@ -351,7 +351,7 @@ bool kernel_page_present(struct page *page)
return false;

p4dp = p4d_offset(pgdp, addr);
- if (p4d_none(READ_ONCE(*p4dp)))
+ if (p4d_none(p4dp_get(p4dp)))
return false;

pudp = pud_offset(p4dp, addr);
diff --git a/arch/arm64/mm/trans_pgd.c b/arch/arm64/mm/trans_pgd.c
index d119119455f1..7afe2beca4ba 100644
--- a/arch/arm64/mm/trans_pgd.c
+++ b/arch/arm64/mm/trans_pgd.c
@@ -99,7 +99,7 @@ static int copy_pud(struct trans_pgd_info *info, p4d_t *dst_p4dp,
unsigned long next;
unsigned long addr = start;

- if (p4d_none(READ_ONCE(*dst_p4dp))) {
+ if (p4d_none(p4dp_get(dst_p4dp))) {
dst_pudp = trans_alloc(info);
if (!dst_pudp)
return -ENOMEM;
@@ -145,7 +145,7 @@ static int copy_p4d(struct trans_pgd_info *info, pgd_t *dst_pgdp,
src_p4dp = p4d_offset(src_pgdp, start);
do {
next = p4d_addr_end(addr, end);
- if (p4d_none(READ_ONCE(*src_p4dp)))
+ if (p4d_none(p4dp_get(src_p4dp)))
continue;
if (copy_pud(info, dst_p4dp, src_p4dp, addr, next))
return -ENOMEM;
--
2.43.0

sashi...@kernel.org

unread,
Nov 5, 2026, 11:34:49 PM (14 days ago) Nov 5
to Anshuman Khandual, catalin...@arm.com, wi...@kernel.org, Ryan Roberts, Mark Rutland, Lorenzo Stoakes, Andrew Morton, David Hildenbrand, Mike Rapoport, Linu Cherian, linu...@kvack.org, kasa...@googlegroups.com
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] arm64/mm: early boot KASAN fault in out-of-line pud_offset_phys()
--

commit eaa67607aabb8af35fe6970c5442e1ab262f0286
Author: Anshuman Khandual <anshuman...@arm.com>

arm64/mm: Use p4dp_get() for P4D accesses

This commit replaces READ_ONCE() with p4dp_get() for P4D accesses across
arm64/mm and moves pud_offset_phys() and pud_offset() out of
arch/arm64/include/asm/pgtable.h into arch/arm64/mm/mmu.c to avoid header
dependency cycles.

> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index afa15b1a8a247..5b5a85a0527d6 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c

[ ... ]

> @@ -2449,3 +2449,17 @@ bool pmdp_test_and_clear_young(struct vm_area_struct *vma, unsigned long address
> return __ptep_test_and_clear_young(vma, address, (pte_t *)pmdp);
> }
> #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG */
> +
> +#if CONFIG_PGTABLE_LEVELS > 3
> +phys_addr_t pud_offset_phys(p4d_t *p4dp, unsigned long addr)
> +{
> + VM_WARN_ON_ONCE(!pgtable_l4_enabled());
> +
> + return p4d_page_paddr(p4dp_get(p4dp)) + pud_index(addr) * sizeof(pud_t);
> +}

[Severity: High]
Does moving pud_offset_phys() out-of-line into arch/arm64/mm/mmu.c without
noinstr or __no_sanitize_address cause an early boot translation fault when
CONFIG_KASAN_GENERIC or CONFIG_KASAN_SW_TAGS is enabled with 4-level or
5-level page tables?

During early boot before start_kernel(), __primary_switched calls
kasan_early_init() in arch/arm64/mm/kasan_init.c to populate the initial
early KASAN shadow page tables:

__primary_switched
-> kasan_early_init()
-> kasan_pgd_populate(..., true)
-> kasan_p4d_populate(..., true)
-> kasan_pud_populate(..., true)
-> kasan_pud_offset(p4dp, addr, node, true)
-> pud_offset_kimg(p4dp, addr)
-> pud_offset_phys(p4dp, addr)

At that point, no KASAN shadow PTEs have been mapped yet, which is why
arch/arm64/mm/Makefile sets:

KASAN_SANITIZE_kasan_init.o := n

Previously, pud_offset_phys() was a static inline helper in
arch/arm64/include/asm/pgtable.h, so it was inlined into kasan_init.o
without KASAN instrumentation.

With pud_offset_phys() moved into arch/arm64/mm/mmu.c, which is compiled with
KASAN sanitization enabled, calling pud_offset_phys() from kasan_early_init()
triggers KASAN shadow memory checks on p4dp before the early shadow memory
is mapped.

Should pud_offset_phys() be marked noinstr (like set_swapper_pgd() in the
same file) or __no_sanitize_address?

--
Sashiko AI review · https://sashiko.dev/#/patchset/20261105043403.332644...@arm.com?part=3
Reply all
Reply to author
Forward
0 new messages