[PATCH v3 1/2] userfaultfd: clear the inherited uffd bit in move_swap_pte()

[PATCH v3 1/2] userfaultfd: clear the inherited uffd bit in move_swap_pte()

[Donggeun Yoo]

UFFDIO_MOVE on a swapped-out page installs the source PTE at the
destination unchanged, so a uffd bit set on a write-protected or
RWP-protected source lands in a destination VMA that was never
registered for either.  Nothing clears it there, and userspace sees:

 - /proc/<pid>/pagemap reports the page as uffd-tracked (bit 57), both
   while it is swapped out and after it is faulted back in.
 - MADV_COLLAPSE fails with EINVAL on a range containing it, because the
   collapse scan bails on a swap entry with the uffd bit set.
 - With CONFIG_PAGE_TABLE_CHECK, faulting the page in warns.
   do_swap_page() carries the bit into the present PTE and, since the
   destination isn't WP-registered, also makes it writable:

   WARNING: mm/page_table_check.c:202 at __page_table_check_ptes_set+0x185/0x1e0
   Call Trace:
    set_ptes+0x67/0xc0
    do_swap_page+0x990/0xfe0
    __handle_mm_fault+0x7d0/0xeb0
    handle_mm_fault+0x9c/0x250
    do_user_addr_fault+0x207/0x650
    exc_page_fault+0x65/0x150
    asm_exc_page_fault+0x26/0x30

Reaching it takes UFFDIO_MOVE out of a WP- or RWP-protected area into
one that isn't, on a page that is swapped out at the time.  It hasn't
been seen in practice.

A resident page doesn't carry the bit, because move_present_ptes() builds
the destination PTE from dst_vma->vm_page_prot.  Clear it on the moved
swap entry as well, then re-arm it if the destination is RWP-registered.

The WP case has been there since v6.8, where UFFDIO_MOVE was added.

Fixes: adef440691ba ("userfaultfd: UFFDIO_MOVE uABI")
Cc: <stable@vger.kernel.org>
Assisted-by: LLM
Signed-off-by: Donggeun Yoo <donggeunyoo.kernel@gmail.com>
---
 mm/userfaultfd.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
index 74f04c323c50f..f39f109f17989 100644
--- a/mm/userfaultfd.c
+++ b/mm/userfaultfd.c
@@ -1449,6 +1449,7 @@ static int move_swap_pte(struct mm_struct *mm, struct vm_area_struct *dst_vma,
 	orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
 	if (pgtable_supports_soft_dirty())
 		orig_src_pte = pte_swp_mksoft_dirty(orig_src_pte);
+	orig_src_pte = pte_swp_clear_uffd(orig_src_pte);
 	/* Re-arm RWP on the moved swap entry if dst_vma is RWP-registered. */
 	if (userfaultfd_rwp(dst_vma))
 		orig_src_pte = pte_swp_mkuffd(orig_src_pte);
-- 
2.53.0

Re: [PATCH v3 1/2] userfaultfd: clear the inherited uffd bit in move_swap_pte()

[Andrew Morton]

On Sat, 26 Sep 2026 21:41:44 +0900 Donggeun Yoo <donggeunyoo.kernel@gmail.com> wrote:

> UFFDIO_MOVE on a swapped-out page installs the source PTE at the
> destination unchanged, so a uffd bit set on a write-protected or
> RWP-protected source lands in a destination VMA that was never
> registered for either.  Nothing clears it there, and userspace sees:
> 
>  - /proc/<pid>/pagemap reports the page as uffd-tracked (bit 57), both
>    while it is swapped out and after it is faulted back in.
>  - MADV_COLLAPSE fails with EINVAL on a range containing it, because the
>    collapse scan bails on a swap entry with the uffd bit set.
>  - With CONFIG_PAGE_TABLE_CHECK, faulting the page in warns.
>    do_swap_page() carries the bit into the present PTE and, since the
>    destination isn't WP-registered, also makes it writable:
> 
>    WARNING: mm/page_table_check.c:202 at __page_table_check_ptes_set+0x185/0x1e0
>    Call Trace:
>     set_ptes+0x67/0xc0
>     do_swap_page+0x990/0xfe0
>     __handle_mm_fault+0x7d0/0xeb0
>     handle_mm_fault+0x9c/0x250
>     do_user_addr_fault+0x207/0x650
>     exc_page_fault+0x65/0x150
>     asm_exc_page_fault+0x26/0x30
> 
> Reaching it takes UFFDIO_MOVE out of a WP- or RWP-protected area into
> one that isn't, on a page that is swapped out at the time.  It hasn't
> been seen in practice.
> 
> A resident page doesn't carry the bit, because move_present_ptes() builds
> the destination PTE from dst_vma->vm_page_prot.  Clear it on the moved
> swap entry as well, then re-arm it if the destination is RWP-registered.
> 
> The WP case has been there since v6.8, where UFFDIO_MOVE was added.

Thanks.  This is triggerable by unprivileged userspace, so can
reviewers please prioritize.

> Fixes: adef440691ba ("userfaultfd: UFFDIO_MOVE uABI")
> Cc: <stable@vger.kernel.org>

This should be whizzed into 7.3-rcX pretty promptly.

It's not really appropriate to give the selftest such treatment - I'm
never really sure how to handle that.  I guess I'll keep things as you
presented them - both for current -rc, [1/2] gets backported, [2/2]
does not.

> --- a/mm/userfaultfd.c
> +++ b/mm/userfaultfd.c
> @@ -1449,6 +1449,7 @@ static int move_swap_pte(struct mm_struct *mm, struct vm_area_struct *dst_vma,
>  	orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
>  	if (pgtable_supports_soft_dirty())
>  		orig_src_pte = pte_swp_mksoft_dirty(orig_src_pte);
> +	orig_src_pte = pte_swp_clear_uffd(orig_src_pte);
>  	/* Re-arm RWP on the moved swap entry if dst_vma is RWP-registered. */
>  	if (userfaultfd_rwp(dst_vma))
>  		orig_src_pte = pte_swp_mkuffd(orig_src_pte);
> -- 
> 2.53.0

Re: [PATCH v3 1/2] userfaultfd: clear the inherited uffd bit in move_swap_pte()

[David Hildenbrand (Arm)]

On 9/27/26 22:54, Andrew Morton wrote:
> On Sat, 26 Sep 2026 21:41:44 +0900 Donggeun Yoo <donggeunyoo.kernel@gmail.com> wrote:
> 
>> UFFDIO_MOVE on a swapped-out page installs the source PTE at the
>> destination unchanged, so a uffd bit set on a write-protected or
>> RWP-protected source lands in a destination VMA that was never
>> registered for either.  Nothing clears it there, and userspace sees:
>>
>>  - /proc/<pid>/pagemap reports the page as uffd-tracked (bit 57), both
>>    while it is swapped out and after it is faulted back in.
>>  - MADV_COLLAPSE fails with EINVAL on a range containing it, because the
>>    collapse scan bails on a swap entry with the uffd bit set.
>>  - With CONFIG_PAGE_TABLE_CHECK, faulting the page in warns.
>>    do_swap_page() carries the bit into the present PTE and, since the
>>    destination isn't WP-registered, also makes it writable:
>>
>>    WARNING: mm/page_table_check.c:202 at __page_table_check_ptes_set+0x185/0x1e0
>>    Call Trace:
>>     set_ptes+0x67/0xc0
>>     do_swap_page+0x990/0xfe0
>>     __handle_mm_fault+0x7d0/0xeb0
>>     handle_mm_fault+0x9c/0x250
>>     do_user_addr_fault+0x207/0x650
>>     exc_page_fault+0x65/0x150
>>     asm_exc_page_fault+0x26/0x30
>>
>> Reaching it takes UFFDIO_MOVE out of a WP- or RWP-protected area into
>> one that isn't, on a page that is swapped out at the time.  It hasn't
>> been seen in practice.
>>
>> A resident page doesn't carry the bit, because move_present_ptes() builds
>> the destination PTE from dst_vma->vm_page_prot.  Clear it on the moved
>> swap entry as well, then re-arm it if the destination is RWP-registered.
>>
>> The WP case has been there since v6.8, where UFFDIO_MOVE was added.
> 
> Thanks.  This is triggerable by unprivileged userspace, so can
> reviewers please prioritize.

I assume most production kernels don't have CONFIG_PAGE_TABLE_CHECK enabled in
the config. And if enabled, not active at runtime.

I'd assume mostly only debug kernels have that config ever enabled. So it might
be less problematic than it seems at first. (at least the WARNING bit)

> 
>> Fixes: adef440691ba ("userfaultfd: UFFDIO_MOVE uABI")
>> Cc: <stable@vger.kernel.org>
> 
> This should be whizzed into 7.3-rcX pretty promptly.
> 
> It's not really appropriate to give the selftest such treatment - I'm
> never really sure how to handle that.  I guess I'll keep things as you
> presented them - both for current -rc, [1/2] gets backported, [2/2]
> does not.

Make sense.

> 
>> --- a/mm/userfaultfd.c
>> +++ b/mm/userfaultfd.c
>> @@ -1449,6 +1449,7 @@ static int move_swap_pte(struct mm_struct *mm, struct vm_area_struct *dst_vma,
>>  	orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
>>  	if (pgtable_supports_soft_dirty())
>>  		orig_src_pte = pte_swp_mksoft_dirty(orig_src_pte);
>> +	orig_src_pte = pte_swp_clear_uffd(orig_src_pte);
>>  	/* Re-arm RWP on the moved swap entry if dst_vma is RWP-registered. */
>>  	if (userfaultfd_rwp(dst_vma))
>>  		orig_src_pte = pte_swp_mkuffd(orig_src_pte);
>> -- 
>> 2.53.0

Acked-by: David Hildenbrand (Arm) <david@kernel.org>

-- 
Cheers,

David

Re: [PATCH v3 1/2] userfaultfd: clear the inherited uffd bit in move_swap_pte()

[Mike Rapoport]

On Sat, Sep 26, 2026 at 09:41:44PM +0900, Donggeun Yoo wrote:
> UFFDIO_MOVE on a swapped-out page installs the source PTE at the
> destination unchanged, so a uffd bit set on a write-protected or
> RWP-protected source lands in a destination VMA that was never
> registered for either.  Nothing clears it there, and userspace sees:
> 
>  - /proc/<pid>/pagemap reports the page as uffd-tracked (bit 57), both
>    while it is swapped out and after it is faulted back in.
>  - MADV_COLLAPSE fails with EINVAL on a range containing it, because the
>    collapse scan bails on a swap entry with the uffd bit set.
>  - With CONFIG_PAGE_TABLE_CHECK, faulting the page in warns.
>    do_swap_page() carries the bit into the present PTE and, since the
>    destination isn't WP-registered, also makes it writable:
> 
>    WARNING: mm/page_table_check.c:202 at __page_table_check_ptes_set+0x185/0x1e0
>    Call Trace:
>     set_ptes+0x67/0xc0
>     do_swap_page+0x990/0xfe0
>     __handle_mm_fault+0x7d0/0xeb0
>     handle_mm_fault+0x9c/0x250
>     do_user_addr_fault+0x207/0x650
>     exc_page_fault+0x65/0x150
>     asm_exc_page_fault+0x26/0x30
> 
> Reaching it takes UFFDIO_MOVE out of a WP- or RWP-protected area into
> one that isn't, on a page that is swapped out at the time.  It hasn't
> been seen in practice.
> 
> A resident page doesn't carry the bit, because move_present_ptes() builds
> the destination PTE from dst_vma->vm_page_prot.  Clear it on the moved
> swap entry as well, then re-arm it if the destination is RWP-registered.
> 
> The WP case has been there since v6.8, where UFFDIO_MOVE was added.
> 
> Fixes: adef440691ba ("userfaultfd: UFFDIO_MOVE uABI")
> Cc: <stable@vger.kernel.org>
> Assisted-by: LLM
> Signed-off-by: Donggeun Yoo <donggeunyoo.kernel@gmail.com>

Acked-by: Mike Rapoport (Microsoft) <rppt@kernel.org>


> ---
>  mm/userfaultfd.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
> index 74f04c323c50f..f39f109f17989 100644
> --- a/mm/userfaultfd.c
> +++ b/mm/userfaultfd.c
> @@ -1449,6 +1449,7 @@ static int move_swap_pte(struct mm_struct *mm, struct vm_area_struct *dst_vma,
>  	orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
>  	if (pgtable_supports_soft_dirty())
>  		orig_src_pte = pte_swp_mksoft_dirty(orig_src_pte);
> +	orig_src_pte = pte_swp_clear_uffd(orig_src_pte);
>  	/* Re-arm RWP on the moved swap entry if dst_vma is RWP-registered. */
>  	if (userfaultfd_rwp(dst_vma))
>  		orig_src_pte = pte_swp_mkuffd(orig_src_pte);
> -- 
> 2.53.0
> 

-- 
Sincerely yours,
Mike.