Re: [syzbot] [mm?] KMSAN: uninit-value in swap_writeout

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On 2025/12/24 08:16, Barry Song wrote:
> On Wed, Dec 24, 2025 at 12:43 PM Pedro Falcato <pfalcato@suse.de> wrote:
>>
>> On Wed, Dec 24, 2025 at 11:46:44AM +1300, Barry Song wrote:
>>>>
>>>> Uninit was created at:
>>>>   __alloc_frozen_pages_noprof+0x421/0xab0 mm/page_alloc.c:5233
>>>>   alloc_pages_mpol+0x328/0x860 mm/mempolicy.c:2486
>>>>   folio_alloc_mpol_noprof+0x56/0x1d0 mm/mempolicy.c:2505
>>>>   shmem_alloc_folio mm/shmem.c:1890 [inline]
>>>>   shmem_alloc_and_add_folio+0xc56/0x1bd0 mm/shmem.c:1932
>>>>   shmem_get_folio_gfp+0xad3/0x1fc0 mm/shmem.c:2556
>>>>   shmem_get_folio mm/shmem.c:2662 [inline]
>>>>   shmem_symlink+0x562/0xad0 mm/shmem.c:4129
>>>>   vfs_symlink+0x42f/0x4c0 fs/namei.c:5514
>>>>   do_symlinkat+0x2ae/0xbb0 fs/namei.c:5541
>>>
>>> +Hugh and Baolin.

Thanks for CCing me.

>>>
>>> This happens in the shmem symlink path, where newly allocated
>>> folios are not cleared for some reason. As a result,
>>> is_folio_zero_filled() ends up reading uninitialized data.
>>>
>>
>> I'm not Hugh nor Baolin, but I would guess that letting
>> is_folio_zero_filled() skip/disable KMSAN would also work. Since all we want
>> is to skip writeout if the folio is zero, whether it is incidentally zero, or not,
>> does not really matter, I think.
> 
> Hi Pedro, thanks! You’re always welcome to chime in.
> 
> You are probably right. However, I still prefer the remaining
> data to be zeroed, as it may be more compression-friendly.
> 
> Random data could potentially lead to larger compressed output,
> whereas a large area of zeros would likely result in much smaller
> compressed data.

Thanks Pedro and Barry. I remember Hugh raised a similar issue before 
(See [1], but I did not investigate further:(). I agree with Hugh's 
point that the uninitialized parts should be zeroed before going the 
outside world.

[1] 
https://lore.kernel.org/all/02a21a55-8fe3-a9eb-f54b-051d75ae8335@google.com/

> Not quite sure if the below can fix the issue:
> 
> diff --git a/mm/shmem.c b/mm/shmem.c
> index ec6c01378e9d..0ca2d4bffdb4 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -4131,6 +4131,7 @@ static int shmem_symlink(struct mnt_idmap *idmap, struct inode *dir,
>   			goto out_remove_offset;
>   		inode->i_op = &shmem_symlink_inode_operations;
>   		memcpy(folio_address(folio), symname, len);
> +		memset(folio_address(folio) + len, 0, folio_size(folio) - len);
>   		folio_mark_uptodate(folio);
>   		folio_mark_dirty(folio);
>   		folio_unlock(folio);

That looks reasonable to me, though I prefer to use the more readable 
helper: folio_zero_range(). Barry, could you send out a formal patch？ 
Thanks.