[lvc-project] [PATCH] ocfs2: fix __counted_by_le() usage in ocfs2_expand_inline_dx_root()

Joseph Qi joseph.qi at linux.alibaba.com
Tue Oct 14 04:50:06 MSK 2025



On 2025/10/14 09:44, Heming Zhao wrote:
> Hi,
> 
> I have a different view.
> 
> On 10/14/25 08:55, Joseph Qi wrote:
>>
>>
>> On 2025/10/13 23:30, Dmitry Antipov wrote:
>>> After annotating 'l_recs[]' VLA member of 'struct ocfs2_extent_list'
>>> with '__counted_by_le(l_count)', 'l_count' should be initialized
>>> before touching 'l_recs[]' with 'memset()' to avoid false positives
>>> from buffer overflow guards when CONFIG_UBSAN_BOUNDS is enabled.
>>>
>>> Fixes: aa545adbe491 ("ocfs2: annotate flexible array members with __counted_by_le()")
>>> Signed-off-by: Dmitry Antipov <dmantipov at yandex.ru>
>>
>> Looks fine.
>> Reviewed-by: Joseph Qi <joseph.qi at linux.alibaba.com>
>>
>> Andrew, could you please help fold this into the origin patch? Thanks.
>>
>>> ---
>>> note: since the commit in subject is not merged to upstream yet, Fixes: reference is taken from
>>> https://web.git.kernel.org/pub/scm/linux/kernel/git/akpm/mm.git/commit/?h=mm-nonmm-unstable&id=aa545adbe491
>>> ---
>>>   fs/ocfs2/dir.c | 4 ++--
>>>   1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/fs/ocfs2/dir.c b/fs/ocfs2/dir.c
>>> index 3c2dacba9b0a..d9a2fcb63bbb 100644
>>> --- a/fs/ocfs2/dir.c
>>> +++ b/fs/ocfs2/dir.c
>>> @@ -4133,10 +4133,10 @@ static int ocfs2_expand_inline_dx_root(struct inode *dir,
>>>       }
>>>         dx_root->dr_flags &= ~OCFS2_DX_FLAG_INLINE;
>>> -    memset(&dx_root->dr_list, 0, osb->sb->s_blocksize -
>>> -           offsetof(struct ocfs2_dx_root_block, dr_list));
>>>       dx_root->dr_list.l_count =
>>>           cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
> 
> If the memset cleans dx_root->dr_list->l_recs[], or "osb->sb->s_blocksize -
> (offsetof(struct ocfs2_dx_root_block, dr_list) + offsetof(struct ocfs2_extent_list, l_recs))" this patch is correct.
> 
> But the subsequent memset() call cleans all the members of dr_list.
> Therefore, dr_list.l_count is reset to ZERO and the data is corrupted.
> 
Oops, my mistake.
Yes, you're right. Thanks for pointing out this.

Joseph




More information about the lvc-project mailing list