Re: [BUG] fs: ext4: possible ABBA deadlock in ext4_inline_data_truncate() and ext4_punch_hole()

[Jia-Ju Bai <baijiaju1990@gmail.com>]

On 2021/12/10 1:00, Theodore Y. Ts'o wrote:
> On Thu, Dec 09, 2021 at 07:10:44PM +0800, Jia-Ju Bai wrote:
>> Hello,
>>
>> My static analysis tool reports a possible ABBA deadlock in the ext4 module
>> in Linux 5.10:
>>
>> ext4_inline_data_truncate()
>>    down_write(&EXT4_I(inode)->i_data_sem); --> Line 1895 (Lock A)
>>    ext4_xattr_ibody_get()
>>      ext4_xattr_inode_get()
>>        ext4_xattr_inode_iget()
>>          inode_lock(inode); --> Line 427 (Lock B)
>>
>> ext4_punch_hole()
>>    inode_lock(inode); --> Line 4018 (Lock B)
>>    ext4_update_disksize_before_punch()
>>      ext4_update_i_disksize()
>>        down_write(&EXT4_I(inode)->i_data_sem); --> Line 3248 (Lock A)
>>
>> When ext4_inline_data_truncate() and ext4_punch_hole() are concurrently
>> executed, the deadlock can occur.
>>
>> I am not quite sure whether this possible deadlock is real and how to fix it
>> if it is real.
> Hi,
>
> Thanks for the report.  I don't believe this is deadlock is possible,
> because the first thing ext4_punch_hole() does is to check to see if
> the inode has inline data --- and if so, it calls
> ext4_convert_inline_data() to convert it to a normal file.  In
> ext4_convert_inline_data(), we take the xattr lock, and then do the
> conversion, and then drop the xattr lock.  So by the time
> ext4_punch_hole() starts doing its work, the inode is not an inline
> data file.
>
> In ext4_inline_data_truncate(), we take the xattr lock, and once we
> have the xattr lock, we check to see if inode is still an inline data
> file.  If it has been converted, we then bail out.
>
> Hence, the ABBA deadlock that your static analysis tool has pointed
> shouldn't happen in practice.

Hi Ted,

Thank you very much for the detailed explanation!
I will improve my static analysis tool for this point.


Best wishes,
Jia-Ju Bai