Re: [Bug report] journal data mode trigger panic in jbd2_journal_commit_transaction

[yangerkun <yangerkun@huawei.com>]

在 2020/11/20 10:54, yangerkun 写道:
> 
> 
> 在 2020/11/19 21:12, Mauricio Oliveira 写道:
>> On Thu, Nov 19, 2020 at 1:25 AM yangerkun <yangerkun@huawei.com> wrote:
>>>
>>>
>>>
>>> 在 2020/11/16 21:50, Mauricio Oliveira 写道:
>>>> Hi Kun,
>>>>
>>>> On Sat, Nov 14, 2020 at 5:18 AM yangerkun <yangerkun@huawei.com> wrote:
>>>>> While using ext4 with data=journal(3.10 kernel), we meet a problem 
>>>>> that
>>>>> we think may never happend...
>>>> [...]
>>>>
>>>> Could you please confirm you mean 5.10-rc* kernel instead of 3.10?
>>>> (It seems so as you mention a recent commit below.)  Thanks!
>>>>
>>>>> For now, what I have seen that can dirty buffer directly is
>>>>> ext4_page_mkwrite(64a9f1449950 ("ext4: data=journal: fixes for
>>>>> ext4_page_mkwrite()")), and runing ext4_punch_hole with keep_size
>>>>> /ext4_page_mkwrite parallel can trigger above warning easily.
>>>> [...]
>>>>
>>>>
>>>
>>> Hi,
>>>
>>> Sorry for the long delay reply... And thanks a lot for your advise! The
>>> bug trigger with a very low probability. So won't trigger with 5.10 can
>>> not prove no bug exist in 5.10.
>>>
>>
>> No worries, and thanks for following up.
>> So I understand that the bug report was indeed on 3.10, and 5.10-rcN
>> is not yet confirmed.
>>
>>> Google a lot and notice that someone before has report the same bug[1].
>>> '3b136499e906 ("ext4: fix data corruption in data=journal mode")' seems
>>> fix the problem. I will try to understand this, and give a analysis
>>> about how to reproduce it!
>>
>> Cool, thanks!
>>
>>> Thanks,
>>> Kun.
>>
>>
>>
> Hi,
> 
> The follow step can reproduce the bug[1] reported before easily. And the 
> bug we meet seems same. Following patch will fix the bug.
> 
> 3b136499e906 ext4: fix data corruption in data=journal mode
> b90197b65518 ext4: use private version of page_zero_new_buffers() for 
> data=journal mode
> 
> 
> 1. mkfs.ext4
> 2. touch $tofile(ino == 12)
> 3. touch $fromfile(ino == 13) and write 4k to fromfile and sync
> 
> mmap $fromfile 4k
> and write 4k
> to $tofile
> 
> ...
> generic_perform_write
>   ext4_write_begin
>    ext4_journal_start
>    (trans 1)
>   if (ino == 12) sleep for 30s
>   ...                           truncate $fromfile
>                                 to 0
>   copied=0,bytes=4k
>   ext4_journalled_write_end
>    page_zero_new_buffers
>     mark_buffer_dirty
>    write_end_fn
>     ...
>     __jbd2_journal_file_buffer
>      test_clear_buffer_dirty
>      __jbd2_journal_temp_unlink_buffer
            this will mark buffer dirty again!
>    ext4_journal_stop
>    (trans 1)
>                                                   trans1 commit
>                                                   ...
>    ext4_truncate_failed_write
>     ...
>     journal_unmap_buffer
>      set_buffer_freed
>                                                   forget list
>                                                    ...
>                                                    clear_buffer_jbddirty
>                                                    ...
>                                                    J_ASSERT_BH(bh,
>                                                    !buffer_dirty(bh))
>                                                    ^^^^^^^^^^^^^^^^^
>                                                    trigger the bug...
> 
> 
> 
> [1]. https://www.spinics.net/lists/linux-ext4/msg56447.html
> 
> Thanks,
> Kun.
> .