We got issue as follows:
[home]# fsck.ext4 -fn ram0yb
e2fsck 1.45.6 (20-Mar-2020)
Pass 1: Checking inodes, blocks, and sizes
Pass 2: Checking directory structure
Symlink /p3/d14/d1a/l3d (inode #3494) is invalid.
Clear? no
Entry 'l3d' in /p3/d14/d1a (3383) has an incorrect filetype (was 7, should be 0).
Fix? no
As symlink file size not match to file content. If symlink data block writback
failed, will call ext4_finish_bio to end io. In this path don't mark buffer
error. When umount do checkpoint can't detect buffer error, then will cleanup
jounral. Actually, correct data maybe in journal area.
To solve this issue, mark buffer error when detect bio error in ext4_finish_bio.
Signed-off-by: Ye Bin <[email protected]>
---
fs/ext4/page-io.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/fs/ext4/page-io.c b/fs/ext4/page-io.c
index 495ce59fb4ad..14695e2b5042 100644
--- a/fs/ext4/page-io.c
+++ b/fs/ext4/page-io.c
@@ -134,8 +134,10 @@ static void ext4_finish_bio(struct bio *bio)
continue;
}
clear_buffer_async_write(bh);
- if (bio->bi_status)
+ if (bio->bi_status) {
+ set_buffer_write_io_error(bh);
buffer_io_error(bh);
+ }
} while ((bh = bh->b_this_page) != head);
spin_unlock_irqrestore(&head->b_uptodate_lock, flags);
if (!under_io) {
--
2.31.1
On 2022/3/21 23:11, Jan Kara wrote:
> Hello Yi!
>
> On Mon 21-03-22 22:38:49, Zhang Yi wrote:
>> On 2022/3/21 19:37, Jan Kara wrote:
>>> On Mon 21-03-22 19:34:08, Ye Bin wrote:
>>>> We got issue as follows:
>>>> [home]# fsck.ext4 -fn ram0yb
>>>> e2fsck 1.45.6 (20-Mar-2020)
>>>> Pass 1: Checking inodes, blocks, and sizes
>>>> Pass 2: Checking directory structure
>>>> Symlink /p3/d14/d1a/l3d (inode #3494) is invalid.
>>>> Clear? no
>>>> Entry 'l3d' in /p3/d14/d1a (3383) has an incorrect filetype (was 7, should be 0).
>>>> Fix? no
>>>>
>>>> As symlink file size not match to file content. If symlink data block
>>>> writback failed, will call ext4_finish_bio to end io. In this path don't
>>>> mark buffer error. When umount do checkpoint can't detect buffer error,
>>>> then will cleanup jounral. Actually, correct data maybe in journal area.
>>>> To solve this issue, mark buffer error when detect bio error in
>>>> ext4_finish_bio.
>>>
>>> Thanks for the patch! Let me rephrase the text a bit:
>>>
>>> As the symlink file size does not match the file content. If the writeback
>>> of the symlink data block failed, ext4_finish_bio() handles the end of IO.
>>> However this function fails to mark the buffer with BH_write_io_error and
>>> so when unmount does journal checkpoint it cannot detect the writeback
>>> error and will cleanup the journal. Thus we've lost the correct data in the
>>> journal area. To solve this issue, mark the buffer as BH_write_io_error in
>>> ext4_finish_bio().
>>>
>>
>> Thinking about this issue in depth, the symlink data block is one kind of
>> metadata, but the page mapping of such block is belongs to the ext4 inode,
>> it's not coordinate to other metadata blocks, e.g. directory block and extents
>> block. This is why we have already fix the same issue of other metadata blocks
>> in commit fcf37549ae19e9 "jbd2: ensure abort the journal if detect IO error
>> when writing original buffer back" but missing the case of symlink data block.
>> So, after Ye Bin's fix, I think it's worth to unify the symlink data block
>> mapping to bdev, any suggestions?
>
> Well, symlink with external block is essentially a case of data=journal
> data block. So even if we would handle symlinks, we would still need to
> deal with other inodes with journalled data. Also we need to keep the> symlink contents in the page cache to make it simple for generic VFS code
> handling symlinks. So I don't see how we could substantially unify
> things...
>
Yeah, this fix is still needed for other regular file's journalled data when we
mounted filesystem with data=jouranl mode. But if we just consider whether if we
could unify the journal mode of ext4's metadata blocks, it seems that using
data=journal mode for symlink's external data block is also complicated and
confused in the creating procedure. Instead, if we use ext4_bread(), it make
things clear, and it seems also has no side effect of reading symlinks. I write
a RFC patch to do this, please take a look at my latest mail "[RFC PATCH] ext4:
convert symlink external data block mapping to bdev".
Thanks,
Yi.