On Tue, Nov 8, 2022 at 7:33 PM Ryusuke Konishi wrote:
> On Tue, Nov 8, 2022 at 3:49 PM Liu Shixin wrote:
> > Then in nilfs_segctor_extend_segments(), we set sb_segnum by prev->sb_nextnum directly,
> > and calculate next sb_segnum by nilfs_sufile_alloc(), since last_alloc is not updated,
> > we will get sb_segnum again.
>
> nilfs_segctor_extend_segments() pre-allocates one or more segments
> large enough to store updated blocks of metadata files that need to be
> written in a series of logs at once, and sets up a chain of segbufs.
> (Sorry for the missing function comment).
>
> sb_segnum is set by prev->sb_nextnum to form a chain of buffers for
> segments. This is expected behavior.
> And, the sb_nextnum (= next sb_segnum) will be given by
> nilfs_sufile_alloc(). This is also expected.
> It looks like the problem is that nilfs_sufile_alloc() here allocates
> the same segnum again.
>
> Because sb_segnum is set by prev->sb_nextnum which is allocated by the
> previous nilfs_sufile_alloc() call,
> this usually does not happen.
>
> A possible anomaly is if the segment pointed by the first nextnum (or
> segnum) was not marked dirty on sufile.
> This may happen if the sufile is corrupted on the test disk image that
> syzbot provided (mounted).
>
> Can you confirm if this is actually happening?
If we can mount the test disk image, the state of sufile can be
confirmed quickly with lssu command:
$ lssu
SEGNUM DATE TIME STAT NBLOCKS
3 2022-11-04 23:23:49 -d- 2048
4 2022-11-04 23:23:50 ad- 103
5 ---------- --:--:-- ad- 0
Here, the flag "d" in STAT means the segment is dirty (in-use) and the
segment of ns_segnum or ns_nextnum is indicated with the "a" flag.
This is an example of a normal disk image.
Or, if it's easy to insert debug code to check, that's fine too.
Regards,
Ryusuke Konishi
