interesting way to break ext3 using loop devices

interesting way to break ext3 using loop devices

[Andi Kleen]

By mistake i found a sick way to break ext3:

Create a sparse (only consisting of a large hole) loop file on a ext3 file system.
losetup a loop device over it
Create a ext3 file system in the loop device
Mount it.
Copy the "parent" file system into the mounted loop fs with cp -rv
Eventually it will work on the copy of the loop file.

When you Ctrl-C it the system becomes quite confused: 
- Syslog is flooded with IO errors
- The loop fs cannot be unmounted anymore
- sync gets stuck

Of course this was clearly an operator error and
copying a file system into itself is clearly a sick thing 
to do, but still it shouldn't cause symptoms like this. Perhaps
it points to a deeper issue.

-Andi

Re: interesting way to break ext3 using loop devices

[Theodore Tso]

On Wed, Nov 15, 2006 at 03:36:20PM +0100, Andi Kleen wrote:
> Of course this was clearly an operator error and
> copying a file system into itself is clearly a sick thing 
> to do, but still it shouldn't cause symptoms like this. Perhaps
> it points to a deeper issue.

Agreed; I'm guessing it's a bug in the loop device at first glance.

						- Ted

Re: interesting way to break ext3 using loop devices

[Bill Rugolsky Jr.]

On Wed, Nov 15, 2006 at 03:36:20PM +0100, Andi Kleen wrote:
> Of course this was clearly an operator error and
> copying a file system into itself is clearly a sick thing 
> to do, but still it shouldn't cause symptoms like this. Perhaps
> it points to a deeper issue.
 
Copying is silly, moving is perhaps not.
That is in fact how the in-place file system conversion tools

   Convertfs - http://directory.fsf.org/sysadmin/misc/ConvertFS.html

and

   anyfs-tools - http://anyfs-tools.sourceforge.net/

work: they move files into a loop mounted filesystem, use FIBMAP to
find out the logical<->physical block map, then umount the lot and move the
blocks of the loop file system to their final physical destination.
The assumption, of course, is that there is enough slack free space to
accommodate the extra meta-data.

Regards,

	Bill Rugolsky