On 02/27/2014 05:29 PM, Matthew Wilcox wrote:
Some distributions use udisks2 to grant permission to local console
users to create new loop devices from files. File systems on these
block devices are then mounted. This is a replacement for several
file systems implemented in user space, and for the users, this is a
good thing because the in-kernel implementations are generally of
higher quality.
Just to be sure I understand; the user owns the file (so can change any
bit in it at will), and the loop device is used to present that file
to the filesystem as a block device to be mounted?
Yes, that's a fair summary.
> Have we fuzz-tested
all the filesystems enough to be sure that's safe? :-)
It raised some eyebrows. But I've looked at some of the userspace
alternatives, and I can see why we ended up with this.
What happens if we have DAX support in the entire stack, and an
enterprising user mounts a file system? Will she be able to fuzz
the file system or binfmt loaders concurrently, changing the bits
while they are being read?
Currently, it appears that the loop device duplicates pages in the
page cache, so this does not seem to be possible, but DAX support
might change this.
I haven't looked at adding DAX support to the loop device, although
that would make sense. At the moment, neither ext2 nor ext4 (our only
DAX-supporting filesystems) use DAX for their metadata, only for user
data. As far as fuzzing the binfmt loaders ... are these filesystems not
forced to be at least nosuid?
The kernel binfmt parser runs as root even without a SUID bit. :)
> I might go so far as to make them noexec.
Oh, that's an interesting idea.
Thanks for thinking about this. I didn't know allowing users to mount
files they owned was something distros actually did. Have we considered
prohibiting the user from modifying the file while it's mounted, eg
forcing its permissions to 0 or pretending it's immutable?
Perhaps like "Text file busy" for executables? How reliable is that in
practice?
Changing file permissions doesn't affected already open descriptors and
might not always be possible (the file system might be mounted
read-only, but still be modifiable beneath).
--
Florian Weimer / Red Hat Product Security Team
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>