February 9, 2013
I've previously discussed how FUSE's nonstandard semantics can cause problems with rsync-based backups. In short, when stat() is called on a FUSE mount owned by another user, the kernel returns EACCES, even though POSIX says EACCES is for when a file's path can't be traversed. This is done to isolate the effects of an unstable or malicious FUSE filesystems to only the user who mounted it.
In my opinion, instead of stretching POSIX by returning EACCES, a better way to isolate FUSE mounts would be to make them invisible to other users. This has been discussed before, first in 2005 with a patch to add "private mounts" to Linux and later in 2006 with a proposal for stat() to return a fake stat structure for FUSE mounts. However, both times the proposals were rejected in favor of using the more general namespace support along with shared subtrees to achieve isolated FUSE mounts.
Unfortunately, while namespaces and shared subtrees are quite powerful, they have not seen widespread adoption, and userspace support for them is limited to some basic command primitives that don't do much on their own. While there is a PAM namespaces module, it's tailored to giving users isolated /tmp directories.
So, I wrote a very simple C program called with-fuse. with-fuse takes a command as its argument and executes that command with gid fuse and in an isolated mount namespace. Any mounts and unmounts performed inside the private namespace are invisible to the rest of the system. At the same time, mount and unmounts performed in the global namespace are immediately visible inside the private namespace. with-fuse can be safely installed setuid-root to give users on the system a means of using FUSE without affecting other users.
$ with-fuse /bin/sh
$ sshfs ...
For with-fuse to work, the following command must be run at system boot (for example, from /etc/rc.local):mount --make-rshared /
Note that with-fuse creates a per-process namespace, not a per-user namespace. That means that the mounts created in one with-fuse namespace will not be visible in another with-fuse namespace, even if both namespaces are owned by the same user. Therefore, the user may wish to run a terminal multiplexer like GNU Screen inside his with-fuse namespace, in order to share the namespace among several shells:
$ with-fuse screen
To ensure that users only use FUSE from within a with-fuse namespace, /dev/fuse should be owned by group fuse and have 660 permissions. No user should be a member of group fuse, as with-fuse will take care of granting that GID.
You can download the source for with-fuse here. It's short and extensively commented if you'd like to learn how it works.
Posted on 2013-02-09 at 06:50:46 UTC | Comments