When opening the slave end of a PTY, it is not possible for userspace to
safely ensure that /dev/pts/$num is actually a slave (in cases where the
mount namespace in which devpts was mounted is controlled by an
untrusted process). In addition, there are several unresolvable
race conditions if userspace were to attempt to detect attacks through
stat(2) and other similar methods [in addition it is not clear how
userspace could detect attacks involving FUSE].
Resolve this by providing an interface for userpace to safely open the
"peer" end of a PTY file descriptor by using the dentry cached by
devpts. Since it is not possible to have an open master PTY without
having its slave exposed in /dev/pts this interface is safe. This
interface currently does not provide a way to get the master pty (since
it is not clear whether such an interface is safe or even useful).
Cc: Christian Brauner <christian.brauner@xxxxxxxxxx>
Cc: Valentin Rothberg <vrothberg@xxxxxxxx>
Signed-off-by: Aleksa Sarai <asarai@xxxxxxx>
Is this going to be documented anywhere? Is there a man page update
that also goes along with this?
I will add one, I didn't know where the man-pages project is hosted /
where patches get pushed? What is the ML?
What userspace program wants to use this?
LXC (Christian is on Cc) will use this, runC will most likely use it,
pending on some design discussions (as well as some future container
runtimes I'm planning on working on). Effectively any container runtime
that wants to safely create terminals and spawn containers inside an
existing container's namespaces will likely want to use this.
[ As an aside, I /would/ argue this is a security fix (it fixes an
interface problem that made doing certain operations securely possible)
but I didn't want to Cc stable@ because it's a feature and not a strict
bugfix. ]
--
Aleksa Sarai
Software Engineer (Containers)
SUSE Linux GmbH
https://www.cyphar.com/
