On 05/07/2023 17:00, Jeff Xu wrote:
On Fri, Jun 30, 2023 at 11:23 AM Mickaël Salaün <mic@xxxxxxxxxxx> wrote:
On 30/06/2023 06:18, Jeff Xu wrote:
On Thu, Jun 29, 2023 at 4:07 AM Mickaël Salaün <mic@xxxxxxxxxxx> wrote:
On 29/06/2023 05:18, Jeff Xu wrote:
resend.
On Wed, Jun 28, 2023 at 12:29 PM Mickaël Salaün <mic@xxxxxxxxxxx> wrote:
On 28/06/2023 19:03, Jeff Xu wrote:
[...]
The sandboxing/Landlock threat model is to restrict a process when it is
sandboxed, but this sandboxing is a request from the same process (or
one of its parent) that happen when it is more trustworthy (or at least
has more privileges) than after it sandbox itself.
The process sandboxing itself can use several kernel features, and one
of it is Landlock. In any case, it should take care of closing file
descriptors that should not be passed to the sandboxed process.
Agree.
The limits of sandboxing are the communication channels from and to
outside the sandbox. The peers talking with sandboxed processes should
then not be subject to confused deputy attacks, which means they must
not enable to bypass the user-defined security policy (from which the
Landlock policy is only a part). Receiving file descriptors should then
not be more important than controlling the communication channels. If a
not-sandboxed process is willing to give more right to a sandboxed
process, by passing FDs or just receiving commands, then this
not-sandboxed process need to be fixed.
This is the rationale to not care about received nor sent file
descriptors. The communication channels and the remote peers must be
trusted to not give more privileges to the sandboxed processes.
If a peer is malicious, it doesn't need to pass a file descriptor to the
sandboxed process, it can just read (data) commands and apply them to
its file descriptors.
I see the reasoning. i.e. sandboxing the process is not more
important than securing communication channels, or securing the peer.
So in a system that let a peer process to pass a socket into a
higher privileged process, when the communication channel or the peer
process is compromised, e.g. swapping the fd/socket into a different
one that the attacker controls, confuse deputy attack can happen. The
recommendation here is to secure peer and communication.
I agree with this approach in general. I need to think about how it
applies to specific cases.
I think the ability to pass file descriptors
should be seen as a way to improve performance by avoiding a user space
process to act as a proxy receiving read/write commands and managing
file descriptors itself. On the other hand, file descriptors could be
used as real capabilities/tokens to manage access, but senders still
need to be careful to only pass the required ones.
All this to say that being able to restrict actions on file descriptors
would be useful for senders/services to send a subset of the file
descriptor capabilities (cf. Capsicum), but not the other way around.
In the Landlock kernel doc:
Similarly to file access modes (e.g. O_RDWR), Landlock access rights
attached to file descriptors are retained even if they are passed
between processes (e.g. through a Unix domain socket). Such access
rights will then be enforced even if the receiving process is not
sandboxed by Landlock. Indeed, this is required to keep a consistent
access control over the whole system, and this avoids unattended
bypasses through file descriptor passing (i.e. confused deputy
attack).
iiuc, the design for file and socket in landlock is different. For
socket, the access rules are applied only to the current process (more
like seccomp), while for file restriction, the rules can be passed
into another un-landlocked process.
The O_RDWR restrictions are enforced by the basic kernel access control,
not Landlock. However, for file truncation, Landlock complements the
basic kernel access rights and behave the same.
There is indeed slight differences between file system and socket
restrictions. For the file system, a file descriptor is a direct access
to a file/data. For the network, we cannot identify for which data/peer
a newly created socket will give access to, we need to wait for a
connect or bind request to identify the use case for this socket. We
could tie the access rights (related to ports) to an opened socket, but
this would not align with the way Landlock access control works for the
file system. Indeed, a directory file descriptor may enable to open
another file (i.e. a new data item), but this opening is restricted by
Landlock. A newly created socket gives access to the network (or a
subset of it), but binding or connecting to a peer (i.e. accessing new
data) is restricted by Landlock. Accesses tied to FDs are those that
enable to get access to the underlying data (e.g. read, write,
truncate). A newly created socket is harmless until it is connected to a
peer, similarly to a memfd file descriptor. A directory opened by a
sandboxed process can be passed to a process outside this sandbox and it
might be allowed to open a relative path/file, which might not be the
case for the sandboxed process.
I think it might be summarize by the difference between underlying FD
data in the case of a regular file (i.e. tied access rights), and
relative new data in the case of a directory or a socket (i.e.
sandboxing policy scope).
