Re: [PATCH net-next 1/3] vsock: add network namespace support

[Stefano Garzarella]

On Wed, 5 Mar 2025 at 10:29, Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
>
> On Wed, Mar 05, 2025 at 10:23:08AM +0100, Stefano Garzarella wrote:
> > On Wed, 5 Mar 2025 at 08:32, Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > >
> > > On Thu, Jan 16, 2020 at 06:24:26PM +0100, Stefano Garzarella wrote:
> > > > This patch adds a check of the "net" assigned to a socket during
> > > > the vsock_find_bound_socket() and vsock_find_connected_socket()
> > > > to support network namespace, allowing to share the same address
> > > > (cid, port) across different network namespaces.
> > > >
> > > > This patch adds 'netns' module param to enable this new feature
> > > > (disabled by default), because it changes vsock's behavior with
> > > > network namespaces and could break existing applications.
> > > > G2H transports will use the default network namepsace (init_net).
> > > > H2G transports can use different network namespace for different
> > > > VMs.
> > >
> > >
> > > I'm not sure I understand the usecase. Can you explain a bit more,
> > > please?
> >
> > It's been five years, but I'm trying!
> > We are tracking this RFE here [1].
> >
> > I also add Jakub in the thread with who I discussed last year a possible
> > restart of this effort, he could add more use cases.
> >
> > The problem with vsock, host-side, currently is that if you launch a VM
> > with a virtio-vsock device (using vhost) inside a container (e.g.,
> > Kata), so inside a network namespace, it is reachable from any other
> > container, whereas they would like some isolation. Also the CID is
> > shared among all, while they would like to reuse the same CID in
> > different namespaces.
> >
> > This has been partially solved with vhost-user-vsock, but it is
> > inconvenient to use sometimes because of the hybrid-vsock problem
> > (host-side vsock is remapped to AF_UNIX).
> >
> > Something from the cover letter of the series [2]:
> >
> >   As we partially discussed in the multi-transport proposal, it could
> >   be nice to support network namespace in vsock to reach the following
> >   goals:
> >   - isolate host applications from guest applications using the same ports
> >     with CID_ANY
> >   - assign the same CID of VMs running in different network namespaces
> >   - partition VMs between VMMs or at finer granularity
> >
> > Thanks,
> > Stefano
> >
> > [1] https://gitlab.com/vsock/vsock/-/issues/2
> > [2] https://lore.kernel.org/virtualization/20200116172428.311437-1-sgarzare@xxxxxxxxxx/
>
>
> Ok so, host side. I get it.

Now that we're talking about it, I also came back to a guest side
case, again related to containers and possible nested VMs.

If you launch a container in a L1 guest, for example to launch a
nested VM, maybe you don't want to have it communicate with the L0
host, so it would be desirable to be able to isolate the virtio-vsock
device from it.

> And the problem with your patches is that
> they affect the guest side. Fix that, basically.

My main problem, IIRC, was making sure to allow the old behavior as
well (but that maybe we had solved with two /dev/vhost-vsock and
/dev/vhost-vsock-netns).

The other problem was really in the guest, on how to tell that the
virtio-vsock device (thus communication with the host) was reachable
from a netnamespace or not.

Stefano