Re: [PATCH] conf, pasta: Add --no-tap option

[Stefano Brivio <sbrivio@redhat.com>]

[Cc'ing Jon for awareness around the part about netlink monitor and
capabilities, four paragraphs down]

On Wed, 7 Jan 2026 16:20:18 +0100
Paul Holzinger <pholzing@redhat.com> wrote:

> On 05/01/2026 22:10, Stefano Brivio wrote:
> > On Mon, 5 Jan 2026 14:48:15 +0100
> > Paul Holzinger <pholzing@redhat.com> wrote:
> >  
> >> Sorry I was out for a while so I didn't had time to clarify on the bug
> >> before.
> >>
> >> On 29/12/2025 10:55, Yumei Huang wrote:  
> >>> This patch introduces a mode where we only forward loopback connections
> >>> and traffic between two namespaces (via the loopback interface, 'lo'),
> >>> without a tap device.
> >>>
> >>> With this, podman can support forwarding ::1 in custom networks when using
> >>> rootlesskit for forwarding ports.  
> >> I guess I didn't really communicate my requirements well.  
> > I guess it's more likely that you actually did, but I mixed up the
> > association between requirements and use cases, sorry for that.
> >
> > In any case, good that we need this anyway, just for another use case.
> > :)
> >  
> >> When we use
> >> rootlessport (rootlesskit) today for custom networks we only do so as
> >> rootless user and it forwards ::1 (by possibly mapping this to v4 inside
> >> the container) fine.  
> > So, wait a moment, is my comment at:
> >
> >    https://github.com/containers/podman/issues/14491#issuecomment-2898191772
> >
> > actually wrong? I don't have time right now to test that but from user
> > reports and some vague memory I thought ::1 forwarding wouldn't work
> > with custom networks regardless of root or rootless, because
> > rootlesskit didn't handle that anyway.  
>
> yes, rootlesskit handles ipv6 just fine, it is just that our 
> rootlessport code remaps that to v4 inside the container.

Actually, at a glance, I don't think that this could be fixed entirely
in the rootlessport implementation, as rootlesskit doesn't seem to look
at the destination address of the original connection at all.

> >> My main point for this feature was using as root (requires further
> >> changes to allow pasta running as root).  
> > ...which should be entirely on Podman side and it's still on my plate,
> > by the way:
> >
> >    https://github.com/containers/podman/issues/17840
> >    https://pad.passt.top/p/Features_2025#L40  
>
> I don't see how this can be fixed on the podman side, the network 
> namespace of a rootful container (not userns=auto) is owned by the root 
> user. If you configure something in there you must have real 
> CAP_NET_ADMIN from the host init userns. So pasta must not drop this 
> privilege before configuring the netns.

Oops, right. My starting point was this change, which is actually
trivial (at least as a test) and something I already tried out, but
then I hit a number of issues in Podman I never really figured out.

So yes, it takes one change in pasta, but the substantial part left for
me to figure out is why Podman didn't just work with it. It's not
necessarily complicated, I spent just a couple of hours on it, so maybe
there's something simple I missed.

> And even then with the future 
> netlink monitor work we would need to keep that privilege level to 
> modify the netns even during runtime?

This just reminded me that, somewhat surprisingly, for netlink
operations, the check on capabilities is not just performed on the
process creating the socket when the socket is created, but also later
*on the sender of the message*.

This is inconsistent with other operations on other types of sockets
where the whole context is checked and assigned at the time of the
creation, and was introduced because of a specific behaviour of Zebra
(the routing daemon) in 2014, see discussion around:

  https://lore.kernel.org/all/87d2g7d9ag.fsf_-_@x220.int.ebiederm.org/#r

and I stumbled upon it a while ago while preparing a seitan demo
replaying nft messages for an unprivileged container:

  https://seitan.rocks/seitan/tree/demo/nft.hjson#n38

So, my blanket answer "we create that socket at the beginning" doesn't
apply here.

However, assuming that this RFC patch from Jon actually works (I haven't
tested it):

  https://archives.passt.top/passt-dev/20251215015441.887736-11-jmaloy@redhat.com/

I would say we're fine with it. Well, there's still the possibility
that it doesn't work if Podman originally detached the network
namespace, I'm not sure.

If it doesn't work, we'll need to retain more capabilities, or even
keep a cloned process around for this kind of stuff. We could also fix
that in the kernel, Zebra doesn't need that quirk anymore.

> >> Because as root podman does
> >> port forwarding via DNAT firewall rules (i.e. custom nftables rules we
> >> add). The kernel however never added support for DNAT on ::1 meaning
> >> clients trying to access that are not getting forwarded. The only way to
> >> support this is using a user space helper. Right now this doesn't work
> >> and we do not use rootlessport for this either so I was just thinking
> >> ahead because we do have these users requests who want ::1 to work as root.
> >>
> >> For the current rootlessport use case we also must bind all ports as
> >> given (i.e. also addresses 0.0.0.0 bind address), just forwarding
> >> loopback to loopback is not what we want or do for security reasons, see
> >> CVE-2021-20199. And logically it would not really work to have another
> >> process bind 0.0.0.0 and this pasta helper bind lo on the same port at
> >> the same time.
> >>
> >> The way I am thinking is bind ports as normal, add the no-tap option and
> >> add two options to give the v4 and v6 namespace (container) side connect
> >> addresses so we never actually connect to lo. Then we also should have a
> >> dynamic way to update the connect addresses at runtime which is required
> >> for podman network connect/disconnect to work which changes the
> >> addresses inside the namespace, see
> >> https://github.com/containers/podman/commit/e88d8dbeae2aebd2d816f16a21891764163afcd4.
> >>
> >> Overall none of this is a blocker for removing rootlessport. I think our
> >> plan was and still is to use the dynamic port forwarding logic David is
> >> working on to replace the rootless custom network port forwarding case
> >> with that.  
> > Regardless of other requirements that are needed as well to support
> > forwarding ::1 for root containers (or rootless with --userns=auto),
> > this feature by itself makes sense as it is and we'll need it as it is,
> > right?
> >
> > By the way we routinely get requests for this feature by pasta (and
> > Podman) users, regardless of any specific Podman integration, so I
> > think the feature is generic enough as to make sense regardless of your
> > plan for root containers.
>
> I am not sure how I would use or integrate a loopback to loopback 
> forwarder in podman so I don't think we would need or can use that as is.

Well, I'm not sure, I just remember that you had in mind some use cases
that could be fixed with this (and even noted them down in the
references from the ticket).

Sorry Yumei, I should have checked more recently, as it looks like this
doesn't currently have as much priority as I thought, at least in
Podman's perspective. In any case it's definitely useful.

By the way, if it's for the root case, we'll still need it the day we
support operation when started as root. If it's to fix up IPv4 / IPv6
loopback mapping in the rootless case, it would be usable right away.

> I think the use case itself is still interesting and if there are end 
> users asking for it sure not objections from me. I guess it could be 
> interesting to expose a service without giving it access to the full 
> internet and without having to deal with complicated firewall rules, 
> i.e. with this we get a container that only could communicate by 
> replying to the forwarded ports.

Right, yes, it might also be one way to implement "isolated" containers
as described in https://bugs.passt.top/show_bug.cgi?id=139 (I still have
to follow up on comments there, and that might take a while, but let me
quickly mention that it has little/nothing to do with local mode).

-- 
Stefano