Re: [PATCH net-next 1/1] rtnetlink: request RTM_GETLINK by pid or fd

[Jiri Benc <jbenc@redhat.com>]

On Mon, 22 Jan 2018 22:23:54 +0100, Christian Brauner wrote:
> That is certainly a good idea and I'm happy to send a follow-up patch
> for that!

Note that I haven't looked into that and I don't know whether it is
easily possible. I'll appreciate if you could try that.

> But there's still value in being able to use
> IFLA_NET_NS_{FD,PID} in scenarios where the network namespace has been
> created by another process. In this case we don't know what its netnsid
> is and not even if it had been assigned one at creation time or not. In
> this case it would be useful to refer to the netns via a pid or fd. A
> more concrete and frequent example is querying a network namespace of a
> (sorry for the buzzword :)) container for all defined network
> interfaces.

That's what spurred my original comment. If you don't know the netnsid
in such case, we're missing something in uAPI but at a different point
than RTM_GETLINK.

When you find yourself in a need to query an interface in another
netns, you had to learn about that interface in the first place.
Meaning you got its ifindex (or ifname, perhaps) somehow. My point is,
you should have learned the netnsid at the same time.

ifindex alone is not an unique identifier of an interface. The
(ifindex, netnsid) pair is. (Also, note that ifindex can change when
moving the interface to a different netns.) So you should never get
ifindex alone when the interface is in another netns than the current
one. If that happens, that is the uAPI that needs to be fixed. You have
to always get the (ifindex, netnsid) pair.

And that's also the way how it should operate in the other direction:
(ifindex, netnsid) is the identifier to query interface in another
netns.

Note that many APIs in networking are based around netnsid - look at
NETLINK_LISTEN_ALL_NSID. It allows you to keep track of interfaces as
they are moved from name spaces to name spaces using a single socket:
you get notifications about interfaces disappearing or appearing in
"watched" name spaces. The name spaces are, of course, referenced by
their netnsid. In order to add another "watched" name space, just
assign it (or, more typically, let the kernel assign) a netnsid.

Btw, we have one missing piece here: when an interface is moved to a
name space that does not have netnsid attached, we want to find out
where the interface was moved to. But there's currently no reliable way
to do it. For veth, the other end can be used to get the netnsid (note
that RTM_GETLINK will return the correct link netnsid even when the
queried veth interface is in a different name space), for openvswitch,
we can now use genetlink, etc., but using different ways for different
interface types is not the best API ever and for standalone interfaces
we have nothing. I'd like to see something added to uAPI to cover this
in a generic way.

But as for this patch, I don't think it's the correct way. We do have
missing pieces in uAPI wrt. netns support but I think they're at
different places. If you have a counter example, please speak up.

Thanks,

 Jiri