Port API: support specifying IP version explicitly ("tcp4", "tcp6")

[AkihiroSuda]

The new version of Docker libnetwork executes two proxy commands: moby/moby#41908 (comment)

rootlesskit-docker-proxy -proto tcp -host-ip 0.0.0.0 -host-port 8080 -container-ip 172.17.0.2 -container-port 80
rootlesskit-docker-proxy -proto tcp -host-ip :: -host-port 8080 -container-ip 172.17.0.2 -container-port 80

rootlesskit-docker-proxy now fails with EADDRINUSE bcause it cause the equivalent of rootlessctl add-ports 0.0.0.0:8080:80/tcp [::]:8080:80/tcp

$ rootlessctl add-ports 0.0.0.0:8080:80/tcp [::]:8080:80/tcp
1
error: listen tcp 0.0.0.0:8080: bind: address already in use

rootlesskit-docker-proxy probably needs to be modified to cause rootlessctl add-ports 0.0.0.0:8080:80/tcp4 [::]:8080:80/tcp6

cc @thaJeztah @arkodg

[AkihiroSuda]

Implementing this is straightforward for builtin port driver (the default of rootless docker), but the problem is that slirp4netns port driver does't support IPv6 yet (rootless-containers/slirp4netns#253).

Should rootlesskit-docker-proxy ignore -host-ip :: when the port driver is set to slirp4netns?
Or can we configure dockerd to skip invoking the proxy with -host-ip ::?

[thaJeztah]

Or can we configure dockerd to skip invoking the proxy with -host-ip ::?

It's tricky; so before docker 20.10, publishing a port to listen on "any" IP (-p 80:80 or -p 0.0.0.0:80:80) implicitly made the port available on both IPv4 and IPv6.

That most likely was an oversight in the original implementation; docker 20.10 "fixed" that, by no longer publishing on IPv6 by default, but there's many people relying on the (faulty?) behaviour, and assume that -p 80:80 makes the port accessible, irregardless if I try IPv4 or IPv6 (which also makes sense).

The libnetwork update attempts to fix the situation, but instead of "implicitly" exposing both on IPv4 and IPv6 now explicitly exposes on both, which shows both in docker inspect of the container, but also (as a "side-effect" ?) starts two proxies.

I'm a bit on the fence what the best solution is (also a bit worried about (by default) starting two proxies now, as there could be a non-neglectible overhead in that)

[AkihiroSuda]

Opened PR #232

When the port driver is set to slirp4netns, docker run -p [v6]:8080:80 skips exposing the port, without showing an error.
This behavior might be weird, but I think we can call it a day. ("slirp4netns" port driver is not used by default, even when the network driver is set to "slirp4netns".)

The default port driver builtin seems now fully functional with IPv6.