expose some Prometheus metrics

[marten-seemann]

This PR exposes some basic TCP transport metrics to Prometheus.

We're using the TCP_INFO socket option to query the kernel for TCP metrics (things like the RTT, congestion window, etc).
In the future, I'm planning to add more metrics, to mirror what we're measuring for QUIC wherever possible.

[Stebalien]

What was the motivation here? Usually, we just pull metrics out of the kernel itself.

[Stebalien]

We're using the TCP_INFO socket option to query the kernel for TCP metrics (things like the RTT, congestion window, etc).

Metrics are usually aggregate. I'm not sure aggregate RTT, congestion, etc. is really all that useful.

[Stebalien]

(maybe aggregate bandwidth would be useful?)

[Stebalien]

Apparently, RTT etc distributions (which will help us pick better timeouts, etc) conn lifetime (may help with connection management choices) etc. So yeah, useful.

[Stebalien]

In terms of making this work with things that aren't tcp.Conn connections. I'm fine forking and maintaining a small fork.

[marten-seemann]

I created a fork of mikioh/tcp that includes mikioh/tcp#2, so we don't have to use any unsafe type castings here. Not my preferred resolution, but until @mikioh merges the PR our options are limited.

[Stebalien]

We have a lot of forks that fall into that category.

[mxinden]

Sorry for the delayed review.

[mxinden]

Looks good to me from the Prometheus perspective!

[Stebalien]

Can we move this to libp2p? Our forks like this have a tendency to live longer than intended.