swarm: implement blackhole detection

[sukunrt]

No description provided.

[sukunrt]

@marten-seemann
For blackhole detection if we want just a binary check like it’s a black hole if the last n consecutive calls are failures we can do away with the sliding window.

This doesn't implement happy eyeballs for TCP since the delay required for that to be effective is about 700ms. That feels too long to me. We can implement happy eyeballs for TCP separately where we introduce a mechanism to notify the worker loop that a connection has been established and security and muxer selection is in progress.

[marten-seemann]

Should we make the changes to the dial prioritization logic in a follow-up PR? Would be nice to disentangle it.

[marten-seemann]

We probably should also add some metrics to our swarm dashboard. What about this: for both IPv6 and UDP, we'd show a row containing

• a timeline view that shows if the swarm thinks that we're behind a blackhole
• the dial success rate
• if behind a blackhole, the next time we'll probe

[sukunrt]

Completely agree with the metrics suggestion.

if behind a blackhole, the next time we'll probe

might be difficult since we probe based on number of requests(every nth request is allowed). It should be fine to make this number of requests till next probe.

[sukunrt]

This is what the dashboard looks like for me:

[sukunrt]

@marten-seemann
An alternative is to allow 5-10 probes every n requests. If we do this we can move the black hole can dial check to swarm.filterKnownUndialables.
This is useful because eventually I'd like to get rid of the backoff check here too(opening a separate issue for that). https://github.com/libp2p/go-libp2p/blob/master/p2p/net/swarm/swarm_dial.go#L392

[sukunrt]

Updated the PR to do address filtering and probing for reevaluating black hole state at a peer level.

[marten-seemann]

Should this go in error log? If we're not noisy about it, we'd miss this bug if we ever introduce it.

[sukunrt]

Added this

[p-shahi]

Add the description in the release issue highlights while its fresh #2326?