Consider dropping connections when the threshold is full

[Stebalien]

That is, consider dropping connections from the underlying TCP socket. Maybe there's an option we can set on the TCP socket to let the kernel do this?

[Stebalien]

go-libp2p-transport-upgrader/listener.go

Lines 116 to 136 in 5ddf5de

	// This records the fact that the connection has been
	// setup and is waiting to be accepted. This call
	// *never* blocks, even if we go over the threshold. It
	// simply ensures that calls to Wait block while we're
	// over the threshold.
	l.threshold.Acquire()
	defer l.threshold.Release()
	select {
	case l.incoming <- conn:
	case <-ctx.Done():
	if l.ctx.Err() == nil {
	// Listener *not* closed but the accept timeout expired.
	log.Warningf("listener dropped connection due to slow accept")
	}
	// Wait on the context with a timeout. This way,
	// if we stop accepting connections for some reason,
	// we'll eventually close all the open ones
	// instead of hanging onto them.
	conn.Close()
	}

[Stebalien]

This doesn't look like the issue. Instead, it looks like a sleep in the error catcher.

[raulk]

We have a lot of piecemeal mechanisms for backpressure scattered across the codebase. At one point we'll need to consider a reusable backpressure gadget, maybe inspired by SEDA.