HTTP server with tower interaction

[lexnv]

This PR allows users to set custom tower ServiceBuilders as middleware
to layer the RPC service.

Exposing such functionality enables users to rely on crates that build upon tower,
such as tower_http to crate a default layered filtering system (ie, enable CORS from
tower_http, or extensive debugging upon receiving a response / request).

Tower HTTP Example

The server is constructed with the tower's ServiceBuilder with the following layer on top:

.layer(
	TraceLayer::new_for_http()
		.on_body_chunk(|chunk: &Bytes, latency: Duration, _: &tracing::Span| {
			tracing::trace!(size_bytes = chunk.len(), latency = ?latency, "sending body chunk")
		})
		.make_span_with(DefaultMakeSpan::new().include_headers(true))
		.on_response(DefaultOnResponse::new().include_headers(true).latency_unit(LatencyUnit::Micros)),
)

Output with RUST_LOG=trace:

2022-07-27T12:48:46.665907Z TRACE request{method=POST
uri=/ version=HTTP/1.1 headers={"content-type": "application/json",
"accept": "application/json", "host": "127.0.0.1:9935", "content-length": "45"}}:
tower_http: sending body chunk size_bytes=38 latency=347.166µs

Next Steps

• Documentation
• Enable service with tower_http

Closes #841

[niklasad1]

great, we should probably check how much overhead tower introduces.

follow-up PR remove the CORS stuff and rely on tower?!

[jsdw]

Is there a benefit to having this inner ServiceData<L> as well as the outer TowerService<L> instead of just having TowerService<L> have all of the props etc from ServiceData in it?

[niklasad1]

orphan rule?

[jsdw]

I don't really follow, but having the inner type isn't a big deal anyway in any case!

[niklasad1]

Look below it implements hyper::service::Service<hyper::Request<hyper::Body>> which can't be done because both the trait and type are not defined by "us".

So one need the newtype to implement that trait AFAIU.

[jsdw]

Do we have to worry about performance at all cloning all of these?

(It looked like a bunch of them were cloned previously so I'm guessing not).

And ah ok, is the inner struct so that you can clone it here? I guess you can just clone &mut self though and have just one struct?

[niklasad1]

I think it's similar as it was before because of the FnMut closure in the hyper service fn

[jsdw]

LGTM, nice work Alex!

My remaining comments are nits and can be ignored!