Restore listenfd suport for `wasmtime serve`

[simolus3]

Feature

I want to serve a wasi:http/service component that I expect won't receive requests very frequently. So instead of keeping a wasmtime serve process running, I'd like to rely on systemd socket activation to start that process when a request comes in and shut it down after a while of inactivity.

#3729 originally added the --listenfd option, but that doesn't appear to do anything anymore. I think it would be a good option to allow wasmtime serve to inherit sockets from a service manager instead of creating and binding one itself.

Benefit

Since wasmtime can launch components quickly, I wouldn't have to keep a process around unless it's actively serving a request.

Implementation

I suppose we can re-use the approach from #3729 but only implement it for wasmtime serve .

FWIW I'm happy to open a PR for this if this sounds like a feature worth adding again.

Alternatives

Using something like systemd-socket-proxyd I guess?

[alexcrichton]

Adding just --listenfd unfortunately probably won't be sufficient here. That was only ever implemented for the wasip1 target which otherwise doesn't support networking. Nowadays you'd likely want component integration (e.g. wasip2 or wasip3), and doing that would require some sort of WIT interface by which the guest could acquire the socket (it can't just be ambiently available -- no preexisting WASI WITs model this capability). That broadens this to a bit of a larger implementation in terms of needing to standardize that or somehow specify that somewhere.

[simolus3]

I don't need the component itself to inherit sockets, I just want to be able to pass the main socket wasmtime serve listens on from an outer context, replacing only this:

wasmtime/src/commands/serve.rs

Lines 618 to 631 in 6680169

	let socket = match &self.addr {
	SocketAddr::V4(_) => tokio::net::TcpSocket::new_v4()?,
	SocketAddr::V6(_) => tokio::net::TcpSocket::new_v6()?,
	};
	// Conditionally enable `SO_REUSEADDR` depending on the current
	// platform. On Unix we want this to be able to rebind an address in
	// the `TIME_WAIT` state which can happen then a server is killed with
	// active TCP connections and then restarted. On Windows though if
	// `SO_REUSEADDR` is specified then it enables multiple applications to
	// bind the port at the same time which is not something we want. Hence
	// this is conditionally set based on the platform (and deviates from
	// Tokio's default from always-on).
	socket.set_reuseaddr(!cfg!(windows))?;
	socket.bind(self.addr)?;

Would it be acceptable to pass just that? Maybe it should be a different flag in that case, e.g. --accept listenfd instead of -S listenfd.

[alexcrichton]

Aha I see yeah, that's much easier! And yeah I think it'd be reasonable to have a wasmtime serve-specific argument for that. If you're up for a PR for that I'd be happy to review. While you're doing that, could you also link documentation for the listenfd semantics? I've never heard of it before myself and I'm not sure where the original implementation came from, but if this is like a systemd convention I think it'd be good to just have a link somewhere about the docs

[simolus3]

Thanks, I've opened a PR. I've also linked this from a doc comment, but the protocol is described here and I think it's pretty much a systemd thing.

On that note, do you think something like a --idle-process-timeout (akin to --idle-instance-timeout) that would shut down wasmtime after a while would also be reasonable? Stopping the wasmtime process after a while of inactivity lets systemd restart it on the next subsequent connection and means a service consumes no resources unless it's in use.