Should AF_XDP's driver running on a seperate thread?

[Lethe10137]

When working on AF_XDP driver, I went into some issue:

rattan/rattan-core/src/metal/io/af_xdp.rs

Lines 62 to 66 in acd3e01

	for packet in packets {
	self.sender
	.blocking_send(packet)
	.map_err(|_| std::io::Error::other("send error"))?;
	}

I believe that the send_bulk quoted above is essential for rattan to achieve the packet forwarding performance comparable to that in the repo camellia 's examples. Yet blocking_send was used there, making it impossible to be used directly in the tokio runtime.

A direct solution is making our veth receiving futures, aka the ones tokio::spawn in VirtualEthernetEgress::new to
be running on a seperate thread, which means using std::thread::spawn. Is this the designed way to use it?

rattan/rattan-core/src/cells/external.rs

Lines 237 to 247 in acd3e01

	for d in driver.into_iter() {
	let notify = AsyncFd::new(d.raw_fd())?;
	tokio::spawn(Self::recv(
	flow_map.clone(),
	notify,
	d.into_receiver(),
	tx.clone(),
	log_tx.clone(),
	base_ts,
	));
	}

@minhuw

[minhuw]

Despite it makes the absolute performance number higher, it may not improve throughput/core, so we need more considerations here.