CHANGELOG: v0.8.0

[nickgarlis]

I would like to give Conn.ReadFunc a go downstream on mdlayher/netlink. It could help us wrap recvmmsg in a custom implementation. I want to validate a theory that this would improve throughput on some netlink families.

[mdlayher]

wfm, though you can always go get x@HEAD too.

[mdlayher]

By the way, curious where you're using this? I have a local experiment with the out-of-tree homa kernel module that needed it, I think!

[nickgarlis]

Thanks for tagging the release !

wfm, though you can always go get x@HEAD too.

I've been experimenting with it already, just wanted to start opening some PRs to get your opinion.

By the way, curious where you're using this? I have a local experiment with the out-of-tree homa kernel module that needed it, I think!

I wanted to use recvmmsg but x/sys does not wrap it yet. It also felt a bit wrong to wrap it in this package instead.

I have a local change for netlink where I do something like this

	err := r.c.ReadFunc(ctx, "recvmmsg", func(fd int) error {
		rn, _, errno := unix.Syscall6(
			unix.SYS_RECVMMSG,
			uintptr(fd),
			uintptr(unsafe.Pointer(&r.hdrs[0])),
			uintptr(len(r.hdrs)),
			unix.MSG_WAITFORONE,
			0,
			0,
		)
		if errno != 0 {
			return errno
		}

		n = int(rn)
		return nil
	})

The intention is to expose a ReceiveBatch method which can increase the throughput for nfqueue and conntrack I think. I am currently testing it with https://github.com/florianl/go-nfqueue and the results seem promising. Although admittedly, the tests are not very scientific, using it does seem to significantly increase throughput even when GSO is used.

I have a local experiment with the out-of-tree homa kernel module that needed it, I think!

Sounds interesting! I am guessing you needed it because the existing wrappers are not compatible with homa?

Edit: ReadFunc also seems useful for users who want to reduce allocations and skip the automatic sender address conversion that the existing wrappers do.