Add support for sendmmsg(2) on linux

[colinmarc]

https://man7.org/linux/man-pages/man2/sendmmsg.2.html

Partially addresses #1156. I would've liked to add recvmmsg in the same PR, but it's actually much more complicated.

[sunfishcode]

Can you comment on why control needs to be a mut reference here?

[colinmarc]

The dumb answer is "because sendmsg requires a mut reference" (as does with_noaddr_msghdr and friends). I assumed some random API modified the cmsg on send, but I don't know which one.

[colinmarc]

I think I addressed all your comments.

[colinmarc]

Hi, anything needed from me on this?

[sunfishcode]

I think this looks good, just a few more review comments:

[colinmarc]

Sorry for pinging and then disappearing - was really sick :)

I rebased and addressed your comments. Please let me know how you'd like to resolve the overlap with #1004. It seems like I could use the updated SockAddrAny from that instead of RawSocketAddr.

[colinmarc]

Rebased on top of #1004 now that it's in main. The test failures look unrelated.

This patch is a lot smaller now!

[colinmarc]

I would like to take a &'a impl SockAddrArg here, but it doesn't seem like the trait ensures that the pointer passed to the closure lives as long as the SockAddrArg itself.

We could obviously immediately call as_any, but that would just be making an explicit clone implicit.

[sunfishcode]

Yes, you're right. Could you add a comment about this? Something like, "this doesn't use SocketAddrArg because that creates a temporary SocketAddrAny that only lives for the duration of the with_sockaddr call, and we need a SocketAddrAny that lives for 'a".

Another possibility would be to add a with_socket_addr_any function to SocketAddrArg like this:

    fn with_socket_addr_any<R>(&self, f: impl FnOnce(&SocketAddrAny) -> R) -> R {
        let any = self.as_any();
        f(&any)
    }

Then users with a family-specific addr could do

    addr.with_socket_addr_any(|addr| {
        sendmmsg(
            &socket,
            &mut [
                MMsgHdr::new_with_addr(
                    &addr,
                    &[IoSlice::new(b"hello")],
                    &mut Default::default(),
                ),
            ],
            SendFlags::empty(),
        )
    });

and then SocketAddrAny's impl SocketAddrArg could skip the as_any() call:

    fn with_socket_addr_any<R>(&self, f: impl FnOnce(&SocketAddrAny) -> R) -> R {
        f(self)
    }

That way users could avoid calling as_any themselves. We could do that, though it's not that much of a simplification.

[colinmarc]

I added a sentence to the docstring pointing users to as_any - let me know what you think or if you think it needs more explication. I tried a few variants and in the end it felt like too much detail for end-users of the library, but maybe it would be useful as an "internal" comment.

[sunfishcode]

Should this use new_wiith_addr, like the v6 version does?

[colinmarc]

I think my idea was to exercise both code paths. I changed it so that both tests exercise both.

[sunfishcode]

Yes, you're right. Could you add a comment about this? Something like, "this doesn't use SocketAddrArg because that creates a temporary SocketAddrAny that only lives for the duration of the with_sockaddr call, and we need a SocketAddrAny that lives for 'a".

Another possibility would be to add a with_socket_addr_any function to SocketAddrArg like this:

    fn with_socket_addr_any<R>(&self, f: impl FnOnce(&SocketAddrAny) -> R) -> R {
        let any = self.as_any();
        f(&any)
    }

Then users with a family-specific addr could do

    addr.with_socket_addr_any(|addr| {
        sendmmsg(
            &socket,
            &mut [
                MMsgHdr::new_with_addr(
                    &addr,
                    &[IoSlice::new(b"hello")],
                    &mut Default::default(),
                ),
            ],
            SendFlags::empty(),
        )
    });

and then SocketAddrAny's impl SocketAddrArg could skip the as_any() call:

    fn with_socket_addr_any<R>(&self, f: impl FnOnce(&SocketAddrAny) -> R) -> R {
        f(self)
    }

That way users could avoid calling as_any themselves. We could do that, though it's not that much of a simplification.

[sunfishcode]

The FreeBSD CI failure is fixed on main in #1322.

[sunfishcode]

Thanks!