Added Socket::set_multicast_if_v4_n to set interface by index (#458)

[boris-n]

Implementation for Socket::set_multicast_if_v4_n (similar to Socket::join_multicast_v4_n) which allows to set interface by index or address (via struct InterfaceIndexOrAddress), for Linux/Android/Fuchsia/FreeBSD/NetBSD. (Other OSes seem to have no support for this.)

[boris-n]

Added feature = "all" as setting multicast interface by index is not supported on Windows and MacOSX.

[boris-n]

Rebased to current master, waiting for review/approval.

[boris-n]

Rebased to current master (again), waiting for review/approval.

[michaelinoue]

Any thoughts on expanding this to set both an address and an if index? It seems like we're pinned to using either or in this PR, but we could potentially want both (e.g. sending multicast traffic from a particular address on the interface).

[Thomasdezeeuw]

I left a comment, but I'm not sure about this pr. I don't use these APIs and currently don't have the time to look into them properly. For this to be merged it would be nice if someone with experience with the options could confirm they work as expected. This also needs a test.

[Thomasdezeeuw]

This doesn't seem right. The enum variant clearly say Index, but we're using an IPv4 address. I think we shouldn't support the index here and recommend to use the Address instead for NetBSD, but I'm not sure.

[boris-n]

The implementation is correct (although I have to admit it is probably confusing). The NetBSD API seems to be a bit "unusual" for setting the multicast interface by index. Let me cite from the NetBSD man page (https://man.netbsd.org/NetBSD-7.0/ip.4):

The IP_MULTICAST_IF option overrides the default for subsequent transmissions from a given socket:
struct in_addr addr;
setsockopt(s, IPPROTO_IP, IP_MULTICAST_IF, &addr, sizeof(addr));
where "addr" is the local IP address of the desired interface [...]. An application may also specify an alternative to the default network interface by index:
struct uint32_t idx = htonl(ifindex);
setsockopt(s, IPPROTO_IP, IP_MULTICAST_IF, &idx, sizeof(idx));
where "ifindex" is an interface index as returned by if_nametoindex(3).

So, while Linux, FreeBSD, Android and Fuchsia use a special struct mreqn (which has a separate field for the interface index) if they want to set the multicast interface by index, NetBSD uses a struct in_addr to set the address by IP and an uint32_t (in network byte order.) to set the address by interface index. struct in_addr and uint32_t both are 4 bytes in size, so the kernel cannot distinguish the two cases from the size parameter. Looking into the kernel code shows that IP addresses in subnet 0.0.0.0/8 are interpreted as interface index (see https://github.com/NetBSD/src/blob/trunk/sys/netinet/ip_output.c#L1691). This is the reason why I converted the interface to an address and used this one with the setsockopt call.

But I agree, the implementation is probably easier to understand if the code aligns to the examples given in the NetBSD manual page. I provided a modified version which does so.

[Thomasdezeeuw]

How about we say that InterfaceIndexOrAddress::Index is not supported on NetBSD (maybe even cfg it out?), and only support the Address variant. In the current implementation both are setting the same option (IP_MULTICAST_IF).

If/when NetBSD supports ip_mreqn we can add support as well.

[boris-n]

But NetBSD has support for setting the multicast interface by index, they documented in the man page that you have to pass it as uint32_t in network byte order (as my code now does).
There is no standardized way for setting the multicast interface by index, and some OSes don't support it. But Linux supports it using struct mreq for IP_MULTICAST_IF, and NetBSD supports it using uint32_t in network byte order for IP_MULTICAST_IF - that's totally fine. So why not support it for NetBSD? Only because NetBSD doesn't do it the Linux way using struct ip_mreqn?
(I agree that the NetBSD way is sub-optimal, the API will not be able to bind to an address in 0.0.0.0/8 address. But that's more an issue with the binding to an address instead of binding to an interface index. And as long as this subnet stays reserved, there is no issue with the NetBSD API.)

[boris-n]

Any thoughts on expanding this to set both an address and an if index? It seems like we're pinned to using either or in this PR, but we could potentially want both (e.g. sending multicast traffic from a particular address on the interface).

This API is not used to set the source address. Even if you use the version with the address, it is only used to select the interface. See details in https://www.man7.org/linux/man-pages/man7/ip.7.html, section about IP_MULTICAST_IF. If you want to set the source address, you have to bind the socket to it.