On Tue, May 31, 2016 at 08:04:24AM +0200, Micha?? Trojnara wrote:

On 31.05.2016 00:42, Konstantin Belousov wrote:

...

I have a following configuration for the outgoing connection [XXX-1] client = yes accept = 127.0.0.1:1564 connect = some-server:9999 local = some-other-address

[cut]

...

The socket was created with INET6 address family, but bind was done for INET AF.

This is pretty much expected if "some-server:9999" is an IPv6 address and "some-other-address" is an IPv4 address. The "local" option indeed cannot handle a mix of IPv4 and IPv6 "connect" addresses.

Usually some-server is the list of addresses, one of them is IPv4, and another is IPv6. This is the only reasonable way to name dual-stack host.

What do you think might be a proper solution?

I would expect that the code which binds the socket for connect(2), matched the address types before binding. A good implementation needs to iterate over the results of getaddrinfo(3) for remote host to try to connect to each returned address. In the same manner, when one outgoing address is attempted to connect to, the getaddrinfo(3) list for the local address would be iterated over, and first matched compatible address selected for binding.

It seems that it is enough to match addrinfo pairs by ai_family/ai_socktype/ai_protocol, but you might also need to e.g. pay some attention to filter out IPv4 mapped to IPv6 entries.

I am not sure how much additional plumbing is required to have local getaddrinfo(3) result in the local_bind().