IPv6 -- getaddrinfo() and bind() ordering with V6ONLY

Recently I ran into an issue that took me a while to sort out, and it is regarding inconsistent behaviour on various OS's with regards to IPv6 sockets (AF_INET6¹) and calling bind(2) after getting the results back from getaddrinfo(3).

A call to getaddrinfo() with the hints set to AF_UNSPEC in ai_family and AI_PASSIVE in ai_flags will return to us 1 or more results that we can bind() to. Sample code for that looks like this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29

struct addrinfo hints, *addrlist;

memset(&hints, 0, sizeof(hints));

// Ask for TCP
hints.ai_socktype = SOCK_STREAM;

// Any family works for us ...
hints.ai_family = AF_UNSPEC;

// Set some hints
hints.ai_flags = 
            AI_PASSIVE    | // We want to use this with bind
            AI_ADDRCONFIG;  // Only return IPv4 or IPv6 if they are configured

int rv;

if ((rv = getaddrinfo(0, "7020", &hints, &addrlist)) != 0) {
    fprintf(stderr, "getaddrinfo: %s", gai_strerror(rv));
    return 1;
}

// Use the list in *addrlist
for (addr = addrlist; addr != 0; addr = addr->ai_next) {
    // use *addr as appropriate
}

// Clean up the memory from getaddrinfo()
freeaddrinfo(addrlist);

On Linux there are two entries returned when the host it is run on has both IPv4 and IPv6 enabled. An AF_INET which was followed by an AF_INET6. Now, it is not said that you are required to use all of the results that are returned, but if you want to listen on all address families it is off course suggested.

Following the steps below for each of the returned results should result in having 1 or more different sockets that are bound to a single port.

1. Create the socket()
2. Set any socket options you want (SO_REUSEADDR for example)
3. Then bind() the socket
4. After that call listen() (followed off course by accept() on the socket)

Only for some unknown reason (and errno is no help) bind() fails when you get to the AF_INET6, which was returned second. Searching online as to why the bind would fail doesn't give you any good results and the thing that is even worse is that if you run the same code on another platform such as FreeBSD, OpenIndiana or Mac OS X no such failure exists. However I started suspecting something was up when I started looking at the output from netstat -lan | grep 7020 on Mac OS X. Where 7020 is the port I passed into getaddrinfo().

tcp46      0      0  *.7020                 *.*  LISTEN     
tcp4       0      0  *.7020                 *.*  LISTEN

Wait a minute ... one of the sockets is on both IPv4 and on IPv6. Some more time spent searching the internet I came across RFC 3493 section 5.3, which is titled "IPV6_V6ONLY option for AF_INET6 Sockets".

As stated in section <3.7 Compatibility with IPv4 Nodes>, AF_INET6 sockets may be used for both IPv4 and IPv6 communications. Some applications may want to restrict their use of an AF_INET6 socket to IPv6 communications only.

This was going down the right route, so I changed my code so that in the steps listed above in number 2 I added the following code if the socket type is AF_INET6:

1
2
3
4
5

if (setsockopt(sockfd, IPPROTO_IPV6, IPV6_V6ONLY, &yes, sizeof(int)) == -1) {
    close(sockfd);
    fprintf(stderr, "setsockopt: %s IPV6_V6ONLY\n", strerror(errno));
    continue;
}

The RFC 3493 section 5.3 also states that this option should be turned off by default, which means that all IPv6 sockets can also communicate over IPv4. Thus technically setting the option manually in code the best way to fix the issue. FreeBSD has had this feature turned on (as in IPv6 sockets can only communicate with IPv6 and NOT IPv4) since 5.x.

The biggest issue is that the remaining operating systems (OS X and OpenIndiana) don't have the same behaviour as Linux which makes troubleshooting this issue more difficult than it should be. The issue is that the RFC doesn't specify what exactly the operating should do when it encounters a request to bind to the same port on IPv4 and IPv6. The only place where I have found this documented is in "IPv6 Network Programming" under "Tips in IPv6 Programming" chapter 4, section 4, appropriately titled "bind(2) Ordering and Conflicts".

If you get a bind() error when attempting to bind to an AF_INET6 socket please make sure that you set the socket option IPV6_V6ONLY on the AF_INET6 socket. The default as required by RFC 3493 is to have that option be off. The default is wrong, and the RFC should have been more specific regarding what the right behaviour is when attempting to bind on an AF_INET6 socket when already bound on an AF_INET while IPV6_V6ONLY is set to false.

The full code that I used for testing, along with a little bit more information is available as a gist on github.