I think udp4 instead of udp should be specified if you expect to use IPv4.

- c, err := net.ListenUDP("udp", &net.UDPAddr{
+ c, err := net.ListenUDP("udp4", &net.UDPAddr{

The behavior is documented here:

I've known about the udp4 workaround, but I don't think that that is optimal for those who want to accept both IPv4 and IPv6 connections and get the correct address versions. With net.IP that wasn't an issue, since the mapped IPv4 addresses were essentially the same as the real, four-byte ones. That isn't the case with netip.Addr.

At the very least, there should be public documentation about that. Having the correct type of address would be better, but I don't know if that's feasible.

With net.IP that wasn't an issue, since the mapped IPv4 addresses were essentially the same as the real, four-byte ones. That isn't the case with netip.Addr.

Can you elaborate this part? I don't understand why netip.Addr is different in this case.

net.ParseIP and similar functions return mapped forms for IPv4 addresses, so the common way to check whether a net.IP contains an IPv4 address was:

if ip4 := ip.To4(); ip4 != nil {
        // Use ip4 for IPv4 code.
} else {
        // Use ip for IPv6 code.
}

Which wasn't entirely correct, since there could be situations where one needs to make a distinction between the real IPv4 addresses and the mapped ones, but those seem to be fairly rare.

netip.Addr makes the distinction clearer with its Is4, but Is4 returns false for the mapped addresses, since there is Is4In6, which returns true for them. Thus, there is a lot of older code which assumes that all mapped addresses are actually just IPv4 addresses, because net.IP essentially has no other way to tell that, since ParseIP and friends always return mapped addresses. And if that code is then rewritten by simply replacing the ip4 != nil check with ip.Is4, which is a perfectly reasonable rewrite at a glance, that code is suddenly broken for addresses returned by the methods from the original post, because the addresses for connections coming over IPv4 returned by them aren't really IPv4 addresses any more.

This is a duplicate of #53554, which has a proposed fix as well.

@dr2chase I think the two issues are different. The proposed fix of #53554 won't affect this issue. IIUC, this issue is more about the documentation.

@dr2chase, I agree with ZekeLu in that this is a separate issue. And I also feel like there may possibly be other places where a naïve conversion could result in code breakage. Perhaps a blog post outlining the dos and don'ts of converting code from net.IP to netip.Addr would clarify the situation. Along with notes in the documentation, of course.

Another example: https://go.dev/play/p/wVDZEuLFdRd

	ta := &net.TCPAddr{
		IP:   net.ParseIP("1.2.3.4"),
		Port: 80,
	}
	fmt.Println(ta.AddrPort())

results in:

[::ffff:1.2.3.4]:80

In retrospect, making https://pkg.go.dev/net/netip#AddrFromSlice not do the automatic unmapping was probably a mistake. I just converted a bunch of code from the original https://pkg.go.dev/inet.af/netaddr#FromStdIP which does do the unmapping and it was pretty tedious, having to sprinkle Unmap calls everywhere. But when the IP is inside an AddrPort like the example in this comment, then fixing them up's even more tedious.

Some options:

• change a bunch of the net package to return 4-byte net.IP slices wherever safe & possible, to not confuse netip.AddrFromSlice (like https://go-review.googlesource.com/c/go/+/415580 and ParseIP)
• change netip.AddrFromSlice to also Unmap. Notably, its API docs aren't very specific on which behavior it implements, so it's note a huge breaking change.
• change the {TCP,UDP}Addr.AddrPort methods to Unmap at least, before the port wrapper

/cc @josharian @rsc @ianlancetaylor @maisem @crawshaw

@bradfitz, just to clarify, in your example the main culprit is probably net.ParseIP, which always returns a 16-byte form, not net.TCPAddr.AddrPort. If you use a “raw” 4-byte IP address or call To4(), it works as intended:

ta := &net.TCPAddr{
	IP:   net.ParseIP("1.2.3.4").To4(),
	Port: 80,
}
fmt.Println(ta.AddrPort())

ta = &net.TCPAddr{
	IP:   net.IP{1, 2, 3, 4},
	Port: 80,
}
fmt.Println(ta.AddrPort())

1.2.3.4:80
1.2.3.4:80

Which is not to say that the problem doesn't exist, of course. I personally would prefer that all functions that are used to convert between net.IP and netip.Addr (and friends) unmap, but I'm not sure how much of a breaking change that would be.

Re-posting #54234 (comment) since it seems to be more relevant here.

I don't think net.TCPAddr.AddrPort and net.UDPAddr.AddrPort should do the unmapping.

The *net.TCPAddr or *net.UDPAddr instance returned by a LocalAddr() or RemoteAddr() call is directly converted from a system socket address type. The length of the IP slice reflects the address family of the socket. If I open an IPv6 socket, I can expect that the length of the returned IP is always 16. Just like how you can expect the address returned by ReadFromUDPAddrPort and ReadMsgUDPAddrPort on an IPv6 socket is either IPv4-mapped IPv6 or IPv6.

IMO this behavior should be preserved, because this is how the underlying socket works. Those who wish to unmap IPv4-mapped IPv6 addresses can always do the unmapping themselves in their own code.

If I open an IPv6 socket, I can expect that the length of the returned IP is always 16

And if there's no address family specified, as in net.Listen("udp", ":1234")? I don't think users expect to randomly get either IPv4 or IPv6 addresses depending on which operating system they're on and how that OS was configured. That's not a great API to not know what you're going to get.

I don't think users expect to randomly get either IPv4 or IPv6 addresses depending on which operating system they're on and how that OS was configured. That's not a great API to not know what you're going to get.

Except it has always been like this. On most systems net.Listen("udp", ":1234") is going to get you a dual-stack IPv6 UDP socket. On OpenBSD and DragonFly BSD this is equivalent to doing net.Listen("udp4", ":1234") on other systems, because these two BSDs do not support dual-stack sockets.

Unless being able to run on OpenBSD and DragonFly BSD is a requirement, one can usually safely assume that net.Listen("udp", ":1234") is going to open a dual-stack IPv6 UDP socket. And in my experience this is usually the way things are being done too.

I guess the question is whether that's the API we want in Go. Seems like a legacy compatibility wart and we could do better than needlessly exposing mapped addresses to users when we have a type that can distinguish what was really meant.

It's not a goal of Go to expose the underlying operating system behavior. Instead we are trying to present an API that works roughly the same on all operating systems. It seems like if we are listening on 127.0.0.1 we should get the IPv4 form of that for net/netip when we call TCPAddr.AddrPort (or UDPAddr.AddrPort). That's more often what people want.

Note that package net has always returned 16-byte forms for everything as the standard form. But net/netip distinguishes them. Given that net has discarded the "IPv4 vs IPv4-in-6" information, we'll get the right answer far more often if we assume that case is "IPv4" rather than "IPv4-in-6".

Do I have that right?

It's not a goal of Go to expose the underlying operating system behavior.

It might not have been an explicit goal, but it has been the actual behavior since the beginning.

Making this change could break existing code that converts the returned net.TCPAddr/UDPAddr to netip.AddrPort and then to the raw sockaddr types by hand and directly makes syscalls on the socket. Right now you can just check netip.Addr.Is4, and you'll know whether to convert it to unix.RawSockaddrInet4 or unix.RawSockaddrInet6. After making this change, this is no longer possible, and now we are back to the pre-netip era where we use the length of the IP slice in net.TCPAddr/UDPAddr directly.

Note that package net has always returned 16-byte forms for everything as the standard form.

Only the parse, lookup functions, and other things that does not speak to sockets. Methods on sockets have always preserved information on the socket's address family.

This proposal has been added to the active column of the proposals project
and will now be reviewed at the weekly proposal review meetings.
— rsc for the proposal review group

There are three conversions using netip.AddrFromSlice in package net. They are:

UDPAddr.AddrPort
TCPAddr.AddrPort
Resolver.LookupNetIP

The decision we have is whether to call netip.AddrFromSlice(...).Unmap() instead of just AddrFromSlice in each of these.

I think we need to understand the impact of this change. We know times when it would help but I don't think we know the full impact well enough. Does anyone want to try to summarize the behavior across different systems and the different API calls?

IPv4-mapped IPv6 addresses don't really make sense in the scope of DNS. It also simplifies things a bit if the returned addresses are guaranteed to be either IPv4 or IPv6. So for Resolver.LookupNetIP it makes sense to call Unmap. As to TCPAddr and UDPAddr, I'm opposed to doing unmap for reasons I mentioned in previous comments.

Does anyone want to try to summarize the behavior across different systems and the different API calls?

I think to move forward on this issue we'd need to see this summary. If no one is interested, maybe we should put this issue on hold.

Placed on hold.
— rsc for the proposal review group

The decision we have is whether to call netip.AddrFromSlice(...).Unmap() instead of just AddrFromSlice in each of these.

I think we need to understand the impact of this change. We know times when it would help but I don't think we know the full impact well enough. Does anyone want to try to summarize the behavior across different systems and the different API calls?

I also met this problem. I send a cl
https://go-review.googlesource.com/c/go/+/443196, can be used as an example to discussed.

I'm sorry about I don't fully understand the impact of this change. But in my case, when i got a ipv4 address, I just want to still get this ipv4 address after calling the netip package.