net: Resolve{TCP,UDP,IP}Addr returns IPv4 address on host without IPv4 connectivity

[bluecmd]

What version of Go are you using (go version)?

$ go version
go version go1.11rc2 linux/amd64

Does this issue reproduce with the latest release?

Yes

What operating system and processor architecture are you using (go env)?

Linux. The issue doesn't depend on CPU arch and has been reproduced on
ARMv5 and x86-64.

What did you do?

A minimal repro is this program:

package main

import (
	"log"
	"net"
)

func main() {
	r, _ := net.ResolveUDPAddr("udp", "ping.sunet.se:0") // IPv4 and IPv6 records
	log.Printf("Result of ResolveUDPAddr: %v", r)

	r2, _ := net.ResolveIPAddr("ip", "ping.sunet.se") // IPv4 and IPv6 records
	log.Printf("Result of ResolveIPAddr: %v", r2)

	r3, _ := net.ResolveIPAddr("ip", "ipv6.sunet.se") // Only IPv6 record
	log.Printf("Result of ResolveIPAddr: %v", r3)
}

This shows the behavior of ResolveUDPAddr and ResolveIPAddr with dual-stacked
hosts and a IPv6 only host as the last entry. This also applies to ResolveTCPAddr.

What did you expect to see?

On hosts that have an IPv6 and an IPv4 address, the IPv6 address is preferred
on hosts that has IPv6 only connectivity.

In the program below, the output should be something like this on an IPv6-only system:

[bluecmd]$ ./ipv6_dial
2018/11/08 16:22:56 Result of ResolveUDPAddr: 2001:6b0:7::14
2018/11/08 16:22:56 Result of ResolveIPAddr: 2001:6b0:7::14
2018/11/08 16:22:56 Result of ResolveIPAddr: 2001:6b0:7::14

What did you see instead?

This is the outcome on an IPv6 only system:

[bluecmd]$ sudo ip ro
[bluecmd]$ sudo ip -6 ro
[..]
default via fe80::200:5eff:fe00:265 dev eth0 proto ra metric 100 pref medium

Yet ResolveIPAddr prefers IPv4:

[bluecmd]$ ./ipv6_dial
2018/11/08 16:22:56 Result of ResolveUDPAddr: 192.36.125.18:0
2018/11/08 16:22:56 Result of ResolveIPAddr: 192.36.125.18
2018/11/08 16:22:56 Result of ResolveIPAddr: 2001:6b0:7::14

This is also not dependent on CGO.

For more logs, see here.

Extra information

Problematic code is most likely this part of ipsock.go:

go/src/net/ipsock.go

Lines 73 to 98 in 7da1f7a

	func (addrs addrList) forResolve(network, addr string) Addr {
	var want6 bool
	switch network {
	case "ip":
	// IPv6 literal (addr does NOT contain a port)
	want6 = count(addr, ':') > 0
	case "tcp", "udp":
	// IPv6 literal. (addr contains a port, so look for '[')
	want6 = count(addr, '[') > 0
	}
	if want6 {
	return addrs.first(isNotIPv4)
	}
	return addrs.first(isIPv4)
	}
	// first returns the first address which satisfies strategy, or if
	// none do, then the first address of any kind.
	func (addrs addrList) first(strategy func(Addr) bool) Addr {
	for _, addr := range addrs {
	if strategy(addr) {
	return addr
	}
	}
	return addrs[0]
	}

My reading is that the code returns IPv4 unless IPv6 was explicitly asked for.
If the host is IPv6 only, then that host is picked on the fallback return in first().

I would suggest changing the code to not prefer IPv4 addresses if there is no IPv4 connectivity.

[ianlancetaylor]

CC @mikioh @iangudger

[mchtech]

Why not prefer to return IPv6 address?

[vlcty]

I'm experiencing the same issue. Any updates on this? I mean IPv6 is the current gen protocol.

[saudiqbal]

Same problem with Wireguard with dual stack hostname. IPv6 should be preferred over IPv4, hope it should be fixed soon.

https://www.reddit.com/r/WireGuard/comments/e05ef1/hostname_in_endpoint_wireguard_uses_ipv4_instead/f8d5fea/

[vlcty]

What needs to be done to re-prioritize this issue? @ianlancetaylor @andybons @rsc

[ianlancetaylor]

@vlcty Someone needs to look at the code and send a fix. I don't personally know offhand how to fix this.

[stapelberg]

AFAICT, Go currently knows about the IP stack’s capabilities (supports IPv4/IPv6 in principle), but not about the actual connectivity (an IPv4/IPv6 address is configured).

In i3status, we determine IPv6 connectivity by creating a UDP socket to a global IPv6 address: https://github.com/i3/i3status/blob/3f27399d730bb9a66bebfed6aff2660828687ca5/src/print_ipv6_addr.c#L82-L93

As UDP sockets are connectionless, this approach does not send anything onto the network.

So, my suggestion for a fix would be to check for IPv6 connectivity, and short-circuit the IPv4 preference in ipsock.go that @bluecmd mentioned.

If we wanted to play it safer, instead of checking for IPv6 connectivity, we could check if IPv4 connectivity is missing.

---

By the way, RFC 6724 (Default Address Selection for Internet Protocol Version 6 (IPv6)) covers this topic in a great amount of detail if anyone wants to read up more. We’re running into this problem mentioned in the RFC’s introduction section:

Furthermore, dual- or hybrid-stack implementations, which support
both IPv6 and IPv4, will very often need to choose between IPv6 and
IPv4 when initiating communication, for example, when DNS name
resolution yields both IPv6 and IPv4 addresses and the network
protocol stack has available both IPv6 and IPv4 source addresses. In
such cases, a simple policy to always prefer IPv6 or always prefer
IPv4 can produce poor behavior.

[miyurusankalpa]

The commit for change that added the code is here a5179bd and it looks like the author had IPv6 disabled in the kernel as per the comments in the issue linked with it.

The culprit for this issue is in the return function.

go/src/net/ipsock.go

Line 86 in 7da1f7a

return addrs.first(isIPv4)

I think it should not return a list IPv4 address by default. It should return the address list as is.

This also looks like the same issue as nodejs/node#6307