On Sun, May 03, 2026 at 09:18:18PM +0200, Jason A. Donenfeld wrote:

On Sat, Mar 21, 2026 at 08:20:53PM +0100, Valentin Spreckels wrote:

quoted

Hi Jason,

On 11/03/2026 23:59, Jason A. Donenfeld wrote:

quoted

Hi Valentin,

On Sun, Feb 08, 2026 at 06:05:45PM +0100, Valentin Spreckels wrote:

quoted

Use the flag introduced in commit 8a321cf7becc6 ("net: add
IFF_NO_ADDRCONF and use it in bonding to prevent ipv6 addrconf")
instead of mangling the addr_gen_mode to prevent ipv6 addrconf.

Can you give some more context here? Why was IFF_NO_ADDRCONF added when
the IN6_ADDR_GEN_MODE_NONE method has been working fine? What's the
difference between these approaches? I don't doubt that your patch is
correct, but I would like to better understand this.

Only wireguard configures addr_gen_mode inside the kernel, otherwise it 
is only set by userspace; userspace is also able to overwrite the 
IFF_NO_ADDRCONF set by wireguard.

Commit 8a321cf7becc ("net: add IFF_NO_ADDRCONF and use it in bonding to 
prevent ipv6 addrconf") introduces the private interface flag 
IFF_NO_ADDRCONF, which isn't accessible by userspace.

Thus use the IFF_NO_ADDRCONF flag in wireguard.


Does that answer your questions? If yes, I will submit a v2 with this as 
commit message.

I applied this here:
https://git.zx2c4.com/wireguard-linux/commit/?id=88427bcbe5bd3711de387b1c1f6540ef6fc05a78

Sorry for the delay! Patch looks good as-is, once I looked into the
internal mechanism.

I'm backing this patch out for now. It seems to break the selftests:

    [+] NS2: ping6 -c 10 -f -W 1 fd00::1
    ping6: connect: Network unreachable

Try it yourself with:

    $ make -C tools/testing/selftests/wireguard/qemu -j$(nproc) 

I assume it's because of:

        case NETDEV_UP:
        case NETDEV_CHANGE:
                if (idev && idev->cnf.disable_ipv6)
                        break;

                if (dev->priv_flags & IFF_NO_ADDRCONF) {
			[...]
                        break;
                }

Feel free to submit a v2 if you think this is fixable or if the tests
themselves are wrong.

Jason