Re: connect() with AF_INET6 freezes on some Debian/unstable machine

[Reco <recoverym4n@gmail.com>]

 Hi.

On Fri, 10 Apr 2015 16:36:48 +0300
Reco <recoverym4n@gmail.com> wrote:

>  Hi.
> 
> On Fri, Apr 10, 2015 at 03:31:11PM +0200, Vincent Lefevre wrote:
> > On 2015-04-10 15:55:38 +0300, Reco wrote:
> > > Out of curiosity, does the problem reproduce if avahi-daemon is
> > > stopped?
> > 
> > Yes. I've purged the avahi-daemon daemon package (I don't think
> > I need it), net.ipv6.conf.all.accept_ra is still set to 0 in
> > /etc/sysctl.conf, and after rebooting the machine:
> > 
> > ypig:~> ip -6 a l dev eth0
> > 2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qlen 1000
> >     inet6 2001:1::21f:29ff:fe04:3efb/64 scope global mngtmpaddr dynamic 
> >        valid_lft 2591997sec preferred_lft 604797sec
> >     inet6 fe80::21f:29ff:fe04:3efb/64 scope link 
> >        valid_lft forever preferred_lft forever
> > 
> > ypig:~> time telnet www.vinc17.net 80
> > Trying 2001:4b98:dc0:45:216:3eff:fe9b:eb2f...
> > ^C
> > zsh: interrupt  telnet www.vinc17.net 80
> > telnet www.vinc17.net 80  0.00s user 0.00s system 0% cpu 22.279 total
> > 
> > Then...
> 
> Then we have a kernel bug here, no doubt.
> 
> I don't recall any userspace which listens RA in Debian, so clearly
> Jessie's kernel must ignore accept_ra. I'll check it in a couple of
> hours, just need to get myself to the place with the right setup.

And, after a couple of experiments, I've confirmed that

net.ipv6.conf.all.accept_ra = 0 

does not seem to do anything indeed.

But, at the same time, 

net.ipv6.conf.default.accept_ra = 0

disables SLAAC just fine in Jessie.

Whenever this can be treated as a kernel bug or not lies in the eye of
the beholder.

I've also not tested whenever systemd will override this setting (its
upstream seem to have a soft spot for ipv6), but from the causal Google
search that seems unlikely.

Still, yet another thing I must not forget to check after upgrade
to jessie.

Reco