On Sun, 2015-12-13 at 12:43 -0800, Eric Dumazet wrote:
> On Sun, 2015-12-13 at 20:20 +0000, Ben Hutchings wrote:
> > On Sun, 2015-12-13 at 18:54 +0000, Luis Henriques wrote:
> > > On Sat, Dec 12, 2015 at 04:18:26AM +0000, Ben Hutchings wrote:
> > > > On Wed, 2015-12-09 at 09:36 +0000, Luis Henriques wrote:
> > > > > 3.16.7-ckt21 -stable review patch.  If anyone has any objections,
> > > > > please let me know.
> > > > > 
> > > > > ------------------
> > > > > 
> > > > > From: Eric Dumazet <edumazet@google.com>
> > > > > 
> > > > > commit 4ece9009774596ee3df0acba65a324b7ea79387c upstream.
> > > > > 
> > > > > sit0 device allocates its percpu storage twice :
> > > > > - One time in ipip6_tunnel_init()
> > > > > - One time in ipip6_fb_tunnel_init()
> > > > > 
> > > > > Thus we leak 48 bytes per possible cpu per network namespace
> > > > > dismantle.
> > > > > 
> > > > > ipip6_fb_tunnel_init() can be much simpler and does not
> > > > > return an error, and should be called after register_netdev()
> > > > [...]
> > > > 
> > > > Doesn't this introduce a race condition when sit is a module?  There
> > > > seems to be nothing to prevent access to the partially initialised
> > > > device after calling register_netdev(), if sit_init_net() is called
> > > > during module loading rather than during namespace creation.
> > > > 
> > > 
> > > This seems to be an upstream issue, not specific to the 3.16.y-ckt
> > > stable kernel.  If that is the case, I guess I'll just keep this patch
> > > and later apply the fix.  Or do you think this race is really likely
> > > to be a worst problem than then issue the patch is trying to fix?
> > 
> > It seems worse than the problem being fixed.
> 
> 1) Sorry Ben, I do not understand the problem you mention.
>    What is a partially initialized device exactly ?

A tunnel device which is registered but hasn't had its private
structure fully initialised yet.

> 2) I have no idea why this patch is even backported to 3.16,
>  since it is fixing a problem added in 3.18 :
> 
> # git describe --contains ebe084aafb7e
> v3.18-rc5~22^2~42^2~1
> 
> If your 3.16 kernel survives this loop without consuming memory like
> crazy, then the backport is not needed.
> 
> modprobe sit
> while :
> do
>  ip netns add ns1
>  ip netns delete ns1
> done

I can't detect a memory leak when doing this.

Ben.

-- 
Ben Hutchings
Life would be so much easier if we could look at the source code.