Re: oops in udpv6_sendmsg

[Eric Dumazet <eric.dumazet@gmail.com>]

On Wed, 2013-04-17 at 07:27 -0700, Eric Dumazet wrote:
> On Wed, 2013-04-17 at 10:11 -0400, Dave Jones wrote:
> > On Tue, Apr 16, 2013 at 07:02:12PM -0700, Eric Dumazet wrote:
> 
> > good news is that with some changes, I was able to make current
> > trinity reproduce this in seconds rather than hours..
> > 
> > ./trinity -q -l off -n -c sendmsg -c connect 
> > 
> > on current tree seems to reliably trigger it for me.
> 
> Good new indeed, I got a crash in 2 seconds
> 
> (have to reproduce it because I lost the console output)
> 
> 

Hmm, sk_dst_get() assumes dst are always freed after RCU grace period,
but it seems not the case with IPv6.

We should atomically set dst->__refcnt to -1 before RCU grace period and
final destruction, then sk_dst_get should do something like :

rcu_read_lock();
dst = rcu_dereference(sk->sk_dst_cache); 
if (dst && !atomic_add_unless(&dst->__refcnt, 1, -1))
	dst = NULL;
rcu_read_unlock();

Ie we should not increment dst->__refcnt if the dst is in dismantle
phase.