Re: Multitude of dst obsolescense race conditions

On Fri, Jun 06, 2014 at 11:57:10AM -0700, Eric Dumazet wrote:

On Fri, 2014-06-06 at 11:12 -0700, Paul E. McKenney wrote:

quoted

It is probably just be me getting lost in the code, but I am not seeing
a synchronize_rcu(), call_rcu(), or synchronize_net() anywhere in
dst_release() or the things that it calls.  If there really isn't such
a call, then I don't see how the above code is safe in the case where
__sk_dst_set() is invoked on one CPU just after sk_dst_get() executes
the rcu_dereference() on some other CPU.

Well, this part is fine, dst_release() do not free dst that are
potentially stored in sk_dst_cache

Only the refcount is decremented.

The bug is elsewhere, we had another thread raising this issue on netdev
this morning.

I am cooking a patch to clear the mess.

Ah, thank you, sorry for the noise!

							Thanx, Paul