Re: [PATCH] inetpeer: fix a race in inetpeer_gc_worker()

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

On Tue, 2012-06-05 at 14:42 +0200, Steffen Klassert wrote:
> On Tue, Jun 05, 2012 at 02:19:12PM +0200, Eric Dumazet wrote:
> > On Tue, 2012-06-05 at 13:56 +0200, Steffen Klassert wrote:
> > > On Tue, Jun 05, 2012 at 11:28:27AM +0200, Eric Dumazet wrote:
> > > > From: Eric Dumazet <edumazet@google.com>
> > > > 
> > > > commit 5faa5df1fa2024 (inetpeer: Invalidate the inetpeer tree along with
> > > > the routing cache) added a race :
> > > > 
> > > > Before freeing an inetpeer, we must respect a RCU grace period, and make
> > > > sure no user will attempt to increase refcnt.
> > > > 
> > > 
> > > As already mentioned in the other mail. In this case, I think
> > > we can just delete the inetpeer once the refcount got zero.
> > > 
> > 
> > Nope, a concurrent lookup can find an entry about to be freed.
> 
> Hm, I agree that we need rcu protection when we remove single entries
> from an inetpeer tree. But in this case we invalidate the entire tree.
> 
> The first lookup after inetpeer_invalidate_tree() was invoked should
> find an empty tree, base->root initialized to peer_avl_empty_rcu.
> 
> Maybe I'm wrong, but I don't see how a lookup should find such an
> old invalidated tree.
> 

You are absolutely wrong yes.

A concurrent lookup can read previous values of the root pointer, even
if you wrote a new value in it. Thats whole RCU point.

Only waiting a rcu grace period make sure all lookups can see the new
root pointer.

I'll send a v2 to avoid atomics in the worker itself.