Re: [PATCH net-next v6 1/8] vxlan: Restructure vxlan socket apis.

[Stephen Hemminger <stephen@networkplumber.org>]

On Mon, 5 Aug 2013 08:49:36 -0700
Pravin Shelar <pshelar@nicira.com> wrote:

> On Sun, Aug 4, 2013 at 2:42 PM, Stephen Hemminger
> <stephen@networkplumber.org> wrote:
> > On Thu,  1 Aug 2013 11:44:40 -0700
> > Pravin B Shelar <pshelar@nicira.com> wrote:
> >
> >> @@ -1642,59 +1651,56 @@ static struct vxlan_sock *vxlan_socket_create(struct net *net, __be16 port)
> >>                        &vxlan_addr.sin_addr, ntohs(vxlan_addr.sin_port), rc);
> >>               sk_release_kernel(sk);
> >>               kfree(vs);
> >> -             return ERR_PTR(rc);
> >> +             return;
> >>       }
> >> +     atomic_set(&vs->refcnt, 0);
> >>
> >>       /* Disable multicast loopback */
> >>       inet_sk(sk)->mc_loop = 0;
> >> +     spin_lock(&vn->sock_lock);
> >> +     hlist_add_head_rcu(&vs->hlist, vs_head(net, port));
> >> +     spin_unlock(&vn->sock_lock);
> >
> > Overall I am fine with this set, except for this.
> > The change causes a socket to be put into the hash list with a ref count
> > of 0 and then you increment the ref count. If some other thread
> > finds the socket and then decrements the ref count, it would go
> > below zero and might leak, crash or delete it prematurely.
> >
> > This concerns me.
> 
> There is no code path which decrements ref-cnt after lookup. Therefore
> above case is not possible.
> But I agree this does not look good, so I will change it so not to
> have refcnt zero vs on hash-table.

I was concerned about:

CPU 0:
   vxlan_socket_create
     Make a vs structure with refcnt == 0
     Add it to hlist

     <preempt or other long IRQ>
                                         CPU 1: Another vxlan being setup (batch script?)
                                                vxlan_init
                                                     vxlan_find_port finds that vs
                                                     atomic_inc(refcnt) now 1
                                          <close or error>
                                                vxlan_uninit
                                                  vxlan_sock_release
                                                     atomic_dec_and_test (now 0)
                                                       delete from hlist
                                                       queue work for deletion

     now decides to continue with vs
     Either vs is already freed, or it exists
     in limbo