netfilter: nf_queue: do not release refcouts until nf_reinject is done

nf_queue is problematic when another NF_QUEUE invocation happens
from nf_reinject().

1. nf_queue is invoked, increments state->sk refcount.
2. skb is queued, waiting for verdict.
3. sk is closed/released.
3. verdict comes back, nf_reinject is called.
4. nf_reinject drops the reference -- refcount can now drop to 0

Instead of get_ref/release_ref pattern, we need to nest the get_ref calls:
    get_ref
       get_ref
       release_ref
     release_ref

So that when we invoke the next processing stage (another netfilter
or the okfn()), we hold at least one reference count on the
devices/socket.

After previous patch, it is now safe to put the entry even after okfn()
has potentially free'd the skb.

Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>

net/netfilter/nf_queue.c

@@ -303,12 +303,10 @@ void nf_reinject(struct nf_queue_entry *entry, unsigned int verdict)
 
 	hooks = nf_hook_entries_head(net, pf, entry->state.hook);
 
-	nf_queue_entry_release_refs(entry);
-
 	i = entry->hook_index;
 	if (WARN_ON_ONCE(!hooks || i >= hooks->num_hook_entries)) {
 		kfree_skb(skb);
-		kfree(entry);
+		nf_queue_entry_free(entry);
 		return;
 	}
 
@@ -347,6 +345,6 @@ next_hook:
 		kfree_skb(skb);
 	}
 
-	kfree(entry);
+	nf_queue_entry_free(entry);
 }
 EXPORT_SYMBOL(nf_reinject);