neighbour: Prevent Race condition in neighbour subsytem

Following Race Condition was detected:

<CPU A, t0>: Executing: __netif_receive_skb() ->__netif_receive_skb_core()
-> arp_rcv() -> arp_process().arp_process() calls __neigh_lookup() which
takes a reference on neighbour entry 'n'.
Moves further along, arp_process() and calls neigh_update()->
__neigh_update(). Neighbour entry is unlocked just before a call to
neigh_update_gc_list.

This unlocking paves way for another thread that may take a reference on
the same and mark it dead and remove it from gc_list.

<CPU B, t1> - neigh_flush_dev() is under execution and calls
neigh_mark_dead(n) marking the neighbour entry 'n' as dead. Also n will be
removed from gc_list.
Moves further along neigh_flush_dev() and calls
neigh_cleanup_and_release(n), but since reference count increased in t1,
'n' couldn't be destroyed.

<CPU A, t3>- Code hits neigh_update_gc_list, with neighbour entry
set as dead.

<CPU A, t4> - arp_process() finally calls neigh_release(n), destroying
the neighbour entry and we have a destroyed ntry still part of gc_list.

Fixes: eb4e8fac("neighbour: Prevent a dead entry from updating gc_list")
Signed-off-by: Chinmay Agarwal <chinagar@codeaurora.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/core/neighbour.c

@@ -131,6 +131,9 @@ static void neigh_update_gc_list(struct neighbour *n)
 	write_lock_bh(&n->tbl->lock);
 	write_lock(&n->lock);
 
+	if (n->dead)
+		goto out;
+
 	/* remove from the gc list if new state is permanent or if neighbor
 	 * is externally learned; otherwise entry should be on the gc list
 	 */
@@ -147,6 +150,7 @@ static void neigh_update_gc_list(struct neighbour *n)
 		atomic_inc(&n->tbl->gc_entries);
 	}
 
+out:
 	write_unlock(&n->lock);
 	write_unlock_bh(&n->tbl->lock);
 }