Because the ipoib_workqueue is not flushed when ipoib interface is
brought down, ipoib_mcast_join() may trigger a join to the broadcast
group after priv->broadcast was set to NULL (during cleanup).  This
will cause the system to be a member of the broadcast group when
interface is down.  As a side effect, this breaks the optimization of
setting the Q_key only when joining the broadcast group.
Signed-off-by: Yossi Etigin <yosefe@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
        if (!priv->broadcast) {
                struct ipoib_mcast *broadcast;
 
+               if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
+                       return;
+
                broadcast = ipoib_mcast_alloc(dev, 1);
                if (!broadcast) {
                        ipoib_warn(priv, "failed to allocate broadcast group\n");