netpoll: Fix RCU usage

The use of RCU in netpoll is incorrect in a number of places:

1) The initial setting is lacking a write barrier.
2) The synchronize_rcu is in the wrong place.
3) Read barriers are missing.
4) Some places are even missing rcu_read_lock.
5) npinfo is zeroed after freeing.

This patch fixes those issues.  As most users are in BH context,
this also converts the RCU usage to the BH variant.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/net/core/netpoll.c b/net/core/netpoll.c
index 19ff660..e9ab4f0 100644
--- a/net/core/netpoll.c
+++ b/net/core/netpoll.c
@@ -261,6 +261,7 @@
 	unsigned long tries;
 	struct net_device *dev = np->dev;
 	const struct net_device_ops *ops = dev->netdev_ops;
+	/* It is up to the caller to keep npinfo alive. */
 	struct netpoll_info *npinfo = np->dev->npinfo;
 
 	if (!npinfo || !netif_running(dev) || !netif_device_present(dev)) {
@@ -810,10 +811,7 @@
 	refill_skbs();
 
 	/* last thing to do is link it to the net device structure */
-	ndev->npinfo = npinfo;
-
-	/* avoid racing with NAPI reading npinfo */
-	synchronize_rcu();
+	rcu_assign_pointer(ndev->npinfo, npinfo);
 
 	return 0;
 
@@ -857,6 +855,16 @@
 
 			if (atomic_dec_and_test(&npinfo->refcnt)) {
 				const struct net_device_ops *ops;
+
+				ops = np->dev->netdev_ops;
+				if (ops->ndo_netpoll_cleanup)
+					ops->ndo_netpoll_cleanup(np->dev);
+
+				rcu_assign_pointer(np->dev->npinfo, NULL);
+
+				/* avoid racing with NAPI reading npinfo */
+				synchronize_rcu_bh();
+
 				skb_queue_purge(&npinfo->arp_tx);
 				skb_queue_purge(&npinfo->txq);
 				cancel_rearming_delayed_work(&npinfo->tx_work);
@@ -864,10 +872,6 @@
 				/* clean after last, unfinished work */
 				__skb_queue_purge(&npinfo->txq);
 				kfree(npinfo);
-				ops = np->dev->netdev_ops;
-				if (ops->ndo_netpoll_cleanup)
-					ops->ndo_netpoll_cleanup(np->dev);
-				np->dev->npinfo = NULL;
 			}
 		}