[NET]: Fix races in net_rx_action vs netpoll.

Keep netpoll/poll_napi from messing with the poll_list.
Only net_rx_action is allowed to manipulate the list.

Signed-off-by: Olaf Kirch <olaf.kirch@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/netdevice.h

@@ -261,6 +261,8 @@ enum netdev_state_t
 	__LINK_STATE_LINKWATCH_PENDING,
 	__LINK_STATE_DORMANT,
 	__LINK_STATE_QDISC_RUNNING,
+	/* Set by the netpoll NAPI code */
+	__LINK_STATE_POLL_LIST_FROZEN,
 };
 
 
@@ -1014,6 +1016,14 @@ static inline void netif_rx_complete(struct net_device *dev)
 {
 	unsigned long flags;
 
+#ifdef CONFIG_NETPOLL
+	/* Prevent race with netpoll - yes, this is a kludge.
+	 * But at least it doesn't penalize the non-netpoll
+	 * code path. */
+	if (test_bit(__LINK_STATE_POLL_LIST_FROZEN, &dev->state))
+		return;
+#endif
+
 	local_irq_save(flags);
 	__netif_rx_complete(dev);
 	local_irq_restore(flags);

net/core/netpoll.c

@@ -124,13 +124,21 @@ static void poll_napi(struct netpoll *np)
 	if (test_bit(__LINK_STATE_RX_SCHED, &np->dev->state) &&
 	    npinfo->poll_owner != smp_processor_id() &&
 	    spin_trylock(&npinfo->poll_lock)) {
+		/* When calling dev->poll from poll_napi, we may end up in
+		 * netif_rx_complete. However, only the CPU to which the
+		 * device was queued is allowed to remove it from poll_list.
+		 * Setting POLL_LIST_FROZEN tells netif_rx_complete
+		 * to leave the NAPI state alone.
+		 */
+		set_bit(__LINK_STATE_POLL_LIST_FROZEN, &np->dev->state);
 		npinfo->rx_flags |= NETPOLL_RX_DROP;
 		atomic_inc(&trapped);
 
 		np->dev->poll(np->dev, &budget);
 
 		atomic_dec(&trapped);
 		npinfo->rx_flags &= ~NETPOLL_RX_DROP;
+		clear_bit(__LINK_STATE_POLL_LIST_FROZEN, &np->dev->state);
 		spin_unlock(&npinfo->poll_lock);
 	}
 }