netdev: Add netdev->select_queue() method.

Devices or device layers can set this to control the queue selection
performed by dev_pick_tx().

This function runs under RCU protection, which allows overriding
functions to have some way of synchronizing with things like dynamic
->real_num_tx_queues adjustments.

This makes the spinlock prefetch in dev_queue_xmit() a little bit
less effective, but that's the price right now for correctness.

Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/netdevice.h

@@ -724,6 +724,9 @@ struct net_device
 	void                    (*poll_controller)(struct net_device *dev);
 #endif
 
+	u16			(*select_queue)(struct net_device *dev,
+						struct sk_buff *skb);
+
 #ifdef CONFIG_NET_NS
 	/* Network namespace this network device is inside */
 	struct net		*nd_net;

net/core/dev.c

@@ -1670,6 +1670,9 @@ static struct netdev_queue *dev_pick_tx(struct net_device *dev,
 {
 	u16 queue_index = 0;
 
+	if (dev->select_queue)
+		queue_index = dev->select_queue(dev, skb);
+
 	skb_set_queue_mapping(skb, queue_index);
 	return netdev_get_tx_queue(dev, queue_index);
 }
@@ -1710,14 +1713,14 @@ int dev_queue_xmit(struct sk_buff *skb)
 	}
 
 gso:
-	txq = dev_pick_tx(dev, skb);
-	spin_lock_prefetch(&txq->lock);
-
 	/* Disable soft irqs for various locks below. Also
 	 * stops preemption for RCU.
 	 */
 	rcu_read_lock_bh();
 
+	txq = dev_pick_tx(dev, skb);
+	spin_lock_prefetch(&txq->lock);
+
 	/* Updates of qdisc are serialized by queue->lock.
 	 * The struct Qdisc which is pointed to by qdisc is now a
 	 * rcu structure - it may be accessed without acquiring