net: Move rcu_barrier from rollback_registered_many to netdev_run_todo.

This patch moves the rcu_barrier from rollback_registered_many
(inside the rtnl_lock) into netdev_run_todo (just outside the rtnl_lock).
This allows us to gain the full benefit of sychronize_net calling
synchronize_rcu_expedited when the rtnl_lock is held.

The rcu_barrier in rollback_registered_many was originally a synchronize_net
but was promoted to be a rcu_barrier() when it was found that people were
unnecessarily hitting the 250ms wait in netdev_wait_allrefs(). Changing
the rcu_barrier back to a synchronize_net is therefore safe.

Since we only care about waiting for the rcu callbacks before we get
to netdev_wait_allrefs() it is also safe to move the wait into
netdev_run_todo.

This was tested by creating and destroying 1000 tap devices and observing
/proc/lock_stat. /proc/lock_stat reports this change reduces the hold
times of the rtnl_lock by a factor of 10. There was no observable
difference in the amount of time it takes to destroy a network device.

Signed-off-by: Eric W. Biederman <[e-mail]>
Signed-off-by: David S. Miller <[e-mail]>

net/core/dev.c

···
 	dev = list_first_entry(head, struct net_device, unreg_list);
 	call_netdevice_notifiers(NETDEV_UNREGISTER_BATCH, dev);
 
-	rcu_barrier();
+	synchronize_net();
 
 	list_for_each_entry(dev, head, unreg_list)
 		dev_put(dev);
···
 	list_replace_init(&net_todo_list, &list);
 
 	__rtnl_unlock();
+
+	/* Wait for rcu callbacks to finish before attempting to drain
+	 * the device list.  This usually avoids a 250ms wait.
+	 */
+	if (!list_empty(&list))
+		rcu_barrier();
 
 	while (!list_empty(&list)) {
 		struct net_device *dev