DSA stacks an Ethernet device on top of an Ethernet device. This can
cause false positive lockdep splats for the transmit queue:
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
=============================================
[ INFO: possible recursive locking detected ]
4.0.0-rc7-01838-g70621a215fc7 #386 Not tainted
---------------------------------------------
kworker/0:0/4 is trying to acquire lock:
 (_xmit_ETHER#2){+.-...}, at: [<
c040e95c>] sch_direct_xmit+0xa8/0x1fc
but task is already holding lock:
 (_xmit_ETHER#2){+.-...}, at: [<
c03f4208>] __dev_queue_xmit+0x4d4/0x56c
other info that might help us debug this:
 Possible unsafe locking scenario:
       CPU0
       ----
  lock(_xmit_ETHER#2);
  lock(_xmit_ETHER#2);
To avoid this, walk the tq queues of the dsa slaves and set a lockdep
class.
Signed-off-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
        return 0;
 }
 
+static struct lock_class_key dsa_slave_netdev_xmit_lock_key;
+static void dsa_slave_set_lockdep_class_one(struct net_device *dev,
+                                           struct netdev_queue *txq,
+                                           void *_unused)
+{
+       lockdep_set_class(&txq->_xmit_lock,
+                         &dsa_slave_netdev_xmit_lock_key);
+}
+
 int dsa_slave_suspend(struct net_device *slave_dev)
 {
        struct dsa_slave_priv *p = netdev_priv(slave_dev);
 
-       netif_device_detach(slave_dev);
-
        if (p->phy) {
                phy_stop(p->phy);
                p->old_pause = -1;
        slave_dev->netdev_ops = &dsa_slave_netdev_ops;
        slave_dev->swdev_ops = &dsa_slave_swdev_ops;
 
+       netdev_for_each_tx_queue(slave_dev, dsa_slave_set_lockdep_class_one,
+                                NULL);
+
        SET_NETDEV_DEV(slave_dev, parent);
        slave_dev->dev.of_node = ds->pd->port_dn[port];
        slave_dev->vlan_features = master->vlan_features;