When
(a) "dev" is a bridge port which the DSA switch tree offloads, but is
    otherwise not a dsa slave (such as a LAG netdev), or
(b) "dev" is the bridge net device itself

then strange things happen to the dev_hold/dev_put pair:
dsa_schedule_work() will still be called with a DSA port that offloads
that netdev, but dev_hold() will be called on the non-DSA netdev.
Then the "if" condition in dsa_slave_switchdev_event_work() does not
pass, because "dev" is not a DSA netdev, so dev_put() is not called.

This results in the simple fact that we have a reference counting
mismatch on the "dev" net device.

This can be seen when we add support for host addresses installed on the
bridge net device.

ip link add br1 type bridge
ip link set br1 address 00:01:02:03:04:05
ip link set swp0 master br1
ip link del br1
[  968.512278] unregister_netdevice: waiting for br1 to become free. Usage count = 5

It seems foolish to do penny pinching and not add the net_device pointer
in the dsa_switchdev_event_work structure, so let's finally do that.
As an added bonus, when we start offloading local entries pointing
towards the bridge, these will now properly appear as 'offloaded' in
'bridge fdb' (this was not possible before, because 'dev' was assumed to
only be a DSA net device):

00:01:02:03:04:05 dev br0 vlan 1 offload master br0 permanent
00:01:02:03:04:05 dev br0 offload master br0 permanent
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>