Function mlx4_en_queue_bond_work() is used in mlx4_en to start a bond
reconfiguration. It gathers data about a new port map setting, takes
a reference on the netdev that triggered the change and queues a work
object on mlx4_en_priv.mdev.workqueue to perform the operation. The
scheduled work is mlx4_en_bond_work() which calls
mlx4_bond()/mlx4_unbond() and consequently mlx4_do_bond().

At the same time, function mlx4_change_port_types() in mlx4_core might
be invoked to change the port type configuration. As part of its logic,
it re-registers the whole device by calling mlx4_unregister_device(),
followed by mlx4_register_device().

The two operations can result in concurrent access to the data about
currently active interfaces on the device.

Functions mlx4_register_device() and mlx4_unregister_device() lock the
intf_mutex to gain exclusive access to this data. The current
implementation of mlx4_do_bond() doesn't do that which could result in
an unexpected behavior. An updated version of mlx4_do_bond() for use
with an auxiliary bus goes and locks the intf_mutex when accessing a new
auxiliary device array.

However, doing so can then result in the following deadlock:
* A two-port mlx4 device is configured as an Ethernet bond.
* One of the ports is changed from eth to ib, for instance, by writing
  into a mlx4_port<x> sysfs attribute file.
* mlx4_change_port_types() is called to update port types. It invokes
  mlx4_unregister_device() to unregister the device which locks the
  intf_mutex and starts removing all associated interfaces.
* Function mlx4_en_remove() gets invoked and starts destroying its first
  netdev. This triggers mlx4_en_netdev_event() which recognizes that the
  configured bond is broken. It runs mlx4_en_queue_bond_work() which
  takes a reference on the netdev. Removing the netdev now cannot
  proceed until the work is completed.
* Work function mlx4_en_bond_work() gets scheduled. It calls
  mlx4_unbond() -> mlx4_do_bond(). The latter function tries to lock the
  intf_mutex but that is not possible because it is held already by
  mlx4_unregister_device().

This particular case could be possibly solved by unregistering the
mlx4_en_netdev_event() notifier in mlx4_en_remove() earlier, but it
seems better to decouple mlx4_en more and break this reference order.

Avoid then this scenario by recognizing that the bond reconfiguration
operates only on a mlx4_dev. The logic to queue and execute the bond
work can be moved into the mlx4_core driver. Only a reference on the
respective mlx4_dev object is needed to be taken during the work's
lifetime. This removes a call from mlx4_en that can directly result in
needing to lock the intf_mutex, it remains a privilege of the core
driver.
Signed-off-by: Petr Pavlu <petr.pavlu@suse.com>
Tested-by: Leon Romanovsky <leonro@nvidia.com>
Reviewed-by: Leon Romanovsky <leonro@nvidia.com>
Acked-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>