On Tue, Nov 16, 2021 at 08:44:42AM -0400, Jason Gunthorpe wrote:

On Tue, Nov 16, 2021 at 07:57:09AM +0100, Jiri Pirko wrote:

quoted

quoted

There is only one place in the entire kernel calling the per-ns
register_netdevice_notifier_dev_net() and it is burred inside another
part of mlx5 for some reason..

Yep. I added it there to solve this deadlock.

I wonder how it can work safely inside a driver, since when are
drivers NS aware?

        uplink_priv->bond->nb.notifier_call = mlx5e_rep_esw_bond_netevent;
        ret = register_netdevice_notifier_dev_net(netdev,
                                                  &uplink_priv->bond->nb,
                                                  &uplink_priv->bond->nn);

Doesn't that just loose events when the user moves netdev to another
namespace?

I don't think so, it looks like holding rtnl_lock is enough.
However, we need all events and not NS-specific ones and it maybe solves
the deadlock, but doesn't solve our issue.

BTW, this makes me wonder how commit 554873e51711 ("net: Do not take net_rwsem in __rtnl_link_unregister()")
aligns with the comment near pernet_ops_rwsem and explode usage in other
places.

    57 /*
    58  * pernet_ops_rwsem: protects: pernet_list, net_generic_ids,
    59  * init_net_initialized and first_device pointer.
    60  * This is internal net namespace object. Please, don't use it
    61  * outside.
    62  */

Thanks