Re: [patch net 3/3] devlink: fix a deadlock with nested instances during namespace remove

[Jakub Kicinski <kuba@kernel.org>]

On Tue,  9 May 2023 12:09:36 +0200 Jiri Pirko wrote:
> The commit 565b4824c39f ("devlink: change port event netdev notifier
> from per-net to global") changed original per-net notifier to be global
> which fixed the issue of non-receiving events of netdev uninit if that
> moved to a different namespace. That worked fine in -net tree.
> 
> However, later on when commit ee75f1fc44dd ("net/mlx5e: Create
> separate devlink instance for ethernet auxiliary device") and
> commit 72ed5d5624af ("net/mlx5: Suspend auxiliary devices only in
> case of PCI device suspend") were merged, a deadlock was introduced
> when removing a namespace with devlink instance with another nested
> instance.
> 
> Here there is the bad flow example resulting in deadlock with mlx5:
> net_cleanup_work -> cleanup_net (takes down_read(&pernet_ops_rwsem) ->
> devlink_pernet_pre_exit() -> devlink_reload() ->
> mlx5_devlink_reload_down() -> mlx5_unload_one_devl_locked() ->
> mlx5_detach_device() -> del_adev() -> mlx5e_remove() ->
> mlx5e_destroy_devlink() -> devlink_free() ->
> unregister_netdevice_notifier() (takes down_write(&pernet_ops_rwsem)

Why don't we have a single, static notifier for all of devlink?
Why the per device/per port notifiers?

We have the devlink port pointer in struct net_device, resolving from
a global event to the correct devlink instance is trivial.