On Tue, Nov 09, 2021 at 06:17:29AM -0800, Jakub Kicinski wrote:
On Tue, 9 Nov 2021 16:12:33 +0200 Leon Romanovsky wrote:
quoted
quoted
You'd need to tell me more about what the notifier is used for (I see
RoCE in the call trace). I don't understand why you need to re-register
a global (i.e. not per netns) notifier when devlink is switching name
spaces.
RDMA subsystem supports two net namespace aware scenarios.
We need global netdev_notifier for shared mode. This is legacy mode where
we listen to all namespaces. We must support this mode otherwise we break
whole RDMA world.
See commit below:
de641d74fb00 ("Revert "RDMA/mlx5: Fix devlink deadlock on net namespace deletion"")
But why re-reg? To take advantage of clean event replay?
IIUC the problem is that the un-reg is called from the reload_down path.
This is how devlink reload was explained to me by Jiri. It suppose to
unload and load whole driver again. In our case, it unloads mlx5_core,
which destroys all ULPs (VDPA, RDMA, e.t.c) and these ULPs are not aware
of devlink at all. Especially they are not aware of the reason of
devlink reload.
This is why I asked you. It is not related to devlink locking directly,
but indirectly I didn't want to work on anything related to locking
without making sure that devlink.c is correct.
Thanks