[net-next,0/5] devlink: hold the instance lock in eswitch callbacks

Message

Jakub Kicinski March 17, 2022, 4:20 a.m. UTC

Series number 2 in the effort to hold the devlink instance lock
in call driver callbacks. We have the following drivers using
this API:

 - bnxt, nfp, netdevsim - their own locking is removed / simplified
   by this series; all of them needed a lock to protect from changes
   to the number of VFs while switching modes, now the VF config bus
   callback takes the devlink instance lock via devl_lock();
 - ice - appears not to allow changing modes while SR-IOV enabled,
   so nothing to do there;
 - liquidio - does not contain any locking;
 - octeontx2/af - is very special but at least doesn't have locking
   so doesn't get in the way either;
 - mlx5 has a wealth of locks - I chickened out and dropped the lock
   in the callbacks so that I can leave the driver be, for now.

The last one is obviously not ideal, but I would prefer to transition
the API already as it make take longer.

LMK if it's an abuse of power / I'm not thinking straight.

Jakub Kicinski (5):
  bnxt: use the devlink instance lock to protect sriov
  devlink: add explicitly locked flavor of the rate node APIs
  netdevsim: replace port_list_lock with devlink instance lock
  netdevsim: replace vfs_lock with devlink instance lock
  devlink: hold the instance lock during eswitch_mode callbacks

 drivers/net/ethernet/broadcom/bnxt/bnxt.c     |  1 -
 drivers/net/ethernet/broadcom/bnxt/bnxt.h     |  6 --
 .../net/ethernet/broadcom/bnxt/bnxt_sriov.c   |  4 +-
 drivers/net/ethernet/broadcom/bnxt/bnxt_vfr.c | 22 ++---
 .../mellanox/mlx5/core/eswitch_offloads.c     | 15 +++-
 .../net/ethernet/netronome/nfp/nfp_devlink.c  |  7 +-
 drivers/net/netdevsim/dev.c                   | 85 +++++++++---------
 drivers/net/netdevsim/netdevsim.h             |  2 -
 include/net/devlink.h                         |  4 +
 net/core/devlink.c                            | 90 ++++++++++++-------
 10 files changed, 128 insertions(+), 108 deletions(-)

Comments

Simon Horman March 18, 2022, 8:12 a.m. UTC | #1

On Wed, Mar 16, 2022 at 09:20:18PM -0700, Jakub Kicinski wrote:
> Series number 2 in the effort to hold the devlink instance lock
> in call driver callbacks. We have the following drivers using
> this API:
> 
>  - bnxt, nfp, netdevsim - their own locking is removed / simplified
>    by this series; all of them needed a lock to protect from changes
>    to the number of VFs while switching modes, now the VF config bus
>    callback takes the devlink instance lock via devl_lock();
>  - ice - appears not to allow changing modes while SR-IOV enabled,
>    so nothing to do there;
>  - liquidio - does not contain any locking;
>  - octeontx2/af - is very special but at least doesn't have locking
>    so doesn't get in the way either;
>  - mlx5 has a wealth of locks - I chickened out and dropped the lock
>    in the callbacks so that I can leave the driver be, for now.
> 
> The last one is obviously not ideal, but I would prefer to transition
> the API already as it make take longer.
> 
> LMK if it's an abuse of power / I'm not thinking straight.

Thanks Jakub,

we have reviewed this from an NFP driver perspective and are happy
with this change - it looks quite straightforward from that perspective.