Re: [PATCH net-next v3 15/16] net/mlx5e: take the rtnl lock when calling netif_set_xps_queue

Quoting Saeed Mahameed (2021-03-12 21:54:18)

On Fri, 2021-03-12 at 16:04 +0100, Antoine Tenart wrote:

quoted

netif_set_xps_queue must be called with the rtnl lock taken, and this
is
now enforced using ASSERT_RTNL(). mlx5e_attach_netdev was taking the
lock conditionally, fix this by taking the rtnl lock all the time.

There is a reason why it is conditional:
we had a bug in the past of double locking here:

[ 4255.283960] echo/644 is trying to acquire lock:

 [ 4255.285092] ffffffff85101f90 (rtnl_mutex){+..}, at:
mlx5e_attach_netdev0xd4/0×3d0 [mlx5_core]

 [ 4255.287264] 

 [ 4255.287264] but task is already holding lock:

 [ 4255.288971] ffffffff85101f90 (rtnl_mutex){+..}, at:
ipoib_vlan_add0×7c/0×2d0 [ib_ipoib]

ipoib_vlan_add is called under rtnl and will eventually call 
mlx5e_attach_netdev, we don't have much control over this in mlx5
driver since the rdma stack provides a per-prepared netdev to attach to
our hw. maybe it is time we had a nested rtnl lock .. 

Not sure we want to add a nested rtnl lock because of xps. I'd like to
see other options first. Could be having a locking mechanism for xps not
relying on rtnl; if that's possible.

As for this series, patches 6, 15 (this one) and 16 are not linked to
and do not rely on the other patches. They're improvement or fixes for
already existing behaviours. The series already gained enough new
patches since v1 and I don't want to maintain it out-of-tree for too
long, so I'll resend it without patches 6, 15 and 16; and then we'll be
able to focus on the xps locking relationship with rtnl.

Antoine