On 7/30/2014 5:14 AM, Benjamin Poirier wrote:

On 2014/07/29 16:42, Amir Vadai wrote:

quoted

On 7/29/2014 3:06 PM, Jean Delvare wrote:

quoted

Le Tuesday 29 July 2014 à 13:44 +0300, Amir Vadai a écrit :

quoted

On 7/29/2014 1:34 PM, Jean Delvare wrote:

quoted

Many thanks for the fast and detailed answer! I'll go backport this
commit to kernel 3.12 (and 3.0, hopefully it doesn't have too many
dependencies?)

I don't think there is any dependency for that commit. If you're
backporting it, also take a small fix to it:
f24f790 net/mlx4_core: Load the Eth driver first

Thanks for the hint. I suppose I should pick 7855bff4 "net/mlx4_core:
Load the IB driver when the device supports IBoE" as well?

Right - missed it.

Thanks for the pointers.
Looking at f24f790f, I'm wondering if request_module(EN_DRV_NAME) ought
to be used instead of _nowait to guarantee the desired effect. Suppose
for example that the process is migrated to a different cpu between the
two request_module calls, the second one could complete before the first
one no?

The issue was addressed after the commit above, through these two
commits in a proper and robust way:
9433c18 IB/mlx4: Invoke UPDATE_QP for proxy QP1 on MAC changes
ce8d9e0 net/mlx4_core: Add UPDATE_QP SRIOV wrapper support

Thanks,
Amir