Re: [dpdk-dev] Problems running netvsc multiq

[Stephen Hemminger <stephen@networkplumber.org>]

On Fri, 07 Dec 2018 13:15:43 +0200
Mohammed Gamal <mgamal@redhat.com> wrote:

> On Wed, 2018-12-05 at 14:32 -0800, Stephen Hemminger wrote:
> > The problem is a regression in 4.20 kernel. Bisecting now.  
> 
> I was bisecting the kernel and the change that seems to introduce this
> regression is this one:
> 
> commit ae6935ed7d424ffa74d634da00767e7b03c98fd3
> Author: Stephen Hemminger <stephen@networkplumber.org>
> Date:   Fri Sep 14 09:10:17 2018 -0700
> 
>     vmbus: split ring buffer allocation from open
>     
>     The UIO driver needs the ring buffer to be persistent(reused)
>     across open/close. Split the allocation and setup of ring buffer
>     out of vmbus_open. For normal usage vmbus_open/vmbus_close there
>     are no changes; only impacts uio_hv_generic which needs to keep
>     ring buffer memory and reuse when application restarts.
>     
>     Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
>     Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
> 

Yes, this is the kernel problem that introduces the problem.
The issue is actually back in the unbind logic.  When device is unbound
from the netvsc driver all the subchannels aren't cleaned up.

Still debugging.