[ovs-dev] [PATCH] netdev: Dynamic per-port Flow API.

Ilya Maximets i.maximets at samsung.com
Mon May 13 12:47:11 UTC 2019


On 13.05.2019 15:33, Ophir Munk wrote:
> Hi,
> 
>> -----Original Message-----
>> From: Ilya Maximets <i.maximets at samsung.com>
>> Sent: Monday, May 13, 2019 12:22 PM
>> To: Ophir Munk <ophirmu at mellanox.com>; ovs-dev at openvswitch.org
>> Cc: Ian Stokes <ian.stokes at intel.com>; Flavio Leitner <fbl at sysclose.org>;
>> Kevin Traynor <ktraynor at redhat.com>; Roni Bar Yanai
>> <roniba at mellanox.com>; Finn Christensen <fc at napatech.com>; Ben Pfaff
>> <blp at ovn.org>; Simon Horman <simon.horman at netronome.com>; Olga
>> Shern <olgas at mellanox.com>; Asaf Penso <asafp at mellanox.com>; Oz
>> Shlomo <ozsh at mellanox.com>
>> Subject: Re: [PATCH] netdev: Dynamic per-port Flow API.
>>
>>> I am recreating this scenario in my setup.
>>
>> I see. Yes, you're right. And I think that this case could be reproduced on
>> current master without any patches. So, it's a bug that we need to fix.
>> Otherwise userspace datapath will try to offload its flows to the unrelated
>> system interfaces. For now we could just forbid offloading to vports in dpif-
>> netdev. I'll prepare the patch. This fix also should be backported.
>>
> 
> Till now there was no vport dpif-netdev offload. I think the bug you mention was minor (getting a redundant log message like "vxlan_sys_4789 could not be found").
> No harm was caused and vport in dpif-system behaved correctly. 
> Is it considered a broken implementation? Not sure. Maybe we can leave it as is in current OVS (no backporting) and have a fix in this series for the new vport dpif-netdev offload.

The issue will happen if "vxlan_sys_4789" will be eventually found.
For example, this could happen if the kernel datapath runs at the
same time and also has vxlan port. In this case dpif-netdev will
try to install completely unrelated flow to the system interface.


More information about the dev mailing list