[ovs-dev] [PATCH V2 1/1] dpif-netlink: Log eth type 0x1234 not offloadable

Ilya Maximets i.maximets at samsung.com
Tue Jul 30 12:19:52 UTC 2019


On 30.07.2019 13:30, Ilya Maximets wrote:
> On 30.07.2019 13:23, Eli Britstein wrote:
>>
>> On 7/30/2019 1:10 PM, Ilya Maximets wrote:
>>> On 03.07.2019 7:58, Eli Britstein wrote:
>>>> Ethernet type 0x1234 is used for testing and not being offloadable. For
>>>> testing offloadable features, log about using this value.
>>>>
>>>> Signed-off-by: Eli Britstein <elibr at mellanox.com>
>>>> Acked-by: Roi Dayan <roid at mellanox.com>
>>>> Signed-off-by: Eli Britstein <elibr at mellanox.com>
>>>> Acked-by: Ben Pfaff <blp at ovn.org>
>>>> ---
>>>>   lib/dpif-netlink.c | 1 +
>>>>   1 file changed, 1 insertion(+)
>>>>
>>>> diff --git a/lib/dpif-netlink.c b/lib/dpif-netlink.c
>>>> index ba80a0079..a0d51ae61 100644
>>>> --- a/lib/dpif-netlink.c
>>>> +++ b/lib/dpif-netlink.c
>>>> @@ -2007,6 +2007,7 @@ parse_flow_put(struct dpif_netlink *dpif, struct dpif_flow_put *put)
>>>>   
>>>>       /* When we try to install a dummy flow from a probed feature. */
>>>>       if (match.flow.dl_type == htons(0x1234)) {
>>>> +        VLOG_INFO_RL(&rl, "eth 0x1234 is special and not offloadable");
>>>>           return EOPNOTSUPP;
>>>>       }
>>> But what is the purpose of this patch?  What is the use case?
>> RH wanted to test that dl_type is offloaded. Coincidentally, they used 
>> 0x1234, and it was not offloaded, and they didn't understand why, and 
>> suggested a log message.
> 
> I'll take a closer look, but it seems that we just need to remove this
> 'if' statement and allow oflloading.

'dpif_probe_feature' always has DPIF_FP_PROBE flag set. Other probing code uses
dpif_execute() which uses DPIF_OP_EXECUTE, hence never calls parse_flow_put().
So, this 'if' statement is wrong and should be deleted as it only forbids
offloading of the real legitimate flows with dl_type 0x1234. Dummy flows never
reaches this code.

> 
>>>
>>> Actually, it looks like we need to just remove above 'if' statement
>>> entirely.  Just a few lines above there is a check if we are probing
>>> or installing real flow:
>>>
>>>     if (put->flags & DPIF_FP_PROBE) {
>>>         return EOPNOTSUPP;
>>>     }
>>>
>>> So, we will never get there while probing.  But why we're restricting
>>> this flow type from being offloaded?  'netdev_flow_put' will refuse
>>> to offload if it doesn't know that flow type, but this shouldn't be
>>> done here.
>>>
>>> In case we have a dummy flow without DPIF_FP_PROBE flag set, we need
>>> to fix upper layers.  Is it the case?
>>
>> I didn't look into it why we restrict it and if there is a real reason 
>> why in this layer. You may be right, but I don't know.
>>
>>>
>>> Best regards, Ilya Maximets.


More information about the dev mailing list