Re: [PATCH v9 1/2] virtio-net: define UDP tunnel segmentation offload feature

[Paolo Abeni <pabeni@redhat.com>]

On 10/28/24 13:26, Willem de Bruijn wrote:
> On Mon, Oct 28, 2024 at 8:08 AM Paolo Abeni <pabeni@redhat.com> wrote:
>> On 10/28/24 04:27, Jason Wang wrote:
>>> As the topic becomes more complicated, I wonder if we can do something
>>> easier, that is focusing on the one checksum offload first. And I
>>> would like to check the current Linux behaviour first for non-GSO UDP
>>> tunnel:
>>>
>>> 1) For TX, as you've mentioned, the csum_start/offset has already
>>> pointed to the inner header (but is this the innermost?)
>>> 2) For RX, what header would csum_start/offset points (considering we
>>> have nested tunnels)
>>>
>>> I'm asking to make sure our proposal doesn't conflict with the
>>> existing implementation.
>>
>> In Linux, csum_start/offset, both for rx and tx, points to the innermost
>> _offloaded_ header.
>>
>> On RX, since GSO partial does not apply, such innermost offloaded header
>> correspond the outermost/only header (when no tunnels are configured) or
>> the 2nd outermost, when tunnels are configured and GRO-ed.
>>
>> On TX, via GSO partial and local csum offload, the csum_start can, in
>> theory point to an header with an header with an arbitrary high level,
>> but currently no driver/tx path allow for GSO partial with more than 1
>> encapsulation level (may change in the future).
>>
>> The proposed specification change in v9 for NEED_CSUM matches the above.
>>
>> Since reaching an agreement WRT DATA_VALID handling looks problematic,
>> would you consider a virtual meeting to discuss the topic?
>>
>> Thanks,
>>
>> Paolo
>>
>> p.s. to be clear, with 'you' I include both Jason and Willem.
> 
> Works for me.

Thanks! I forgot to mention that it would be great if you could share
(even off-list) a few timeslots that could fit your schedule, and I'll
try to find a match.

Cheers,

Paolo