On 27/08/14 15:43, Tom Herbert wrote:
> On Tue, Aug 26, 2014 at 11:13 PM, Anton Ivanov
> <[email protected]> wrote:
>> On 26/08/14 22:42, Tom Herbert wrote:
>>> On Tue, Aug 26, 2014 at 2:09 PM, Anton Ivanov
>>> <[email protected]> wrote:
>>>> On 26/08/14 21:47, Tom Herbert wrote:
>>>>> Remote checksum offload is intended to be used to get checksum offload
>>>>> of encapsulated packets to work with legacy NICs (those that don't
>>>>> understand encapsulation protocols). I will be posting an
>>>>> implementation to Linux shortly for GUE and/or Geneve.
>>>> This one was floated on the mailing list during the IETF-89 (and
>>>> immediately thereabouts) discussions around geneve, fixed/floating
>>>> headers, etc a while back.
>>>>
>>>> We have had a working implementation for the GRE case for a while now
>>>> (we have it in a limited release product based around UML in customer
>>>> labs). I need to get around to contribute it upstream (in fact, I was
>>>> going to work on the patchsets this week).
>>>>
>>> You extended GRE with fields for checksum offload?
>> GRE has an optional checksum already as per RFC 2784 - nothing to extend.
>>
>> If it is computed at source and checked by hardware or hypervisor at
>> destination you can compute the payload checksum from the GRE checksum
>> and pass it to the guest. It is an excellent reference example for your
>> idea because it can be implemented today using IP_HDRINCL which is not
>> yet possible for the UDP encaps as there is no sane way to get to the
>> header and get the csum from it for those without changing the kernel.
>> If nvGRE had the checksum specified (which it presently does not) it
>> could have been subjected to the same technique.
>>
> I don't think we're talking about same thing. I think you're referring
> to an offloading technique for receive side, which yes, can be done
> without protocol change pretty easily. Remote checksum offload is
> meant to offload the transmit checksum computation for encapsulation,
> which I believe does need some protocol assistance.

Either that or OOB control plane configuration which says for this 
Virtual Network Identifier and/or source compute all checksums, the 
frames are coming without ones. F.E. for gre key 0x012345678 compute the 
checksums before re-xmitting.

IMHO it is not strictly necessary to stick this into the protocol.

A.

>
>
_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3

Reply via email to