[ovs-dev] [PATCH 2/2] [RFC] tunnelling: stt: Prototype Implementation

[Jesse Gross]

On Thu, Apr 19, 2012 at 7:39 PM, Simon Horman <horms at verge.net.au> wrote:
> On Tue, Apr 17, 2012 at 07:58:33PM -0700, Jesse Gross wrote:
>>  * It should always be possible to encapsulate any Ethernet frame.
>> Things that don't have explicit support like QinQ can still be sent
>> they just won't get the benefit of offloading (similar to a NIC that
>> doesn't support a particular header format).
>
> My reading of the spec was that it required the inner frame to
> be untagged ethernet. I'll correct my code as per your description above.

It can be anything, you just have to handle the offloads in software
before encapsulation.

>>  * There are a lot of atomic ops and other expensive things like division...
>
> Is there anywhere in particular you are noticing theses?

Just generally sprinkled around - atomic inc when computing the ID,
modulo for the source port, atomic ops for reassembly.  However, I
agree that we should focus on correctness first.