Date:	Tue, 4 Nov 2014 09:33:39 -0800
From:	Jesse Gross <jesse@...ira.com>
To:	Tom Herbert <therbert@...gle.com>
Cc:	David Miller <davem@...emloft.net>, netdev <netdev@...r.kernel.org>
Subject: Re: [PATCH net-next 0/7] gue: Remote checksum offload

On Mon, Nov 3, 2014 at 4:59 PM, Tom Herbert <therbert@...gle.com> wrote:
> On Mon, Nov 3, 2014 at 4:19 PM, Jesse Gross <jesse@...ira.com> wrote:
>> On Mon, Nov 3, 2014 at 2:39 PM, Tom Herbert <therbert@...gle.com> wrote:
>>> On Mon, Nov 3, 2014 at 1:26 PM, Jesse Gross <jesse@...ira.com> wrote:
>>>> On Sat, Nov 1, 2014 at 3:57 PM, Tom Herbert <therbert@...gle.com> wrote:
>>>>> This patch set implements remote checksum offload for
>>>>> GUE, which is a mechanism that provides checksum offload of
>>>>> encapsulated packets using rudimentary offload capabilities found in
>>>>> most Network Interface Card (NIC) devices. The outer header checksum
>>>>> for UDP is enabled in packets and, with some additional meta
>>>>> information in the GUE header, a receiver is able to deduce the
>>>>> checksum to be set for an inner encapsulated packet. Effectively this
>>>>> offloads the computation of the inner checksum. Enabling the outer
>>>>> checksum in encapsulation has the additional advantage that it covers
>>>>> more of the packet than the inner checksum including the encapsulation
>>>>> headers.
>>>>
>>>> Tom, I have a pretty hard time squaring this with your previous
>>>> comments on hardware offload. This looks almost identical to a
>>>> protocol-specific hardware offload to me in terms of the implications
>>>> on the stack. It actually is more invasive and less likely to scale
>>>> across protocols, so the relative cost/benefit doesn't really add up
>>>> in my mind.
>>>
>>> With this patch and checksum-unnecessary conversion we can provide
>>> checksum offload for encapsulation on millions of already deployed
>>> NICs without any HW or FW change.  Why do you think this is not a good
>>> cost/benefit tradeoff?
>>
>> I just don't see how this is consistent with your previously stated
>> goal of keeping protocol-specific offload code out of the core stack.
>> Can you explain how this is different?
>
> I think my request was more to avoid putting protocol-specific HW
> offload code in the core stack when existing mechanisms could be used.
> For instance, ntuple filtering is a more generic interface to tell a
> device about special processing for a UDP port than adding an port
> registration mechanism that needs to account for each possible
> encapsulation protocol.

Yes, haven't forgotten about the previous discussion. Hopefully, we'll
be able to spend some time working on this soon and see how it pans
out.

> In these patches we do modify __skb_udp_tunnel_segment which I assume
> is what you're referring to in touching the core stack. There are two
> parts to this: 1) Allowing checksum offload of outer UDP header is
> applicable to any UDP encapsulation protocol 2) When doing remote
> checksum we need to avoid touching the inner checksum.  The latter is
> indicated by SKB_GSO_TUNNEL_REMCSUM being set by the encapsulation
> layer. As I mention in the I-D, remote checksum offload can be
> implemented by any encapsulation protocol that supports some
> reasonable extension (for instance, this is probably something that
> could be implemented in geneve). SKB_GSO_TUNNEL_REMCSUM is a generic
> interface by that definition.

I understand that in theory that this could be applied to other
protocols but in practice I think that is relatively unlikely in most
use cases. Optimizations that result in externally visible change are
usually a no-no (I know this is funny coming from me given STT but the
intention was always that it would be a stopgap until hardware support
was available.)

> But, if you really have a strong objection, I suppose we can start
> using at gso_segment in udp_offloads and put the remote checksum
> offload processing for GSO in a GUE specific segment function.

I thought about this and while it does seem a little unfortunate, I
think it may be the best solution since I suspect that this won't be
the last instance of something like this. It also mirrors what we have
on receive with GRO.

I think if we did that we could also change SKB_GSO_TUNNEL_REMCSUM to
SKB_GSO_SW_ONLY (maybe there is a better name) to indicate that this
is something don't expect hardware to implement. It seems like this
could be something that could be generally useful in the future as
well.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists