On Sun, Dec 6, 2015 at 1:30 PM, Tom Herbert [off-list ref] wrote:

On Sun, Dec 6, 2015 at 10:44 AM, Alexander Duyck
[off-list ref] wrote:

quoted

On Sun, Dec 6, 2015 at 8:31 AM, Tom Herbert [off-list ref] wrote:

quoted

quoted

The only spot I think you and I disagreed on was the approach.  I
don't know if the hard push back does anything but punish the users by
delaying the time needed to find a reasonable solution.  I really
think if we are going to get the hardware vendors to change their
behavior we have to create a market demand for it.  Having a bit of
marketable data showing the folly of this approach versus the 1's
compliment checksum would probably do more to encourage and/or shame
them into it than simply pushing for this based on engineering
opinion.

I don't know what "marketable data" means. But I do know that we're
like 70 postings into this thread, into the third patch set regarding
this, yet nobody has bothered to contribute any data on what these
patches do and what the quantifiable benefits are with HW offload of
these protocols. I would test this stuff myself, but I don't have
access to any NICs with necessary support. If someone else can start
testing and providing meaningful data it would be most helpful...

Here is an example of something kind of like what I am talking about:
http://www.mellanox.com/related-docs/whitepapers/CB_Intel_XL710.pdf

I have seen evidence of the gains first hand.  The biggest gain ends
up being the result of GRO, and you cannot make use of GRO without
some form of Rx checksum offload.

Right, but we recoup the gains of GRO simply by enabling the UDP
checksum. This works for all the UDP encapsulations, and probably
about all NICs in deployment. You don't need protocol specific
offloads for this. I have posted performance data many times on this,
it is a clear win.

That works for Linux to Linux, but what about the cases where you have
a non-Linux endpoint on the other end such as something like a Cisco
switch?  That is where having the protocol specific offload is useful
as long as the hardware has sufficient capabilities to support it.

As far as trying to get the vendors to give up their protocol parsing,
it will probably never happen.  It wouldn't surprise me if it is due
to product requirements from Microsoft in order to support things like
RSS, RSC, and filtering on inner header fields.  If Linux doesn't care
about that we can drop support for it, but it still doesn't mean they
can drop those bits from the hardware design so they would likely
interpret this as a request to add a new feature instead of fixing or
replacing the existing checksum approach.

- Alex