[xdp-hints] Re: Interest in TX-side XDP metadata hints from Intel (+ Red Hat)

[Stanislav Fomichev <sdf@google.com>]

On Thu, Sep 21, 2023 at 1:17 AM Song, Yoong Siang
<yoong.siang.song@intel.com> wrote:
>
> On Thursday, September 21, 2023 12:36 PM, Stanislav Fomichev <sdf@google.com> wrote:
> >On Wed, Sep 20, 2023 at 7:10 PM Song, Yoong Siang
> ><yoong.siang.song@intel.com> wrote:
> >>
> >> On Friday, September 15, 2023 5:06 AM, Stanislav Fomichev <sdf@google.com>
> >wrote:
> >> >On Wed, Sep 13, 2023 at 6:22 PM Song, Yoong Siang
> >> ><yoong.siang.song@intel.com> wrote:
> >> >>
> >> >> On Saturday, September 9, 2023 12:38 AM Stanislav Fomichev
> >> ><sdf@google.com> wrote:
> >> >> >On Fri, Sep 8, 2023 at 7:18 AM Jesper Dangaard Brouer
> >> >> ><hawk@kernel.org>
> >> >> >wrote:
> >> >> >>
> >> >> >> Hi Stanislav,
> >> >> >>
> >> >> >> Sharing that Intel (+Red Hat) have an interest in getting XDP
> >> >> >> metadata hints working on smaller devices used for "Edge
> >> >> >> Controls for
> >> >Industrial"
> >> >> >> (ECI)[1].
> >> >> >>
> >> >> >> What are your plans for your patchset[2] ?
> >> >> >
> >> >> >Sorry, I was OOO for two weeks and then got distracted by some
> >> >> >internal work :- ( I'll probably post another respin next week. I
> >> >> >think I've addressed most of the Maciej's feedback, but the only
> >> >> >thing that we need to move this forward is to support another
> >> >> >device besides mlx5. Maciej was suggesting ice, but I don't have
> >> >> >the hw :-( I have some idfp-capable devices, so I'll probably try
> >> >> >to play with it instead (but
> >> >the idpf driver is still not pulled).
> >> >> >
> >> >> >So if you want to help speed it up, we should probably focus first
> >> >> >on getting some of the intel nics support tx metadata.
> >> >> >To proceed, maybe I should post my next iteration and we can
> >> >> >collaborate on the ice driver changes?
> >> >> >Who should I CC explicitly?
> >> >> >
> >> >> Hi Stanislav,
> >> >>
> >> >> I can help to speed it up. I don't have ice HW on my hand as well,
> >> >> so maybe we can start with igc drivers, followed by stmmac drivers.
> >> >>
> >> >> Please post your next iteration, then we can collaborate based on that.
> >> >
> >> >Awesome, thanks! The v2 is out, you're on the CC. Feel free to reach
> >> >out directly if you have any questions or suggestions.
> >>
> >> Hi Stanislav,
> >>
> >> Thanks for the v2. I do have one question. What is the code base which you are
> >using?
> >> When I try to apply your v2 patch set on top of latest net-next, there
> >> are few merge conflicts, and it seem like Rx Metadata is not included in your
> >base.
> >
> >Anything prior to my recent commit "bpf: make it easier to add new metadata
> >kfunc" should work. That's most likely the reason it conflicts.
> >So try to cherry pick it on top of the following and let me know if you still have an
> >issue; I can rebase and repost:
> >5c04433daf9e - bpf: Charge modmem for struct_ops trampoline (6 days
> >ago) <Song Liu>
>
> This base work for me. Thanks.
> Btw, how we should collaborate?
> After done coding and testing on my side, I send you my patches,
> then you will include them into your patchset v3?

Yeah, that works for me. Unless you have better ideas :-)

> >Regarding rx metadata: it's separate from af_xdp and it's more generically
> >implemented in the xdp layer, see:
> >Documentation/networking/xdp-rx-metadata.rst
> >
> >Also see tools/testing/selftests/bpf/prog_tests/xdp_metadata.c for an example
> >on how to use both rx and tx sides.
> >
> >> Thanks & Regards
> >> Siang
> >>
> >> >
> >> >I also have to abaon my idea of playing with idpf since the driver
> >> >that's posted on the list doesn't have af_xdp support :-(
> >> >
> >> >
> >> >> Thanks & Regards
> >> >> Siang
> >> >>
> >> >> >> My suggestion (to Intel) is that we implement the TX-timestamp
> >> >> >> (plus
> >> >> >> LaunchTime) on top of Stanslav's patches[2].  For drivers
> >> >> >> igc+igb+stmmac which I think will cover the tests[3] need by
> >> >> >> Arun/Costin (?). And then post it upstream for review.
> >> >> >>
> >> >> >> The Intel test tool[3] isn't open, but it is an AF_XDP based
> >> >> >> Real-Time Compute Performance (RTCP) application that show-case
> >> >> >> a real-time control loop application.  The current code uses
> >> >> >> out-of-tree[4]
> >> >"TXTIME"
> >> >> >> LaunchTime feature, that want not upstreamed.
> >> >> >>
> >> >> >> Lets us instead implement a proper TX LaunchTime feature on top of [2].
> >> >> >>
> >> >> >> --Jesper
> >> >> >> cc. XDP-hints ML as anyone is free to code this up, code help
> >> >> >> here[5]
> >> >> >>
> >> >> >>
> >> >> >> [1] https://eci.intel.com/
> >> >> >> [2]
> >> >> >> https://lore.kernel.org/all/20230809165418.2831456-1-sdf@google.
> >> >> >> com / [3] https://github.com/OTCShare2/rtcp-xdp
> >> >> >> [4] https://lore.kernel.org/all/20210621230635.3a83851c@carbon/
> >> >> >> [5]
> >> >> >> https://github.com/xdp-project/xdp-project/blob/master/areas/tsn
> >> >> >> /co
> >> >> >> de0
> >> >> >> 1_follow_qdisc_TSN_offload.org#tx-time-to-hardware-driver-igc