Re: XDP multi-buffer incl. jumbo-frames (Was: [RFC V1 net-next 1/1] net: ena: implement XDP drop support)

[Willem de Bruijn <willemdebruijn.kernel@gmail.com>]

On Wed, Jun 26, 2019 at 11:01 AM Toke Høiland-Jørgensen <toke@redhat.com> wrote:
>
> Jesper Dangaard Brouer <brouer@redhat.com> writes:
>
> > On Wed, 26 Jun 2019 13:52:16 +0200
> > Toke Høiland-Jørgensen <toke@redhat.com> wrote:
> >
> >> Jesper Dangaard Brouer <brouer@redhat.com> writes:
> >>
> >> > On Tue, 25 Jun 2019 03:19:22 +0000
> >> > "Machulsky, Zorik" <zorik@amazon.com> wrote:
> >> >
> >> >> On 6/23/19, 7:21 AM, "Jesper Dangaard Brouer" <brouer@redhat.com> wrote:
> >> >>
> >> >>     On Sun, 23 Jun 2019 10:06:49 +0300 <sameehj@amazon.com> wrote:
> >> >>
> >> >>     > This commit implements the basic functionality of drop/pass logic in the
> >> >>     > ena driver.
> >> >>
> >> >>     Usually we require a driver to implement all the XDP return codes,
> >> >>     before we accept it.  But as Daniel and I discussed with Zorik during
> >> >>     NetConf[1], we are going to make an exception and accept the driver
> >> >>     if you also implement XDP_TX.
> >> >>
> >> >>     As we trust that Zorik/Amazon will follow and implement XDP_REDIRECT
> >> >>     later, given he/you wants AF_XDP support which requires XDP_REDIRECT.
> >> >>
> >> >> Jesper, thanks for your comments and very helpful discussion during
> >> >> NetConf! That's the plan, as we agreed. From our side I would like to
> >> >> reiterate again the importance of multi-buffer support by xdp frame.
> >> >> We would really prefer not to see our MTU shrinking because of xdp
> >> >> support.
> >> >
> >> > Okay we really need to make a serious attempt to find a way to support
> >> > multi-buffer packets with XDP. With the important criteria of not
> >> > hurting performance of the single-buffer per packet design.
> >> >
> >> > I've created a design document[2], that I will update based on our
> >> > discussions: [2] https://github.com/xdp-project/xdp-project/blob/master/areas/core/xdp-multi-buffer01-design.org
> >> >
> >> > The use-case that really convinced me was Eric's packet header-split.

Thanks for starting this discussion Jesper!

> >> >
> >> >
> >> > Lets refresh: Why XDP don't have multi-buffer support:
> >> >
> >> > XDP is designed for maximum performance, which is why certain driver-level
> >> > use-cases were not supported, like multi-buffer packets (like jumbo-frames).
> >> > As it e.g. complicated the driver RX-loop and memory model handling.
> >> >
> >> > The single buffer per packet design, is also tied into eBPF Direct-Access
> >> > (DA) to packet data, which can only be allowed if the packet memory is in
> >> > contiguous memory.  This DA feature is essential for XDP performance.
> >> >
> >> >
> >> > One way forward is to define that XDP only get access to the first
> >> > packet buffer, and it cannot see subsequent buffers. For XDP_TX and
> >> > XDP_REDIRECT to work then XDP still need to carry pointers (plus
> >> > len+offset) to the other buffers, which is 16 bytes per extra buffer.
> >>
> >> Yeah, I think this would be reasonable. As long as we can have a
> >> metadata field with the full length + still give XDP programs the
> >> ability to truncate the packet (i.e., discard the subsequent pages)
> >
> > You touch upon some interesting complications already:
> >
> > 1. It is valuable for XDP bpf_prog to know "full" length?
> >    (if so, then we need to extend xdp ctx with info)
>
> Valuable, quite likely. A hard requirement, probably not (for all use
> cases).

Agreed.

One common validation use would be to drop any packets whose header
length disagrees with the actual packet length.

> >  But if we need to know the full length, when the first-buffer is
> >  processed. Then realize that this affect the drivers RX-loop, because
> >  then we need to "collect" all the buffers before we can know the
> >  length (although some HW provide this in first descriptor).
> >
> >  We likely have to change drivers RX-loop anyhow, as XDP_TX and
> >  XDP_REDIRECT will also need to "collect" all buffers before the packet
> >  can be forwarded. (Although this could potentially happen later in
> >  driver loop when it meet/find the End-Of-Packet descriptor bit).

Yes, this might be quite a bit of refactoring of device driver code.

Should we move forward with some initial constraints, e.g., no
XDP_REDIRECT, no "full" length and no bpf_xdp_adjust_tail?

That already allows many useful programs.

As long as we don't arrive at a design that cannot be extended with
those features later.

> >
> >
> > 2. Can we even allow helper bpf_xdp_adjust_tail() ?
> >
> >  Wouldn't it be easier to disallow a BPF-prog with this helper, when
> >  driver have configured multi-buffer?
>
> Easier, certainly. But then it's even easier to not implement this at
> all ;)
>
> >  Or will it be too restrictive, if jumbo-frame is very uncommon and
> >  only enabled because switch infra could not be changed (like Amazon
> >  case).

Header-split, LRO and jumbo frame are certainly not limited to the Amazon case.

> I think it would be preferable to support it; but maybe we can let that
> depend on how difficult it actually turns out to be to allow it?
>
> >  Perhaps it is better to let bpf_xdp_adjust_tail() fail runtime?
>
> If we do disallow it, I think I'd lean towards failing the call at
> runtime...

Disagree. I'd rather have a program fail at load if it depends on
multi-frag support while the (driver) implementation does not yet
support it.