Re: [PATCH bpf-next 2/4] bpf: Craft non-linear skbs in BPF_PROG_TEST_RUN

[Paul Chaignon <paul.chaignon@gmail.com>]

On Thu, Sep 04, 2025 at 09:27:58AM -0700, Amery Hung wrote:
> 
> 
> On 9/4/25 9:02 AM, Daniel Borkmann wrote:
> > On 9/4/25 5:56 PM, Alexei Starovoitov wrote:
> > > On Thu, Sep 4, 2025 at 5:11 AM Paul Chaignon
> > > <paul.chaignon@gmail.com> wrote:
> > > > 
> > > > This patch adds support for crafting non-linear skbs in BPF test runs
> > > > for tc programs, via a new flag BPF_F_TEST_SKB_NON_LINEAR. When this
> > > > flag is set, only the L2 header is pulled in the linear area.
> > > 
> > > ...
> > > 
> > > > +               /* eth_type_trans expects the Ethernet header in
> > > > the linear area. */
> > > > +               __pskb_pull_tail(skb, ETH_HLEN);
> > > 
> > > Looks useful, but only L2 ? Is it realistic ?
> > > I don't recall any driver that would do L2 only.
> > > Is L2 only enough to cover all corner cases in your progs ?
> > > Should the linear size be a configurable parameter for prog_run() ?

I think it's enough for Cilium. It's a bit of a worst case for us AFAIU.
In any case, I'm definitely not against making it configurable.

> > 
> > Yeah perhaps we could make this configurable. The ETH_HLEN is a common
> > case
> > we've seen and also what virtual drivers pull in at min, but with NICs
> > doing
> > header/data split its probably better to let the user define this as part
> > of the testing. Then we're more flexible.
> > 
> 
> How about letting users specify the linear size through ctx->data_end? I am
> working on a set that introduces a kfunc, bpf_xdp_pull_data(). A part of is
> to support non-linear xdp_buff in test_run_xdp, and I am doing it through
> ctx->data_end. Is it something reasonable for test_run_skb?

Oh, nice! That was next on my list :)

Why use data_end though? I guess it'd work for skb, but can't we just
add a new field to the anonymous struct for BPF_PROG_TEST_RUN?

> 
> > Cheers,
> > Daniel
>