Re: [PATCH net v4] net: xdp: account for layer 3 packets in generic skb handler

[Jakub Kicinski <kuba@kernel.org>]

On Fri, 14 Aug 2020 08:56:48 +0200 Jason A. Donenfeld wrote:
> On Thu, Aug 13, 2020 at 11:01 PM Jakub Kicinski <kuba@kernel.org> wrote:
> > > I had originally dropped this patch, but the issue kept coming up in
> > > user reports, so here's a v4 of it. Testing of it is still rather slim,
> > > but hopefully that will change in the coming days.  
> >
> > Here an alternative patch, untested:  
> 
> Funny. But come on now... Why would we want to deprive our users of
> system consistency?

We should try for consistency between xdp and cls_bpf instead.

> Doesn't it make sense to allow users to use the same code across
> interfaces? You actually want them to rewrite their code to use a
> totally different trigger point just because of some weird kernel
> internals between interfaces?

We're not building an abstraction over the kernel stack so that users
won't have to worry how things work. Users need to have a minimal
understanding of how specific hooks integrate with the stack and what
they are for. And therefore why cls_bpf is actually more efficient to
use in L3 tunnel case.

> Why not make XDP more useful and more generic across interfaces? It's
> very common for systems to be receiving packets with a heavy ethernet
> card from the current data center, in addition to receiving packets
> from a tunnel interface connected to a remote data center, with a need
> to run the same XDP program on both interfaces. Why not support that
> kind of simplicity?
> 
> This is _actually_ something that's come up _repeatedly_. This is a
> real world need from real users who are doing real things. Why not
> help them?

I'm sure it comes up repeatedly because we don't return any errors,
so people waste time investigating why it doesn't work.

> It's not at the expense of any formal consistency, or performance, or
> even semantic perfection. It costs very little to support these
> popular use cases.