Re: [PATCH net-next v4 2/3] virtio_net: add optional flow dissection in virtio_net_hdr_to_skb

在 2021/6/10 下午12:05, Alexei Starovoitov 写道:

On Wed, Jun 9, 2021 at 8:53 PM Jason Wang [off-list ref] wrote:

quoted

It works but it couples virtio with bpf.

Jason,

I think your main concern is that it makes virtio_net_hdr into uapi?

Actually not. Note that, vnet header is already a part of uAPI 
(include/uapi/linux/virtio_net.h).

The concern is that is may confuse the userspace since we will have two 
sets of vnet header uapi (and they are already out of sync).

That's not the case. __sk_buff is uapi, but pointers to sockets
and other kernel data structures are not.
Yes. It's a bit weird that uapi struct has a pointer to kernel internal,
but I don't see it as a deal breaker.

Yes, but looking at the existing fields of __sk_buff. All are pretty 
generic fields that are device agnostic. This patch breaks this.

Tracing progs have plenty of such cases.
In networking there is tcp-bpf where everything is kernel internal and non-uapi.
So after this patch virtio_net_hdr is free to change without worrying about bpf
progs reading it.

So I wonder why not simply use helpers to access the vnet header like 
how tcp-bpf access the tcp header?

Thanks