On Mon, 23 Feb 2026 18:11:54 +0100 Lorenzo Bianconi wrote:
> > Off the top of my head drivers prefer reporting UNNECESSARY when they
> > have both, and reserve COMPLETE for cases where L4 could not be found
> > or is incorrect. Why don't we report both? We're using 3 args, we still
> > have 3 to go. We could turn ip_summed into a bitmap and have explicit
> > output args for both level and csum complete value?  
> 
> Ack, thx for the explanation. Just for sake of understanding, is there
> any NIC capable of reporting both csum_value and csum for the same packet
> in the DMA descriptor? Or is this change needed to be future-proof?

Both nfp and fbnic definitely can. Off the top of my head - mlx5 also
can, but I haven't double checked.

> > One more thing I'd like us to at least have a plan for at this stage
> > is how to deal with COMPLETE + modified packet + XDP_PASS.
> > Right now some drivers discard COMPLETE when XDP is attached since
> > they can't be sure if XDP modifies the packet. Other drivers don't
> > and we end up with bad csum splat. Do we have a recommendation on
> > the correct behavior? If not - should we have a kfunc to adjust /
> > discard csum complete explicitly?  
> 
> At the moment there is no way to store the csum value we got running
> bpf_xdp_metadata_rx_checksum() in order to be consumed during
> xdp_buff/xdp_frame to skb conversion (this info can just be consumed in the
> ebpf program bound to the NIC) but

I think the scope here is much narrower than the xdp_buf to xdp_frame
to skb conversion. We are just pass information between the program and
driver which owns xdp_buff. Very similar to your new xmo.

We could either tell the driver to discard the csum complete or even
add a helper to "adjust" the the csum value. Similar to the helper
we have to adjust the csum in TC / skb context.

> I guess the issue you pointed out can be solved in the verifier
> during program load time. What do you think?

It could, but at the verifier level we'd probably have to be fairly
coarse-grained. Any write to the packet data would mean csum complete
cannot be trusted, that's not too hard. But also any tail call / fentry?
I'm not really up to date on the latest in program chaining in BPF but
I think a lot of real-life deployments would use either chaining or
fentry. So in practice it may be a lot of complexity for having csum
complete always disabled w/ XDP, in practice.

Up to you. I'm totally okay to just say** that drivers should never
report csum complete with XDP (until appropriate API is built).
Perhaps this will force those who care about XDP+csum_complete to
tell us what their requirements are?

[**] "just say" == document and add driver kselftest that validates it

Reply via email to