[xdp-hints] Re: Questions about XDP hints

[Stanislav Fomichev <stfomichev@gmail.com>]

On 10/12, tianmuyang wrote:
> Hi all:
> 	There has been some discussions about adding checksum hint in AF_XDP such as this thread[1]. Now, we also plan to add checksum hint. My questions are:
> 	1. In this msg[2], is it appropriate if xdp_csum_status only includes 4 enums/macros(CHECKSUM_NONE...CHECKSUM_PARTIAL in skbuff.h)? Thus it becomes more generic. Also, in this msg[3] we can simply pass skb->ip_summed to csum_status in veth_xdp_rx_csum().
> 	2. What should be taken care of if I want to add a new hint? IOW, what is acceptable to add a new hint?

There is no clear guidance on what's acceptable and what's not. Each
hint it evaluated case by case. IIRC, last time rx csum discussion
stalled due to disagreement about the level of details which should
be exposed from the generic kfuncs. Feel free to revive the discussion
with another patchset.

Regarding (1): the consensus seems to be (IIRC) is to expose tree cases
only: no-csum, csum-unnecessary, csum-complete+csum. Anything else
gets too device specific and too convoluted to handle on the xdp prog
side.


> Thanks!
> 
> [1] https://lore.kernel.org/bpf/CAADnVQJPgpo7J0qVTQJYYocZ=Jnw=O5GfN2=PyAQ55+WWG_DVg@mail.gmail.com/
> [2] https://lore.kernel.org/bpf/20230728173923.1318596-13-larysa.zaremba@intel.com/
> [3] https://lore.kernel.org/bpf/20230728173923.1318596-18-larysa.zaremba@intel.com/
>