Re: [RFC PATCH bpf-next v1 0/7] xdp: RX checksum metadata hint and checksum assertion over redirect

[Jesper Dangaard Brouer <hawk@kernel.org>]

On 02/07/2026 16.52, Lorenzo Bianconi wrote:
[...]
>>
>> What's left that is genuinely separate is the "assertion" half -- a non-dev-bound
>> bpf_xdp_assert_rx_csum() that preserves the HW verdict across a
>> cpumap/redirect: it sets a flag on the xdp_buff that rides into the
>> xdp_frame and becomes skb->ip_summed = CHECKSUM_UNNECESSARY in
>> __xdp_build_skb_from_frame().
>>
>> Should I resend that as a small standalone series (v2, assert-only)?
>> It also looks like a PoC that you and Jakub discussed on v3 [1].
> 
> I think we should address the 'CHECKSUM_COMPLETE' use case for it, let's work
> on the problem later, even Jesper is interested in it ;)
>

Yes, I'm very interested in this use-case of transferring HW offload
info when XDP_REDIRECT'ing into CPUMAP or veth.  In my patchset[1] I've
added rx_hash, rx_vlan_tag, and rx_timestamp, but not csum as it was not
available.  My short term use-case is: RX-hash as setting it earlier
allows the GRO engine to hash flows (see dev_gro_receive[2]), plus it
carries info on packet protocol type TCP/UDP IPv4/IPv6.

- [1] 
https://lore.kernel.org/all/175146824674.1421237.18351246421763677468.stgit@firesoul/#r
- [2] https://elixir.bootlin.com/linux/v7.1.2/source/net/core/gro.c#L477

The long term[3] vision behind the patchset is that xdp_frame becomes
the "mini-SKB" that carries enough offload info to avoid allocating SKBs
in the driver.  The ICE driver is very close to this vision. See [4]
where it calls xdp_build_skb_from_buff() and later fills out SKB fields
with HW offloads (ice_process_skb_fields).

What I'm asking is that we generalize this idea of transferring or
storing HW offloads, that are relevant to the SKB.  My patchset makes
info avail to both xdp_buff and xdp_frame by storing data[5] in the
packet top headroom (extending xdp_frame). And like your patch[6] use
xdp_buff_flags to tell with is active. You patch only needed a single
flag as you only handle CHECKSUM_UNNECESSARY, but Lorenzo's patch also
have CHECKSUM_COMPLETE, for which we need to store a csum.

I cannot remember if it was Kuba or Eric that suggested (at netdevconf)
that instead of using a fixed struct, that I should use the same scheme
as SKB extensions.  We could call it SKB core-extensions, to signal the
possibility to move existing SKB member into this area.

--Jesper

- [3] 
https://lore.kernel.org/all/fbb026f9-54cf-49ba-b0dc-0df0f54c6961@kernel.org/
- [4] 
https://elixir.bootlin.com/linux/v7.1.2/source/drivers/net/ethernet/intel/ice/ice_txrx.c#L1044-L1060
- [5] 
https://lore.kernel.org/all/175146829944.1421237.13943404585579626611.stgit@firesoul/
- [6] https://lore.kernel.org/all/20260630191510.81402-1-deliran@verdict.gg/