On Fri, 27 Feb 2026 14:21:44 +0100 Lorenzo Bianconi wrote: > > > 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. > > IIUC, for the CSUM_COMPLETE case, we want to add a kfunc used to update (or > invalidate) the checksum value (if the packet has been modified by the eBPF > program bounded to the NIC) and report the updated checksum to the driver if > the XDP verdict is XDP_PASS. Correct? > > I guess we could have two approaches here: > - Write the new checksum value into the xdp_metadata area (if available) > where the driver can load it and update the checksum value before > allocating the skb. > The main downside of this approach is we need modify each driver. > - Add a new xmo callback used to set the checksum value and report it > from the eBPF program into a specific memory area provided by the driver > (e.g. DMA descriptor) that is used to build the skb. > > What do you think?
Exactly. The invalidation is easier 'cause using a single bit in the flags should be uncontroversial. If we want to be able to repair / provide the csum complete then we have to pick one of the two options you outlined. As you may suspect from previous discussions I favor the latter. But we'd probably have to have a PoC with either one and see where the consensus falls. Actually, thinking about it more, I guess this is not just a CSUM_COMPLETE issue. XDP_PASS will also risk reporting invalid UNNECESSARY to the stack (e.g. when XDP stripped a UDP tunnel which which the NIC compute the UNNECESSARY but the packet inside the tunnel has an invalid csum). > Moreover, since we already have this issue upstream, do you think > this new feature must be part this series or can we do it with a > follow-up patch/series? We don't have to add the kfunc to adjust / invalidate the csum. But we should document how the drivers are expected to behave until such kfunc exists and we should add a selftest that checks the documented expectation.
