On Mon, 2025-01-27 at 10:04 -0800, Jakub Kicinski wrote:

On Fri, 24 Jan 2025 12:45:42 +0100 Toke Høiland-Jørgensen wrote:

quoted

quoted

quoted

I think there is no simple fix for that. That needs some discussion
around the "expectations" to the headroom / meta data area in front of
the actual packet data.  

By 'simple' you mean without some new UAPI to signal the size of that
'reserved area' by the driver? I don't see any other easy way out as well :-/  

Yeah, I don't think we can impose UAPI restrictions on the metadata area
at this point. I guess the best we can do is to educate users that they
should call the timestamp kfunc before they modify the metadata?

I may be misunderstanding the discussion, but I think the answer 
is that the driver must be fixed. The metadata-in-prepend problem
also exists for simple adjust head use case, so it existed since
early days of BPF. The driver should copy out (or parse) the metadata
before it invokes the XDP prog. The nfp driver does that.

That would have to happen for each packet, without affecting ZC
performance. How can that be achieved?

So we have at least two drivers with that problem, igc + nfp. 

My main point: Enabling and implementing ZC (zero copy) mode at one
hand, but then starting to copy the meta data for each packet doesn't
sound reasonable.