On Thu, 30 Jun 2022 13:53:11 +0200 Maciej Fijalkowski wrote:

quoted

quoted

Would it be possible to move the marking to when the queue is getting
bound instead of the recv/send paths?   

I mean when socket is getting bound.  

So Bjorn said that it was the design choice to follow the standard
sockets' approach. I'm including a dirty diff for a discussion which
allows me to get napi_id at bind() time. But, this works for ice as this
driver during the XDP prog/XSK pool load only disables the NAPI, so we are
sure that napi_id stays the same. That might not be the case for other
AF_XDP ZC enabled drivers though, they might delete the NAPI and this
approach wouldn't work...or am I missing something?

Possible, but IDK if we're changing anything substantially in that
regard. The existing code already uses sk_mark_napi_id_once() so
if the NAPI ID changes during the lifetime of the socket the user
will be out of luck already. But no strong feelings.

I'd prefer the diff below though as it simplifies the data path, but I
can't say if it's safe to do so. We would have to be sure about drivers
keeping their NAPI struct. This would also allow us to drop napi_id from
xdp_rxq_info.