Re: [PATCH bpf 2/2] veth: update mem type in xdp_buff

[Maciej Fijalkowski <maciej.fijalkowski@intel.com>]

On Tue, Oct 07, 2025 at 06:11:53PM -0700, Jakub Kicinski wrote:
> On Tue, 7 Oct 2025 16:59:21 +0200 Maciej Fijalkowski wrote:
> > > My thinking was that we should try to bake the rxq into "conversion"
> > > APIs, draft diff below, very much unfinished and I'm probably missing
> > > some cases but hopefully gets the point across:  
> > 
> > That is not related IMHO. The bugs being fixed have existing rxqs. It's
> > just the mem type that needs to be correctly set per packet.
> > 
> > Plus we do *not* convert frame to buff here which was your initial (on
> > point) comment WRT onstack rxqs. Traffic comes as skbs from peer's
> > ndo_start_xmit(). What you're referring to is when source is xdp_frame (in
> > veth case this is when ndo_xdp_xmit or XDP_TX is used).
> 
> I guess we're slipping into a philosophical discussion but I'd say 
> that the problem is that rxq stores part of what is de facto xdp buff
> state. It is evacuated into the xdp frame when frame is constructed,
> as packet is detached from driver context. We need to reconstitute it
> when we convert frame (skb, or anything else) back info an xdp buff.

So let us have mem type per xdp_buff then. Feels clunky anyways to change
it on whole rxq on xdp_buff basis. Maybe then everyone will be happy?

> 
> xdp_convert_buff_to_frame() and xdp_convert_frame_to_buff() should be
> a mirror image of each other, to put it more concisely.
> 
> > However the problem pointed out by AI (!) is something we should fix as
> > for XDP_{TX,REDIRECT} xdp_rxq_info is overwritten and mem type update is
> > lost.
> 
> > > +/* Initialize an xdp_buff from an skb.
> > > + *
> > > + * Note: if skb has frags skb_cow_data_for_xdp() must be called first,
> > > + * or caller must otherwise guarantee that the frags come from a page pool
> > > + */
> > > +static inline
> > > +void xdp_convert_skb_to_buff(const struct xdp_frame *frame,
> > > +			     struct xdp_buff *xdp, struct xdp_rxq_info *rxq)  
> > 
> > I would expect to get skb as an input here
> 
> Joł. Don't nit pick my draft diff :D It's not meant as a working patch.

Polish sneaked in so this got really under your skin :D

>