Re: [PATCH v5 bpf 1/2] xdp: introduce xdp_convert_skb_to_buff()

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

On Thu, Oct 30, 2025 at 07:05:11PM -0700, Jakub Kicinski wrote:
> On Thu, 30 Oct 2025 21:22:34 +0100 Maciej Fijalkowski wrote:
> > > > Why do you say so?
> > > > 
> > > > netif_receive_generic_xdp()
> > > > 	netif_skb_check_for_xdp()
> > > > 	skb_cow_data_for_xdp() failed
> > > > 		go through skb linearize path
> > > > 			returned skb data is backed by kmalloc, not page_pool,
> > > > 			means mem type for this particular xdp_buff has to be
> > > > 			MEM_TYPE_PAGE_SHARED
> > > > 
> > > > Are we on the same page now?  
> > > 
> > > No, I think I already covered this, maybe you disagreed and I missed it.
> > > 
> > > The mem_type set here is expected to be used only for freeing pages. 
> > > XDP can only free fagments (when pkt is trimmed), it cannot free the
> > > head from under the skb. So only fragments matter here, we can ignore
> > > the head.  
> > 
> > ...and given that linearize path would make skb a frag-less one...okay -
> > I'm buying this! :D I have some other thoughts, but I would like to
> > finally close this pandora's box, you probably have similar feelings.
> > 
> > So plain assignment like:
> > xdp->rxq->mem.type = MEM_TYPE_PAGE_POOL;
> 
> Yes, LGTM.
> 
> > would be fine for you? Plus AI reviewer has kicked me in the nuts on veth
> > patch so have to send v6 anyways.
> 
> The veth side unfortunately needs more work than Mr Robot points out.
> For some reason veth tries to turn skb into an xdp_frame..

That is beyond the scope of the fix that I started doing as you're
undermining overall XDP support in veth, IMHO.

I can follow up on this on some undefined future but right now I will
have to switch to some other work.

If you disagree and insist on addressing skb->xdp_frame in veth within
this patchset then I'm sorry but I will have to postpone my activities
here.

> 
> Either we have to make it not do that - we could probably call
> xdp_do_generic_redirect() and for Tx .. figure out the right incantation 
> to give the frame back to the peer veth.
> 
> Or, if we didn't hit CoW, you need to actually add the incantation we
> removed here, there:
> 
> 	xdp->rxq->mem.type = skb->pp_recycle && page_pool_page_is_pp(..) ?
> 		MEM_TYPE_PAGE_POOL : MEM_TYPE_PAGE_SHARED;
> 
> Or CoW the head retroactively if we hit the Tx/Redir path.
> 
> My intuition is that first option (making the handling as similar to
> XDP generic as possible) is going to be least buggy long term.