Re: [PATCH v2 bpf 1/2] xdp: update xdp_rxq_info's mem type in XDP generic hook

[Jakub Kicinski <kuba@kernel.org>]

On Wed, 22 Oct 2025 13:22:25 +0200 Maciej Fijalkowski wrote:
> > > veth's pp works on order-0 pages well, however I agree it would be better
> > > to use virt_to_head_page() here.  
> > 
> > In this case the mem.type update is for consuming frags only, right?
> > We can't free the head itself since the skb is attached to it.
> > So running the predicates on xdp->data is probably wrong.  
> 
> See the veth patch where we bump refcount of related pages in order to
> keep the data as skb is consumed.

That one is wrong too. In veth's case since we care about refs on
individual frags you _should_ actually use skb->pp_recycle for the
condition. And you need to test whether each individual fragment
is from the PP. Basically call skb_pp_frag_ref() like
skb_try_coalesce() does?

BTW it is _not_ legal to call get_page() on slab pages any more
so all of this code should probably fail when head_frag=0.
But let's leave that problem to someone else.

> > Is it possible to get to bpf_prog_run_generic_xdp() (with frags)
> > and without going thru netif_skb_check_for_xdp() ? If no then
> > frags must have come from skb_pp_cow(). 
> > And the type is always MEM_TYPE_PAGE_POOL ?  
> 
> We have a fallback path netif_skb_check_for_xdp() that linearizes skb, in
> case COW code failed. But bare in mind that bpf_prog_run_generic_xdp() has
> other callsites, besides generic XDP itself (cpumap and devmap). That is
> why I wouldn't like to base this helper on assumptions such as frags
> presence -> mem_type is pp.

I assumed cpumap and devmap must be hit from a generic XDP handler.
Ergo we also much have gone thru the COW.