Re: [PATCH net-next] xdp: add multi-buff support for xdp running in generic mode

[Lorenzo Bianconi <lorenzo@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 1629 bytes --]

> On Tue, 28 Nov 2023 23:27:29 +0100 Lorenzo Bianconi wrote:
> > > Yes, don't we allow writes to fragments in XDP based on the assumption
> > > that it runs on Rx so that paged data must not be zero copy?
> > > bpf_xdp_store_bytes() doesn't seem to have any checks which would
> > > stop it from writing fragments, as far as I can see.  
> > 
> > do you mean in the skb use-case we could write to fragments (without copying
> > them) if the skb is not cloned and the paged area is not 'zero-copied'?
> 
> The zero-copy thing is a red herring. If application uses
> sendpage/sendfile/splice the frag may be a page cache page
> of a file. Or something completely read only.

ack, thx for pointing this out. It is clear now :)

> 
> IIUC you're trying to avoid the copy if the prog is mbuf capable.
> So I was saying that can't work for forms of XDP which actually 
> deal with skbs. But that wasn't really your question, sorry :)
> 
> > With respect to this patch it would mean we can rely on pskb_expand_head() to
> > reallocate the skb and to covert it to a xdp_buff and we do not need to explicitly
> > reallocate fragments as we currently do for veth in veth_convert_skb_to_xdp_buff() [0].
> > Is my understanding correct or am I missing something?
> 
> The difference is that pskb_expand_head() will give you a linear skb,
> potentially triggering an order 5 allocation. Expensive and likely to
> fail under memory pressure.

ack

> 
> veth_convert_skb_to_xdp_buff() tries to allocate pages, and keep
> the skb fragmented.

I will rework the patch using this approach.

Regards,
Lorenzo

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]