Re: [PATCH net] virtio-net: drop the multi-buffer XDP packet in zerocopy

[Bui Quang Minh]

On 6/9/25 23:58, Jakub Kicinski wrote:
> On Fri, 6 Jun 2025 22:48:53 +0700 Bui Quang Minh wrote:
> > > > But currently, if a multi-buffer packet arrives, it will not go through
> > > > XDP program so it doesn't increase the stats but still goes to network
> > > > stack. So I think it's not a correct behavior.
> > > Sounds fair, but at a glance the normal XDP path seems to be trying to
> > > linearize the frame. Can we not try to flatten the frame here?
> > > If it's simply to long for the chunk size that's a frame length error,
> > > right?
> > Here we are in the zerocopy path, so the buffers for the frame to fill
> > in are allocated from XDP socket's umem. And if the frame spans across
> > multiple buffers then the total frame size is larger than the chunk
> > size.
> Is that always the case? Can the multi-buf not be due to header-data
> split of the incoming frame? (I'm not familiar with the virtio spec)

Ah, maybe I cause some confusion :) zerocopy here means zerocopy if the 
frame is redirected to XDP socket. In this zerocopy mode, XDP socket 
will provide buffers to virtio-net, the frame from vhost will be placed 
in those buffers. If the bind XDP program in virtio-net returns 
XDP_REDIRECT to that XDP socket, then the frame is zerocopy. In case 
XDP_PASS is returned, the frame's data is copied to newly created skb 
and the frame's buffer is returned to XDP socket. AFAIK, virtio-net has 
not supported header-data split yet.

> > Furthermore, we are in the zerocopy so we cannot linearize by
> > allocating a large enough buffer to cover the whole frame then copy the
> > frame data to it. That's not zerocopy anymore. Also, XDP socket zerocopy
> > receive has assumption that the packet it receives must from the umem
> > pool. AFAIK, the generic XDP path is for copy mode only.
> Generic XDP == do_xdp_generic(), here I think you mean the normal XDP
> patch in the virtio driver? If so then no, XDP is very much not
> expected to copy each frame before processing.

Yes, I mean generic XDP = do_xdp_generic(). I mean that we can linearize 
the frame if needed (like in netif_skb_check_for_xdp()) in copy mode for 
XDP socket but not in zerocopy mode.

> This is only slightly related to you patch but while we talk about
> multi-buf - in the netdev CI the test which sends ping while XDP
> multi-buf program is attached is really flaky :(
> https://netdev.bots.linux.dev/contest.html?executor=vmksft-drv-hw&test=ping-py.ping-test-xdp-native-mb&ld-cases=1

metal-drv-hw means the NETIF is the real NIC, right?

Thanks,
Quang Minh.