On 6/11/25 03:37, Jakub Kicinski wrote:
On Tue, 10 Jun 2025 22:18:32 +0700 Bui Quang Minh wrote:
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.
Okay, I meant the copies in the driver - virtio calls
xdp_linearize_page() in a few places, for normal XDP.
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?
The "metal" in the name refers to the AWS instance type that hosts
the runner. The test runs in a VM over virtio, more details:
https://github.com/linux-netdev/nipa/wiki/Running-driver-tests-on-virtio
I've figured out the problem. When the test fails, in mergeable_xdp_get_buf
xdp_room = SKB_DATA_ALIGN(XDP_PACKET_HEADROOM +
sizeof(struct skb_shared_info));
if (*len + xdp_room > PAGE_SIZE)
return NULL;
*len + xdp_room > PAGE_SIZE and NULL is returned, so the packet is
dropped. This case happens when add_recvbuf_mergeable is called when XDP
program is not loaded, so it does not reserve space for
XDP_PACKET_HEADROOM and struct skb_shared_info. But when the vhost uses
that buffer and send back to virtio-net, XDP program is loaded. The code
has the assumption that XDP frag cannot exceed PAGE_SIZE which I think
is not correct anymore. Due to that assumption, when the frame data +
XDP_PACKET_HEADROOM + sizeof(struct skb_shared_info) > PAGE_SIZE, the
code does not build xdp_buff but drops the frame. xdp_linearize_page has
the same assumption. As I don't think the assumption is correct anymore,
the fix might be allocating a big enough buffer to build xdp_buff.
Thanks,
Quang Minh.
