Re: [PATCH 2/4] virtio-net: add support of UDP segmentation (USO) on the host

On Wed, May 12, 2021 at 2:56 PM Yuri Benditovich
[off-list ref] wrote:

On Wed, May 12, 2021 at 5:33 PM Willem de Bruijn
[off-list ref] wrote:

quoted

On Wed, May 12, 2021 at 2:10 AM Yuri Benditovich
[off-list ref] wrote:

quoted

On Tue, May 11, 2021 at 8:48 PM Willem de Bruijn
[off-list ref] wrote:

quoted

On Tue, May 11, 2021 at 12:43 AM Yuri Benditovich
[off-list ref] wrote:

quoted

Large UDP packet provided by the guest with GSO type set to
VIRTIO_NET_HDR_GSO_UDP_L4 will be divided to several UDP
packets according to the gso_size field.

Signed-off-by: Yuri Benditovich <redacted>
---
 include/linux/virtio_net.h | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/include/linux/virtio_net.h b/include/linux/virtio_net.h
index b465f8f3e554..4ecf9a1ca912 100644
--- a/include/linux/virtio_net.h
+++ b/include/linux/virtio_net.h

@@ -51,6 +51,11 @@ static inline int virtio_net_hdr_to_skb(struct sk_buff *skb,
                        ip_proto = IPPROTO_UDP;
                        thlen = sizeof(struct udphdr);
                        break;
+               case VIRTIO_NET_HDR_GSO_UDP_L4:
+                       gso_type = SKB_GSO_UDP_L4;
+                       ip_proto = IPPROTO_UDP;
+                       thlen = sizeof(struct udphdr);
+                       break;

If adding a new VIRTIO_NET_HDR type I suggest adding separate IPv4 and
IPv6 variants, analogous to VIRTIO_NET_HDR_GSO_TCPV[46]. To avoid
having to infer protocol again, as for UDP fragmentation offload (the
retry case below this code).

Thank you for denoting this important point of distinguishing between v4 and v6.
Let's try to take a deeper look to see what is the correct thing to do
and please correct me if I'm wrong:
1. For USO we do not need to guess the protocol as it is used with
VIRTIO_NET_HDR_F_NEEDS_CSUM (unlike UFO)

Enforcing that is a good start. We should also enforce that
skb->protocol is initialized to one of htons(ETH_P_IP) or
htons(ETH_P_IPV6), so that it does not have to be inferred by parsing.

As this feature is new and is not used in any public release of any
misbehaving driver, probably it is enough to state in the spec that
VIRTIO_NET_HDR_F_NEEDS_CSUM is required for USO packets.
The spec states that the USO feature requires checksumming feature.

The spec is not sufficient. These rules need to be enforced in the
kernel code, too.

quoted

These requirements were not enforced for previous values, and cannot
be introduced afterwards, which has led to have to add that extra code
to handle these obscure edge cases.

I agree that with well behaved configurations, the need for separate
_V4 and _V6 variants is not needed.

quoted

and the USO packets
transmitted by the guest are under the same clause as both
VIRTIO_NET_HDR_GSO_TCP, i.e. under if (hdr->flags &
VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2. If we even define VIRTIO_NET_HDR_GSO_UDPv4_L4 and
VIRTIO_NET_HDR_GSO_UDPv6_L4 - both will be translated to
SKB_GSO_UDP_L4, so this information is immediately lost (the code will
look like:
case VIRTIO_NET_HDR_GSO_UDP4_L4: case VIRTIO_NET_HDR_GSO_UDP6_L4
    gso_type = SKB_GSO_UDP;

3. When we will define the respective guest features (like
VIRTIO_NET_F_HOST_USO4 VIRTIO_NET_F_HOST_USO6) we will need to

This is my typo: VIRTIO_NET_F_GUEST_USO4...

quoted

quoted

recreate the virtio_net header from the skb when both v4 and v6 have
the same SKB_GSO_UDP_L4, (see virtio_net_hdr_from_skb) and I'm not
sure whether somebody needs the exact v4 or v6 information on guest RX
path.

FWIW, it is good to keep in mind that virtio_net_hdr is also used
outside virtio, in both ingress and egress paths.

Can you please elaborate in which scenarios we do not have any virtio
device in path but need virtio_net_hdr?

Packet sockets, tuntap.

quoted

quoted

4. What is completely correct is that when we will start working with
the guest RX path we will need to define something like NETIF_F_USO4
and NETIF_F_USO6 and configure them according to exact guest offload
capabilities.
Do you agree?

I don't immediately see the need for advertising this device feature
on a per-protocol basis. Can you elaborate?

Separate offload setting (controlled by the guest) for v4 and v6 in
guest RX path is mandatory, at least Windows always requires this for
any offload.
In this case it seems easy to have also virtio-net device features to
be indicated separately (the TAP/TUN should report its capabilities).

Ah, ok.