vm networking: add flag vnet_hdr

When segmentation offload is enabled, and unsegmented packets are sent
to a VM (i.e. when running a container in the root netns), the kernel
will detect that packets are larger than expected and proceed.

That's not the case for containers (i.e. when running a container with
its own netns, and a veth pair). In that case, packets reach the
virtio-net interface, are forwarded to the bridge, and then to the
appropriate veth.

Unsegmented packets with GSO fields unset are dropped by the kernel
either at the bridge or at the veth level. That may be due to the
correct network topology where the vnet interface is attached to a
bridge.

In that case, we need to tell libkrun that the network backend sends /
receives virtio_net_hdr structs with the packets, and the backend need
to preserve GSO fields for VM-to-VM connections, or populate them for
host-to-VM connections.

Signed-off-by: Albin Kerouanton <albin.kerouanton@docker.com>

Author: akerouanton

docs/vm-networking.md

@@ -49,6 +49,8 @@ that take the following fields:
   VFKIT magic sequence after connecting to the `socket`. Accept any of `1, t, T,
   TRUE, true, True, 0, f, F, FALSE, false, False`. Any other value is invalid and
   will produce an error.
+- `vnet_hdr` (optional, defaults to false): Indicate whether the VMM includes
+  virtio-net headers along with Ethernet frames.
 
 Note that the first network specified will be used as the default gateway.
 

internal/shim/task/networking_unix.go

@@ -30,6 +30,11 @@ import (
 	"github.com/containerd/nerdbox/internal/vm"
 )
 
+const (
+	NET_FLAG_VFKIT = 1 << iota // See https://github.com/containers/libkrun/blob/357ec63fee444b973e4fc76d2121fd41631f121e/include/libkrun.h#L271C9-L271C23
+	NET_FLAG_INCLUDE_VNET_HEADER
+)
+
 type networksProvider struct {
 	nws []network
 }
@@ -43,6 +48,7 @@ type network struct {
 	addr6    netip.Prefix     // addr6 is the IPv6 address + subnet mask of the network interface
 	features uint32           // features is a bitmask of virtio-net features enabled on this network endpoint
 	vfkit    bool             // vfkit is a boolean flag indicating whether libkrun must send the VFKIT magic sequence after connecting to the socket.
+	vnetHdr  bool             // vnetHdr is a boolean flag indicating whether libkrun must include virtio-net headers along with Ethernet frames.
 }
 
 const (
@@ -57,6 +63,7 @@ const (
 	addrField     = "addr"
 	featuresField = "features" // features is a bitwise-OR separated list of virtio-net features. See https://docs.oasis-open.org/virtio/virtio/v1.3/csd01/virtio-v1.3-csd01.html#x1-2370003
 	vfkitField    = "vfkit"    // vfkit is a boolean flag indicating whether libkrun must send the VFKIT magic sequence after connecting to the socket.
+	vnetHdrField  = "vnet_hdr"
 
 	nwModeUnixgram   = "unixgram"
 	nwModeUnixstream = "unixstream"
@@ -149,6 +156,12 @@ func parseNetwork(annotation string) (network, error) {
 				return network{}, fmt.Errorf("parsing vfkit field: %w", err)
 			}
 			n.vfkit = vfkit
+		case vnetHdrField:
+			vnetHdr, err := strconv.ParseBool(value)
+			if err != nil {
+				return network{}, fmt.Errorf("parsing vnet_hdr field: %w", err)
+			}
+			n.vnetHdr = vnetHdr
 		default:
 			return network{}, fmt.Errorf("unknown network field: %s", key)
 		}
@@ -180,7 +193,10 @@ func (p *networksProvider) SetupVM(ctx context.Context, vmi vm.Instance) error {
 
 		var flags uint32
 		if nw.vfkit {
-			flags = 1 // See https://github.com/containers/libkrun/blob/357ec63fee444b973e4fc76d2121fd41631f121e/include/libkrun.h#L271C9-L271C23
+			flags = NET_FLAG_VFKIT
+		}
+		if nw.vnetHdr {
+			flags |= NET_FLAG_INCLUDE_VNET_HEADER
 		}
 
 		if err := vmi.AddNIC(ctx, nw.endpoint, nw.mac, nwMode, nw.features, flags); err != nil {