On 6/19/25 6:02 PM, Akihiko Odaki wrote:

On 2025/06/18 1:12, Paolo Abeni wrote:

quoted

@@ -2426,7 +2453,17 @@ static int tun_xdp_one(struct tun_struct *tun,
  	if (metasize > 0)
  		skb_metadata_set(skb, metasize);
  
-	if (tun_vnet_hdr_to_skb(tun->flags, skb, gso)) {
+	/*
+	 * Assume tunnel offloads are enabled if the received hdr is large
+	 * enough.
+	 */
+	if (READ_ONCE(tun->vnet_hdr_sz) >= TUN_VNET_TNL_SIZE &&
+	    xdp->data - xdp->data_hard_start >= TUN_VNET_TNL_SIZE)
+		features = NETIF_F_GSO_UDP_TUNNEL |
+			   NETIF_F_GSO_UDP_TUNNEL_CSUM;

xdp->data - xdp->data_hard_start may not represent the header size.

struct tun_xdp_hdr is filled in vhost_net_build_xdp() in 
drivers/vhost/net.c. This function sets the two fields with 
xdp_prepare_buff(), but the arguments passed to xdp_prepare_buff() does 
not seem to represent the exact size of the header.

Indeed the xdp->data - xdp->data_hard_start range additionally contains
some padding and eventually the xdp specific headroom. The problem is
that both info are vhost_net specific and tun can't (or at least shoul)
not be aware of them.

The only IMHO feasible refinement could be using xdp->data_meta instead
of xdp->data when available. Alternatively I could drop entirely the test.

/P