Some devices supporting USO have only ipv4 checksum enabled,
but USO gets disabled because both ipv4 and ipv6 functionality is checked.

IPv4 or IPv6 or both checksums, can be utilized with UDP segmentation
offload. Separate checks for both IPv4 and IPv6 checksums.

Fixes: 2b2bc3bab158 ("net: Make USO depend on CSUM offload")
Signed-off-by: Radharapu Rakesh <rakesh.radharapu@amd.com>
---
 net/core/dev.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 1cb134ff7327..a22f8f6e2ed1 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -10465,11 +10465,13 @@ static void netdev_sync_lower_features(struct net_device *upper,
 
 static bool netdev_has_ip_or_hw_csum(netdev_features_t features)
 {
-	netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
-	bool ip_csum = (features & ip_csum_mask) == ip_csum_mask;
+	netdev_features_t ipv4_csum_mask = NETIF_F_IP_CSUM;
+	netdev_features_t ipv6_csum_mask = NETIF_F_IPV6_CSUM;
+	bool ipv4_csum = (features & ipv4_csum_mask) == ipv4_csum_mask;
+	bool ipv6_csum = (features & ipv6_csum_mask) == ipv6_csum_mask;
 	bool hw_csum = features & NETIF_F_HW_CSUM;
 
-	return ip_csum || hw_csum;
+	return ipv4_csum || ipv6_csum || hw_csum;
 }
 
 static netdev_features_t netdev_fix_features(struct net_device *dev,
-- 
2.34.1

On 3/12/25 12:54 PM, Radharapu Rakesh wrote:
>  net/core/dev.c | 8 +++++---
>  1 file changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 1cb134ff7327..a22f8f6e2ed1 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -10465,11 +10465,13 @@ static void netdev_sync_lower_features(struct net_device *upper,
>  
>  static bool netdev_has_ip_or_hw_csum(netdev_features_t features)
>  {
> -	netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
> -	bool ip_csum = (features & ip_csum_mask) == ip_csum_mask;
> +	netdev_features_t ipv4_csum_mask = NETIF_F_IP_CSUM;
> +	netdev_features_t ipv6_csum_mask = NETIF_F_IPV6_CSUM;
> +	bool ipv4_csum = (features & ipv4_csum_mask) == ipv4_csum_mask;
> +	bool ipv6_csum = (features & ipv6_csum_mask) == ipv6_csum_mask;
>  	bool hw_csum = features & NETIF_F_HW_CSUM;
>  
> -	return ip_csum || hw_csum;
> +	return ipv4_csum || ipv6_csum || hw_csum;
>  }

The above will additionally affect TLS offload, and will likely break
i.e. USO over IPv6 traffic landing on devices supporting only USO over
IPv4, unless such devices additionally implement a suitable
ndo_features_check().

Such situation will be quite bug prone, I'm unsure we want this kind of
change - even without looking at the TLS side of it.

/P