nfp: fix esp-tx-csum-offload doesn't take effect

[ Upstream commit 1cf78d4 ]

When esp-tx-csum-offload is set to on, the protocol stack shouldn't
calculate the IPsec offload packet's csum, but it does. Because the
callback `.ndo_features_check` incorrectly masked NETIF_F_CSUM_MASK bit.

Fixes: 57f273a ("nfp: add framework to support ipsec offloading")
Signed-off-by: Huanhuan Wang <huanhuan.wang@corigine.com>
Signed-off-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>

drivers/net/ethernet/netronome/nfp/nfp_net_common.c

@@ -38,6 +38,7 @@
 #include <net/tls.h>
 #include <net/vxlan.h>
 #include <net/xdp_sock_drv.h>
+#include <net/xfrm.h>
 
 #include "nfpcore/nfp_dev.h"
 #include "nfpcore/nfp_nsp.h"
@@ -1897,6 +1898,9 @@ nfp_net_features_check(struct sk_buff *skb, struct net_device *dev,
 			features &= ~NETIF_F_GSO_MASK;
 	}
 
+	if (xfrm_offload(skb))
+		return features;
+
 	/* VXLAN/GRE check */
 	switch (vlan_get_protocol(skb)) {
 	case htons(ETH_P_IP):