Before commit 00d066a4d4ed ("netdev_features: convert NETIF_F_LLTX to
dev->lltx"), NETIF_F_LLTX was set unconditionally in both
__gre_tunnel_init() and ip6gre_tnl_init_features() alongside
GRE_FEATURES:

    dev->features |= GRE_FEATURES | NETIF_F_LLTX;

When that commit converted NETIF_F_LLTX to the dev->lltx flag, it
placed 'dev->lltx = true' after the SEQ/CSUM early returns instead
of before them. This causes GRE/GRETAP/ip6gre tunnels with SEQ or
CSUM+encap to lose lockless TX, reintroducing _xmit_lock acquisition
around their ndo_start_xmit. Since GRE xmit re-enters the stack via
ip_tunnel_xmit(), holding _xmit_lock risks ABBA deadlock with the
underlay device.

  CPU0                        CPU1
  ----                        ----
  lock(&qdisc_xmit_lock_key#6);
                              lock(&qdisc_xmit_lock_key#3);
                              lock(&qdisc_xmit_lock_key#6);
  lock(&qdisc_xmit_lock_key#3);

Fix by moving dev->lltx = true before the early returns in both
functions, restoring the original unconditional behavior.

Fixes: 00d066a4d4ed ("netdev_features: convert NETIF_F_LLTX to dev->lltx")
Signed-off-by: Yun Zhou <yun.zhou@windriver.com>
---
v2:
  - also fix the same issue for ipv6

 net/ipv4/ip_gre.c  | 4 ++--
 net/ipv6/ip6_gre.c | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/net/ipv4/ip_gre.c b/net/ipv4/ip_gre.c
index 9fbff16cda1d..8c5ad8ec8d09 100644
--- a/net/ipv4/ip_gre.c
+++ b/net/ipv4/ip_gre.c
@@ -1018,6 +1018,8 @@ static void __gre_tunnel_init(struct net_device *dev)
 	dev->features		|= GRE_FEATURES;
 	dev->hw_features	|= GRE_FEATURES;
 
+	dev->lltx = true;
+
 	/* TCP offload with GRE SEQ is not supported, nor can we support 2
 	 * levels of outer headers requiring an update.
 	 */
@@ -1029,8 +1031,6 @@ static void __gre_tunnel_init(struct net_device *dev)
 
 	dev->features |= NETIF_F_GSO_SOFTWARE;
 	dev->hw_features |= NETIF_F_GSO_SOFTWARE;
-
-	dev->lltx = true;
 }
 
 static int ipgre_tunnel_init(struct net_device *dev)
diff --git a/net/ipv6/ip6_gre.c b/net/ipv6/ip6_gre.c
index 7c09a269b352..b843116e9b70 100644
--- a/net/ipv6/ip6_gre.c
+++ b/net/ipv6/ip6_gre.c
@@ -1455,6 +1455,8 @@ static void ip6gre_tnl_init_features(struct net_device *dev)
 	dev->features		|= GRE6_FEATURES;
 	dev->hw_features	|= GRE6_FEATURES;
 
+	dev->lltx = true;
+
 	/* TCP offload with GRE SEQ is not supported, nor can we support 2
 	 * levels of outer headers requiring an update.
 	 */
@@ -1466,8 +1468,6 @@ static void ip6gre_tnl_init_features(struct net_device *dev)
 
 	dev->features |= NETIF_F_GSO_SOFTWARE;
 	dev->hw_features |= NETIF_F_GSO_SOFTWARE;
-
-	dev->lltx = true;
 }
 
 static int ip6gre_tunnel_init_common(struct net_device *dev)
-- 
2.43.0

On Mon, Jul 13, 2026 at 11:09:45PM +0800, Yun Zhou wrote:
> Before commit 00d066a4d4ed ("netdev_features: convert NETIF_F_LLTX to
> dev->lltx"), NETIF_F_LLTX was set unconditionally in both
> __gre_tunnel_init() and ip6gre_tnl_init_features() alongside
> GRE_FEATURES:
> 
>     dev->features |= GRE_FEATURES | NETIF_F_LLTX;
> 
> When that commit converted NETIF_F_LLTX to the dev->lltx flag, it
> placed 'dev->lltx = true' after the SEQ/CSUM early returns instead
> of before them. This causes GRE/GRETAP/ip6gre tunnels with SEQ or
> CSUM+encap to lose lockless TX, reintroducing _xmit_lock acquisition
> around their ndo_start_xmit. Since GRE xmit re-enters the stack via
> ip_tunnel_xmit(), holding _xmit_lock risks ABBA deadlock with the
> underlay device.
> 
>   CPU0                        CPU1
>   ----                        ----
>   lock(&qdisc_xmit_lock_key#6);
>                               lock(&qdisc_xmit_lock_key#3);
>                               lock(&qdisc_xmit_lock_key#6);
>   lock(&qdisc_xmit_lock_key#3);
> 
> Fix by moving dev->lltx = true before the early returns in both
> functions, restoring the original unconditional behavior.
> 
> Fixes: 00d066a4d4ed ("netdev_features: convert NETIF_F_LLTX to dev->lltx")
> Signed-off-by: Yun Zhou <yun.zhou@windriver.com>

Reviewed-by: Ido Schimmel <idosch@nvidia.com>