[PATCH net-next v3 0/5] net: centralize master device offload feature computation

[PATCH net-next v3 0/5] net: centralize master device offload feature computation

[Hangbin Liu]

Currently, master devices (bonding, bridge, team) manually call
netdev_compute_master_upper_features() scattered throughout their port
add/remove operations. This approach requires each driver to remember
to update features at the right times and leads to code duplication.

The series adds a new ndo_update_offloads callback that is automatically
invoked during feature updates when upper/lower device relationships change.
This centralizes the feature computation flow and removes the burden
from individual drivers.

---
Changes in v3:
- ndo_set_features may be skipped when no feature change, fallback to
  use ndo_update_offloads for offload computation (Simon Horman)
- Link to v2: https://lore.kernel.org/r/20260313-offload_compute-v2-0-ffbc8ce5d50c@gmail.com

Changes in v2:
- Fix macsec Security Entity uninitialized issue (Sabrina Dubroca)
  - Tested with macsec-offload.sh, rtnetlink.sh, link_netns.py, all pssed
- Link to v1: https://lore.kernel.org/r/20260310-offload_compute-v1-0-3df79c09ea65@gmail.com

---
Hangbin Liu (5):
      net: add ndo_update_offloads for offload computation
      net: use ndo_update_offloads to set offload features for bonding/bridge/team
      macsec: move netdev_upper_dev_link() after macsec_changelink_common()
      failover: use ndo_update_offloads for failover offload compute
      net: no need to disable LRO specifically

 drivers/net/bonding/bond_main.c | 13 ++++----
 drivers/net/macsec.c            | 23 +++++++-------
 drivers/net/net_failover.c      | 66 +++++------------------------------------
 drivers/net/team/team_core.c    | 14 ++++-----
 include/linux/netdevice.h       |  7 +++++
 include/net/net_failover.h      |  7 -----
 net/8021q/vlan.c                |  2 --
 net/bridge/br_device.c          |  6 ++++
 net/bridge/br_if.c              |  6 ----
 net/core/dev.c                  | 11 +++++--
 net/hsr/hsr_slave.c             |  1 -
 11 files changed, 55 insertions(+), 101 deletions(-)
---
base-commit: 5446b8691eb8278f10deca92048fad84ffd1e4d5
change-id: 20260310-offload_compute-4c0bafa2e022

Best regards,
-- 
Hangbin Liu <liuhangbin@gmail.com>

[PATCH net-next v3 1/5] net: add ndo_update_offloads for offload computation

[Hangbin Liu]

Add a new ndo_update_offloads callback to net_device_ops that allows
devices to compute and update their offload features during feature
updates.

This callback enables master devices to recompute their features
based on current slave device configuration. This is particularly
useful for bonding, bridging, team, and failover devices that need
to aggregate features from their lower devices.

The callback is optional and only implemented by devices that need
dynamic offload feature computation.

Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
---
 include/linux/netdevice.h | 7 +++++++
 net/core/dev.c            | 3 +++
 2 files changed, 10 insertions(+)

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index ae269a2e7f4d..acaec0340266 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -1281,6 +1281,12 @@ struct netdev_net_notifier {
  *	constraints, and returns the resulting flags. Must not modify
  *	the device state.
  *
+ * void (*ndo_update_offloads)(struct net_device *dev);
+ *	Called during feature update to allow device to compute and update
+ *	offload features (like vlan_features, hw_enc_features) based on
+ *	current lower device configuration. Typically used by master
+ *	devices to aggregate features from slave devices.
+ *
  * int (*ndo_set_features)(struct net_device *dev, netdev_features_t features);
  *	Called to update device configuration to new features. Passed
  *	feature set might be less than what was returned by ndo_fix_features()).
@@ -1558,6 +1564,7 @@ struct net_device_ops {
 							struct sock *sk);
 	netdev_features_t	(*ndo_fix_features)(struct net_device *dev,
 						    netdev_features_t features);
+	void			(*ndo_update_offloads)(struct net_device *dev);
 	int			(*ndo_set_features)(struct net_device *dev,
 						    netdev_features_t features);
 	int			(*ndo_neigh_construct)(struct net_device *dev,
diff --git a/net/core/dev.c b/net/core/dev.c
index 200d44883fc1..bbd532aa6a1b 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -11011,6 +11011,9 @@ int __netdev_update_features(struct net_device *dev)
 	ASSERT_RTNL();
 	netdev_ops_assert_locked(dev);
 
+	if (dev->netdev_ops->ndo_update_offloads)
+		dev->netdev_ops->ndo_update_offloads(dev);
+
 	features = netdev_get_wanted_features(dev);
 
 	if (dev->netdev_ops->ndo_fix_features)

-- 
Git-155)

[PATCH net-next v3 2/5] net: use ndo_update_offloads to set offload features for bonding/bridge/team

[Hangbin Liu]

Convert bonding, bridge, and team drivers to use the new
ndo_update_offloads callback instead of manually calling
netdev_compute_master_upper_features() during port add/remove operations.

This change centralizes the feature computation flow:

Before:
- netdev_master_upper_dev_link()
- netdev_compute_master_upper_features()  <- called manually
  - compute offload features
  - netdev_change_features()
    - __netdev_update_features()
      - update other features

After:
- netdev_master_upper_dev_link()
  - __netdev_upper_dev_link()
    - netdev_change_features()
      - __netdev_update_features()
        - ndo_update_offloads()
          - netdev_compute_master_upper_features()
        - update other features

This ensures offload features are computed automatically when
upper/lower device links change, removing the need for manual
feature computation calls in the driver code.

The netdev_change_features() call in team_uninit() is also removed
since it calls team_port_del() for each port, which now triggers
feature updates automatically.

Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
---
 drivers/net/bonding/bond_main.c | 10 +++++++---
 drivers/net/team/team_core.c    | 11 +++++++----
 net/bridge/br_device.c          |  6 ++++++
 net/bridge/br_if.c              |  4 ----
 net/core/dev.c                  |  8 ++++++--
 5 files changed, 26 insertions(+), 13 deletions(-)

diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
index 90b71bd53040..ea6bb20f9348 100644
--- a/drivers/net/bonding/bond_main.c
+++ b/drivers/net/bonding/bond_main.c
@@ -1509,6 +1509,11 @@ static netdev_features_t bond_fix_features(struct net_device *dev,
 	return features;
 }
 
+static void bond_update_offloads(struct net_device *dev)
+{
+	netdev_compute_master_upper_features(dev, true);
+}
+
 static int bond_header_create(struct sk_buff *skb, struct net_device *bond_dev,
 			      unsigned short type, const void *daddr,
 			      const void *saddr, unsigned int len)
@@ -2273,7 +2278,6 @@ int bond_enslave(struct net_device *bond_dev, struct net_device *slave_dev,
 	}
 
 	bond->slave_cnt++;
-	netdev_compute_master_upper_features(bond->dev, true);
 	bond_set_carrier(bond);
 
 	/* Needs to be called before bond_select_active_slave(), which will
@@ -2528,7 +2532,6 @@ static int __bond_release_one(struct net_device *bond_dev,
 		call_netdevice_notifiers(NETDEV_RELEASE, bond->dev);
 	}
 
-	netdev_compute_master_upper_features(bond->dev, true);
 	if (!(bond_dev->features & NETIF_F_VLAN_CHALLENGED) &&
 	    (old_features & NETIF_F_VLAN_CHALLENGED))
 		slave_info(bond_dev, slave_dev, "last VLAN challenged slave left bond - VLAN blocking is removed\n");
@@ -4026,7 +4029,7 @@ static int bond_slave_netdev_event(unsigned long event,
 	case NETDEV_FEAT_CHANGE:
 		if (!bond->notifier_ctx) {
 			bond->notifier_ctx = true;
-			netdev_compute_master_upper_features(bond->dev, true);
+			netdev_change_features(bond->dev);
 			bond->notifier_ctx = false;
 		}
 		break;
@@ -5953,6 +5956,7 @@ static const struct net_device_ops bond_netdev_ops = {
 	.ndo_add_slave		= bond_enslave,
 	.ndo_del_slave		= bond_release,
 	.ndo_fix_features	= bond_fix_features,
+	.ndo_update_offloads	= bond_update_offloads,
 	.ndo_features_check	= passthru_features_check,
 	.ndo_get_xmit_slave	= bond_xmit_get_slave,
 	.ndo_sk_get_lower_dev	= bond_sk_get_lower_dev,
diff --git a/drivers/net/team/team_core.c b/drivers/net/team/team_core.c
index b7282f5c9632..affbcc6fc130 100644
--- a/drivers/net/team/team_core.c
+++ b/drivers/net/team/team_core.c
@@ -1247,7 +1247,6 @@ static int team_port_add(struct team *team, struct net_device *port_dev,
 	port->index = -1;
 	list_add_tail_rcu(&port->list, &team->port_list);
 	team_port_enable(team, port);
-	netdev_compute_master_upper_features(team->dev, true);
 	__team_port_change_port_added(port, !!netif_oper_up(port_dev));
 	__team_options_change_check(team);
 
@@ -1337,7 +1336,6 @@ static int team_port_del(struct team *team, struct net_device *port_dev, bool un
 	}
 	kfree_rcu(port, rcu);
 	netdev_info(dev, "Port device %s removed\n", portname);
-	netdev_compute_master_upper_features(team->dev, true);
 
 	return 0;
 }
@@ -1645,7 +1643,6 @@ static void team_uninit(struct net_device *dev)
 	team_mcast_rejoin_fini(team);
 	team_notify_peers_fini(team);
 	team_queue_override_fini(team);
-	netdev_change_features(dev);
 }
 
 static void team_destructor(struct net_device *dev)
@@ -1972,6 +1969,11 @@ static netdev_features_t team_fix_features(struct net_device *dev,
 	return features;
 }
 
+static void team_update_offloads(struct net_device *dev)
+{
+	netdev_compute_master_upper_features(dev, true);
+}
+
 static int team_change_carrier(struct net_device *dev, bool new_carrier)
 {
 	struct team *team = netdev_priv(dev);
@@ -2007,6 +2009,7 @@ static const struct net_device_ops team_netdev_ops = {
 	.ndo_add_slave		= team_add_slave,
 	.ndo_del_slave		= team_del_slave,
 	.ndo_fix_features	= team_fix_features,
+	.ndo_update_offloads	= team_update_offloads,
 	.ndo_change_carrier     = team_change_carrier,
 	.ndo_features_check	= passthru_features_check,
 };
@@ -2944,7 +2947,7 @@ static int team_device_event(struct notifier_block *unused,
 	case NETDEV_FEAT_CHANGE:
 		if (!port->team->notifier_ctx) {
 			port->team->notifier_ctx = true;
-			netdev_compute_master_upper_features(port->team->dev, true);
+			netdev_change_features(port->team->dev);
 			port->team->notifier_ctx = false;
 		}
 		break;
diff --git a/net/bridge/br_device.c b/net/bridge/br_device.c
index f7502e62dd35..137cad13379b 100644
--- a/net/bridge/br_device.c
+++ b/net/bridge/br_device.c
@@ -289,6 +289,11 @@ static netdev_features_t br_fix_features(struct net_device *dev,
 	return br_features_recompute(br, features);
 }
 
+static void br_update_offloads(struct net_device *dev)
+{
+	netdev_compute_master_upper_features(dev, false);
+}
+
 #ifdef CONFIG_NET_POLL_CONTROLLER
 static void br_poll_controller(struct net_device *br_dev)
 {
@@ -456,6 +461,7 @@ static const struct net_device_ops br_netdev_ops = {
 	.ndo_add_slave		 = br_add_slave,
 	.ndo_del_slave		 = br_del_slave,
 	.ndo_fix_features        = br_fix_features,
+	.ndo_update_offloads	 = br_update_offloads,
 	.ndo_fdb_add		 = br_fdb_add,
 	.ndo_fdb_del		 = br_fdb_delete,
 	.ndo_fdb_del_bulk	 = br_fdb_delete_bulk,
diff --git a/net/bridge/br_if.c b/net/bridge/br_if.c
index d39571e13744..030248bc94c5 100644
--- a/net/bridge/br_if.c
+++ b/net/bridge/br_if.c
@@ -680,8 +680,6 @@ int br_add_if(struct net_bridge *br, struct net_device *dev,
 
 	br_mtu_auto_adjust(br);
 
-	netdev_compute_master_upper_features(br->dev, false);
-
 	kobject_uevent(&p->kobj, KOBJ_ADD);
 
 	return 0;
@@ -734,8 +732,6 @@ int br_del_if(struct net_bridge *br, struct net_device *dev)
 	if (changed_addr)
 		call_netdevice_notifiers(NETDEV_CHANGEADDR, br->dev);
 
-	netdev_compute_master_upper_features(br->dev, false);
-
 	return 0;
 }
 
diff --git a/net/core/dev.c b/net/core/dev.c
index bbd532aa6a1b..8bcbf7c22830 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -8923,6 +8923,9 @@ static int __netdev_upper_dev_link(struct net_device *dev,
 	__netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
 				    priv);
 
+	/* re-compute all features after adding link */
+	netdev_change_features(upper_dev);
+
 	return 0;
 
 rollback:
@@ -9015,6 +9018,9 @@ static void __netdev_upper_dev_unlink(struct net_device *dev,
 	__netdev_update_lower_level(upper_dev, priv);
 	__netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
 				    priv);
+
+	/* re-compute all features after removing link */
+	netdev_change_features(upper_dev);
 }
 
 /**
@@ -12880,8 +12886,6 @@ void netdev_compute_master_upper_features(struct net_device *dev, bool update_he
 
 	netif_set_tso_max_segs(dev, tso_max_segs);
 	netif_set_tso_max_size(dev, tso_max_size);
-
-	netdev_change_features(dev);
 }
 EXPORT_SYMBOL(netdev_compute_master_upper_features);
 

-- 
Git-155)

[PATCH net-next v3 3/5] macsec: move netdev_upper_dev_link() after macsec_changelink_common()

[Hangbin Liu]

After calling netdev_change_features() in __netdev_upper_dev_link(),
the call stack looks like:

- netdev_upper_dev_link
  - __netdev_upper_dev_link
    - netdev_change_features
      - notifier_call_chain
        - rtnetlink_event
          - rtmsg_ifinfo_event
            - rtmsg_ifinfo_build_skb

In macsec, we call netdev_upper_dev_link() before
macsec_changelink_common(), which causes the fields of the MACsec
Security Entity to be uninitialized. Later, macsec_fill_info() returns
-EMSGSIZE and triggers WARN_ON() in rtmsg_ifinfo_build_skb.

Fix this by moving netdev_upper_dev_link() after
macsec_changelink_common(), and return 0 if secy->key_len is not
initialized yet.

Suggested-by: Sabrina Dubroca <sd@queasysnail.net>
Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
---
 drivers/net/macsec.c | 23 ++++++++++++-----------
 1 file changed, 12 insertions(+), 11 deletions(-)

diff --git a/drivers/net/macsec.c b/drivers/net/macsec.c
index f6cad0746a02..6bb38084bc1e 100644
--- a/drivers/net/macsec.c
+++ b/drivers/net/macsec.c
@@ -4161,10 +4161,6 @@ static int macsec_newlink(struct net_device *dev,
 	lockdep_set_class(&dev->addr_list_lock,
 			  &macsec_netdev_addr_lock_key);
 
-	err = netdev_upper_dev_link(real_dev, dev, extack);
-	if (err < 0)
-		goto unregister;
-
 	/* need to be already registered so that ->init has run and
 	 * the MAC addr is set
 	 */
@@ -4177,12 +4173,12 @@ static int macsec_newlink(struct net_device *dev,
 
 	if (rx_handler && sci_exists(real_dev, sci)) {
 		err = -EBUSY;
-		goto unlink;
+		goto unregister;
 	}
 
 	err = macsec_add_dev(dev, sci, icv_len);
 	if (err)
-		goto unlink;
+		goto unregister;
 
 	if (data) {
 		err = macsec_changelink_common(dev, data);
@@ -4190,6 +4186,10 @@ static int macsec_newlink(struct net_device *dev,
 			goto del_dev;
 	}
 
+	err = netdev_upper_dev_link(real_dev, dev, extack);
+	if (err < 0)
+		goto del_dev;
+
 	/* If h/w offloading is available, propagate to the device */
 	if (macsec_is_offloaded(macsec)) {
 		const struct macsec_ops *ops;
@@ -4200,7 +4200,7 @@ static int macsec_newlink(struct net_device *dev,
 			ctx.secy = &macsec->secy;
 			err = macsec_offload(ops->mdo_add_secy, &ctx);
 			if (err)
-				goto del_dev;
+				goto unlink;
 
 			macsec->insert_tx_tag =
 				macsec_needs_tx_tag(macsec, ops);
@@ -4209,7 +4209,7 @@ static int macsec_newlink(struct net_device *dev,
 
 	err = register_macsec_dev(real_dev, dev);
 	if (err < 0)
-		goto del_dev;
+		goto unlink;
 
 	netdev_update_features(dev);
 	netif_stacked_transfer_operstate(real_dev, dev);
@@ -4219,10 +4219,10 @@ static int macsec_newlink(struct net_device *dev,
 
 	return 0;
 
-del_dev:
-	macsec_del_dev(macsec);
 unlink:
 	netdev_upper_dev_unlink(real_dev, dev);
+del_dev:
+	macsec_del_dev(macsec);
 unregister:
 	unregister_netdevice(dev);
 	return err;
@@ -4337,7 +4337,8 @@ static int macsec_fill_info(struct sk_buff *skb,
 		csid = secy->xpn ? MACSEC_CIPHER_ID_GCM_AES_XPN_256 : MACSEC_CIPHER_ID_GCM_AES_256;
 		break;
 	default:
-		goto nla_put_failure;
+		WARN_ON_ONCE(1);
+		return 0;
 	}
 
 	if (nla_put_sci(skb, IFLA_MACSEC_SCI, secy->sci,

-- 
Git-155)

[PATCH net-next v3 4/5] failover: use ndo_update_offloads for failover offload compute

[Hangbin Liu]

Convert net_failover to use the ndo_update_offloads callback and calls
netdev_compute_master_upper_features() to compute new offload flags
during slave registration/unregistration.

This simplifies the failover code significantly by removing the custom
feature computation logic and relying on the centralized feature update
mechanism. The callback is automatically invoked during feature updates
when upper/lower device links change.

Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
---
 drivers/net/net_failover.c | 66 +++++-----------------------------------------
 include/net/net_failover.h |  7 -----
 2 files changed, 7 insertions(+), 66 deletions(-)

diff --git a/drivers/net/net_failover.c b/drivers/net/net_failover.c
index d0361aaf25ef..85c238b88b85 100644
--- a/drivers/net/net_failover.c
+++ b/drivers/net/net_failover.c
@@ -300,6 +300,11 @@ static int net_failover_vlan_rx_kill_vid(struct net_device *dev, __be16 proto,
 	return 0;
 }
 
+static void failover_update_offloads(struct net_device *dev)
+{
+	netdev_compute_master_upper_features(dev, true);
+}
+
 static const struct net_device_ops failover_dev_ops = {
 	.ndo_open		= net_failover_open,
 	.ndo_stop		= net_failover_close,
@@ -312,6 +317,7 @@ static const struct net_device_ops failover_dev_ops = {
 	.ndo_vlan_rx_kill_vid	= net_failover_vlan_rx_kill_vid,
 	.ndo_validate_addr	= eth_validate_addr,
 	.ndo_features_check	= passthru_features_check,
+	.ndo_update_offloads	= failover_update_offloads,
 };
 
 #define FAILOVER_NAME "net_failover"
@@ -373,61 +379,6 @@ static rx_handler_result_t net_failover_handle_frame(struct sk_buff **pskb)
 	return RX_HANDLER_ANOTHER;
 }
 
-static void net_failover_compute_features(struct net_device *dev)
-{
-	netdev_features_t vlan_features = FAILOVER_VLAN_FEATURES &
-					  NETIF_F_ALL_FOR_ALL;
-	netdev_features_t enc_features  = FAILOVER_ENC_FEATURES;
-	unsigned short max_hard_header_len = ETH_HLEN;
-	unsigned int dst_release_flag = IFF_XMIT_DST_RELEASE |
-					IFF_XMIT_DST_RELEASE_PERM;
-	struct net_failover_info *nfo_info = netdev_priv(dev);
-	struct net_device *primary_dev, *standby_dev;
-
-	primary_dev = rcu_dereference(nfo_info->primary_dev);
-	if (primary_dev) {
-		vlan_features =
-			netdev_increment_features(vlan_features,
-						  primary_dev->vlan_features,
-						  FAILOVER_VLAN_FEATURES);
-		enc_features =
-			netdev_increment_features(enc_features,
-						  primary_dev->hw_enc_features,
-						  FAILOVER_ENC_FEATURES);
-
-		dst_release_flag &= primary_dev->priv_flags;
-		if (primary_dev->hard_header_len > max_hard_header_len)
-			max_hard_header_len = primary_dev->hard_header_len;
-	}
-
-	standby_dev = rcu_dereference(nfo_info->standby_dev);
-	if (standby_dev) {
-		vlan_features =
-			netdev_increment_features(vlan_features,
-						  standby_dev->vlan_features,
-						  FAILOVER_VLAN_FEATURES);
-		enc_features =
-			netdev_increment_features(enc_features,
-						  standby_dev->hw_enc_features,
-						  FAILOVER_ENC_FEATURES);
-
-		dst_release_flag &= standby_dev->priv_flags;
-		if (standby_dev->hard_header_len > max_hard_header_len)
-			max_hard_header_len = standby_dev->hard_header_len;
-	}
-
-	dev->vlan_features = vlan_features;
-	dev->hw_enc_features = enc_features | NETIF_F_GSO_ENCAP_ALL;
-	dev->hard_header_len = max_hard_header_len;
-
-	dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
-	if (dst_release_flag == (IFF_XMIT_DST_RELEASE |
-				 IFF_XMIT_DST_RELEASE_PERM))
-		dev->priv_flags |= IFF_XMIT_DST_RELEASE;
-
-	netdev_change_features(dev);
-}
-
 static void net_failover_lower_state_changed(struct net_device *slave_dev,
 					     struct net_device *primary_dev,
 					     struct net_device *standby_dev)
@@ -550,7 +501,6 @@ static int net_failover_slave_register(struct net_device *slave_dev,
 	}
 
 	net_failover_lower_state_changed(slave_dev, primary_dev, standby_dev);
-	net_failover_compute_features(failover_dev);
 
 	call_netdevice_notifiers(NETDEV_JOIN, slave_dev);
 
@@ -621,8 +571,6 @@ static int net_failover_slave_unregister(struct net_device *slave_dev,
 
 	dev_put(slave_dev);
 
-	net_failover_compute_features(failover_dev);
-
 	netdev_info(failover_dev, "failover %s slave:%s unregistered\n",
 		    slave_is_standby ? "standby" : "primary", slave_dev->name);
 
@@ -736,7 +684,7 @@ struct failover *net_failover_create(struct net_device *standby_dev)
 	/* Don't allow failover devices to change network namespaces. */
 	failover_dev->netns_immutable = true;
 
-	failover_dev->hw_features = FAILOVER_VLAN_FEATURES |
+	failover_dev->hw_features = MASTER_UPPER_DEV_VLAN_FEATURES |
 				    NETIF_F_HW_VLAN_CTAG_TX |
 				    NETIF_F_HW_VLAN_CTAG_RX |
 				    NETIF_F_HW_VLAN_CTAG_FILTER;
diff --git a/include/net/net_failover.h b/include/net/net_failover.h
index b12a1c469d1c..f0f038d113a0 100644
--- a/include/net/net_failover.h
+++ b/include/net/net_failover.h
@@ -30,11 +30,4 @@ struct net_failover_info {
 struct failover *net_failover_create(struct net_device *standby_dev);
 void net_failover_destroy(struct failover *failover);
 
-#define FAILOVER_VLAN_FEATURES	(NETIF_F_HW_CSUM | NETIF_F_SG | \
-				 NETIF_F_FRAGLIST | NETIF_F_ALL_TSO | \
-				 NETIF_F_HIGHDMA | NETIF_F_LRO)
-
-#define FAILOVER_ENC_FEATURES	(NETIF_F_HW_CSUM | NETIF_F_SG | \
-				 NETIF_F_RXCSUM | NETIF_F_ALL_TSO)
-
 #endif /* _NET_FAILOVER_H */

-- 
Git-155)

[PATCH net-next v3 5/5] net: no need to disable LRO specifically

[Hangbin Liu]

In commit "net: use ndo_update_offloads to set offload features for
bonding/bridge/team" we called netdev_change_features() in
__netdev_upper_dev_link(), which will disable LRO automatically on lower
device if upper device doesn't have LRO enabled. So we don't need to
dev_disable_lro() again after netdev_upper_dev_link().

Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
---
 drivers/net/bonding/bond_main.c | 3 ---
 drivers/net/team/team_core.c    | 3 ---
 net/8021q/vlan.c                | 2 --
 net/bridge/br_if.c              | 2 --
 net/hsr/hsr_slave.c             | 1 -
 5 files changed, 11 deletions(-)

diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
index ea6bb20f9348..f4f6ae02f7ca 100644
--- a/drivers/net/bonding/bond_main.c
+++ b/drivers/net/bonding/bond_main.c
@@ -2221,9 +2221,6 @@ int bond_enslave(struct net_device *bond_dev, struct net_device *slave_dev,
 	}
 #endif
 
-	if (!(bond_dev->features & NETIF_F_LRO))
-		dev_disable_lro(slave_dev);
-
 	res = netdev_rx_handler_register(slave_dev, bond_handle_frame,
 					 new_slave);
 	if (res) {
diff --git a/drivers/net/team/team_core.c b/drivers/net/team/team_core.c
index affbcc6fc130..4f18205dd30d 100644
--- a/drivers/net/team/team_core.c
+++ b/drivers/net/team/team_core.c
@@ -1191,9 +1191,6 @@ static int team_port_add(struct team *team, struct net_device *port_dev,
 		goto err_enable_netpoll;
 	}
 
-	if (!(dev->features & NETIF_F_LRO))
-		dev_disable_lro(port_dev);
-
 	err = netdev_rx_handler_register(port_dev, team_handle_frame,
 					 port);
 	if (err) {
diff --git a/net/8021q/vlan.c b/net/8021q/vlan.c
index 2b74ed56eb16..fda3a80e9340 100644
--- a/net/8021q/vlan.c
+++ b/net/8021q/vlan.c
@@ -193,8 +193,6 @@ int register_vlan_dev(struct net_device *dev, struct netlink_ext_ack *extack)
 	vlan_group_set_device(grp, vlan->vlan_proto, vlan_id, dev);
 	grp->nr_vlan_devs++;
 
-	netdev_update_features(dev);
-
 	return 0;
 
 out_unregister_netdev:
diff --git a/net/bridge/br_if.c b/net/bridge/br_if.c
index 030248bc94c5..0aa653a1e651 100644
--- a/net/bridge/br_if.c
+++ b/net/bridge/br_if.c
@@ -620,8 +620,6 @@ int br_add_if(struct net_bridge *br, struct net_device *dev,
 	if (err)
 		goto err5;
 
-	dev_disable_lro(dev);
-
 	list_add_rcu(&p->list, &br->port_list);
 
 	nbp_update_port_count(br);
diff --git a/net/hsr/hsr_slave.c b/net/hsr/hsr_slave.c
index 44f83c8c56a7..4b6ab185392b 100644
--- a/net/hsr/hsr_slave.c
+++ b/net/hsr/hsr_slave.c
@@ -169,7 +169,6 @@ static int hsr_portdev_setup(struct hsr_priv *hsr, struct net_device *dev,
 	res = netdev_rx_handler_register(dev, hsr_handle_frame, port);
 	if (res)
 		goto fail_rx_handler;
-	dev_disable_lro(dev);
 
 	return 0;
 

-- 
Git-155)

Re: [PATCH net-next v3 3/5] macsec: move netdev_upper_dev_link() after macsec_changelink_common()

[Sabrina Dubroca]

2026-03-16, 12:26:11 +0800, Hangbin Liu wrote:
> After calling netdev_change_features() in __netdev_upper_dev_link(),
> the call stack looks like:
> 
> - netdev_upper_dev_link
>   - __netdev_upper_dev_link
>     - netdev_change_features
>       - notifier_call_chain
>         - rtnetlink_event
>           - rtmsg_ifinfo_event
>             - rtmsg_ifinfo_build_skb
> 
> In macsec, we call netdev_upper_dev_link() before
> macsec_changelink_common(), which causes the fields of the MACsec
> Security Entity to be uninitialized. Later, macsec_fill_info() returns
> -EMSGSIZE and triggers WARN_ON() in rtmsg_ifinfo_build_skb.
> 
> Fix this by moving netdev_upper_dev_link() after
> macsec_changelink_common(), and return 0 if secy->key_len is not
> initialized yet.
> 
> Suggested-by: Sabrina Dubroca <sd@queasysnail.net>
> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
> ---
>  drivers/net/macsec.c | 23 ++++++++++++-----------
>  1 file changed, 12 insertions(+), 11 deletions(-)

I'm not sure I'll manage to look at the rest of the series before it's
applied, but for this patch:

Reviewed-by: Sabrina Dubroca <sd@queasysnail.net>

Thanks Hangbin.

-- 
Sabrina

Re: [PATCH net-next v3 1/5] net: add ndo_update_offloads for offload computation

[Sabrina Dubroca]

2026-03-16, 12:26:09 +0800, Hangbin Liu wrote:
> Add a new ndo_update_offloads callback to net_device_ops that allows
> devices to compute and update their offload features during feature
> updates.
> 
> This callback enables master devices to recompute their features
> based on current slave device configuration. This is particularly
> useful for bonding, bridging, team, and failover devices that need
> to aggregate features from their lower devices.
> 
> The callback is optional and only implemented by devices that need
> dynamic offload feature computation.

Maybe a dumb idea (and sorry to suggest this quite late in your
submissions): since all implementations of this callback are only
calling netdev_compute_master_upper_features(), does this need to be a
new ndo, or could this be some kind of flag within struct net_device
(2 flags since netdev_compute_master_upper_features takes a bool
argument) that changes __netdev_update_features()'s behavior? Since
the goal for the netdev_compute_master_upper_features() work was to
make this code more common, not introducing a new ndo where random
hacks can accumulate separately in each driver would maybe make more
sense?

-- 
Sabrina

Re: [PATCH net-next v3 1/5] net: add ndo_update_offloads for offload computation

[Hangbin Liu]

On Tue, Mar 17, 2026 at 04:14:35PM +0100, Sabrina Dubroca wrote:
> 2026-03-16, 12:26:09 +0800, Hangbin Liu wrote:
> > Add a new ndo_update_offloads callback to net_device_ops that allows
> > devices to compute and update their offload features during feature
> > updates.
> > 
> > This callback enables master devices to recompute their features
> > based on current slave device configuration. This is particularly
> > useful for bonding, bridging, team, and failover devices that need
> > to aggregate features from their lower devices.
> > 
> > The callback is optional and only implemented by devices that need
> > dynamic offload feature computation.
> 
> Maybe a dumb idea (and sorry to suggest this quite late in your
> submissions): since all implementations of this callback are only
> calling netdev_compute_master_upper_features(), does this need to be a
> new ndo, or could this be some kind of flag within struct net_device

Ideally all dev with IFF_MASTER should re-compute the offload. But at present
some master devices do have this flag, or have their own offload
implementation.

Do you mean add a new private flag, like IFF_RECOMPUTE_OFFLOAD? For the second
parameter, maybe we can pass false for bridge specifically.

Thanks
Hangbin

> (2 flags since netdev_compute_master_upper_features takes a bool
> argument) that changes __netdev_update_features()'s behavior? Since
> the goal for the netdev_compute_master_upper_features() work was to
> make this code more common, not introducing a new ndo where random
> hacks can accumulate separately in each driver would maybe make more
> sense?
> 
> -- 
> Sabrina

Re: [PATCH net-next v3 1/5] net: add ndo_update_offloads for offload computation

[Sabrina Dubroca]

2026-03-18, 01:15:20 +0000, Hangbin Liu wrote:
> On Tue, Mar 17, 2026 at 04:14:35PM +0100, Sabrina Dubroca wrote:
> > 2026-03-16, 12:26:09 +0800, Hangbin Liu wrote:
> > > Add a new ndo_update_offloads callback to net_device_ops that allows
> > > devices to compute and update their offload features during feature
> > > updates.
> > > 
> > > This callback enables master devices to recompute their features
> > > based on current slave device configuration. This is particularly
> > > useful for bonding, bridging, team, and failover devices that need
> > > to aggregate features from their lower devices.
> > > 
> > > The callback is optional and only implemented by devices that need
> > > dynamic offload feature computation.
> > 
> > Maybe a dumb idea (and sorry to suggest this quite late in your
> > submissions): since all implementations of this callback are only
> > calling netdev_compute_master_upper_features(), does this need to be a
> > new ndo, or could this be some kind of flag within struct net_device
> 
> Ideally all dev with IFF_MASTER should re-compute the offload. But at present
> some master devices do have this flag, or have their own offload
> implementation.
> 
> Do you mean add a new private flag, like IFF_RECOMPUTE_OFFLOAD? For the second
> parameter, maybe we can pass false for bridge specifically.

I hadn't thought specifically about where to store that flag. As a
private flag, why not (this should be internal to the kernel, so not a
uapi flag like IFF_MASTER), but priv_flags is marked as "hotpath" now,
so maybe something similar to needs_free_netdev or
netns_immutable. Either way, then devices are free to do their own
magic, or request the core to call
netdev_compute_master_upper_features() for them in some common
locations. Does that makes sense?

-- 
Sabrina

Re: [PATCH net-next v3 1/5] net: add ndo_update_offloads for offload computation

[Hangbin Liu]

On Thu, Mar 19, 2026 at 12:29:25AM +0100, Sabrina Dubroca wrote:
> 2026-03-18, 01:15:20 +0000, Hangbin Liu wrote:
> > On Tue, Mar 17, 2026 at 04:14:35PM +0100, Sabrina Dubroca wrote:
> > > 2026-03-16, 12:26:09 +0800, Hangbin Liu wrote:
> > > > Add a new ndo_update_offloads callback to net_device_ops that allows
> > > > devices to compute and update their offload features during feature
> > > > updates.
> > > > 
> > > > This callback enables master devices to recompute their features
> > > > based on current slave device configuration. This is particularly
> > > > useful for bonding, bridging, team, and failover devices that need
> > > > to aggregate features from their lower devices.
> > > > 
> > > > The callback is optional and only implemented by devices that need
> > > > dynamic offload feature computation.
> > > 
> > > Maybe a dumb idea (and sorry to suggest this quite late in your
> > > submissions): since all implementations of this callback are only
> > > calling netdev_compute_master_upper_features(), does this need to be a
> > > new ndo, or could this be some kind of flag within struct net_device
> > 
> > Ideally all dev with IFF_MASTER should re-compute the offload. But at present
> > some master devices do have this flag, or have their own offload
> > implementation.
> > 
> > Do you mean add a new private flag, like IFF_RECOMPUTE_OFFLOAD? For the second
> > parameter, maybe we can pass false for bridge specifically.
> 
> I hadn't thought specifically about where to store that flag. As a
> private flag, why not (this should be internal to the kernel, so not a
> uapi flag like IFF_MASTER), but priv_flags is marked as "hotpath" now,
> so maybe something similar to needs_free_netdev or
> netns_immutable. Either way, then devices are free to do their own
> magic, or request the core to call
> netdev_compute_master_upper_features() for them in some common
> locations. Does that makes sense?

This looks better than private flags. I just feel this magic hide a little
deep for a common master feature.

Thanks
Hangbin

Re: [net-next,v3,2/5] net: use ndo_update_offloads to set offload features for bonding/bridge/team

[Paolo Abeni]

This is an AI-generated review of your patch. The human sending this
email has considered the AI review valid, or at least plausible.

For local reproduction steps see https://netdev-ai.bots.linux.dev/ai-local.html
---
net: use ndo_update_offloads to set offload features for bonding/bridge/team

This patch converts bonding, bridge, and team drivers to use the new
ndo_update_offloads callback, centralizing offload feature computation
during port add/remove operations.

> diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
> index 90b71bd53040..ea6bb20f9348 100644
> --- a/drivers/net/bonding/bond_main.c
> +++ b/drivers/net/bonding/bond_main.c

[ ... ]

> diff --git a/drivers/net/team/team_core.c b/drivers/net/team/team_core.c
> index b7282f5c9632..affbcc6fc130 100644
> --- a/drivers/net/team/team_core.c
> +++ b/drivers/net/team/team_core.c
> @@ -1247,7 +1247,6 @@ static int team_port_add(struct team *team, struct net_device *port_dev,
>  	port->index = -1;
>  	list_add_tail_rcu(&port->list, &team->port_list);
>  	team_port_enable(team, port);
> -	netdev_compute_master_upper_features(team->dev, true);
>  	__team_port_change_port_added(port, !!netif_oper_up(port_dev));
>  	__team_options_change_check(team);

[ ... ]

> diff --git a/net/bridge/br_if.c b/net/bridge/br_if.c
> index d39571e13744..030248bc94c5 100644
> --- a/net/bridge/br_if.c
> +++ b/net/bridge/br_if.c
> @@ -680,8 +680,6 @@ int br_add_if(struct net_bridge *br, struct net_device *dev,
>
>  	br_mtu_auto_adjust(br);
>
> -	netdev_compute_master_upper_features(br->dev, false);
> -
>  	kobject_uevent(&p->kobj, KOBJ_ADD);
>
>  	return 0;

[ ... ]

> diff --git a/net/core/dev.c b/net/core/dev.c
> index bbd532aa6a1b..8bcbf7c22830 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -8923,6 +8923,9 @@ static int __netdev_upper_dev_link(struct net_device *dev,
>  	__netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
>  				    priv);
>
> +	/* re-compute all features after adding link */
> +	netdev_change_features(upper_dev);
> +
>  	return 0;
>
>  rollback:

Could this netdev_change_features() call happen too early for team and
bridge drivers?

The call triggers __netdev_update_features(), which invokes both
ndo_update_offloads() and ndo_fix_features(). While ndo_update_offloads()
works with the netdev lower device list (already updated by
__netdev_adjacent_dev_link_neighbour()), the ndo_fix_features() callbacks
iterate driver-specific port lists.

For team_port_add():

drivers/net/team/team_core.c:team_port_add() {
    ...
    team_upper_dev_link(team, port, extack);  /* line ~1231 */
        netdev_upper_dev_link()
            __netdev_upper_dev_link()
                netdev_change_features()  /* <-- triggers feature update */
                    __netdev_update_features()
                        ops->ndo_fix_features()  /* team_fix_features() */

    list_add_tail_rcu(&port->list, &team->port_list);  /* line 1247 */
    ...
}

team_fix_features() iterates team->port_list to compute features, but the
new port hasn't been added to that list yet.

Similarly for br_add_if():

net/bridge/br_if.c:br_add_if() {
    ...
    netdev_master_upper_dev_link(dev, br->dev, NULL, NULL, extack);  /* line ~628 */
        __netdev_upper_dev_link()
            netdev_change_features()  /* <-- triggers feature update */
                __netdev_update_features()
                    ops->ndo_fix_features()  /* br_fix_features() */

    list_add_rcu(&p->list, &br->port_list);  /* line 634 */
    ...
}

br_fix_features() calls br_features_recompute() which iterates
br->port_list, but the new port isn't in that list yet.

If the new port lacks support for a feature the master currently has
enabled (like NETIF_F_SG, NETIF_F_IP_CSUM, or NETIF_F_TSO), that feature
would remain incorrectly enabled. The old code path called
netdev_compute_master_upper_features() AFTER the port was added to the
driver list, so the new port's capabilities were included in the feature
computation.

The bonding driver isn't affected because bond_fix_features() uses
bond_for_each_slave(), which maps to netdev_for_each_lower_private() and
iterates the netdev infrastructure's lower device list rather than a
separate driver list.

Re: [PATCH net-next v3 0/5] net: centralize master device offload feature computation

[Paolo Abeni]

On 3/16/26 5:26 AM, Hangbin Liu wrote:
> Currently, master devices (bonding, bridge, team) manually call
> netdev_compute_master_upper_features() scattered throughout their port
> add/remove operations. This approach requires each driver to remember
> to update features at the right times and leads to code duplication.
> 
> The series adds a new ndo_update_offloads callback that is automatically
> invoked during feature updates when upper/lower device relationships change.
> This centralizes the feature computation flow and removes the burden
> from individual drivers.

I'm sorry for the late feedback, but I think the driver cleanup and
simplification is not worthy the core complexity and the new NDO added.

The AI reported issue would probably need very non trivial changes to
both team and bond driver to be addressed.

Most of the cleanups belong to netdev_compute_master_upper_features()
introduction in the failover driver. Factoring out that single change
would be IMHO a better option, if possible.

The new ndo looks controversial. We already have 2 different ndo at
__netdev_update_features() time, with slightly different semantic.
Adding another one feels like a design issue.

/P

Re: [PATCH net-next v3 0/5] net: centralize master device offload feature computation

[Hangbin Liu]

On Thu, Mar 19, 2026 at 10:52:34AM +0100, Paolo Abeni wrote:
> I'm sorry for the late feedback, but I think the driver cleanup and
> simplification is not worthy the core complexity and the new NDO added.
> 
> The AI reported issue would probably need very non trivial changes to
> both team and bond driver to be addressed.
> 
> Most of the cleanups belong to netdev_compute_master_upper_features()
> introduction in the failover driver. Factoring out that single change
> would be IMHO a better option, if possible.
> 
> The new ndo looks controversial. We already have 2 different ndo at
> __netdev_update_features() time, with slightly different semantic.
> Adding another one feels like a design issue.

Thanks, I will check if we can call netdev_compute_master_upper_features()
in failover driver directly.

Hangbin

Re: [PATCH net-next v3 0/5] net: centralize master device offload feature computation

[Jakub Kicinski]

On Thu, 19 Mar 2026 10:52:34 +0100 Paolo Abeni wrote:
> On 3/16/26 5:26 AM, Hangbin Liu wrote:
> > Currently, master devices (bonding, bridge, team) manually call
> > netdev_compute_master_upper_features() scattered throughout their port
> > add/remove operations. This approach requires each driver to remember
> > to update features at the right times and leads to code duplication.
> > 
> > The series adds a new ndo_update_offloads callback that is automatically
> > invoked during feature updates when upper/lower device relationships change.
> > This centralizes the feature computation flow and removes the burden
> > from individual drivers.  
> 
> I'm sorry for the late feedback, but I think the driver cleanup and
> simplification is not worthy the core complexity and the new NDO added.

+1 FWIW, I didn't want to be negative but I felt the same way..
We are always one step away from an inf loop propagating the features
it's a delicate space :S