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.