Re: [PATCH net-next 4/5] bridge: Linearize skb once the ND message type is validated

Nikolay Aleksandrov <[email protected]> Mon, 20 Jul 2026 12:26:05 +0300
Newsgroups dev.linux.lists.bridge,org.kernel.vger.linux-kernel,org.kernel.vger.netdev
Message-ID <[email protected]>
On 19/07/2026 16:34, Danielle Ratson wrote:
> br_nd_send() parses ND options from ns->opt[] and therefore needs the skb
> to be linear. Commit a01aee7cafc5 ("bridge: br_nd_send: linearize skb
> before parsing ND options") ensured that by linearizing inside
> br_nd_send() itself.
> 
> Move the linearization up into br_is_nd_neigh_msg(), right after
> ndisc_check_ns_na() has validated the message as an NS/NA. This makes a
> linear buffer a property of every recognized ND message, so that this and
> any future ND message handling operate on a linear skb and cannot
> reintroduce that class of bug by forgetting to linearize.
> 
> Since the skb is now linear by the time br_nd_send() runs, drop the
> linearization there and derive ns from the transport header set by
> ndisc_check_ns_na(), instead of recomputing it from the network header.
> 
> If linearization fails under memory pressure, br_is_nd_neigh_msg() returns
> NULL and the packet falls back to normal forwarding rather than being
> suppressed.
> 
> Reviewed-by: Petr Machata <[email protected]>
> Signed-off-by: Danielle Ratson <[email protected]>
> ---
>   net/bridge/br_arp_nd_proxy.c | 8 +++++---
>   1 file changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/net/bridge/br_arp_nd_proxy.c b/net/bridge/br_arp_nd_proxy.c
> index 445c930ed59b..46779d9fad61 100644
> --- a/net/bridge/br_arp_nd_proxy.c
> +++ b/net/bridge/br_arp_nd_proxy.c
> @@ -240,6 +240,9 @@ struct nd_msg *br_is_nd_neigh_msg(struct sk_buff *skb)
>   	if (ndisc_check_ns_na(skb))
>   		return NULL;
>   
> +	if (skb_linearize(skb))
> +		return NULL;
> +
>   	return (struct nd_msg *)skb_transport_header(skb);
>   }

This one is a bit weird - I wouldn't expect the check and validation to also
linearize the skb. I'd rename this helper to show that it also linearizes the skb.
Other than that the patch looks good.

>   
> @@ -259,7 +262,7 @@ static void br_nd_send(struct net_bridge *br, struct net_bridge_port *p,
>   	bool dad;
>   	u16 pvid;
>   
> -	if (!dev || skb_linearize(request))
> +	if (!dev)
>   		return;
>   
>   	len = LL_RESERVED_SPACE(dev) + sizeof(struct ipv6hdr) +
> @@ -276,8 +279,7 @@ static void br_nd_send(struct net_bridge *br, struct net_bridge_port *p,
>   	skb_set_mac_header(reply, 0);
>   
>   	daddr = eth_hdr(request)->h_source;
> -	ns = (struct nd_msg *)(skb_network_header(request) +
> -			       sizeof(struct ipv6hdr));
> +	ns = (struct nd_msg *)skb_transport_header(request);
>   
>   	/* Do we need option processing ? */
>   	ns_olen = request->len - (skb_network_offset(request) +