Re: [PATCH net-next v2 1/5] udp_tunnel: properly deal with xfrm gro encap.

[Sabrina Dubroca <sd@queasysnail.net>]

2025-03-21, 12:52:52 +0100, Paolo Abeni wrote:
> The blamed commit below does not take in account that xfrm
> can enable GRO over UDP encapsulation without going through
> setup_udp_tunnel_sock().
> 
> At deletion time such socket will still go through
> udp_tunnel_cleanup_gro(), and the failed GRO type lookup will
> trigger the reported warning.
> 
> Add the GRO accounting for XFRM tunnel when GRO is enabled, and
> adjust the known gro types accordingly.
> 
> Note that we can't use setup_udp_tunnel_sock() here, as the xfrm
> tunnel setup can be "incremental" - e.g. the encapsulation is created
> first and GRO is enabled later.
> 
> Also we can not allow GRO sk lookup optimization for XFRM tunnels, as
> the socket could match the selection criteria at enable time, and
> later on the user-space could disconnect/bind it breaking such
> criteria.
> 
> Reported-by: syzbot+8c469a2260132cd095c1@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=8c469a2260132cd095c1
> Fixes: 311b36574ceac ("udp_tunnel: use static call for GRO hooks when possible")
> Signed-off-by: Paolo Abeni <pabeni@redhat.com>
> ---
> v1 -> v2:
>  - do proper account for xfrm, retain the warning
> ---
>  net/ipv4/udp.c         | 5 +++++
>  net/ipv4/udp_offload.c | 4 +++-
>  2 files changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
> index db606f7e41638..79efbf465fb04 100644
> --- a/net/ipv4/udp.c
> +++ b/net/ipv4/udp.c
> @@ -2903,10 +2903,15 @@ static void set_xfrm_gro_udp_encap_rcv(__u16 encap_type, unsigned short family,
>  {
>  #ifdef CONFIG_XFRM
>  	if (udp_test_bit(GRO_ENABLED, sk) && encap_type == UDP_ENCAP_ESPINUDP) {
> +		bool old_enabled = !!udp_sk(sk)->gro_receive;
> +
>  		if (family == AF_INET)
>  			WRITE_ONCE(udp_sk(sk)->gro_receive, xfrm4_gro_udp_encap_rcv);
>  		else if (IS_ENABLED(CONFIG_IPV6) && family == AF_INET6)
>  			WRITE_ONCE(udp_sk(sk)->gro_receive, ipv6_stub->xfrm6_gro_udp_encap_rcv);
> +
> +		if (!old_enabled && udp_sk(sk)->gro_receive)
> +			udp_tunnel_update_gro_rcv(sk, true);

We're not under any lock at this point, so this is a bit racy. I think
we'll "only" end up leaking a ref on cur->count if this happens. Not
ideal, but much better than the current situation.

>  	}
>  #endif
>  }
> diff --git a/net/ipv4/udp_offload.c b/net/ipv4/udp_offload.c
> index 088aa8cb8ac0c..02365b818f1af 100644
> --- a/net/ipv4/udp_offload.c
> +++ b/net/ipv4/udp_offload.c
> @@ -37,9 +37,11 @@ struct udp_tunnel_type_entry {
>  	refcount_t count;
>  };
>  
> +/* vxlan, fou and xfrm have 2 different gro_receive hooks each */
>  #define UDP_MAX_TUNNEL_TYPES (IS_ENABLED(CONFIG_GENEVE) + \
>  			      IS_ENABLED(CONFIG_VXLAN) * 2 + \
> -			      IS_ENABLED(CONFIG_NET_FOU) * 2)
> +			      IS_ENABLED(CONFIG_NET_FOU) * 2 + \
> +			      IS_ENABLED(CONFIG_XFRM) * 2)
>  
>  DEFINE_STATIC_CALL(udp_tunnel_gro_rcv, dummy_gro_rcv);
>  static DEFINE_STATIC_KEY_FALSE(udp_tunnel_static_call);
> -- 
> 2.48.1
> 
> 

-- 
Sabrina