> On Tue, Jun 09, 2026 at 12:28:08PM +0200, Lorenzo Bianconi wrote:
> > On Jun 09, Pablo Neira Ayuso wrote:
> > > Hi Lorenzo,
> > > 
> > > On Mon, Jun 08, 2026 at 07:06:52PM +0200, Lorenzo Bianconi wrote:
> > > > Switch nf_flow_ip6_tunnel_proto() from skb_header_pointer() to
> > > > pskb_may_pull() for header validation, aligning it with the approach
> > > > used in nf_flow_ip4_tunnel_proto().
> > > > Move ctx->offset update inside the IPPROTO_IPV6 conditional block since
> > > > it should only be adjusted when a tunnel is actually detected.
> > > > While at it, use nexthdr instead of the hardcoded IPPROTO_IPV6 constant
> > > > when setting ctx->tun.proto.
> > > > 
> > > > Fixes: d98103575dcdd ("netfilter: flowtable: Add IP6IP6 rx sw acceleration")
> > > > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> > > > ---
> > > >  net/netfilter/nf_flow_table_ip.c | 10 +++++-----
> > > >  1 file changed, 5 insertions(+), 5 deletions(-)
> > > > 
> > > > diff --git a/net/netfilter/nf_flow_table_ip.c b/net/netfilter/nf_flow_table_ip.c
> > > > index 9c05a50d6013..2946399ab715 100644
> > > > --- a/net/netfilter/nf_flow_table_ip.c
> > > > +++ b/net/netfilter/nf_flow_table_ip.c
> > > > @@ -347,15 +347,15 @@ static bool nf_flow_ip6_tunnel_proto(struct nf_flowtable_ctx *ctx,
> > > >  				     struct sk_buff *skb)
> > > >  {
> > > >  #if IS_ENABLED(CONFIG_IPV6)
> > > > -	struct ipv6hdr *ip6h, _ip6h;
> > > > +	struct ipv6hdr *ip6h;
> > > >  	__be16 frag_off;
> > > >  	u8 nexthdr;
> > > >  	int hdrlen;
> > > >  
> > > > -	ip6h = skb_header_pointer(skb, ctx->offset, sizeof(*ip6h), &_ip6h);
> > > > -	if (!ip6h)
> > > > +	if (!pskb_may_pull(skb, sizeof(*ip6h) + ctx->offset))
> > > >  		return false;
> > > >  
> > > > +	ip6h = (struct ipv6hdr *)(skb_network_header(skb) + ctx->offset);
> > > >  	if (ip6h->hop_limit <= 1)
> > > >  		return false;
> > > 
> > > Not shown in the patch, but is there still a corner case here that
> > > needs to be covered?
> > > 
> > > ipv6_skip_exthdr() uses skb_header_pointer() internal, then another
> > > pskb_may_pull() is needed to make sure no other IPv6 extension header
> > > sits between the outer and the inner IPPROTO_IPV6 header, allowing to
> > > be in a non-linear area of the skb?        
> > > 
> > > > @@ -367,9 +367,9 @@ static bool nf_flow_ip6_tunnel_proto(struct nf_flowtable_ctx *ctx,
> > > >  
> > > 
> > > I mean:
> > > 
> > >         if (!pskb_may_pull(skb, hdrlen))
> > >                 return false;
> > > 
> > > where hdrlen is what ipv6_skip_exthdr() returns.
> > > 
> > > Then, I think it should be safe to call skb_pull() on
> > > ctx->tun.hdr_size.
> > > 
> > > Let me know, thanks.
> > 
> > I think you are right, here we need to run:
> > 
> > 	if (!pskb_may_pull(skb, hdrlen))
> > 		return false;
> > 
> > in order to be sure we can pull ctx->tun.hdr_size in nf_flow_ip_tunnel_pop().
> > Doing so, we can roll-back to the original skb_header_pointer() to access the
> > outer ip6 header here. What do you think?
> 
> Yes, initial skb_header_pointer() then pskb_may_pull(skb, hdrlen) to
> ensure the entire should be fine.
> 
> I think this need one more fix: This needs to resort to classic path
> if there are intermediate extension headers sitting in between the
> outer and inner headers in IP6IP6, ie. ipv6_ext_hdr() == true. Those
> extensions need to be handled by the IPv6 stack.

In my setup we have just a single Destination Option extension header (60)
between the outer and the inner IPV6 headers. In order to check if we have
other extensions headers other than Destination Option (and if so, send the
packet the networking stack) I guess we need to implement something similar
to ipv6_skip_exthdr(), agree?

> 
> nf_flow_ip6_tunnel_proto() needs to be fixed to deal with this. 

Do you want to do it with a dedicated patch or do you prefer to do it in this
one?

> And I suspect nf_flow_ip4_tunnel_proto() with IP options have the same
> problem, the flowtable need to resort to the classic stack path.

In the IPv4 case, if the packet has options nf_flow_ip4_tunnel_proto() will
return false and so the packet will be sent to the networking stack.

Regards,
Lorenzo

> 
> Thanks.