Re: [PATCH net v2] fou: ensure GUE headers have enough headroom

[Ido Schimmel <idosch@nvidia.com>]

On Fri, Aug 21, 2026 at 10:59:08PM +0800, Chengfeng Ye wrote:
> ipgre_changelink() installs GUE encapsulation before it publishes the
> new GRE header length and updates dev->needed_headroom.  The transmit
> path does not serialize with RTNL, so it can interleave as follows:
> 
>   CPU 0 (ipgre_changelink)        CPU 1 (ipgre_xmit)
>   install GUE encapsulation
>                                   reserve the old needed_headroom
>   publish larger GRE flags
>   update tunnel->tun_hlen
>                                   push the larger GRE header
>                                   push the GUE and UDP headers
>   update dev->needed_headroom
> 
> With REMCSUM, the new layout can push 16 bytes of GRE and 20 bytes of
> GUE/UDP headers into an skb with only 32 bytes of actual headroom.  The
> final UDP push writes four bytes before skb->head.
> 
> With the update window widened, the kernel reported:
> 
>   skbuff: skb_under_panic: ... len:128 put:8 ... dev:gre0poc
>   kernel BUG at net/core/skbuff.c:214!
>   Oops: invalid opcode: 0000 [#1] SMP KASAN NOPTI
>   Call Trace:
>    skb_push
>    fou_build_udp
>    gue_build_header
>    ip_tunnel_xmit
>    __gre_xmit
>    ipgre_xmit
> 
> Make __gue_build_header() ensure space for both the GUE header it is
> about to push and the UDP header that follows.  On normally sized skbs
> the check is a no-op.  If configuration changes race with transmission,
> skb_cow_head() expands the head before either GUE write, or returns an
> error without modifying the packet.
> 
> skb_cow_head() may move skb->head.  Refresh skb-derived inner_iph
> pointers after encapsulation, and read the tunnel fragment and TTL
> fields before ip_tunnel_encap() so ip_tunnel_xmit() does not dereference
> tnl_params after a possible skb head reallocation.

It is up to the caller of ip_tunnel_encap() to make sure that there is
enough headroom in the packet. Otherwise, what is the point of the
encap_hlen() callback? Also, adding skb_cow_head() in the build_header()
callback forces every caller to refresh pointers to skb->head.

There's a WIP patch to fix the same problem in IPv6 [1]. I suggest doing
something similar in ip_tunnel_xmit() and ip_md_tunnel_xmit(). Note that
IPv6 reserves the headroom before the encap push, so the snapshot is
enough there. IPv4 reserves it after, so we also need an skb_cow_head()
before ip_tunnel_encap().

[1] https://lore.kernel.org/netdev/cover.1786088695.git.petalzu987@gmail.com/