On Tue, Feb 20, 2024 at 12:21:56PM +0000, Tom Parkin wrote:

l2tp_ip6_sendmsg needs to avoid accounting for the transport header
twice when splicing more data into an already partially-occupied skbuff.

To manage this, we check whether the skbuff contains data using
skb_queue_empty when deciding how much data to append using
ip6_append_data.

However, the code which performed the calculation was incorrect:

     ulen = len + skb_queue_empty(&sk->sk_write_queue) ? transhdrlen : 0;

...due to C operator precedence, this ends up setting ulen to
transhdrlen for messages with a non-zero length, which results in
corrupted packets on the wire.

Add parentheses to correct the calculation in line with the original
intent.

Fixes: 9d4c75800f61 ("ipv4, ipv6: Fix handling of transhdrlen in __ip{,6}_append_data()")
Cc: David Howells <dhowells@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Tom Parkin <redacted>

Reviewed-by: Simon Horman <horms@kernel.org>

---
This issue was uncovered by Debian build-testing for the
golang-github-katalix-go-l2tp package[1].

It seems 9d4c75800f61 has been backported to the linux-6.1.y stable
kernel (and possibly others), so I think this fix will also need
backporting.

The bug is currently seen on at least Debian Bookworm, Ubuntu Jammy, and 
Debian testing/unstable.

In that case perhaps this is appropriate - citing the patch that 9d4c75800f61
tried to fix?

	Fixes: a32e0eec7042 ("l2tp: introduce L2TPv3 IP encapsulation support for IPv6")

Unfortunately tests using "ip l2tp" and which focus on dataplane
transport will not uncover this bug: it's necessary to send a packet
using an L2TPIP6 socket opened by userspace, and to verify the packet on
the wire.  The l2tp-ktest[2] test suite has been extended to cover this.

[1]. https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1063746
[2]. https://github.com/katalix/l2tp-ktest

...