Re: [PATCH ipsec] xfrm: Ensure consistent address families when resolving templates

[Steffen Klassert <steffen.klassert@secunet.com>]

On Tue, Apr 25, 2023 at 01:34:44PM +0800, Herbert Xu wrote:
> On Mon, Apr 24, 2023 at 03:23:02PM +0200, Tobias Brunner wrote:
> > xfrm_state_find() uses `encap_family` of the current template with
> > the passed local and remote addresses to find a matching state.
> > This check makes sure that there is no mismatch and out-of-bounds
> > read in mixed-family scenarios where optional tunnel or BEET mode
> > templates were skipped that would have changed the addresses to
> > match the current template's family.
> > 
> > This basically enforces the same check as validate_tmpl(), just at
> > runtime when one or more optional templates might have been skipped.
> > 
> > Signed-off-by: Tobias Brunner <tobias@strongswan.org>
> > ---
> >  net/xfrm/xfrm_policy.c | 5 +++++
> >  1 file changed, 5 insertions(+)
> 
> I'm confused.  By skipping, you're presumably referring to IPcomp.
> 
> For IPcomp, skipping should only occur on inbound, but your patch
> is changing a code path that's only invoked for outbound.  What's
> going on?

The problem is, that you can configure it for outbound too.
Even though, it does not make much sense. syzbot reported
a stack-out-of-bounds issue with intermediate optional
templates that change the address family:

https://www.spinics.net/lists/netdev/msg890567.html

I tried to fix this by rejecting such a configuration:

https://lore.kernel.org/netdev/ZCZ79IlUW53XxaVr@gauss3.secunet.de/T/

This broke some strongswan configurations.

Tobias patch is the next attempt to fix that.