On Tue, May 26, 2020 at 2:33 PM Theo de Raadt <dera...@openbsd.org> wrote:
>
> Jason A. Donenfeld <ja...@zx2c4.com> wrote:
>
> > Hey Klemens,
> >
> > On Tue, May 26, 2020 at 9:13 AM Klemens Nanni <k...@openbsd.org> wrote:
> > > I worked with the patches from the wireguard-openbsd repository after
> > > version one of this diff on tech@ became a bit old.
> > >
> > > That was until yesterday; the kernel would panic due to memory
> > > alignment issues in various spots, `amd64# ping6 tunnel-ip.sparc64'
> > > would receive a single ICMP6 echo reply from the sparc64 peer before
> > > panicing its kernel very early in noise init code.
> > >
> > > I've had fixed a few bugs and ran into different panics during later
> > > code paths.
> > >
> > > Fortunately, using "rev. 2" of your diff on top of a clean checkout just
> > > works so far on sparc64 without any additional patches required - this
> > > is a very pleasnt suprise and I can start looking at it under production
> > > use cases more thoroughly now.
> >
> > That's good news that it's working for you now, but I didn't change
> > anything within the last 24 hours (you mentioned "yesterday") that
> > would seem related to this. So perhaps don't get too excited just yet.
> > I've got a decent MIPS setup here and a hacked up qemu for having a
> > bit more detail on alignment traps. I'll see if I can reproduce any
> > issues.
>
> I suspect problems around forcing __packed.
>
> __packed is very dangerous.
>
> If you know the sizes of the objects are correctly sized and placed to
> not have gap-pads, then not using __packed might be better.
Right, most times, uses of packed are wrong and can be avoided. I'll
work through that all and see what can be minimized. In a cursory look
I did see code to the effect of `struct { u32 a; u64 b; } __packed`,
which could trap.
Jason
