On Wed, May 27, 2020 at 2:12 AM Jason A. Donenfeld <ja...@zx2c4.com> wrote:
> Hi again Klemens,
> On Tue, May 26, 2020 at 5:42 PM Jason A. Donenfeld <ja...@zx2c4.com> wrote:
> >
> > On Tue, May 26, 2020 at 4:52 PM Jason A. Donenfeld <ja...@zx2c4.com> wrote:
> > > With regards to your crash, though, that's a bit more puzzling, and
> > > I'd be interested to learn more details. Because these structs are
> > > already naturally aligned, the __packed attribute, even with the odd
> > > nesting Matt had prior, should have produced all entirely aligned
> > > accesses. That makes me think your kaboom was coming from someplace
> > > else. One possibility is that you were running the git tree on the two
> > > days that I was playing with uint128_t, only to find out that some of
> > > openbsd's patches to clang miscalculate stack sizes when they're in
> > > use, so that work was shelved for another day and the commits removed;
> > > perhaps you were just unlucky? Or you hit some other bug that's
> > > lurking. Either way, output from ddb's `bt` would at least be useful.
> >
> > Do you know off hand if we're able to assume any type of alignment
> > with mbuf->m_data? mtod just casts without any address fixup, which
> > means if mbuf->m_data isn't aligned by some other mechanism, we're in
> > trouble. But I would assume there _is_ some alignment imposed, since
> > the rest of the stack appears to parse tcp headers and such directly
> > without byte-by-byte copies being made.
> After a day of TCG-run compilation, I've got a working sparc64 setup
> and can confirm the issue. Working on a fix.

The latest git master should work well now:

solar# arch -s
solar# wg-quick up demo
[#] ifconfig wg0 create
[#] wg setconf wg0 /dev/fd/63
[#] ifconfig wg0 inet alias
[#] ifconfig wg0 mtu 1420
[#] ifconfig wg0 up
[#] cp /etc/resolv.conf /etc/resolv.conf.wg-quick-backup.demo
[#] printf nameserver %s\n
[#] route -q -n add -inet -iface
[#] route -q -n add -inet -iface
[#] route -q -n delete -inet
[#] route -q -n add -inet -gateway
[+] Backgrounding route monitor
solar# curl zx2c4.com/ip

One interesting quirk in doing this on qemu is that the 6.7 and
-current kernel both crash:

Loading FCode image...
Loaded 6882 bytes
entry point is 0x4000
Evaluating FCode...
OpenBSD IEEE 1275 Bootblock 2.0
Unhandled Exception 0x0000000000000030
PC = 0x00000000ffd0f3ac NPC = 0x00000000ffd0f3b0
Stopping execution

Luckily it works fine on 6.6, so that's where I debugged this issue.
But this might be a bug worth looking into. Otto's recent bootblk
patch is a possible culprit, so I've CC'd him.


Reply via email to