From: Crystal Kolipe <kolipe.c@exoticsilicon.com>
Subject: Re: vmd(8): Use 32-bit direct kernel launch for both amd64 and i386
To: Dave Voutila <dv@sisu.io>
Cc: Mike Larkin <mlarkin@nested.page>, tech@openbsd.org
Date: Wed, 20 Aug 2025 21:28:21 +0100

On Wed, Aug 20, 2025 at 03:36:51PM -0400, Dave Voutila wrote:
> Crystal Kolipe <kolipe.c@exoticsilicon.com> writes:
> > As noted previously, this is 7.7-release with local patches unrelated to vmd,
> > and loadfile_elf.c updated to HEAD, with hshoexer@'s patch applied.
> >
> > Note: with just the previous patch by bluhm@ applied, (I.E. loadfile_elf.c at
> > HEAD), vmd works as expected, (no regressions).
> 
> Can't quite grok this note. Does this mean with this diff plus a
> separate diff it's working on this machine?
> 
> I can't tell if things are now working or still broken with the diff.

Here is a summary:

* We have various OpenBSD vms running on various vmd hosts, almost all of them
  boot the OpenBSD kernel directly and do not make use of a bios image.

* Both the hosts and the vms run -release with various local modifications to
  the kernel and userland, (mostly things that are not expected to conflict
  with any work on vmd, E.G. wscons and softraid stuff)

* Since we are in the minority by:
  * not using a bios image for booting
  * running modified kernels
  * running some vms with large memory allocations, ( > 16 Gb)

  ... it seemed useful to test this patch and look for any breakage.

From a quick read of the code, I naively assumed that it would 'just work',
as the changes were minimal.

* Updated loadfile_elf.c to HEAD, (I.E. to v1.51), and recompiled vmd.

  ... everything works as before

* Applied hshoexer@'s patch on top of v1.51

  ... all vms, (both amd64 and i386), fail to start, with vm_resetcpu: failed

* Looked at vm_resetcpu(), and saw no obvious cause, so posted my observation
  to the list, with the assumption that I was missing some other update from
  -current that I would need to backport.

* Read the code again with more care, and realised that the new code was
  not a no-op for i386 compared with -release as I had originally thought, but
  sets ET | PE, (which I overlooked)

* Deleted just that line from loadfile_elf.c, effectivly making the changes a
  no-op for i386 vms.

  ... result: i386 vms boot, amd64 vms do not, (as expected). [1]

Honestly, this looks like my bad for trying to apply the diff to a modified
-release, which I was aware of right from the start, but decided to test
anyway for the reasons mentioned above.

[1] Although at this point, amd64 vms just hang with no output rather than
    report vm_resetcpu: failed.