Skipping tests that are unimplemented in 32-bit emulation

[Mark Johnston]

On Mon, Aug 06, 2018 at 02:18:35PM -0600, Alan Somers wrote:
> On Mon, Aug 6, 2018 at 1:14 PM, Mark Johnston <markj at freebsd.org> wrote:
> 
> > On Sun, Jul 29, 2018 at 11:23:33AM -0600, Alan Somers wrote:
> > > I recently tried running the i386 test suite in a chroot on an amd64
> > > system.  162 tests failed, and 33 were broken.  Some of the failures were
> > > due to system calls that haven't been implemented in 32-bit emulation.
> > > setfib(2) is an example.  I think it's unlikely that anybody will ever
> > need
> > > 32-bit emulation for setfib(2), so perhaps we should just skip the test?
> > > What's the best way to do that?  I can come up with two ways:
> > >
> > > 1) At runtime, check the hw.machine sysctl and see if it matches some
> > > compile time preprocessor constant.  I don't know what constant to use,
> > > though.  Checking __amd64__ would only work for i386 binaries on amd64
> > > kernels, and not something else like mips binaries on mips64 kernels (I
> > > don't know if we support that, but I don't want to rule it out).
> > >
> > > 2) At buildtime, put an "allowed_architectures=i386" metadata property
> > into
> > > the Kyuafile for that test program.  This would require support in
> > > /usr/share/mk/bsd.test.mk.  It would also require patching Kyua itself,
> > > because currently "Kyua config" returns the architecture for which it was
> > > built, not the one on which it's running.
> > >
> > > Thoughts?
> >
> > I don't have any particular suggestions, but I'd personally rather avoid
> > a solution that requires tests to opt-in to running under 32-bit
> > emulation, which I think excludes 2).  I'd be happy to help annotate
> > any failing tests as required.  It bugs me that the test suite currently
> > doesn't cover such relatively complicated functions as
> > freebsd32_copy_msg_out().
> 
> I don't think that 2 would necessarily be opt-in, because an undefined
> value for allowed_architectures is interpreted as meaning "all".  It could
> be opt-out instead.  But it could still be a little awkward.  Option 1
> could be accomplished for atf-c testcases by comparing the value of
> __LP64__ to a hardcoded list of known 64-bit processors as returned by
> uname(3).  But I don't know how to implement 1 for atf-sh programs.  An
> atf-sh program would need to know the architecture of any binary that it
> might invoke.  Is there anything in /etc indicating what architecture the
> image was built for?  Should we just use "file /lib/libc.so.*"?

Could we instead build and install a /usr/tests32 suite on systems that
can support it, and use a top-level Kyuafile and kyua.conf that overrides
the "architecture" variable?