Re: [PATCH] tests/*/Makefile: make sure group.list DIRT exists before install

["Darrick J. Wong" <djwong@kernel.org>]

On Sun, Feb 25, 2024 at 07:01:48PM +0100, Martin Jansa wrote:
> On Sun, Feb 25, 2024 at 5:35 PM Zorro Lang <zlang@redhat.com> wrote:
> >
> > On Thu, Feb 08, 2024 at 11:52:41PM +0100, Martin Jansa wrote:
> > > * sometimes make install was failing with:
> > >   cp: cannot stat 'group.list': No such file or directory
> > >   and bunch of non-fatal messages:
> > >   mv: failed to preserve ownership for 'group.list': Invalid argument
> >
> > I'm wondering ... isn't that expected?
> >
> > Why not let "make install" fail directly, if there's not a completed "make" ?
> > Isn't that more familiar?
> 
> It might if it fails reliably, but in OpenEmbedded builds where I was
> seeing this in some CI builds it failed in about 1% of the builds, which is
> enough to make the failure annoying and if this is meant to be expected
> behavior than not failing reliably, for context see:
> https://git.openembedded.org/meta-openembedded/commit/?id=e8a1c9823b7e55dd2ca7f19a3fd3a05ae676bfe2

Aha, that explains the weird language in the commit message about
"removing tests/*/group.list" and "adding some sleep" even though the
patch doesn't actually do that.

> so here it calls make in do_compile task (defined in autotools-brokensep),
> then make install in do_install task and do_install was failing
> _sometimes_, and I haven't seen single failure since this change was
> applied.
> 
> I'm not very familiar with xfstests, I wasn't even building it
> intentionally (it's just part of one CI job which builds everything
> available), so I might be missing something.

Huh, that looks like the result of make trying to run the 'default' and
'install' targets in parallel because there's no "install: default"
clause in the Makefile.

Soooo... are you folks using 'make -jXX default install' in your CI
system?

--D

> Regards,
>