Re: Git 2.8.1 fails test 32 of t7300-clean.sh, breaks profile build

From: Jeff King <peff@peff.net>
Date: May 3, 2016, 19:00 UTC
Message-ID: <20160503190053.GC30530@sigill.intra.peff.net>
In-Reply-To: <CAGZ79kYxSCsXGLBoBTAOLfe_pOu7YNBQoccs_fA0btEzpnU=-w@mail.gmail.com>

On Tue, May 03, 2016 at 11:53:36AM -0700, Stefan Beller wrote:

Show 10 quoted lines

> > is related or worth doing on top. But I don't think so. That code is
> > just trying to convert some error-cases into "let's err on the side of
> > assuming it is a repo". Doing that for all values of gitfile_error is
> > definitely the wrong thing (it would treat a totally non-existent
> > ".git" file as "yes, it's there", which is clearly bogus).
> 
> The proposed change is overly eager indeed.
> What if we get back a READ_GITFILE_ERR_STAT_FAILED ?
> I would think that is a reasonable indicator of a submodule being there?
> (The stat failure may be transient ENOMEM Out of memory (i.e., kernel memory).)

That would certainly be wrong with read_gitfile_gently() as it is today; it does not distinguish various values of errno for stat(), so that would get the "there's not even a .git file here at all" case wrong.

So the first step would be to have read_gitfile_gently() start looking for ENOENT versus other errors. I don't know if that's worth the trouble; we're pretty cavalier about treating stat failure as "file does not exist" in the rest of the code.

-Peff