Re: Minor bug, git ls-files -o aborts because of broken submodules

On Fri, Jan 22, 2016 at 04:17:29PM +0700, Duy Nguyen wrote:

$ git init abc
$ cd abc
$ mkdir def
$ echo 'gitdir: blah blah' >def/.git
$ git ls-files -o
fatal: Not a git repository: def/blah blah

If some directory looks like a submodule but turns out not, that's not
a fatal error. The stack trace is something like this. I suspect
do_submodule_path should use the gently version..

#1  0x0000000000588a78 in die
#2  0x0000000000558ded in read_gitfile_gently
#3  0x000000000051e2f6 in do_submodule_path
#4  0x000000000051e484 in git_pathdup_submodule
#5  0x00000000005340ac in resolve_gitlink_ref_recursive
#6  0x00000000005342cf in resolve_gitlink_ref
#7  0x00000000004dd20d in treat_directory
#8  0x00000000004dd760 in treat_one_path
#9  0x00000000004dd971 in treat_path
#10 0x00000000004de038 in read_directory_recursive

I think we'd have to change the semantics going up the stack, as
do_submodule_path has no way to report failure.

But I think this is another case of

  http://thread.gmane.org/gmane.comp.version-control.git/265560/focus=281253

There the question was about performance (lots of these clog up the
linear ref_cache list), but I think the root cause is the same: going
too far into ref resolution without realizing we don't have a real
submodule.

Andreas posted some patches, but they didn't make it to the list. Here
are my replies, which did:

  http://article.gmane.org/gmane.comp.version-control.git/282028

  http://article.gmane.org/gmane.comp.version-control.git/282029

  http://thread.gmane.org/gmane.comp.version-control.git/282030

However, from going over them, I think there was a problem in the series;
we really do need to know the sha1 in some of these cases. So I think
maybe the simplest thing would be to catch this case in
resolve_gitlink_ref(), before we go deeper. Let me see if I can come up
with something.

-Peff