Re: 'pu' broken at t5304 tonight

From: Brandon Williams <bmwill@google.com>
Date: Jun 12, 2017, 18:03 UTC
Message-ID: <20170612180357.GA154599@google.com>
In-Reply-To: <CAP8UFD0nDaBCiWiho7zJtUZ3MpgPY97NQ0AXnGq7parQBi=sMw@mail.gmail.com>

On 06/11, Christian Couder wrote:

Show 30 quoted lines

> On Sat, Jun 10, 2017 at 9:05 PM, Kevin Daudt <me@ikke.info> wrote:
> > On Sat, Jun 10, 2017 at 02:48:36PM +0200, Kevin Daudt wrote:
> 
> >> For me, this bisects to the latest merge:
> >>
> >> 2047eebd3 (Merge branch 'bw/repo-object' into pu, 2017-06-10), but
> >> neither of the parent of the merge break this test, so it looks like
> >> it's because of an interaction between the repo-object topic and another
> >> topic.
> >
> > Merging the repo-object with different other topic branches reveals this
> > topic to cause the bad interaction:
> >
> > b56c91004 (Merge branch 'nd/prune-in-worktree' into pu, 2017-06-10)
> >
> > Still investigating why it happens.
> 
> Yeah, 9570b25a97 (revision.c: --indexed-objects add objects from all
> worktrees, 2017-04-19) adds the following test to t5304-prune.sh but
> this fails if nd/prune-in-worktree is rebased on top of
> bw/repo-object:
> 
> test_expect_success 'prune: handle index in multiple worktrees' '
>        git worktree add second-worktree &&
>        echo "new blob for second-worktree" >second-worktree/blob &&
>        git -C second-worktree add blob &&
>        git prune --expire=now &&
>        git -C second-worktree show :blob >actual &&
>        test_cmp second-worktree/blob actual
> '

I think I discovered what is going on here (and its mostly my fault :D).

(e7a6a3b15 revision.c: --indexed-objects add objects from all worktrees) queries each worktree's index to add object which are pending in the per-worktree index's. To do this is uses 'worktree_git_path' in order to construct a path to the worktree's index file. My series introduced a change to how 'adjust_git_path' functioned, instead of checking if the index file came from the environment before replacing with the result of 'get_index_file()', i change it to unconditionally do this replacing which breaks the ability to get a path to a worktree's index.

So this breakage is definitely my fault, but you would still run into this exact same issue if you had set 'GIT_INDEX_FILE' as no worktree's would be able to find their index files as they would be overridden by the envvar.

-- 
Brandon Williams