Re: Regression: submodule worktrees can clobber core.worktree config
- Date: Thu, 10 Jan 2019 00:57:42 +0100
- From: Tomasz Śniatowski <tsniatowski@xxxxxxxx>
- Subject: Re: Regression: submodule worktrees can clobber core.worktree config
On Wed, 9 Jan 2019 at 18:42, Stefan Beller <sbeller@xxxxxxxxxx> wrote:
> On Tue, Jan 8, 2019 at 2:16 PM Tomasz Śniatowski <tsniatowski@xxxxxxxx> wrote:
> > After upgrading to 2.20.1 I noticed in some submodule+worktree scenarios git
> > will break the submodule configuration. Reproducible with:
> > git init a && (cd a; touch a; git add a; git commit -ma)
> > git init b && (cd b; git submodule add ../a; git commit -mb)
> > git -C b worktree add ../b2
> > git -C b/a worktree add ../../b2/a
> > git -C b status
> > git -C b2 submodule update
> > git -C b status
> > The submodule update in the _worktree_ puts an invalid core.worktree value in
> > the _original_ repository submodule config (b/.git/modules/a/config), causing
> > the last git status to error out with:
> > fatal: cannot chdir to '../../../../../../b2/a': No such file or directory
> > fatal: 'git status --porcelain=2' failed in submodule a
> > Looking at the config file itself, the submodule update operation applies the
> > following change (the new path is invalid):
> > - worktree = ../../../a
> > + worktree = ../../../../../../b2/a
> > This worked fine on 2.19.2 (no config change, no error), and was useful to have
> > a worktree with (large) submodules that are also worktrees.
> Thanks for reporting the issue!
> > Bisects down to:
> > 74d4731da1 submodule--helper: replace connect-gitdir-workingtree by
> > ensure-core-worktree
> So this would need to update the worktree config, not the generic config.
> We'd need to replace the line
> cfg_file = repo_git_path(&subrepo, "config");
> in builtin/submodule--helper.c::ensure_core_worktree()
> to be a worktree specific call.
> Or the other way round we'd want to make repo_git_path to
> be worktree specific and introduce repo_common_path for
> the main working tree.
> Looking at Duys tree,
> is pretty much what we need.
> Reverting that topic that introduced this (4d6d6e,
> Merge branch 'sb/submodule-update-in-c'), might be possible but
> that would conflict with another followup that fixes issues in
> that series
> (see sb/submodule-unset-core-worktree-when-worktree-is-lost
> so I'd rather just cherry-pick the commit from Duy.
I had a look at https://gitlab.com/pclouds/git/commits/submodules-in-worktrees,
and it doesn't seem to be quite all okay.
The submodule update step of the repro (that breaks the config on 2.20) emits
an error message instead, and leaves the config unchanged:
git -C b2 submodule update
fatal: could not set 'core.worktree' to '../../../../../../b2/a'
It looks a bit like it's still trying to do the wrong thing, but errors out
during the attempt (repo_config_set_worktree_gently returns false).
Curiously, even though it says "fatal", it will then perform the actual
submodule update if it's required.
Same behavior on master with a subset of that branch cherry-picked, that is:
along with two others it needed to build: