Skip to contentSkip to conversation

git-new-workdir submodules interact poorly with core.worktree

DISCUSSION Edward Z. Yang 2 participants 2 messages Jun 15, 2016

lore
Share & tools
lore
#0 lore raw

tl;dr You can't git-new-workdir checkouts which use core.worktree. This is unfortunate because 'git submodule init' uses core.worktree by default, which means you can't recursively git-new-workdir without a hack.

In the beginning, the Developer created the remote Git repository and the submodule.

mkdir -p remote/sub
(cd remote/sub && git init && touch a && git add a && git commit -m "sub init") mkdir remote/top
cd remote/top
git init
git submodule add ../sub
git commit -m "top init"
cd ../..

And the Developer said, "Let there be a local clone and submodule", and lo, there was a local clone and submodule:

git clone remote/top top
(cd top && git submodule init && git submodule update)

the Developer blessed the working copy, and said "Be fruitful and increase in number with git-new-workdir":

git-new-workdir top worktop

Unfortunately, this workdir didn't have the submodules initialized.

    $ ls worktop/sub/
    $

Now, the Developer could have run:

    $ (cd worktop && git submodule init && git submodule update)

but the resulting submodule would not have been shared with the original submodule, in the same way that git-new-workdir shared the Git metadata.

The Developer sought to create the submodule in its own likeness, but it did not work:

    $ rmdir worktop/sub && git-new-workdir top/sub worktop/sub
    fatal: Could not chdir to '../../../sub': No such file or directory

What was the Developer's fall from grace? A glance at the config of the original and new submodule shed light on the matter:

    $ cat top/sub/.git
    gitdir: ../.git/modules/sub
    $ cat top/.git/modules/sub/config
    [core]
            repositoryformatversion = 0
            filemode = true
            bare = false
            logallrefupdates = true
            worktree = ../../../sub
    $ cat worktop/sub/.git/config
    [core]
            repositoryformatversion = 0
            filemode = true
            bare = false
            logallrefupdates = true
            worktree = ../../../sub

git-new-workdir sought to reuse the config of top/sub/.git, but this configuration had core.worktree set. For the original checkout, this worked fine, since its location was .git/modules/sub; but for the new workdir, this relative path was nonsense.

I do not think there is really a way to make this work with core.worktree. Our saving grace, however, is there is a hack that can make this work: we just need to use the pre-501770e1bb5d132ae4f79aa96715f07f6b84e1f6 style of cloning submodules:

git clone remote/top oktop
git clone remote/sub oktop/sub
(cd oktop && git submodule init && git submodule update)

Now recursive git-new-workdir will work.

What's the upshot? I propose two new features:

1. A flag for git submodule update which reverts to the old behavior of making a seperate .git directory rather than collecting them together in the top-level .git/modules

2. Teach git-new-workdir to complain if core.worktree is set in the source config, and how to recursively copy submodules.

What do peopl think?

Thanks,
Edward

#1 lore raw

Am 12.09.2014 um 15:58 schrieb Edward Z. Yang:

Quoted 77 lines

tl;dr You can't git-new-workdir checkouts which use core.worktree. This is unfortunate because 'git submodule init' uses core.worktree by default, which means you can't recursively git-new-workdir without a hack.

In the beginning, the Developer created the remote Git repository and the submodule.

mkdir -p remote/sub
(cd remote/sub && git init && touch a && git add a && git commit -m "sub init") mkdir remote/top
cd remote/top
git init
git submodule add ../sub
git commit -m "top init"
cd ../..

And the Developer said, "Let there be a local clone and submodule", and lo, there was a local clone and submodule:

git clone remote/top top
(cd top && git submodule init && git submodule update)

the Developer blessed the working copy, and said "Be fruitful and increase in number with git-new-workdir":

git-new-workdir top worktop

Unfortunately, this workdir didn't have the submodules initialized.

     $ ls worktop/sub/
     $

Now, the Developer could have run:

     $ (cd worktop && git submodule init && git submodule update)

but the resulting submodule would not have been shared with the original submodule, in the same way that git-new-workdir shared the Git metadata.

The Developer sought to create the submodule in its own likeness, but it did not work:

     $ rmdir worktop/sub && git-new-workdir top/sub worktop/sub
     fatal: Could not chdir to '../../../sub': No such file or directory

What was the Developer's fall from grace? A glance at the config of the original and new submodule shed light on the matter:

     $ cat top/sub/.git
     gitdir: ../.git/modules/sub
     $ cat top/.git/modules/sub/config
     [core]
             repositoryformatversion = 0
             filemode = true
             bare = false
             logallrefupdates = true
             worktree = ../../../sub
     $ cat worktop/sub/.git/config
     [core]
             repositoryformatversion = 0
             filemode = true
             bare = false
             logallrefupdates = true
             worktree = ../../../sub

git-new-workdir sought to reuse the config of top/sub/.git, but this configuration had core.worktree set. For the original checkout, this worked fine, since its location was .git/modules/sub; but for the new workdir, this relative path was nonsense.

I do not think there is really a way to make this work with core.worktree. Our saving grace, however, is there is a hack that can make this work: we just need to use the pre-501770e1bb5d132ae4f79aa96715f07f6b84e1f6 style of cloning submodules:

Quoted 5 lines

git clone remote/top oktop
git clone remote/sub oktop/sub
(cd oktop && git submodule init && git submodule update)

Now recursive git-new-workdir will work.

Thanks for the report and a nice summary.

Quoted 5 lines

What's the upshot? I propose two new features:

1. A flag for git submodule update which reverts to the old behavior of making a seperate .git directory rather than collecting them together in the top-level .git/modules

That would play bad with the upcoming recursive submodule update (which needs .git/modules to safely remove a submodule work tree), so I wouldn't want to do that step backwards.

Quoted 2 lines

2. Teach git-new-workdir to complain if core.worktree is set in the source config, and how to recursively copy submodules.

I'd prefer pursuing this approach.