[PATCH 0/3] worktree: add --recurse-submodules support to git worktree add

[PATCH 0/3] worktree: add --recurse-submodules support to git worktree add

[Jimmy Aguilar Mena]

This series implements the native --recurse-submodules flag for
"git worktree add" discussed in the earlier RFC thread.

The approach follows Phillip Wood's and Junio's feedback: each linked
worktree gets its own per-worktree submodule gitdir under
$GIT_COMMON_DIR/worktrees/<id>/modules/<name>/, so HEAD, refs, and
the index are independent per worktree while pack files and loose
objects are shared via hardlinks.  The gitdir isolation is the same
model git worktree already uses for the superproject.

Patch 1 adds the --recurse-submodules flag to builtin/worktree.c and
calls "git submodule update --init --recursive" from within the new
worktree after checkout.

Patch 2 teaches clone_submodule() in builtin/submodule--helper.c to
detect when the main worktree already has the submodule cloned and
reuse it via "git clone --local --no-checkout --separate-git-dir"
instead of fetching from the remote URL.  This avoids redundant
network access and disk use: the objects are already present locally.

Patch 3 adds tests to t2405-worktree-submodule.sh covering both the
happy path and the gitdir-isolation invariant.

Cleanup is automatic: submodule gitdirs under worktrees/<id>/modules/
are removed when "git worktree remove" calls remove_dir_recursively()
on the worktree entry, exactly as with the superproject's per-worktree
state.

Changes since the RFC:
- Replaced the shell-script cp -al prototype with a native C
   implementation in builtin/worktree.c and builtin/submodule--helper.c.
- Per-worktree gitdir isolation is now handled by the existing
   submodule_name_to_gitdir() path; no extra plumbing is needed.
- Added t2405 tests verifying both behaviour and gitdir placement.

Jimmy Aguilar Mena (3):
   worktree: add --recurse-submodules flag to worktree add
   submodule--helper: reuse main-worktree gitdir in linked worktrees
   t2405: add tests for worktree add --recurse-submodules

  builtin/submodule--helper.c   | 54 +++++++++++++++++++++++++++++++++++
  builtin/worktree.c            | 23 +++++++++++++++
  t/t2405-worktree-submodule.sh | 24 +++++++++++++++-
  3 files changed, 100 insertions(+), 1 deletion(-)

Re: [PATCH 0/3] worktree: add --recurse-submodules support to git worktree add

[Junio C Hamano]

Jimmy Aguilar Mena <kratsbinovish@gmail.com> writes:

> The approach follows Phillip Wood's and Junio's feedback: each linked
> worktree gets its own per-worktree submodule gitdir under
> $GIT_COMMON_DIR/worktrees/<id>/modules/<name>/, so HEAD, refs, and
> the index are independent per worktree while pack files and loose
> objects are shared via hardlinks.  The gitdir isolation is the same
> model git worktree already uses for the superproject.

I do not quite follow.  The point of git-native worktree support
(which improved a lot compared to its precursor, "git-new-workdir",
is that it can work well in a hardlink-challenged platforms.  You
shouldn't worry about "hardlinking" yourself at all.

After the superproject successfully did "submodule init", you can
move the submodule's repository with "absorbgitdirs" to
$GIT_DIR/modules/<submodule>/ of the superproject.  The primary
motivation behind this feature was that you can switch to a commit
in the superproject that does *not* have the submodule bound to it
at all (and obviously you do not want to lose the submodule
repository only because you tentatively switch to such a commit and
have to re-download when you switch back), but I think it gives the
single instance of submodule repository that you can share across
worktrees of the submodule.  Because the single directory created
with "absorbgitdirs" looks like a bare repository, you should be
able to create two worktrees off of that, with their own HEAD etc.

Re: [PATCH 0/3] worktree: add --recurse-submodules support to git worktree add

[Phillip Wood]

On 16/04/2026 18:05, Junio C Hamano wrote:
> Jimmy Aguilar Mena <kratsbinovish@gmail.com> writes:
> 
>> The approach follows Phillip Wood's and Junio's feedback: each linked
>> worktree gets its own per-worktree submodule gitdir under
>> $GIT_COMMON_DIR/worktrees/<id>/modules/<name>/, so HEAD, refs, and
>> the index are independent per worktree while pack files and loose
>> objects are shared via hardlinks.  The gitdir isolation is the same
>> model git worktree already uses for the superproject.
> 
> I do not quite follow.  The point of git-native worktree support
> (which improved a lot compared to its precursor, "git-new-workdir",
> is that it can work well in a hardlink-challenged platforms.  You
> shouldn't worry about "hardlinking" yourself at all.
> 
> After the superproject successfully did "submodule init", you can
> move the submodule's repository with "absorbgitdirs" to
> $GIT_DIR/modules/<submodule>/ of the superproject.  The primary
> motivation behind this feature was that you can switch to a commit
> in the superproject that does *not* have the submodule bound to it
> at all (and obviously you do not want to lose the submodule
> repository only because you tentatively switch to such a commit and
> have to re-download when you switch back), but I think it gives the
> single instance of submodule repository that you can share across
> worktrees of the submodule.  Because the single directory created
> with "absorbgitdirs" looks like a bare repository, you should be
> able to create two worktrees off of that, with their own HEAD etc.

I haven't thought much about it but that would mean that "git worktree 
remove" ought to remove the submodule's worktree when the worktree 
containing the submodule is removed. Worktrees avoid hardlinks by 
creating a "commondir" file in the worktree's gitdir which contains the 
relative path to "$GIT_COMMON_DIR". I think we could probably do the 
same here and create 
"$GIT_COMMON_DIR/worktrees/<id>/modules/<name>/commondir" containing 
"../../../../modules/<name>" if we want to store the submodule's gitdir 
under the worktree's gitdir. That way removing a worktree's gitdir 
removes all the gitdirs of its submodules without any extra effort. 
There are probably other tradeoffs between the two approaches that I've 
not thought of.

Thanks

Phillip

Re: [PATCH 0/3] worktree: add --recurse-submodules support to git worktree add

[Phillip Wood]

On 16/04/2026 19:38, Phillip Wood wrote:
> On 16/04/2026 18:05, Junio C Hamano wrote:
>> Jimmy Aguilar Mena <kratsbinovish@gmail.com> writes:
>>
>>> The approach follows Phillip Wood's and Junio's feedback: each linked
>>> worktree gets its own per-worktree submodule gitdir under
>>> $GIT_COMMON_DIR/worktrees/<id>/modules/<name>/, so HEAD, refs, and
>>> the index are independent per worktree while pack files and loose
>>> objects are shared via hardlinks.  The gitdir isolation is the same
>>> model git worktree already uses for the superproject.
>>
>> I do not quite follow.  The point of git-native worktree support
>> (which improved a lot compared to its precursor, "git-new-workdir",
>> is that it can work well in a hardlink-challenged platforms.  You
>> shouldn't worry about "hardlinking" yourself at all.
>>
>> After the superproject successfully did "submodule init", you can
>> move the submodule's repository with "absorbgitdirs" to
>> $GIT_DIR/modules/<submodule>/ of the superproject.  The primary
>> motivation behind this feature was that you can switch to a commit
>> in the superproject that does *not* have the submodule bound to it
>> at all (and obviously you do not want to lose the submodule
>> repository only because you tentatively switch to such a commit and
>> have to re-download when you switch back), but I think it gives the
>> single instance of submodule repository that you can share across
>> worktrees of the submodule.  Because the single directory created
>> with "absorbgitdirs" looks like a bare repository, you should be
>> able to create two worktrees off of that, with their own HEAD etc.
> 
> I haven't thought much about it but that would mean that "git worktree 
> remove" ought to remove the submodule's worktree when the worktree 
> containing the submodule is removed. Worktrees avoid hardlinks by 
> creating a "commondir" file in the worktree's gitdir which contains the 
> relative path to "$GIT_COMMON_DIR". I think we could probably do the 
> same here and create "$GIT_COMMON_DIR/worktrees/<id>/modules/<name>/ 
> commondir" containing "../../../../modules/<name>" if we want to store 
> the submodule's gitdir under the worktree's gitdir. That way removing a 
> worktree's gitdir removes all the gitdirs of its submodules without any 
> extra effort. There are probably other tradeoffs between the two 
> approaches that I've not thought of.

I've realized that creating the submodule's gitdir under the worktree's 
gitdir means that "git gc" running in the submodule repository wont see 
the per-worktree refs and index file and will happily prune those 
objects. Junio's suggestion avoids that problem.

Thanks

Phillip