Git Submodules

May 3, 2024

Have you ever worked on a project that depends on other libraries, be that first party or third party? Chances are, unless you are just getting started in the tech industry, that you are. There are many ways to manage dependencies, from utilizing dependency management systems (like pip, npm), to offloading them to build systems (such as Bazel, Maven, etc). In some cases, it’s desirable to let the source control system handle dependencies, and this post will cover how to do it in Git using Git submodules.

Git is the most popular DVCS used today. I’d say it has a steep learning curve for basic stuff (pull, push, commit, branching, merging), and then steep for learning advanced stuff (importing subtrees, rewriting history, etc). One of those advanced stuff is working with Git submodules and understanding it. As with many Git stuff, once you understand how things work “under the hood”, everything starts to make sense. I believe there’s a reason why the git manual says: “git - the stupid content tracker”.

Background about my recent work

In 2023, I led a project to migrate the Chromium browser codebase from a bespoke dependency solution to Git submodules. The transition was not smooth - we discovered bugs throughout our codebase, but also Git (discussion; both are now fully fixed). Moreover, we introduced a completely new user experience and users weren’t familiar with it.

What are Git submodules?

Git submodules are Git’s way to include other Git projects into another project. For example, if you depend on a library called “foo”, you could import foo as a submodule. The submodule has its history, and it’s completely unaware of projects that embed them (those are often called superprojects). To initialize, add, and inspect submodules, users can use git-submodule. man git-submodule covers that in great detail. This post will focus more on low-level details and how Git tracks submodules.

To have a proper Git submodule, you need an entry in the .gitmodules file, and a gitlink - a special entry in the Git database that contains information about the commit itself. Let’s see an example of the source code of my website:

$ cat .gitmodules
[submodule "themes/paper"]
        path = themes/paper
        url = https://github.com/sokac/hugo-paper

$ git ls-tree HEAD -- themes/paper
160000 commit 7c94235bf3e597d0d6cc61a4261123eb7dea441c  themes/paper

Those two pieces of information are sufficient for Git to know how to include themes/paper submodule into my website project (or superproject, since it has submodules). Note that a submodule can have more information (e.g. branch name), but those three bits of information are required.

If you enter a submodule, you can run the same Git commands as you can in a regular Git repository. For example, you can checkout to a different revision, commit new changes, etc.

From a superproject perspective, there are two important states that a submodule can be: in-sync or off-sync (e.g. git status shows new commits). Moreover, Git, by default, reports some additional information about submodules: untracked files and modified files (but not committed).

In-sync state means that a git submodule is checked out to a commit hash that superproject expects it (again, this is stored as a gitlink). Off-sync state means the opposite. To move from off-sync to in-sync state, the user can either use the git submodule update operation, or manually checkout to the desired commit hash.

Getting started with submodules

If you have a git repository to which you want to add a new submodule, the easiest way to do so is by using the git submodule command. Let’s say you have a repository https://example.org/repository.git (to clone, you would use git clone https://example.org/repository.git). To add that as a submodule to the existing repository, you can run: git submodule add https://example.org/repository.git path_you_want. That command will perform necessary changes to the Git database (that is, Git index) and .gitmodules files, as well as clone the repository.

If you cd into that directory, it behaves as a regular git repository. However, git repositories that are submodules have different .git/ paths from regular clones - the actual content of .git/ is stored inside the superproject git repository.

Contributing code to a submodule is straightforward - the same as contributing to any other repository. E.g. you can checkout different branches and commit code.

Working on a superproject

Working on a superproject gets complicated. In the previous paragraph, I mentioned we can contribute code to the submodule repository. Well, such changes can be reflected in the superproject. For example, if you checkout to a different branch (which points to a different commit), the superproject will show an off-sync state. If you forget you did that, and use git commit -a to commit all changes, you will include submodule change too. What is worse - those submodule changes may only exist on your local workstation. If such a bad commit gets pushed to remote, your teammates won’t be able to sync submodules unless the offending commit is reverted, or you push changes from the submodule in question. And on top of that, if you use a code review system that creates a new commit hash on submit (e.g. rebase-always strategy, or adds some metadata to the commit message), you will be forced to revert the change in the superproject.

I believe it’s a good practice to stage files first (using git add) and commit staged files only (i.e. don’t use -a). But, that’s easier said than done. With that, let’s see what other options we have when working with submodules.

Preventing accidental submodule commits

One way to prevent accidental submodule commits was discussed above. If you still like to stage all your changes, Git, as of this year (2024), supports excluding submodules when adding files using pathspecs. So, instead of using git add -a, you would use:

git add ':(exclude,attr:builtin_objectmode=160000)'

Since it’s a long command, you should have an alias for that. See the chromium submodule documentation on how to do that.

Another way to prevent accidental commits is via git hooks. A pre-commit hook allows you to inspect what goes into a commit. From there, a hook can instruct Git not to move forward with the commit creation by exiting with non-zero code. Or, the hook can unstage submodule changes (i.e. gitlinks) and continue with the commit flow. The Chromium project uses the latter hook, and it’s open-sourced (link). If you wish to keep it simple, you can use this simple bash script (store it in .git/pre-commit and make sure it’s executable):

#!/usr/bin/bash
set -ex

# Skip hook if SKIP_HOOK is 1, e.g. `SKIP_HOOK=1 git commit`
if [[ "$SKIP_HOOK" == "1" ]]; then
  exit 0
fi

diff=$(git diff-index --cached --ignore-submodules=dirty HEAD | grep "^:160000 160000")
if [[ -n "$diff" ]]; then
  echo "gitlink change detected in diff, aborting commit."
  exit 1
fi
exit 0

Bonus: Conflicts in submodules

This is probably the most confusing problem with submodules. Let’s say you add a commit that updates a submodule. And so does someone else who pushes their code before you. You are forced to pull the latest from origin, and you get a merge conflict. Git doesn’t know what should be done and needs you to figure that out. This can be frustrating, but let’s see what is happening under the hood.

Let’s assume there was a change A which had a submodule pointing to commit1. You created change B that points to commit2. Your teammate created change C that points to commit3. When you rebase (or merge) your change on top of commit 3, Git sees that submodule was updated from commit 1 to commit 2 in your commit, and from commit 1 to commit 3. So, which one is valid? Git certainly doesn’t know that.

When this happens, git won’t do anything with the affected submodule. If you just use git add submodule_path, it will use whatever commit hash is currently checked out locally. This may or may not be what you want. In most cases, you want to enter the submodule, checkout to the correct branch or commit hash, go back to the superproject and git add submodule_path. If you notice a frequent problem with submodule conflicts, it’s likely time to invest in tooling to reduce the frequency of conflicts, change processes, or build small tooling that helps resolve the problem.

Conclusion

As I mentioned, Git has a learning curve and that is true with submodules. If your build system or dependency system supports external dependencies, I would go down that route. If those are not options, Git submodules can be a good way to solve the problem - but become familiar with submodules, how they work under the hood, and how to effectively resolve conflicts - define best practices and write playbooks on what to do in what cases. Your teammates and stakeholders will also need to be familiar with it, so: speak up and share your plans early.

Note: Did I miss something? Feel free to reach out to me and I’ll do my best to update this post.