7.11 Git Tools - Submodules

The simplest model of using submodules in a project would be if you were simply consuming a subproject and wanted to get updates from it from time to time but were not actually modifying anything in your checkout. Let’s walk through a simple example there.

If you want to check for new work in a submodule, you can go into the directory and run git fetch and git merge the upstream branch to update the local code.

Now if you go back into the main project and run git diff --submodule you can see that the submodule was updated and get a list of commits that were added to it. If you don’t want to type --submodule every time you run git diff, you can set it as the default format by setting the diff.submodule config value to “log”.

If you commit at this point then you will lock the submodule into having the new code when other people update.

There is an easier way to do this as well, if you prefer to not manually fetch and merge in the subdirectory. If you run git submodule update --remote, Git will go into your submodules and fetch and update for you.

This command will by default assume that you want to update the checkout to the master branch of the submodule repository. You can, however, set this to something different if you want. For example, if you want to have the DbConnector submodule track that repository’s “stable” branch, you can set it in either your .gitmodules file (so everyone else also tracks it), or just in your local .git/config file. Let’s set it in the .gitmodules file:

If you leave off the -f .gitmodules it will only make the change for you, but it probably makes more sense to track that information with the repository so everyone else does as well.

When we run git status at this point, Git will show us that we have “new commits” on the submodule.

If you set the configuration setting status.submodulesummary, Git will also show you a short summary of changes to your submodules:

At this point if you run git diff we can see both that we have modified our .gitmodules file and also that there are a number of commits that we’ve pulled down and are ready to commit to our submodule project.

This is pretty cool as we can actually see the log of commits that we’re about to commit to in our submodule. Once committed, you can see this information after the fact as well when you run git log -p.

Git will by default try to update all of your submodules when you run git submodule update --remote. If you have a lot of them, you may want to pass the name of just the submodule you want to try to update.

Let’s now step into the shoes of your collaborator, who has their own local clone of the MainProject repository. Simply executing git pull to get your newly committed changes is not enough:

By default, the git pull command recursively fetches submodules changes, as we can see in the output of the first command above. However, it does not update the submodules. This is shown by the output of the git status command, which shows the submodule is “modified”, and has “new commits”. What’s more, the brackets showing the new commits point left (<), indicating that these commits are recorded in MainProject but are not present in the local DbConnector checkout. To finalize the update, you need to run git submodule update:

Note that to be on the safe side, you should run git submodule update with the --init flag in case the MainProject commits you just pulled added new submodules, and with the --recursive flag if any submodules have nested submodules.

If you want to automate this process, you can add the --recurse-submodules flag to the git pull command (since Git 2.14). This will make Git run git submodule update right after the pull, putting the submodules in the correct state. Moreover, if you want to make Git always pull with --recurse-submodules, you can set the configuration option submodule.recurse to true (this works for git pull since Git 2.15). This option will make Git use the --recurse-submodules flag for all commands that support it (except clone).

There is a special situation that can happen when pulling superproject updates: it could be that the upstream repository has changed the URL of the submodule in the .gitmodules file in one of the commits you pull. This can happen for example if the submodule project changes its hosting platform. In that case, it is possible for git pull --recurse-submodules, or git submodule update, to fail if the superproject references a submodule commit that is not found in the submodule remote locally configured in your repository. In order to remedy this situation, the git submodule sync command is required:

It’s quite likely that if you’re using submodules, you’re doing so because you really want to work on the code in the submodule at the same time as you’re working on the code in the main project (or across several submodules). Otherwise you would probably instead be using a simpler dependency management system (such as Maven or Rubygems).

So now let’s go through an example of making changes to the submodule at the same time as the main project and committing and publishing those changes at the same time.

So far, when we’ve run the git submodule update command to fetch changes from the submodule repositories, Git would get the changes and update the files in the subdirectory but will leave the sub-repository in what’s called a “detached HEAD” state. This means that there is no local working branch (like master, for example) tracking changes. With no working branch tracking changes, that means even if you commit changes to the submodule, those changes will quite possibly be lost the next time you run git submodule update. You have to do some extra steps if you want changes in a submodule to be tracked.

In order to set up your submodule to be easier to go in and hack on, you need to do two things. You need to go into each submodule and check out a branch to work on. Then you need to tell Git what to do if you have made changes and then git submodule update --remote pulls in new work from upstream. The options are that you can merge them into your local work, or you can try to rebase your local work on top of the new changes.

First of all, let’s go into our submodule directory and check out a branch.

Let’s try updating our submodule with the “merge” option. To specify it manually, we can just add the --merge option to our update call. Here we’ll see that there was a change on the server for this submodule and it gets merged in.

If we go into the DbConnector directory, we have the new changes already merged into our local stable branch. Now let’s see what happens when we make our own local change to the library and someone else pushes another change upstream at the same time.

Now if we update our submodule we can see what happens when we have made a local change and upstream also has a change we need to incorporate.

If you forget the --rebase or --merge, Git will just update the submodule to whatever is on the server and reset your project to a detached HEAD state.

If this happens, don’t worry, you can simply go back into the directory and check out your branch again (which will still contain your work) and merge or rebase origin/stable (or whatever remote branch you want) manually.

If you haven’t committed your changes in your submodule and you run a submodule update that would cause issues, Git will fetch the changes but not overwrite unsaved work in your submodule directory.

If you made changes that conflict with something changed upstream, Git will let you know when you run the update.

You can go into the submodule directory and fix the conflict just as you normally would.

Now we have some changes in our submodule directory. Some of these were brought in from upstream by our updates and others were made locally and aren’t available to anyone else yet as we haven’t pushed them yet.

If we commit in the main project and push it up without pushing the submodule changes up as well, other people who try to check out our changes are going to be in trouble since they will have no way to get the submodule changes that are depended on. Those changes will only exist on our local copy.

In order to make sure this doesn’t happen, you can ask Git to check that all your submodules have been pushed properly before pushing the main project. The git push command takes the --recurse-submodules argument which can be set to either “check” or “on-demand”. The “check” option will make push simply fail if any of the committed submodule changes haven’t been pushed.

As you can see, it also gives us some helpful advice on what we might want to do next. The simple option is to go into each submodule and manually push to the remotes to make sure they’re externally available and then try this push again. If you want the check behavior to happen for all pushes, you can make this behavior the default by doing git config push.recurseSubmodules check.

The other option is to use the “on-demand” value, which will try to do this for you.

As you can see there, Git went into the DbConnector module and pushed it before pushing the main project. If that submodule push fails for some reason, the main project push will also fail. You can make this behavior the default by doing git config push.recurseSubmodules on-demand.

If you change a submodule reference at the same time as someone else, you may run into some problems. That is, if the submodule histories have diverged and are committed to diverging branches in a superproject, it may take a bit of work for you to fix.

If one of the commits is a direct ancestor of the other (a fast-forward merge), then Git will simply choose the latter for the merge, so that works fine.

Git will not attempt even a trivial merge for you, however. If the submodule commits diverge and need to be merged, you will get something that looks like this:

So basically what has happened here is that Git has figured out that the two branches record points in the submodule’s history that are divergent and need to be merged. It explains it as “merge following commits not found”, which is confusing but we’ll explain why that is in a bit.

To solve the problem, you need to figure out what state the submodule should be in. Strangely, Git doesn’t really give you much information to help out here, not even the SHA-1s of the commits of both sides of the history. Fortunately, it’s simple to figure out. If you run git diff you can get the SHA-1s of the commits recorded in both branches you were trying to merge.

So, in this case, eb41d76 is the commit in our submodule that we had and c771610 is the commit that upstream had. If we go into our submodule directory, it should already be on eb41d76 as the merge would not have touched it. If for whatever reason it’s not, you can simply create and checkout a branch pointing to it.

What is important is the SHA-1 of the commit from the other side. This is what you’ll have to merge in and resolve. You can either just try the merge with the SHA-1 directly, or you can create a branch for it and then try to merge that in. We would suggest the latter, even if only to make a nicer merge commit message.

So, we will go into our submodule directory, create a branch named “try-merge” based on that second SHA-1 from git diff, and manually merge.

We got an actual merge conflict here, so if we resolve that and commit it, then we can simply update the main project with the result.

It can be a bit confusing, but it’s really not very hard.

Interestingly, there is another case that Git handles. If a merge commit exists in the submodule directory that contains both commits in its history, Git will suggest it to you as a possible solution. It sees that at some point in the submodule project, someone merged branches containing these two commits, so maybe you’ll want that one.

This is why the error message from before was “merge following commits not found”, because it could not do this. It’s confusing because who would expect it to try to do this?

If it does find a single acceptable merge commit, you’ll see something like this:

The suggested command Git is providing will update the index as though you had run git add (which clears the conflict), then commit. You probably shouldn’t do this though. You can just as easily go into the submodule directory, see what the difference is, fast-forward to this commit, test it properly, and then commit it.

This accomplishes the same thing, but at least this way you can verify that it works and you have the code in your submodule directory when you’re done.

There is a foreach submodule command to run some arbitrary command in each submodule. This can be really helpful if you have a number of submodules in the same project.

For example, let’s say we want to start a new feature or do a bugfix and we have work going on in several submodules. We can easily stash all the work in all our submodules.

Then we can create a new branch and switch to it in all our submodules.

You get the idea. One really useful thing you can do is produce a nice unified diff of what is changed in your main project and all your subprojects as well.

Here we can see that we’re defining a function in a submodule and calling it in the main project. This is obviously a simplified example, but hopefully it gives you an idea of how this may be useful.

You may want to set up some aliases for some of these commands as they can be quite long and you can’t set configuration options for most of them to make them defaults. We covered setting up Git aliases in Git Aliases, but here is an example of what you may want to set up if you plan on working with submodules in Git a lot.

This way you can simply run git supdate when you want to update your submodules, or git spush to push with submodule dependency checking.

For instance, switching branches with submodules in them can also be tricky with Git versions older than Git 2.13. If you create a new branch, add a submodule there, and then switch back to a branch without that submodule, you still have the submodule directory as an untracked directory:

Removing the directory isn’t difficult, but it can be a bit confusing to have that in there. If you do remove it and then switch back to the branch that has that submodule, you will need to run submodule update --init to repopulate it.

Again, not really very difficult, but it can be a little confusing.

Newer Git versions (Git >= 2.13) simplify all this by adding the --recurse-submodules flag to the git checkout command, which takes care of placing the submodules in the right state for the branch we are switching to.

Using the --recurse-submodules flag of git checkout can also be useful when you work on several branches in the superproject, each having your submodule pointing at different commits. Indeed, if you switch between branches that record the submodule at different commits, upon executing git status the submodule will appear as “modified”, and indicate “new commits”. That is because the submodule state is by default not carried over when switching branches.

This can be really confusing, so it’s a good idea to always git checkout --recurse-submodules when your project has submodules. For older Git versions that do not have the --recurse-submodules flag, after the checkout you can use git submodule update --init --recursive to put the submodules in the right state.

Luckily, you can tell Git (>=2.14) to always use the --recurse-submodules flag by setting the configuration option submodule.recurse: git config submodule.recurse true. As noted above, this will also make Git recurse into submodules for every command that has a --recurse-submodules option (except git clone).

The other main caveat that many people run into involves switching from subdirectories to submodules. If you’ve been tracking files in your project and you want to move them out into a submodule, you must be careful or Git will get angry at you. Assume that you have files in a subdirectory of your project, and you want to switch it to a submodule. If you delete the subdirectory and then run submodule add, Git yells at you:

You have to unstage the CryptoLibrary directory first. Then you can add the submodule:

Now suppose you did that in a branch. If you try to switch back to a branch where those files are still in the actual tree rather than a submodule – you get this error:

You can force it to switch with checkout -f, but be careful that you don’t have unsaved changes in there as they could be overwritten with that command.

Then, when you switch back, you get an empty CryptoLibrary directory for some reason and git submodule update may not fix it either. You may need to go into your submodule directory and run a git checkout . to get all your files back. You could run this in a submodule foreach script to run it for multiple submodules.

It’s important to note that submodules these days keep all their Git data in the top project’s .git directory, so unlike much older versions of Git, destroying a submodule directory won’t lose any commits or branches that you had.

With these tools, submodules can be a fairly simple and effective method for developing on several related but still separate projects simultaneously.