My First Contribution to the Git Project

This tutorial assumes you’re already fairly familiar with using Git to manage source code. The Git workflow steps will largely remain unexplained.

This tutorial aims to summarize the following documents, but the reader may find useful additional context:

If you get stuck, you can seek help in the following places.

Git is mirrored in a number of locations. Clone the repository from one of them; https://git-scm.com/downloads suggests one of the best places to clone from is the mirror on GitHub.

To build Git from source, you need to have a handful of dependencies installed on your system. For a hint of what’s needed, you can take a look at INSTALL, paying close attention to the section about Git’s dependencies on external programs and libraries. That document mentions a way to "test-drive" our freshly built Git without installing; that’s the method we’ll be using in this tutorial.

Make sure that your environment has everything you need by building your brand new clone of Git from the above step:

In this tutorial, we will add a new command, git psuh, short for “Pony Saying ‘Um, Hello”’ - a feature which has gone unimplemented despite a high frequency of invocation during users' typical daily workflow.

(We’ve seen some other effort in this space with the implementation of popular commands such as sl.)

Let’s start by making a development branch to work on our changes. Per Documentation/SubmittingPatches, since a brand new command is a new feature, it’s fine to base your work on master. However, in the future for bugfixes, etc., you should check that document and base it on the appropriate branch.

For the purposes of this document, we will base all our work on the master branch of the upstream project. Create the psuh branch you will use for development like so:

We’ll make a number of commits here in order to demonstrate how to send a topic with multiple patches up for review simultaneously.

Lots of the subcommands are written as builtins, which means they are implemented in C and compiled into the main git executable. Implementing the very simple psuh command as a built-in will demonstrate the structure of the codebase, the internal API, and the process of working together as a contributor with the reviewers and maintainer to integrate this change into the system.

Built-in subcommands are typically implemented in a function named "cmd_" followed by the name of the subcommand, in a source file named after the subcommand and contained within builtin/. So it makes sense to implement your command in builtin/psuh.c. Create that file, and within it, write the entry point for your command in a function matching the style and signature:

We’ll also need to add the declaration of psuh; open up builtin.h, find the declaration for cmd_pull, and add a new line for psuh immediately before it, in order to keep the declarations alphabetically sorted:

Be sure to #include "builtin.h" in your psuh.c.

Go ahead and add some throwaway printf to that function. This is a decent starting point as we can now add build rules and register the command.

Let’s try to build it. Open Makefile, find where builtin/pull.o is added to BUILTIN_OBJS, and add builtin/psuh.o in the same way next to it in alphabetical order. Once you’ve done so, move to the top-level directory and build simply with make. Also add the DEVELOPER=1 variable to turn on some additional warnings:

Great, now your new command builds happily on its own. But nobody invokes it. Let’s change that.

The list of commands lives in git.c. We can register a new command by adding a cmd_struct to the commands[] array. struct cmd_struct takes a string with the command name, a function pointer to the command implementation, and a setup option flag. For now, let’s keep mimicking push. Find the line where cmd_push is registered, copy it, and modify it for cmd_psuh, placing the new line in alphabetical order (immediately before cmd_pull).

The options are documented in builtin.h under "Adding a new built-in." Since we hope to print some data about the user’s current workspace context later, we need a Git directory, so choose RUN_SETUP as your only option.

Go ahead and build again. You should see a clean build, so let’s kick the tires and see if it works. There’s a binary you can use to test with in the bin-wrappers directory.

Check it out! You’ve got a command! Nice work! Let’s commit this.

git status reveals modified Makefile, builtin.h, and git.c as well as untracked builtin/psuh.c and git-psuh. First, let’s take care of the binary, which should be ignored. Open .gitignore in your editor, find /git-pull, and add an entry for your new command in alphabetical order:

Checking git status again should show that git-psuh has been removed from the untracked list and .gitignore has been added to the modified list. Now we can stage and commit:

You will be presented with your editor in order to write a commit message. Start the commit with a 50-column or less subject line, including the name of the component you’re working on, followed by a blank line (always required) and then the body of your commit message, which should provide the bulk of the context. Remember to be explicit and provide the "Why" of your change, especially if it couldn’t easily be understood from your diff. When editing your commit message, don’t remove the Signed-off-by trailer which was added by -s above.

Go ahead and inspect your new commit with git show. "psuh:" indicates you have modified mainly the psuh command. The subject line gives readers an idea of what you’ve changed. The sign-off line (-s) indicates that you agree to the Developer’s Certificate of Origin 1.1 (see the Documentation/SubmittingPatches [[dco]] header).

For the remainder of the tutorial, the subject line only will be listed for the sake of brevity. However, fully-fleshed example commit messages are available on the reference implementation linked at the top of this document.

It’s probably useful to do at least something besides printing out a string. Let’s start by having a look at everything we get.

Modify your cmd_psuh implementation to dump the args you’re passed, keeping existing printf() calls in place:

Build and try it. As you may expect, there’s pretty much just whatever we give on the command line, including the name of our command. (If prefix is empty for you, try cd Documentation/ && ../bin-wrappers/git psuh). That’s not so helpful. So what other context can we get?

Add a line to #include "config.h". Then, add the following bits to the function body:

git_config() will grab the configuration from config files known to Git and apply standard precedence rules. git_config_get_string_tmp() will look up a specific key ("user.name") and give you the value. There are a number of single-key lookup functions like this one; you can see them all (and more info about how to use git_config()) in Documentation/technical/api-config.txt.

You should see that the name printed matches the one you see when you run:

Great! Now we know how to check for values in the Git config. Let’s commit this too, so we don’t lose our progress.

Still, it’d be nice to know what the user’s working context is like. Let’s see if we can print the name of the user’s current branch. We can mimic the git status implementation; the printer is located in wt-status.c and we can see that the branch is held in a struct wt_status.

wt_status_print() gets invoked by cmd_status() in builtin/commit.c. Looking at that implementation we see the status config being populated like so:

But as we drill down, we can find that status_init_config() wraps a call to git_config(). Let’s modify the code we wrote in the previous commit.

Be sure to include the header to allow you to use struct wt_status:

Then modify your cmd_psuh implementation to declare your struct wt_status, prepare it, and print its contents:

Run it again. Check it out - here’s the (verbose) name of your current branch!

Let’s commit this as well.

Now let’s see if we can get some info about a specific commit.

Luckily, there are some helpers for us here. commit.h has a function called lookup_commit_reference_by_name to which we can simply provide a hardcoded string; pretty.h has an extremely handy pp_commit_easy() call which doesn’t require a full format object to be passed.

Add the following includes:

Then, add the following lines within your implementation of cmd_psuh() near the declarations and the logic, respectively.

The struct strbuf provides some safety belts to your basic char*, one of which is a length member to prevent buffer overruns. It needs to be initialized nicely with STRBUF_INIT. Keep it in mind when you need to pass around char*.

lookup_commit_reference_by_name resolves the name you pass it, so you can play with the value there and see what kind of things you can come up with.

pp_commit_easy is a convenience wrapper in pretty.h that takes a single format enum shorthand, rather than an entire format struct. It then pretty-prints the commit according to that shorthand. These are similar to the formats available with --pretty=FOO in many Git commands.

Build it and run, and if you’re using the same name in the example, you should see the subject line of the most recent commit in origin/master that you know about. Neat! Let’s commit that as well.

Awesome! You’ve got a fantastic new command that you’re ready to share with the community. But hang on just a minute - this isn’t very user-friendly. Run the following:

Your new command is undocumented! Let’s fix that.

Take a look at Documentation/git-*.txt. These are the manpages for the subcommands that Git knows about. You can open these up and take a look to get acquainted with the format, but then go ahead and make a new file Documentation/git-psuh.txt. Like with most of the documentation in the Git project, help pages are written with AsciiDoc (see CodingGuidelines, "Writing Documentation" section). Use the following template to fill out your own manpage:

The most important pieces of this to note are the file header, underlined by =, the NAME section, and the SYNOPSIS, which would normally contain the grammar if your command took arguments. Try to use well-established manpage headers so your documentation is consistent with other Git and UNIX manpages; this makes life easier for your user, who can skip to the section they know contains the information they need.

Now that you’ve written your manpage, you’ll need to build it explicitly. We convert your AsciiDoc to troff which is man-readable like so:

or

While this isn’t as satisfying as running through git help, you can at least check that your help page looks right.

You can also check that the documentation coverage is good (that is, the project sees that your command has been implemented as well as documented) by running make check-docs from the top-level.

Go ahead and commit your new documentation change.

Try and run ./bin-wrappers/git psuh -h. Your command should crash at the end. That’s because -h is a special case which your command should handle by printing usage.

Take a look at Documentation/technical/api-parse-options.txt. This is a handy tool for pulling out options you need to be able to handle, and it takes a usage string.

In order to use it, we’ll need to prepare a NULL-terminated array of usage strings and a builtin_psuh_options array.

Add a line to #include "parse-options.h".

At global scope, add your array of usage strings:

Then, within your cmd_psuh() implementation, we can declare and populate our option struct. Ours is pretty boring but you can add more to it if you want to explore parse_options() in more detail:

Finally, before you print your args and prefix, add the call to parse-options():

This call will modify your argv parameter. It will strip the options you specified in options from argv and the locations pointed to from options entries will be updated. Be sure to replace your argc with the result from parse_options(), or you will be confused if you try to parse argv later.

It’s worth noting the special argument --. As you may be aware, many Unix commands use -- to indicate "end of named parameters" - all parameters after the -- are interpreted merely as positional arguments. (This can be handy if you want to pass as a parameter something which would usually be interpreted as a flag.) parse_options() will terminate parsing when it reaches -- and give you the rest of the options afterwards, untouched.

Now that you have a usage hint, you can teach Git how to show it in the general command list shown by git help git or git help -a, which is generated from command-list.txt. Find the line for git-pull so you can add your git-psuh line above it in alphabetical order. Now, we can add some attributes about the command which impacts where it shows up in the aforementioned help commands. The top of command-list.txt shares some information about what each attribute means; in those help pages, the commands are sorted according to these attributes. git psuh is user-facing, or porcelain - so we will mark it as "mainporcelain". For "mainporcelain" commands, the comments at the top of command-list.txt indicate we can also optionally add an attribute from another list; since git psuh shows some information about the user’s workspace but doesn’t modify anything, let’s mark it as "info". Make sure to keep your attributes in the same style as the rest of command-list.txt using spaces to align and delineate them:

Build again. Now, when you run with -h, you should see your usage printed and your command terminated before anything else interesting happens. Great!

Go ahead and commit this one, too.

The tests in Git live in t/ and are named with a 4-digit decimal number using the schema shown in the Naming Tests section of t/README.

Since this a toy command, let’s go ahead and name the test with t9999. However, as many of the family/subcmd combinations are full, best practice seems to be to find a command close enough to the one you’ve added and share its naming space.

Create a new file t/t9999-psuh-tutorial.sh. Begin with the header as so (see "Writing Tests" and "Source test-lib.sh" in t/README):

Tests are framed inside of a test_expect_success in order to output TAP formatted results. Let’s make sure that git psuh doesn’t exit poorly and does mention the right animal somewhere:

Indicate that you’ve run everything you wanted by adding the following at the bottom of your script:

Make sure you mark your test script executable:

You can get an idea of whether you created your new test script successfully by running make -C t test-lint, which will check for things like test number uniqueness, executable bit, and so on.

Let’s try and run locally:

You can run the full test suite and ensure git-psuh didn’t break anything:

Go ahead and commit this change, as well.

Before you can send your patch off to be reviewed using GitGitGadget, you will need to fork the Git project and upload your changes. First thing - make sure you have a GitHub account.

Head to the GitHub mirror and look for the Fork button. Place your fork wherever you deem appropriate and create it.

To upload your branch to your own fork, you’ll need to add the new fork as a remote. You can use git remote -v to show the remotes you have added already. From your new fork’s page on GitHub, you can press "Clone or download" to get the URL; then you need to run the following to add, replacing your own URL and remote name for the examples provided:

or to use the HTTPS URL:

Run git remote -v again and you should see the new remote showing up. git fetch remotename (with the real name of your remote replaced) in order to get ready to push.

Next, double-check that you’ve been doing all your development in a new branch by running git branch. If you didn’t, now is a good time to move your new commits to their own branch.

As mentioned briefly at the beginning of this document, we are basing our work on master, so go ahead and update as shown below, or using your preferred workflow.

Finally, you’re ready to push your new topic branch! (Due to our branch and command name choices, be careful when you type the command below.)

Now you should be able to go and check out your newly created branch on GitHub.

In order to have your code tested and formatted for review, you need to start by opening a Pull Request against gitgitgadget/git. Head to https://github.com/gitgitgadget/git and open a PR either with the "New pull request" button or the convenient "Compare & pull request" button that may appear with the name of your newly pushed branch.

Review the PR’s title and description, as it’s used by GitGitGadget as the cover letter for your change. When you’re happy, submit your pull request.

If it’s your first time using GitGitGadget (which is likely, as you’re using this tutorial) then someone will need to give you permission to use the tool. As mentioned in the GitGitGadget documentation, you just need someone who already uses it to comment on your PR with /allow <username>. GitGitGadget will automatically run your PRs through the CI even without the permission given but you will not be able to /submit your changes until someone allows you to use the tool.

If the CI fails, you can update your changes with git rebase -i and push your branch again:

In fact, you should continue to make changes this way up until the point when your patch is accepted into next.

Now that your CI is passing and someone has granted you permission to use GitGitGadget with the /allow command, sending out for review is as simple as commenting on your PR with /submit.

Skip ahead to Responding to Reviews for information on how to reply to review comments you will receive on the mailing list.

Once you have your branch again in the shape you want following all review comments, you can submit again:

Next, go look at your pull request against GitGitGadget; you should see the CI has been kicked off again. Now while the CI is running is a good time for you to modify your description at the top of the pull request thread; it will be used again as the cover letter. You should use this space to describe what has changed since your previous version, so that your reviewers have some idea of what they’re looking at. When the CI is done running, you can comment once more with /submit - GitGitGadget will automatically add a v2 mark to your changes.

Configuration for send-email can vary based on your operating system and email provider, and so will not be covered in this tutorial, beyond stating that in many distributions of Linux, git-send-email is not packaged alongside the typical git install. You may need to install this additional package; there are a number of resources online to help you do so. You will also need to determine the right way to configure it to use your SMTP server; again, as this configuration can change significantly based on your system and email setup, it is out of scope for the context of this tutorial.

Sending emails with Git is a two-part process; before you can prepare the emails themselves, you’ll need to prepare the patches. Luckily, this is pretty simple:

The command will make one patch file per commit. After you run, you can go have a look at each of the patches with your favorite text editor and make sure everything looks alright; however, it’s not recommended to make code fixups via the patch file. It’s a better idea to make the change the normal way using git rebase -i or by adding a new commit than by modifying a patch.

Check and make sure that your patches and cover letter template exist in the directory you specified - you’re nearly ready to send out your review!

In addition to an email per patch, the Git community also expects your patches to come with a cover letter, typically with a subject line [PATCH 0/x] (where x is the number of patches you’re sending). Since you invoked format-patch with --cover-letter, you’ve already got a template ready. Open it up in your favorite editor.

You should see a number of headers present already. Check that your From: header is correct. Then modify your Subject: to something which succinctly covers the purpose of your entire topic branch, for example:

Make sure you retain the “[PATCH 0/X]” part; that’s what indicates to the Git community that this email is the beginning of a review, and many reviewers filter their email for this type of flag.

You’ll need to add some extra parameters when you invoke git send-email to add the cover letter.

Next you’ll have to fill out the body of your cover letter. This is an important component of change submission as it explains to the community from a high level what you’re trying to do, and why, in a way that’s more apparent than just looking at your diff. Be sure to explain anything your diff doesn’t make clear on its own.

Here’s an example body for psuh:

The template created by git format-patch --cover-letter includes a diffstat. This gives reviewers a summary of what they’re in for when reviewing your topic. The one generated for psuh from the sample implementation looks like this:

Finally, the letter will include the version of Git used to generate the patches. You can leave that string alone.

At this point you should have a directory psuh/ which is filled with your patches and a cover letter. Time to mail it out! You can send it like this:

After you run the command above, you will be presented with an interactive prompt for each patch that’s about to go out. This gives you one last chance to edit or quit sending something (but again, don’t edit code this way). Once you press y or a at these prompts your emails will be sent! Congratulations!

Awesome, now the community will drop everything and review your changes. (Just kidding - be patient!)

This section will focus on how to send a v2 of your patchset. To learn what should go into v2, skip ahead to Responding to Reviews for information on how to handle comments from reviewers.

We’ll reuse our psuh topic branch for v2. Before we make any changes, we’ll mark the tip of our v1 branch for easy reference:

Refine your patch series by using git rebase -i to adjust commits based upon reviewer comments. Once the patch series is ready for submission, generate your patches again, but with some new flags:

The --range-diff master..psuh-v1 parameter tells format-patch to include a range-diff between psuh-v1 and psuh in the cover letter (see git-range-diff(1)). This helps tell reviewers about the differences between your v1 and v2 patches.

The -v2 parameter tells format-patch to output your patches as version "2". For instance, you may notice that your v2 patches are all named like v2-000n-my-commit-subject.patch. -v2 will also format your patches by prefixing them with "[PATCH v2]" instead of "[PATCH]", and your range-diff will be prefaced with "Range-diff against v1".

Afer you run this command, format-patch will output the patches to the psuh/ directory, alongside the v1 patches. Using a single directory makes it easy to refer to the old v1 patches while proofreading the v2 patches, but you will need to be careful to send out only the v2 patches. We will use a pattern like "psuh/v2-.patch" (not "psuh/.patch", which would match v1 and v2 patches).

Edit your cover letter again. Now is a good time to mention what’s different between your last version and now, if it’s something significant. You do not need the exact same body in your second cover letter; focus on explaining to reviewers the changes you’ve made that may not be as visible.

You will also need to go and find the Message-Id of your previous cover letter. You can either note it when you send the first series, from the output of git send-email, or you can look it up on the mailing list. Find your cover letter in the archives, click on it, then click "permalink" or "raw" to reveal the Message-Id header. It should match:

Your Message-Id is <foo.12345.author@example.com>. This example will be used below as well; make sure to replace it with the correct Message-Id for your previous cover letter - that is, if you’re sending v2, use the Message-Id from v1; if you’re sending v3, use the Message-Id from v2.

While you’re looking at the email, you should also note who is CC’d, as it’s common practice in the mailing list to keep all CCs on a thread. You can add these CC lines directly to your cover letter with a line like so in the header (before the Subject line):

Now send the emails again, paying close attention to which messages you pass in to the command:

In some cases, your very small change may consist of only one patch. When that happens, you only need to send one email. Your commit message should already be meaningful and explain at a high level the purpose (what is happening and why) of your patch, but if you need to supply even more context, you can do so below the --- in your patch. Take the example below, which was generated with git format-patch on a single commit, and then edited to add the content between the --- and the diffstat.

After a few days, you will hopefully receive a reply to your patchset with some comments. Woohoo! Now you can get back to work.

It’s good manners to reply to each comment, notifying the reviewer that you have made the change suggested, feel the original is better, or that the comment inspired you to do something a new way which is superior to both the original and the suggested change. This way reviewers don’t need to inspect your v2 to figure out whether you implemented their comment or not.

Reviewers may ask you about what you wrote in the patchset, either in the proposed commit log message or in the changes themselves. You should answer these questions in your response messages, but often the reason why reviewers asked these questions to understand what you meant to write is because your patchset needed clarification to be understood.

Do not be satisfied by just answering their questions in your response and hear them say that they now understand what you wanted to say. Update your patches to clarify the points reviewers had trouble with, and prepare your v2; the words you used to explain your v1 to answer reviewers' questions may be useful thing to use. Your goal is to make your v2 clear enough so that it becomes unnecessary for you to give the same explanation to the next person who reads it.

If you are going to push back on a comment, be polite and explain why you feel your original is better; be prepared that the reviewer may still disagree with you, and the rest of the community may weigh in on one side or the other. As with all code reviews, it’s important to keep an open mind to doing something a different way than you originally planned; other reviewers have a different perspective on the project than you do, and may be thinking of a valid side effect which had not occurred to you. It is always okay to ask for clarification if you aren’t sure why a change was suggested, or what the reviewer is asking you to do.

Make sure your email client has a plaintext email mode and it is turned on; the Git list rejects HTML email. Please also follow the mailing list etiquette outlined in the Maintainer’s Note, which are similar to etiquette rules in most open source communities surrounding bottom-posting and inline replies.

When you’re making changes to your code, it is cleanest - that is, the resulting commits are easiest to look at - if you use git rebase -i (interactive rebase). Take a look at this overview from O’Reilly. The general idea is to modify each commit which requires changes; this way, instead of having a patch A with a mistake, a patch B which was fine and required no upstream reviews in v1, and a patch C which fixes patch A for v2, you can just ship a v2 with a correct patch A and correct patch B. This is changing history, but since it’s local history which you haven’t shared with anyone, that is okay for now! (Later, it may not make sense to do this; take a look at the section below this one for some context.)

The Git project has four integration branches: seen, next, master, and maint. Your change will be placed into seen fairly early on by the maintainer while it is still in the review process; from there, when it is ready for wider testing, it will be merged into next. Plenty of early testers use next and may report issues. Eventually, changes in next will make it to master, which is typically considered stable. Finally, when a new release is cut, maint is used to base bugfixes onto. As mentioned at the beginning of this document, you can read Documents/SubmittingPatches for some more info about the use of the various integration branches.

Back to now: your code has been lauded by the upstream reviewers. It is perfect. It is ready to be accepted. You don’t need to do anything else; the maintainer will merge your topic branch to next and life is good.

However, if you discover it isn’t so perfect after this point, you may need to take some special steps depending on where you are in the process.

If the maintainer has announced in the "What’s cooking in git.git" email that your topic is marked for next - that is, that they plan to merge it to next but have not yet done so - you should send an email asking the maintainer to wait a little longer: "I’ve sent v4 of my series and you marked it for next, but I need to change this and that - please wait for v5 before you merge it."

If the topic has already been merged to next, rather than modifying your patches with git rebase -i, you should make further changes incrementally - that is, with another commit, based on top of the maintainer’s topic branch as detailed in https://github.com/gitster/git. Your work is still in the same topic but is now incremental, rather than a wholesale rewrite of the topic branch.

The topic branches in the maintainer’s GitHub are mirrored in GitGitGadget, so if you’re sending your reviews out that way, you should be sure to open your PR against the appropriate GitGitGadget/Git branch.

If you’re using git send-email, you can use it the same way as before, but you should generate your diffs from <topic>..<mybranch> and base your work on <topic> instead of master.